JP2023531345A - Improved discourse parsing - Google Patents

Abstract

The systems, devices, and methods of the present invention include discourse trees. In some aspects, the system creates a discourse tree by identifying basic discourse units in the text. A discourse tree contains nodes, each non-terminal node representing a rhetorical relationship between two basic discourse units, and each terminal node being associated with a basic discourse unit. The system identifies type elaboration or joint rhetorical relations in the reference sentences of the discourse tree. The system selects the candidate sentence with the highest syntactic generalization score from the set of syntactic generalization scores. The system identifies semantic relationships that correspond to candidate sentences. Semantic relationships correspond to words in a candidate sentence and define roles in the candidate sentence. The system replaces rhetorical relations in the discourse tree with updated rhetorical relations corresponding to semantic relations, thereby creating an updated discourse tree.

Description

The present disclosure relates generally to linguistics. More specifically, the present disclosure relates to improved discourse tree generation.

Background Linguistics is the scientific study of language. One aspect of linguistics is the application of computer science to natural human languages such as English. Computer applications of linguistics are increasing due to the tremendous increase in processor speed and memory capacity. For example, computer-aided analysis of linguistic discourse facilitates numerous applications such as automated agents that can answer questions from users. However, such applications fail to leverage rich discourse-related information to answer questions, perform dialog management, or provide recommendation systems.

Overview In general, the systems, devices, and methods of the present invention relate to improved discourse trees.

In some aspects, a method of improving accuracy of a discourse tree includes creating a discourse tree from a text by identifying elementary discourse units in the text, the discourse tree comprising nodes, each non-terminal node of a node in the discourse tree representing a rhetorical relationship between two elementary discourse units, each terminal node of a node in the discourse tree being associated with an elementary discourse unit, the method further comprising identifying a rhetorical relationship of type elaboration or joint in the discourse tree. wherein the rhetorical relation relates the first elementary discourse unit and the second elementary discourse unit, the first elementary discourse unit and the second elementary discourse unit forming a reference sentence, the method further comprising determining a syntactic generalization score for each candidate sentence of the set of candidate sentences, each candidate sentence having a corresponding semantic relationship, determining identifying one or more common entities between the candidate sentence and the reference sentence; calculating a syntactic generalization score equal to the number of one or more common entities in the discourse tree, the method further comprising selecting the candidate sentence having the highest syntactic generalization score among the syntactic generalization scores; identifying semantic relationships corresponding to the candidate sentences, the semantic relationships corresponding to words in the candidate sentences and defining roles in the candidate sentences; , including creating an updated discourse tree.

In some aspects, creating a discourse tree from the text includes providing the text to a classification model and using the classification model to identify a first elementary discourse unit, a second elementary discourse unit, and a rhetorical relation.

In some aspects, the rhetorical relation updated is one of an end, a means, a cause, or a temporal sequence.

In some aspects, each of the one or more common entities shares a common portion of speech between the candidate sentence and the reference sentence.

In some aspects, the method further includes forming a response from the updated discourse tree and outputting the response to an external device.

In some aspects, the method includes forming a first syntactic parsetree from each candidate sentence and forming a second syntactic parsetree from the reference sentence, wherein identifying one or more common entities between the candidate sentence and the reference sentence includes, for each common entity, identifying common entities in the first syntactic partree and the second syntactic partree.

In some aspects, the method includes forming a communicative discourse tree from the updated discourse tree by matching each fragment having a verb to a verb signature, identifying that the text contains discourse by applying a classification model trained to detect discourse to the communicative discourse tree, forming a response from the text, and outputting the response to an external device.

In some aspects, the method includes forming a communicative discourse tree from the updated discourse tree by matching each fragment having a verb to a verb signature; identifying that the text contains an argument corresponding to an assertion by applying a classification model trained to detect arguments to the communicative discourse tree; , an assertion term and a domain definition clause, and a variable portion including a set of defeasible rules from the communicative discourse tree and facts from the communicative behavior of the communicative discourse tree, the method further comprising forming a text response from the text and outputting the text response to an external device in response to determining that the evaluated consistency is greater than the threshold.

The methods described above may be implemented as a tangible computer-readable medium and/or operating within a computer processor and associated memory.

FIG. 4 illustrates an exemplary discourse tree environment according to one aspect; FIG. 3 illustrates an example of a discourse tree according to one aspect; FIG. 4 illustrates yet another example of a discourse tree according to one aspect; FIG. 4 illustrates an exemplary schema according to one aspect; FIG. 4 illustrates a node-link representation of a hierarchical binary tree according to one aspect; 6 shows an exemplary indented text encoding for the representation in FIG. 5, according to one aspect; FIG. FIG. 12 illustrates an example discourse tree for an example claim regarding property taxes, according to one aspect; FIG. 8 illustrates exemplary responses to the question presented in FIG. 7; FIG. 11 shows a discourse tree for a formal answer according to one aspect; FIG. 4 illustrates a discourse tree for raw answers according to one aspect; FIG. 10 illustrates a communicative discourse tree for a first agent's claims, according to one aspect; FIG. 10 illustrates a communicative discourse tree for a second agent's claims, according to one aspect; FIG. 10 illustrates a communicative discourse tree for a third agent's claims, according to one aspect; FIG. 4 illustrates perspective intersection according to one aspect; FIG. 2 illustrates an exemplary process for building a discourse tree for communication, according to one aspect; FIG. 3 shows a discourse tree and scenario graph, according to one aspect; FIG. 4 illustrates forming a request-response pair, according to one aspect; FIG. 4 illustrates a discourse tree for maximal common sub-communications, according to one aspect; FIG. 4 illustrates a tree in kernel learning format for communication discourse trees, according to one aspect; FIG. 4 illustrates an exemplary process used to implement a rhetorical match classifier, according to one aspect; FIG. 10 illustrates a chatbot commenting on a post, according to one aspect; FIG. 10 illustrates a chatbot commenting on a post, according to one aspect; FIG. 10 illustrates a discourse tree for algorithmic text, according to one aspect; FIG. 4 illustrates an annotated sentence according to one aspect; FIG. 4 illustrates an annotated sentence according to one aspect; FIG. 4 illustrates discourse actions of a dialog according to one aspect; FIG. 4 illustrates discourse actions of a dialog according to one aspect; 1 illustrates an exemplary communicative discourse tree according to one aspect; 1 illustrates an exemplary communicative discourse tree according to one aspect; 1 illustrates an exemplary communicative discourse tree according to one aspect; 1 illustrates an exemplary communicative discourse tree according to one aspect; 1 illustrates an exemplary communicative discourse tree according to one aspect; 1 illustrates an exemplary communicative discourse tree according to one aspect; 1 illustrates an exemplary communicative discourse tree according to one aspect; 1 illustrates an exemplary communicative discourse tree according to one aspect; 1 illustrates an exemplary process for determining arguments using machine learning, according to one aspect; 4 is a fragment of a discourse tree according to one aspect; 4 illustrates a discourse tree for boundary review according to one aspect; 1 illustrates a discourse tree of sentences illustrating a complex semantic approach to sentiment analysis, according to one aspect; 1 illustrates an exemplary method for validating an argument, according to one aspect; 1 illustrates an exemplary communicative discourse tree of an argument according to one aspect; 1 illustrates an exemplary method for validating an argument using reversible logic programming, according to one aspect. 1 illustrates an exemplary dialectical tree according to one aspect; 1 illustrates a discourse tree and a semantic tree, according to one aspect; 1 illustrates a discourse tree and a semantic tree, according to one aspect; 1 is a flowchart of an exemplary process for generating an improved discourse tree, according to one aspect; FIG. 4 illustrates generalization of sentences and templates with known semantic relationships, according to one aspect. FIG. 11 illustrates a match between two sentences, according to one aspect; FIG. 1 shows a simplified diagram of a distributed system for implementing one of the aspects; FIG. 1 is a simplified block diagram of components of a system environment in which services provided by components of aspect systems may be provided as cloud services, according to one aspect; FIG. 1 illustrates an exemplary computer system in which various aspects of the invention may be implemented;

DETAILED DESCRIPTION Aspects disclosed herein provide technical improvements to the field of computer-implemented linguistics. More specifically, the disclosed solution generates an improved discourse tree by determining updated rhetorical relations of the discourse tree from the semantic relations of the semantic expressions of the text. Improved discourse trees can enable improved applications using discourse trees, such as dialog management, reasoning, argument detection, retrieval, and navigation. The improved discourse tree can be augmented with communicative behaviors, thereby forming a communicative discourse tree (“CDT”). Communicative behavior is collaborative behavior undertaken by individuals based on mutual deliberation and discussion.

Accordingly, technical advantages of some aspects include discourse trees that more accurately represent source texts and improved autonomous agents, such as autonomous agents that can verify arguments within texts using CDT. For example, CDT can be used to determine agreement between sentences, or to detect or verify arguments within a text. A valid argument is, for example, an argument that is logically consistent and whose text supports the premises of the argument.

More specifically, by incorporating labels that identify communicative behaviors, learning of communicative discourse trees can occur across feature sets that are richer than the purely rhetorical relationships and syntax of elementary discourse units (EDUs). With such a feature set, additional techniques such as classification can be used to determine the level of rhetorical agreement between questions and answers or between request-response pairs, detect arguments in text, and validate arguments in text, thereby enabling improved automated agents. In doing so, computing systems enable autonomous agents that can intelligently answer questions.

In another example, a rhetorical classification application executing on a computing device receives a question from a user. A rhetorical taxonomy application generates a communicative discourse tree for a question. A communicative discourse tree is a discourse tree containing communicative behaviors. A rhetorical classification application accesses a database of possible answers to questions. Using predictive models, the rhetorical matching application determines the level of complementarity between the question and each possible answer. In response to determining that the level of complementarity is above the threshold, the rhetorical match classifier provides an answer to the user, eg, via a display device.

In another example, a rhetorical classification application generates a communicative discourse tree (CDT) from input text and uses machine learning to validate arguments in a subset of the text. Rhetorical taxonomy applications create logic programs by extracting facts and retractable rules from a communicative discourse tree, and provide facts and retractable rules to logic systems such as defeasible logic programming (DeLP). The logic system then accesses the fixed rules and domain-specific definition clauses and solves the logic program, thereby determining whether the argument is valid (e.g., the argument supports the claim) or invalid (e.g., the argument does not support the claim).

Specific Definitions "Rhetorical structure theory", as used herein, is a field of research and research that has provided a logical foundation that can enable the analysis of discourse coherence.

A "discourse tree" or "DT", as used herein, refers to a structure that expresses rhetorical relationships about sentences forming part of a sentence.

"Rhetoric relation", "rhetorical relation", "consistency relation" or "discourse relation" as used herein describes how two segments of discourse are logically connected to each other. Examples of rhetorical relationships include elaboration, contrast and attribution.

A "sentence fragment" or "fragment" as used herein is a portion of a sentence that can be separated from the rest of the sentence. Fragments are basic discourse units. For example, in the sentence "Dutch accident investigators say that evidence points to pro-Russian rebels as being responsible for shooting down the plane," the two fragments are "Dutch accident investigators say that evidence points to pro-Russian rebels" and "as being responsible for shooting down the plane." A fragment may, but need not, contain a verb.

A "signature" or "frame" as used herein refers to a verb characteristic in a fragment. Each signature may contain one or more thematic roles. For example, for the fragment "Dutch accident investigators say that evidence points to pro-Russian rebels", the verb may be "say" and the signature for this particular use of the verb "say" may be "agent verb topic." In this case, "investigators" are agents and "evidence" are topics.

"Thematic role", as used herein, refers to the component of a signature used to describe the role of one or more words. Continuing with the example above, "agent" and "topic" are subject roles.

"Nuclearity" as used herein refers to which text segment, fragment or span is closer to the writer's center of purpose. The core is the more central span and the satellites are less central.

"Coherency", as used herein, refers to what links two rhetorical relationships together.

A "communicative verb," as used herein, is a verb that indicates communication. For example, the verb "deny" is a communicative verb.

A "communicative action" as used herein describes an action performed by one or more agents and the subject matter of the agents.

As used herein, a "claim" is a true statement of something. For example, a claim may be "I am not responsible for paying rent this month" or "I am late on rent."

As used herein, an "argument" is a reason or set of reasons stated to support a claim. An exemplary argument for the above claim is "necessary repairs were not completed."

As used herein, "argument validity" or "effectiveness" refers to whether the arguments supporting a claim are internal and consistent. Internal consistency refers to whether an argument is consistent with itself, eg, does not contain two conflicting statements. External consistency refers to whether an argument is consistent with known facts and rules.

As used herein, a "logical system" or "logical program" is a set of instructions, rules, facts, and other information that can represent an argument for a particular argument. Solving the logic system yields a determination of whether the argument is valid.

As used herein, a "dialectic tree" is a tree that represents individual arguments. Solve the dialectical tree to determine the truth or falsity of the claims supported by each argument. Evaluating the dialectical tree involves determining the validity of individual arguments.

FIG. 1 illustrates an exemplary discourse tree environment according to one aspect. FIG. 1 shows computing device 101 , input text 130 and discussion indicator 165 . Computing device 101 includes one or more of application 102 , discourse parser 104 , answer database 105 , rhetorical match classifier 120 , and training data 125 . Examples of computing devices include devices 4902, 4904, 4906 and 4908 shown in Figures 49 and 50, respectively, as well as cloud computing device 5002 and client devices 5004, 5006, 5008.

In some examples, the application 102 generates discourse trees with higher quality and/or accuracy than with previous solutions. In one example, discourse parser 104 generates a discourse tree from input text 130 . Application 102 analyzes discourse trees and generates semantic representations, such as abstract semantic representation (AMR) graphs. AMR is a semantic expression language. AMR graphs are full-text, rooted, labeled, directed, acyclic graphs (DAGs). From the AMR graph, using the techniques disclosed herein, application 102 generates an improved discourse tree, which can then be used to perform discourse analysis. An example process for creating an improved discourse tree is discussed with respect to FIG.

In another example, application 102 answers questions received via chat sessions. Input text 130 can be a single question or a stream of questions. Application 102 creates a discourse tree for question communication from input text 130 and selects one or more candidate answers. Answers can be obtained from existing databases, such as answer database 105 . Input text 130 may be generated by any mobile device such as a mobile phone, smart phone, tablet, laptop, smartwatch. Mobile devices can communicate with computing device 101 over a data network. In this way, the mobile device can provide the computing device 101 with questions from the user, for example.

Continuing the example, from the candidate answers, application 102 determines the most suitable answer. Different methods can be used. In one aspect, application 102 may create a discourse tree for candidate answer communication for each candidate answer and compare the discourse tree for question communication to each candidate discourse tree. Application 102 identifies the best match between the discourse tree for question communication and the discourse tree for candidate answer communication. Application 102 then accesses or queries the database for text from the optimal communicative discourse tree. Application 102 then sends the text associated with the second communication discourse tree to the mobile device.

In another example, application 102 creates a discourse tree for answer communication for each answer candidate. Application 102 then creates a question-answer pair that includes input text 130 and a candidate answer for each candidate answer. Application 102 provides question-answer pairs to a predictive model, such as rhetorical match classifier 120 . Using the trained rhetorical match classifier 120, the application 102 determines whether a question-answer pair exceeds a threshold level of match, eg, indicating whether the answer addresses the question. Otherwise, application 102 continues to analyze additional pairs containing questions and different answers until a suitable answer is found. By using a communicative discourse tree, rhetorical agreement and communicative behavior between questions and answers can be accurately modeled.

In a further example, application 102 uses rhetorical match classifier 120 to determine whether an argument is present or absent in input text 130 . For example, the rhetorical classification application 102 accesses input text 130 that reads: "[t]he rent was properly refused.... The landlord contacted me, the tenant, and the rent was requested. However, I refused the rent since I demanded repair to be done." I reminded the landlord about necessary repairs, but the landlord issued the three-day notice confirming that the rent was overdue.Regretfully, the property still stayed unrepaired." However, I refused the rent since I demanded repair to be done. I reminded the landlord about necessary repairs, but the landlord issued the three-day notice confirming that the rent was overdue. Regretfully, the property still stayed unrepaired."

To detect arguments, application 102 determines a communicative discourse tree from input text 130 and provides the communicative discourse tree to a trained classifier, such as rhetorical matching classifier 120 . Application 102 receives a prediction of whether an argument exists from rhetorical match classifier 120 . Application 102 provides that prediction as discussion indicator 165 . The rhetorical match classifier 120 compares the communicative discourse trees to those identified as positive (argument) or negative (no argument) in the training set. An exemplary process is discussed with respect to FIG.

In yet another aspect, application 102 can validate arguments present in input text 130 . An exemplary process is discussed with respect to FIG. In one example, application 102 determines the existence of an argument by using rhetorical match classifier 120, for example. Application 102 can then determine whether the detected argument is valid or invalid. Reversible logic programming can be used. One exemplary process is discussed with respect to FIG. Application 102 may output an argument indicator 165 that may indicate whether an argument has been detected, and if so, indicate whether the argument is valid or invalid.

Rhetorical Structural Theory and Discourse Trees Linguistics is the scientific study of language. For example, linguistics can include the structure of sentences (syntax), e.g., subject-verb-object, the meaning of sentences (semantics), e.g., "man bites dog" versus "dog bites man," and what speakers do during conversation, i.e., discourse analysis or analysis of language beyond sentences.

The theoretical basis for discourse (Rhetoric Structure Theory (RST)) can be attributed to Mann, William and Thompson, Sandra, "Rhetorical structure theory: A Theory of Text organization" (Text-Interdisciplinary Journal for the Study of Discourse) 8(3):243-281: 1988). Just as the syntax and semantics of programming language theory have helped enable modern software compilers, RST has helped enable the analysis of discourse. More specifically, RST envisions building blocks on at least two levels. The two levels include a first level such as nuclearity and rhetorical relations and a second level of structure or schema. A discourse parser or other computer software can parse the text into a discourse tree.

Rhetorical structure theory models the logical organization of a text (the structure used by the writer) in dependence on the relationships between parts of the text. RST simulates textual coherence by forming a hierarchical, connected structure of text via discourse trees. Rhetorical relations are divided into peer classes and subclasses. These relationships are maintained across two or more text spans, thus providing consistency. These text spans are called elementary discourse units (EDUs). Clauses within a sentence and sentences within a text are logically connected by the author. The meaning of a given sentence is related to the meaning of the previous sentence and the next sentence. This logical relationship between clauses is called the coherence structure of the text. RST is one of the most popular theories of discourse, based on a tree-like discourse structure, the discourse tree (DT). The leaves of DT correspond to EDUs (contiguous atomic text spans). Adjacent EDUs are connected by coherence relations (eg, attributes, sequences) that form higher-level discourse units. These units are also subordinate to this relationship link. EDUs linked by relationships are further differentiated based on their relative importance. The core is the core part of the relationship and the satellite is the peripheral part. As noted above, topical and rhetorical matches are analyzed together to determine the correct request-response pair. When a speaker answers a question such as a phrase or sentence, the speaker's answer must address the topic of the question. If the question is implicitly formulated, then through the seed text of the message, we expect an appropriate answer that not only maintains the topic but also matches the generalized state of awareness about this seed.

Rhetorical Relations As noted above, some aspects described in this specification use discourse trees for communication. Rhetorical relations can be explained in various ways. For example, Mann and Thompson describe 23 possible relationships. There is "Rhetorical Structure Theory: A Theory of Text Organization" by C. Mann, William & Thompson, Sandra (1987) ("Mann and Thompson") Several other relationships are possible.

Some empirical studies assume that the majority of texts are constructed using nuclear-satellite relationships. See Mann and Thompson. However, other relationships do not involve a finite selection of nuclei. Examples of such relationships are shown below.

FIG. 2 shows an example discourse tree according to one aspect. FIG. 2 includes discourse tree 200 . The discourse tree includes text spans 201 , text spans 202 , text spans 203 , relations 210 and relations 228 . The numbers in FIG. 2 correspond to three text spans. FIG. 3 corresponds to the following text example with three text spans numbered 1, 2, 3:

1. Honolulu, Hawaii will be site of the 2017 Conference on Hawaiian History.

2. It is expected that 200 historians from the U.S. and Asia will attend.

3. The conference will be concerned with how the Polynesians sailed to Hawaii.

For example, relationship 210 or specification describes the relationship between text span 201 and text span 202 . Relationship 228 indicates the relationship (detailed) between text span 203 and text span 204 . As shown, text spans 202 and 203 elaborate on text span 201 . In the example above, text span 1 is the kernel, assuming the purpose is to notify the reader of a meeting. Text spans 2 and 3 provide more details about the meeting. In FIG. 2, the horizontal numbers (eg, 1-3, 1, 2, 3) cover spans of text (possibly composed of further spans), and the vertical lines indicate a kernel or kernels. The curve represents the rhetorical relationship (detailed), the direction of the arrow pointing from the satellite to the nucleus. If only text spans acted as satellites instead of cores, removing the satellites would still leave the text consistent. If we remove the kernel from FIG. 2, text spans 2 and 3 become difficult to understand.

FIG. 3 shows a further example of a discourse tree according to one aspect. FIG. 3 includes components 301 and 302 , text spans 305 - 307 , relation 310 and relation 328 . Relationships 310 show relationships 310 (enablings) between components 306 and 305 and between components 307 and 305 . Figure 3 points to the following text spans:
1. The new Tech Report abstracts are now in the journal area of the library near the abridged dictionary.

2. Please sign your name by any means that you would be interested in seeing.

3. Last day for sign-ups is 31 May.
As can be seen, relationship 328 indicates the relationship, or enablement, between entity 307 and entity 306 . FIG. 3 illustrates that multiple kernels can be nested, but there is only one text span with the most kernels.

Building Discourse Trees Discourse trees can be generated using a variety of methods. A simple example of how to build a DT bottom up is:
(1) Divide the discourse text into multiple units according to the following (a) and (b).

(a) unit size may vary depending on the purpose of analysis;
(b) typically the unit is a clause;

(2) Examine each unit and each adjacent unit. Are there relationships between them?
(3) Mark the relationship if it holds.

(4) If the relationship does not hold, the unit may be on the border of a higher level relationship. Note the relationships held between larger units (spans).

(5) Continue until all units in the text are captured.
Mann and Thompson also describe a second level of building block structures called schema applications. In RST, rhetorical relations are not mapped directly onto the text, they are fitted onto constructs called schema applications, which are further fitted onto the text. Schema applications derive from a simpler structure called a schema (as illustrated by FIG. 4). Each schema indicates how a particular unit of text can be decomposed into smaller units of text. A rhetorical structure tree or DT is a hierarchical system of schema applications. Schema applications link several consecutive text spans to create complex text spans. Complex text spans can also be linked by higher level schema applications. RST argues that the structure of all coherent discourse can be described by a single rhetorical structure tree, whose top-level schema creates spans that encompass the entire discourse.

FIG. 4 shows an exemplary schema according to one aspect. FIG. 4 shows that the joint schema is a list of items that consist of nuclei but do not contain satellites. FIG. 4 shows schemas 401-406. Schema 401 shows the contextual relationship between text spans 410 and text spans 428 . Scheme 402 shows the sequence relationship between text spans 420 and 421 and between text spans 421 and 422 . Schema 403 shows a contrasting relationship between text span 430 and text span 431 . Schema 404 shows the joint relationship between text span 440 and text span 441 . Schema 405 shows the motivation relationship between 450 and 451 and the enablement relationship between 452 and 451 . Schema 406 shows the joint relationship between text span 460 and text span 462 . An example joint scheme is shown in FIG. 4 for the following three text spans.

1. Skies will be partly sunny in the New York metropolitan area today.

2. It will be more humid, with temperatures in the middle 80's.

3. Tonight will be mostly cloudy, with the low temperature between 65 and 70.

2-4 show discourse trees in several graphical representations, other representations are possible.

FIG. 5 shows a node-linked representation of a hierarchical binary tree according to one aspect. As can be seen from FIG. 5, the leaves of the DT correspond to contiguous but non-overlapping text spans called elementary discourse units (EDUs). Adjacent EDUs are connected by relations (eg, descriptions, attributes...) and form larger discourse units connected by relations. "Discourse analysis in RST includes two subtasks: discourse segmentation is the task of identifying EDUs, and discourse parsing is the task of linking discourse units into labeled trees." Joty, Shafiq R and Giuseppe Carenini, Raymond T Ng, and Yashar Mehdad (2013), "Combining intra-and multi-sentential rhetorical parsing for document-level discourse analysis" (ACL (1), pages 486-496).

FIG. 5 shows text spans that are leaves or terminal nodes on the tree, numbered in the order in which they appear throughout the text shown in FIG. FIG. 5 includes tree 500 . Tree 500 includes, for example, nodes 501-507. Nodes show relationships. The nodes may be non-terminal nodes such as node 501 or terminal nodes such as nodes 502-507. As can be seen, nodes 503 and 504 are related by a joint relationship. Nodes 502, 505, 506 and 508 are kernels. Dotted lines indicate that branches or text spans are satellites. These relationships are the nodes in the gray boxes.

FIG. 6 shows exemplary indented text encoding for the representation in FIG. 5, according to one aspect. FIG. 6 includes text 600 and text sequences 602-604. Text 600 is represented in a manner more amenable to computer programming. Text sequence 602 corresponds to node 502 . Sequence 603 corresponds to node 503 . Sequence 604 corresponds to node 504 . In FIG. 6, "N" indicates the nucleus and "S" indicates the satellite.

Discourse Parser Example Automatic discourse segmentation can be performed in a variety of ways. For example, given a sentence, the segmentation model identifies boundaries of complex elementary discourse units by predicting whether a boundary should be inserted before each particular token in the sentence. For example, one framework considers each token in a sentence independently in succession. In this framework, the segmentation model scans sentence token by token and uses binary classification such as support vector machines or logistic regression to predict whether it is appropriate to insert a boundary before the token being examined. In another example, the task is a continuous labeling problem. Once the text is segmented into basic discourse units, sentence-level discourse parsing can be performed to build a discourse tree. Machine learning techniques can be used.

In one aspect of the invention, two Rhetorical Structure Theory (RST) discourse parsers are used, the CoreNLPProcessor, which relies on component syntax, and the FastNLPProcessor, which uses dependency syntax. See Surdeanu, Mihai & Hicks, Thomas & Antonio Valenzuela-Escarcega, Marco, "Two Practical Rhetorical Structure Theory Parsers" (2015).

In addition, the two discourse parsers mentioned above, CoreNLPProcessor and FastNLPProcessor, use Natural Language Processing (NLP) for parsing. For example, Stanford CoreNLP presents basic shapes for multiwords that are parts of speech, whether or not they are the names of companies, people, etc., normalizes dates, times and numerical quantities, marks the structure of sentences in terms of phrase and syntactic dependence, and indicates which noun phrases refer to the same entity. In practice, RST is still a theory that can work in many cases of discourse, but it may not work in some cases. There are many variables including, but not limited to, what EDUs are in the coherent text, i.e. what discourse segmenters are used, what inventory of relations is used, what relations are selected for EDUs, the corpus of documents used for training and testing, and even what parsers are used. Thus, for example, in the above-mentioned paper by Surdeanu et al., "Two Practical Rhetorical Structure Theory Parsers," tests must be run on a particular corpus with specialized metrics to determine which parser gives better performance. Thus, unlike computer language parsers, which yield predictable results, discourse parsers (and segmenters) can yield unpredictable results, depending on the training and/or testing text corpus. Thus, discourse trees are a mixture of predictable techniques (e.g., compilers) and unpredictable techniques (e.g., chemistry that requires experimentation to determine which combinations can produce desired results).

To objectively judge how good discourse analysis is, a set of metrics is used, for example, the Precision/Recall/F1 metric by Daniel Marcu, "The Theory and Practice of Discourse Parsing and Summarization" (MIT Press) (2000). Accuracy, or positive predictive value, is the fraction of relevant instances among retrieved instances, while recall (also known as sensitivity) is the fraction of relevant instances retrieved over the total amount of relevant instances. Therefore, both precision and recall are based on understanding and criteria for relevance. Suppose a computer program for recognizing dogs in pictures identifies 8 dogs in a picture containing 12 dogs and some cats. Of the 8 dogs identified, 5 are actually dogs (true positives) and the rest are cats (false positives). The accuracy of the program is 5/8 and its recall is 5/12. If a search engine returns 30 pages, but only 20 of them are relevant, and does not return an additional 40 relevant pages, its accuracy is 20/30=2/3 and its recall is 20/60=1/3. So, in this case, precision is "how useful are the search results" and recall is "how complete are the results?" The F1 score (F-score or F-criterion) is a measure of test accuracy. It takes into account both precision and recall of the test to calculate the score. F1=2×(precision×recall)/(precision+recall)), which is the harmonic mean of precision and recall. The F1 score reaches its optimum value at 1 (perfect precision and recall) and its worst value at 0.

Autonomous Agents or Chatbots A conversation between person A and person B is a form of discourse. For example, there are applications such as FaceBook Messenger, WhatsApp, Slack, SMS, etc., and conversations between A and B typically may be via messages in addition to more traditional email and voice conversations. Chatbots (sometimes referred to as intelligent bots or virtual assistants, etc.) are "intelligent" machines that, for example, replace person B and mimic a conversation between two people to varying degrees. An example ultimate goal is for person A to be unable to distinguish whether B is a person or a machine (Turning test developed by Alan Turing in 1950). Artificial intelligence and natural language processing, including discourse analysis, machine learning, have come a long way toward the long-term goal of passing the Turing test. Of course, computers are also increasingly capable of searching and processing vast repositories of data and performing complex analyzes on the data, including predictive analytics, and the long-term goal is to combine human-like chatbots with computers.

For example, users can interact with the intelligent bot platform through conversational exchanges. This interaction, called a conversational user interface (UI), is a dialogue between an end-user and a chatbot, just like between two people. This could be as simple as the end-user saying "Hello" to the chatbot, the chatbot replying "Hi", and the chatbot asking the user what the business is about, or it could be a transactional interaction with a banking chatbot, such as transferring money from one account to another, or an exchange of information in a HR chatbot, such as checking a vacation balance, or asking an FAQ in a retail chatbot, such as how to handle returns. can be Natural language processing (NLP) and machine learning (ML) algorithms combined with other approaches can be used to classify end-user intent. A high-level intent is what the end-user wants to accomplish (eg, get an account balance and make a purchase). An intent is essentially a mapping of customers entered into units of work that the backend should perform. Thus, based on the phrases uttered by the user in the chatbot, these map to specific distinct use cases or units of work, e.g., for balance inquiries, money transfers, and tracking expenses, all "use cases" that the chatbot should be able to support and resolve which units of work must be triggered from free text entry that the end user inputs in natural language.

The underlying principle behind making AI chatbots respond like humans is that the human brain is able to formulate and understand requests, as well as being able to respond better to human requests, much better than machines. Therefore, if Person B is imitated, the chatbot's request/response should be significantly improved. So the first part of the problem is how the human brain formulates and understands requests. Models are used for imitation. RST and DT allow a formal and repeatable way to do this.

At a high level, there are typically two types of requests. Specifically, (1) a request to perform some action, and (2) a request for information (eg, a question). The first type has a response in which a unit of work is created. The second type has responses to questions (ie, good answers, for example). Answers may, for example, in some aspects take the form of AI constructing answers from extensive knowledge bases or from matching the best existing answers by searching the internet or intranet or other publicly or privately available data sources.

Communicative Discourse Trees and Rhetorical Classifiers Aspects of the present disclosure build and use communicative discourse trees to analyze whether the rhetorical structure of a request or question matches an answer. More specifically, aspects described herein create representations of request-response pairs, learn these representations, and associate pairs with classes of valid or invalid pairs. In this manner, an autonomous agent can receive a question from a user, process the question, for example, by retrieving multiple answers, determine the best answer among the multiple answers, and provide the answer to the user.

More specifically, the aspects described herein employ rhetorical relations and speech actions (or communicative actions) to express the linguistic features of text. A rhetorical relation is a relation between parts of a sentence, typically obtained from a discourse tree. Speech actions are obtained as verbs from a verb resource such as VerbNet. By using both rhetorical relationships and communicative behaviors, the aspects described herein can accurately recognize valid request-response pairs. To do so, multiple aspects correlate the syntactic structure of the question with the syntactic structure of the answer. By using the structure, a better answer can be determined.

For example, if an autonomous agent receives an indication from a person that he or she wishes to sell an item with a particular characteristic, the autonomous agent should provide search results that not only contain that characteristic but also indicate an intent to purchase. In this manner, the autonomous agent has determined the user's intent. Similarly, if an autonomous agent receives a request from someone to share information about a particular item, the search results should include an intent to receive recommendations. When a person asks an autonomous agent about its opinion on a topic, the autonomous agent shares its opinion on the subject rather than soliciting another's opinion.

Analysis of Request-Response Pairs FIG. 7 illustrates an exemplary discourse tree for a request example regarding property taxes, according to one aspect. Node labels are relations and lines with arrows point to satellites. Nuclei are solid lines. Figure 7 shows the following text.

Request: “My husbands’ grandmother gave him his grandfather’s truck. She signed the title over but due to my husband having unpaid fines on his license, he was not able to get the truck put in his name. I wanted to put in my name and paid the property tax and got insurance for the truck. refund. I am just wondering that since I am not going to have a tag on this truck, is it possible to get the property tax refunded?(My husband's grandmother gave him his grandfather's truck. She transferred the ownership but my husband had an outstanding fine against his license so he couldn't put the truck in his name. I wanted it in my name so I paid the property tax and got insurance for the truck. I got the property tax refunded and got the property tax refunded. Sometimes I didn't have the money to do so.Now, depending on the circumstances, I won't be able to afford to own the truck.I have visited the insurance agent and have been denied a refund.I'm wondering if I can get my property tax refunded if I can't put an identification tag on this truck."

Response: “The property tax is assessed on property that you own. Just because you chose to not register it does not mean that you don't own it, so the tax is not refundable. Even if you have not titled the vehicle yet, you still own it within the boundaries of the tax district, so the tax is payable. Note that all states give you a limited amount of time to transfer title and pay the use tax. same time, but you absolutely need to title it within the period of time stipulated in state law (property tax is on property you own. Just because you chose not to register it, this doesn't mean you don't own it. So you can't get the tax refunded. Even if you haven't been entitled to the vehicle yet, you still own it within the boundaries of the taxable lot and will have to pay the tax eventually.) Please keep in mind that some states also limit the time available to you for transferring ownership and paying the use tax.If you file late, you will be subject to fines in addition to the regular tax and regular fee.You do not have to register the vehicle at the same time, but you must ensure that you are entitled to it within the time stipulated by state law."
As can be seen from FIG. 7, analyzing the above text yields the following results. "My husbands' grandmother gave him his grandfather's truck" is elaborated by the expression "She signed the title over but due to my husband" elaborated by the expressions "I wanted to put in my name", "and paid the property tax", and "having unpaid fines on his license, he was not able to get the truck put in his name".

"My husbands' grandmother gave him his grandfather's truck. She signed the title over but due to my husband having unpaid fines on his license, he was not able to get the truck put in his name. I wanted to put in my name and paid the property tax and got insurance for the truck."

“My husbands’ grandmother gave him his grandfather’s truck. She signed the title over but due to my husband having unpaid fines on his license, he was not able to get the truck put in his name. I wanted to put in my name and the property tax and got insurance for the truck. Contrast with "Now, due to circumstances," detailed by "the truck".

“My husbands’ grandmother gave him his grandfather’s truck. She signed the title over but due to my husband having unpaid fines on his license, he was not able to get the truck put in his name. I wanted to put in my name and the property tax and got insurance for the truck. is detailed in "I am just wondering that since I am not going to have a tag on this truck, is it possible to get the property tax refunded?"

"I am just wondering" belongs to the same unit "that" as "is it possible to get the property tax refunded?" with the condition "since I am not going to have a tag on this truck".

As mentioned above, the main subject of the topic is "property tax on automobiles". The question contains the contradiction that, on the one hand, all property is taxable, on the other hand, ownership is somewhat imperfect. A suitable response should address the topic of the question and clarify any discrepancies. Because of this, respondents are making even stronger arguments about the need to pay taxes on everything owned regardless of registration status. An example of this is Yahoo! A member of the positive training set from the Answers evaluation domain. The main subject of the topic is "property tax on motor vehicles". The question contains the contradiction that, on the one hand, all property is taxable, while on the other hand, ownership is somewhat imperfect. A suitable answer/response should address the topic of the question and clarify any discrepancies. The reader may realize that since the question contains a rhetorical relation of contrast, the answer must match this question in a similar relation to be convincing. In other cases, the answer will appear incomplete even to people who are not experts in the field.

FIG. 8 shows exemplary responses to the questions presented in FIG. 7, according to certain aspects of the invention. The core is "The property tax is assessed on property" detailed by "that you own". "The property tax is assessed on property that you own" is also the core detailed by "Just because you chose to not register it does not mean that you don't own it, so the tax is not refundable. Even if you have not titled the vehicle yet, you still own it within the boundaries of the tax district, so the tax is payable. Note that all states give you a limited amount of time to transfer title and pay the use tax."

The core, "The property tax is assessed on property that you own. Just because you chose to not register it does not mean that you don't own it, so the tax is not refundable. Even if you have not titled the vehicle yet, you still own it within the boundaries of the tax district, so the tax is payable. Note that all states give you a limited amount of time to transfer title and pay the use tax." fees”. This is further elaborated by the contrast "but you absolutely need to title it within the period of time stipulated in state law" and "You don't need to register it at the same time".

By comparing the DT of FIG. 7 with the DT of FIG. 8, it is possible to determine how well the response (FIG. 8) matches the request (FIG. 7). In some aspects of the invention, the framework described above is used, at least in part, to determine DTs for request/response and rhetorical agreement between DTs.

In another example, the question "What does The Investigative Committee of the Russian Federation do" has at least two answers, for example an official answer or an actual answer.

FIG. 9 shows a discourse tree for a formal answer according to one aspect. As shown in Figure 9, the official response or statement is "The Investigative Committee of the Russian Federation is the main federal investigating authority which operates as Russia's Anti-corruption agency and has statutory responsibility for inspecting the police forces, combating police corruption and police misconduct, is responsible for conducting investigations into local authorities and federal governmental bodies." We have a statutory responsibility to eradicate the behavior and are responsible for conducting investigations of local and federal government bodies."

FIG. 10 shows a discourse tree for raw answers according to one aspect. As shown in Figure 10, another, possibly more honest, answer is as follows. “Investigative Committee of the Russian Federation is supposed to fight corruption. However, top-rank officers of the Investigative Committee of the Russian Federation are charged with creation of a criminal community. "However, top-ranking officials of the Russian Federation's Commission of Inquiry have a role in founding criminal gangs. Not only that, but it has been reported that these officials have been implicated in large-scale bribery, money laundering, obstruction of justice, abuse of power, extortion and racketeering. Dozens of high-profile cases, including those involving big criminals, have ultimately been undermined through the actions of these officials."
Choice of answer depends on the context. The rhetorical structure allows us to distinguish between 'official', 'politically correct' template-based answers and 'actual', 'raw', 'reports from the field' or 'controversial' answers. See FIGS. 9 and 10. FIG. Sometimes the question itself can give hints as to what category of answer is expected. Answers in the first category are appropriate if the question is formulated as a question with a quasi-factual or defining character without secondary meaning. In other cases, if the question has the meaning "tell me what it really is," the second category is appropriate. In general, after extracting the rhetorical structure from the question, it is easier to select appropriate answers that will have a similar rhetorical structure, a matching rhetorical structure, or a complementary rhetorical structure.

Official responses are based on elaborations and joints that are neutral in terms of the controversies the text may contain (see figure). At the same time, the raw responses contain contrasting relationships. This relationship between phrases about what the agent is expected to do and phrases about what the agent was found to do is extracted.

Classification of Request-Response Pairs Application 102 can determine whether a given answer or response, such as an answer obtained from answer database 105 or a public database, is in response to a given question or request. More specifically, application 102 analyzes whether a request and response pair is accurate or inaccurate by determining one or both of (i) relevance or (ii) rhetorical agreement between the request and response. Rhetorical agreement can be analyzed without taking into account relationships that can be processed orthogonally.

Application 102 may use a variety of methods to determine similarity between question-answer pairs. For example, application 102 can determine the level of similarity between individual questions and individual answers. Alternatively, application 102 can determine a measure of similarity between a first pair of questions and answers and a second pair of questions and answers.

For example, application 102 employs a rhetorical match classifier 120 trained to predict matching or non-matching answers. Application 102 can process two pairs at a time, eg <q1, a1> and <q2, a2>. Application 102 compares q1 to q2 and a1 to a1 to generate a combined similarity score. Such a comparison allows determining whether an unknown question/answer pair contains the correct answer by evaluating the distance from another question/answer pair with a known label. In particular, the unlabeled pair <q2,a2> can be treated such that both q2 and a2 can be compared to their corresponding components q1 and a2 of the labeled pair <q2,a2> on the basis of such words or structures, rather than "guessing" correctness based on words or structures shared by q2 and a2. Since this method aims to classify answers without relying on domains, it can only exploit the structural connectivity between questions and answers, and cannot exploit the "meaning" of the answers.

In one aspect, application 102 uses training data 125 to train rhetorical match classifier 120 . In this manner, the rhetorical match classifier 120 is trained to determine the similarity between question and answer pairs. This is a classification problem. Training data 125 may include a positive training set and a negative training set. The training data 125 includes matching request-response pairs in the positive data set and arbitrary or less relevant or relevant request-response pairs in the negative data set. For positive datasets, different domains are selected with distinct acceptance criteria that indicate whether the answer or response is suitable for the question.

Each training data set contains a set of training pairs. Each training set includes a discourse tree for question communication that expresses questions and a discourse tree for answer communication that expresses answers and the expected level of complementarity between questions and answers. Using an iterative process, application 102 provides training pairs to rhetorical match classifier 120 and receives levels of complementarity from the model. Application 102 calculates a loss function by determining the difference between the determined level of complementarity and the expected level of complementarity for a particular training pair. Based on the loss function, application 102 adjusts the internal parameters of the classification model to minimize the loss function.

Acceptance criteria may vary depending on the application. For example, acceptance criteria may be low for community question-answers, automated question-answers, automated customer support systems, manual customer support systems, social network communications, and writings by individuals, such as consumers, about their experiences with the product, such as surveys and complaints. RR acceptance criteria are likely to be high in scientific and technical documents, professional journals, health and legal documents in FAQ format, and professional social networks such as "stackoverflow".

Communicative discourse tree (CDT)
Application 102 can create, analyze, and compare discourse trees for communication. Communicative discourse trees are designed to combine rhetorical information with speech action structure. The CDT contains arcs labeled with expressions for communicative behavior. By combining communicative behavior, CDT allows modeling of RST relationships and communicative behavior. The CDT is the epitome of Peirce intersection. See Galitsky, B. Ilvovsky, D. Kuznetsov SO, Rhetoric Map of an Answer to Compound Queries Knowledge Trail Inc. ACL 2015, 681-686 (Galitsky (2015)). Speech trees can be learned over a richer set of features than the syntax and proper rhetorical relations of the Basic Discourse Units (EDUs).

In one example, a dispute between three parties regarding the cause of the downing of a commercial airliner, Malaysia Airlines Flight 17, is analyzed. An RST representation of the argument being exchanged is constructed. In this example, three competing agents, Dutch Investigators, the Russian Federation Commission of Inquiry and the self-proclaimed Donetsk People's Republic, are exchanging their views on the matter. In the contentious conflict that this example illustrates, each party may all blame their respective counterparts. To appear more persuasive, each party formulates its response not only to make its own case, but to reject the other party's. To this end, each party attempts to match the other's style and discourse.

FIG. 11 shows a communicative discourse tree for a first agent's claims, according to one aspect. FIG. 11 shows a communicative discourse tree 100 representing the following text. “Dutch accident investigators say that evidence points to pro-Russian rebels as being responsible for shooting down plane. The report indicates where the missile was fired from and identifies who was in control of the territory and pins the downing of MH17 on the pro-Russian rebels. Identifying whether they were in control of the territory and blaming pro-Russian insurgents for shooting down MH17."
As can be seen from FIG. 11, the non-terminal nodes of the CDT are rhetorical relations, and the terminal nodes are the basic discourse units (phrases, sentence fragments) that are the subject of these relations. Several arcs of the CDT are labeled with representations about communicative behaviors, including agents who are actors and the subject of these behaviors (what is being communicated). For example, the kernel node for detailed relations (on the left) is labeled say (Dutch, evidence) and the satellites are labeled responsible (rebels, shooting down). These labels are not intended to represent that the subject of the EDU is evidence and shooting down, but are intended to match this CDT with others for the purpose of finding similarities between this CDT and others. In this case, simply linking these communicative behaviors by rhetorical relations rather than providing information in communicative discourse is too restrictive to express structure about what is being communicated and how. The requirement for RR pairs to have the same rhetorical relation or adjusted rhetorical relation is too weak, requiring matching CDT labels for arcs in addition to matching nodes.

The straight edges of this graph are syntactic relations, and the curved arcs are discourse relations such as anaphora, same-entity, sub-entity, rhetorical relation and communicative behavior. This graph contains much richer information than just combining partrees for individual sentences. In addition to CDTs, parse intersections can be generalized at the level of words, relationships, phrases and sentences. Speech actions are logical predicates that represent the agents involved in each speech action and their subject. Arguments for logical predicates are formed according to their semantic roles, as suggested by frameworks such as VerbNet. See "A Large-scale Classification of English Verbs" by Karin Kipper, Anna Korhonen, Neville Ryant and Martha Palmer (Language Resources and Evaluation Journal, 42(1), 21-40, Springer Netherland, 2008). and/or See "VerbNet overview, extensions, mappings and apps" by Karin Kipper Schuler, Anna Korhonen, and Susan W. Brown (Tutorial, NAACL-HLT: 2009, Boulder, Colorado).

FIG. 12 shows a communicative discourse tree for a second agent's claims, according to one aspect. FIG. 12 shows a communicative discourse tree 1200 representing the following text. "The Investigative Committee of the Russian Federation believes that the plane was hit by a missile, which was not produced in Russia. The committee cites an investigation that established the type of the missile."
FIG. 13 shows a communicative discourse tree for a third agent's claims, according to one aspect. FIG. 13 shows a communicative discourse tree 1300 representing the following text. “Rebels, the self-proclaimed Donetsk People's Republic, deny that they controlled the territory from which the missile was allegedly fired. It became possible only after three months after the tragedy to say if rebels controlled one or another town. It could only be published after menstruation."
As can be seen from the communication discourse trees 1100-1300, responses are not arbitrary. The response mentions the same entity as the original text. For example, communication discourse trees 1200 and 1300 are related to communication discourse tree 1100 . The responses confirm discrepancies with assumptions and feelings about these entities and about the behavior of these entities.

More specifically, the involved agent's reply should reflect the communicative discourse of the first seed message. As a simple observation, as the first agent uses attributes to convey its claims, the other agents either provide their own attributes or attack the validity of the advocates' attributes, or do both, according to the set. Each CDT pair can be studied to capture a wide variety of characteristics of how the communicative structure of the seed message should be preserved in successive messages.

Discourse relations or speech behaviors (communicative behaviors) alone are often insufficient to verify request-response matching. As can be seen from the examples shown in FIGS. 11-13, the discourse structure and types of interactions between agents are useful. However, it is not necessary to analyze the domain of dialogue (eg military conflict or politics) or the subject matter (ie entities) of these dialogues.

Representing Rhetorical Relationships and Communication Behavior Two approaches are often used to compute similarities between abstract structures. (1) represent these structures in numerical space and represent similarities as numbers (statistical learning approaches); or (2) use structural representations such as trees and graphs rather than numerical spaces to represent similarities as the largest common substructure. Representing similarity as the largest common substructure is called generalization.

Learning communicative behavior aids in the presentation and understanding of arguments. Computational verb lexicons support the retrieval of entities for actions and help provide a rule-based format for denoting their meaning. Verbs represent the semantics of the event being described as well as the relational information between the participants in that event, projecting syntactic structures that encode that information. Verbs, especially communicative action verbs, are highly variable and can display a rich range of semantic behaviors. Accordingly, verb taxonomy helps the learning system to address this complexity by organizing verbs that share core semantic properties into groups.

VerbNet is one such lexicon that identifies the semantic roles and syntactic pattern features of verbs within each class, and makes explicit the connections between syntactic patterns and the underlying semantic relationships that can be inferred for all members of the class. See "Language Resources and Evaluation" by Karin Kipper, Anna Korhonen, Neville Ryant and Martha Palmer, Vol. 42, No. 1 (March 2008) 21). Each syntactic frame or verb signature for a class has a corresponding semantic expression that details the semantic relationships between event participants over the course of the event.

For example, the verb "amuse" is part of a cluster of similar verbs with similar structures of argumentation (semantic role) such as amaze, anger, arouse, disturb, irritate. The demonstrative roles of these communicative behaviors are Experiencer (usually a living entity), Stimulus and Result. Each verb can have a semantic class that is distinguished by syntactic characteristics of how the verb appears within a sentence or frame. For example, the frame for "amuse" uses the following main noun phrase (NP), noun (N), communicative action (V), verb phrase (VP), adverb (ADV):

NP V NP. For example: "The teacher amused the children". Syntax: Stimulus V Experiencer. Clauses: amuse(Stimulus, E, Emotion, Experiencer), cause(Stimulus, E), emotional_state(Result(E), Emotion, Experiencer).

NP V ADV-Middle: Example: "Small children amuse quickly". Syntax: Experiencer V ADV. Clauses: amuse(Experiencer, Prop):-, property(Experiencer, Prop), adv(Prop).

NP V NP-PRO-ARB. Example "The teacher amused". Syntax: Stimulus V. amuse(Stimulus, E, Emotion, Experiencer): cause(Stimulus, E), emotional_state(Result(E), Emotion, Experiencer).

NP cause V NP. Example "The teacher's dolls amused the children". Syntax: Stimulus <+genitive>('s) V Experiencer. amuse(Stimulus, E, Emotion, Experiencer): cause(Stimulus, E), emotional_state(during(E), Emotion, Experiencer).

NP V NP ADJ. For example "This performance bored me totally". Syntax: Stimulus V Experiencer Result. amuse (Stimulus, E, Emotion, Experiencer). cause(Stimulus,E), emotional_state(result(E), Emotion, Experiencer), Pred(result(E), Experiencer).

Communication behavior can be characterized into multiple clusters. For example, verbs with predicative complements (appoint, characterize, dub, declare, conjecture, masquerade, orphan, captain, consider, classify); perceptual verbs (see, sight, peer); mental state verbs (amuse). verbs of desire (want, long); verbs of judgment (judgment); verbs of evaluation (assess, estimate); verbs of exploration (hunt, search, stalk, investigate, rum). verbs of social dialogue (correspond, marry, meet, battle); verbs of communication (transfer(message), inquire, interrogate, tell, speaking, talk, chat, say, comp). lain, advise, confess, lecture, overstate, promise; avoidance verbs; measure verbs (register, cost, fit, price, bill); relative verbs (begin, complete, continue, stop) stop), establish, sustain).

The aspects described herein have advantages over statistical learning models. In contrast to statistical solutions, aspects using classification systems can provide verbs or verb-like structures that are determined to yield desired characteristics (such as rhetorical matching). For example, statistical machine learning models represent similarities as numbers, which can make interpretation difficult.

Representing Request-Response Pairs Representing request-response pairs facilitates taxonomy-based operations based on certain pairs. In one example, a request-response pair can be represented as a parse intersection. A parse intersection is a representation of a parse tree for two or more sentences, using the discourse-level relationships between the words and parts of the sentences in a graph. See Galitsky (2015). Topic similarities between questions and answers can be represented as a common subgraph of perspective intersections. The higher the number of common graph nodes, the higher the similarity.

FIG. 14 illustrates Peirce intersection according to one aspect. FIG. 14 shows a parse intersection 1400 that includes a parsetree for a request 1401 and a corresponding parsetree for a response 1402 .

The question represented by the parse tree 1401 is as follows. "I just had a baby and it looks more like the husband I had my baby with. However it does not look like me at all and I am scared that he was cheating on me with another lady and I had her kid. This child is the best thing that has ever happened to me and I cannot imagine giving my baby to the real mom."I'm afraid she's taking care of her child, which is the best thing that's ever happened to me, and I can't imagine giving a baby to a real mother."
The response represented by response 1402 is as follows. “Marital therapists advise on dealing with a child being born from an affair as follows. One option is for the husband to avoid contact but just have the basic legal and financial commitments. Another option is to have the wife fully involved and have the baby fully integrated into the family just like a child from a previous marriage. option is to get the wife fully involved so that the baby is fully integrated into the family, like a child from a previous marriage."

FIG. 14 represents a greedy approach to representing linguistic information about a paragraph of text. The straight edges of this graph are syntactic relations, and the curved arcs are discourse relations such as anaphora, same-entity, sub-entity, rhetorical relation and communicative behavior. Solid arcs are for same-entity/sub-entity/anaphora relationships and dashed arcs are for rhetorical relationships and communicative behaviors. Elliptical labels on straight edges indicate syntactic relationships. The lemma is written in the box for the node, and the lemma form is written to the right of the node.

Parse intersection 1400 contains much richer information than just combining parse trees for individual sentences. Navigation through this graph along edges for syntactic relations and arcs for discourse relations allows a given parse intersection to be transformed into a semantically equivalent form for matching with other parse intersections to perform text similarity evaluation tasks. As many links as possible are presented to form a complete formal representation of the paragraph. Each discourse arc produces a pair of parse phrases that can be potential matches.

Topic similarities between seeds (requests) and responses are represented as common subgraphs of perspective intersections. These are visualized as connected clouds. The higher the number of common graph nodes, the higher the similarity. For rhetorical matching, common subgraphs do not need to grow if present in a given text. However, the rhetorical relationship and communicative behavior of seeds and responses are interrelated and require correspondence.

Generalizations on Communicative Behaviors The similarity between two communicative behaviors A ₁ and A ₂ is defined as abstract verbs possessing features that are common between A ₁ and A ₂ . Defining the similarity of two verbs as an abstract verb-like structure supports inductive learning tasks such as assessing rhetorical agreement. In one example, the similarity between two common verbs agree and disagree can be generalized as follows. agree ^ disagree = verb(Interlocutor, Proposed_action, Speaker). In this case, the Interlocutor is the person who proposed the Proposed_action to the Speaker to whom the Speaker is communicating his response. In addition, proposed_action is the action that the Speaker would perform if the request or proposal were to be accepted or rejected, and the Speaker is the person to whom the particular action is proposed and the person who responds to the request or proposal made.

In yet another example, the similarity between the verbs agree and explain is: agree ^ explain = verb(Interlocutor, *, Speaker). The subject of communicative behavior is generalized in the context of communicative behavior, but not in other "physical" behaviors. Aspects thus generalize individual occurrences of communicative behavior with corresponding subjects.

Additionally, a sequence of communicative actions representing a dialog can be compared to other such sequences of similar dialogs. In this manner, the dynamic discourse structure of the dialogue as well as the meaning of individual communicative behaviors are represented (as opposed to their static structure reflected by rhetorical relations). Generalization is a complex structural expression that occurs at each level. The lemma of communicative behavior is generalized in the lemma, whose semantic roles are generalized with their respective semantic roles.

Communicative behavior is used by text authors to indicate the structure of a dialogue or conflict. See "Speech acts: an essay in the philosophy of language" by Searle, J. R. (1969), Cambridge University Press. The subject is generalized in the context of these actions and not other "physical" actions. Individual occurrences of communicative behavior are thus generalized as discourse 'steps' in their subjects and also in their pairs.

Generalization of communication behavior can also be considered from the viewpoint of matching with verb frames such as VerbNet. Communicative links reflect discourse structures that are related by participation (or mention) rather than by a single agent in the text. A link forms a sequence that connects words (a verb or a number of words that imply a person's communicative intent) of a communicative action.

A communicative behavior includes an actor, one or more agents working with those actors, and a phrase that describes the characteristics of this behavior. Communicative behavior can be described as a function of form: verb (agent, subject, cause). In this case, the verb characterizes some type of interaction between the agents involved (e.g., explain, confirm, remind, disagree, deny, etc.). Subject refers to the information conveyed or the object of the statement. Cause refers to the motivation or explanation for the subject.

Scenarios (labeled directed graphs) are subgraphs of Perth intersection G=(V,A). where V={action ₁ , action _{2 ,} .

Each arc action _i , action _j εA _sequence corresponds to two actions v _i , ag _i , s _i , c _i and v _j , ag _j , s _j , c _j that refer to the same subject (eg s _j =s _i or different themes). Each arc action _i , action _j εA _cause corresponds to an attack relationship between action _i and action _j indicating that the cause of action _i is in conflict with the subject or cause of action _j .

The perspective intersection subgraphs associated with interaction scenarios between agents have several salient features. For example: (1) all vertices are time-ordered, so there is one incoming arc and one outgoing arc for every vertex (except for the initial and terminal vertices). (2) For A _sequence arcs, at most only one incoming and one outgoing arc is allowed. (3) For A _cause arcs, there may be many incoming arcs as well as many outgoing arcs from a given vertex. The vertices involved may be associated with different agents or the same agent (ie if this agent contradicts itself). To compute the similarity between perspective intersections and their communication behavior, the induced subgraphs, the subgraphs with similar labels and the same configuration of arcs and the strict correspondences of vertices are analyzed.

The following similarities exist by analyzing the communicative behavioral arcs of Peirce crossovers. (1) one topic communication action from T1 against another topic communication action from T2 (communication action arc is not used) and (2) one pair of topic communication action from T1 (communication action arc is used) compared to another pair of communication action from T2.

A generalization of two different communicative behaviors is based on their attributes. See Galitsky et al. (2013). 14, one communicative action from T1, namely cheating(husband, wife, another lady), can be compared with a second communicative action from T2, namely avoid(husband, contact(husband, another lady). The generalization results in communicative_action(husband,*), whereby if a given agent (=husband) is mentioned in Q as the subject of CA, then he (she) is (possibly A constraint of the form that it should also be the subject of CA is introduced for A. Two communicative behaviors can always be generalized, but this is not the case with their subjects: if the result of their generalization is empty, then the result of generalizing the communicative behavior with these subjects is also empty.

Generalization of RST Relations Some relations between discourse trees can be generalized, such as arcs representing the same type of relations (presentation relations such as contrasts, subject relations such as conditions, and polynuclear relations such as lists). A nucleus or a situation denoted by a nucleus is denoted by "N". A satellite or a situation indicated by a satellite is indicated by an 'S'. "W" indicates the writer. "R" indicates a reader (listener). Contexts are proposals, completed or ongoing actions, and communicative actions and states (including beliefs, desires, approve, explain, reconcile, etc.). A generalization of the two RST relations with the above parameters is expressed as follows:
rst1(N1,S1,W1,R1)^rst2(N2,S2,W2,R2)=
(rst1^rst2)(N1^N2,S1^S2,W1^W2,R1^R2).
The text in N1, S1, W1, R1 is generalized as phrases. For example, rst1 ^ rst2 can be generalized to (1) The generalization is empty if relation_type(rst1) != relation_type(rst2). (2) Otherwise, rhetorical signatures are generalized as sentences.

sentence(N1,S1,W1,R1) ^ sentence(N2,S2,W2,R2)
See Iruskieta, Mikel, Iria da Cunha and Maite Taboada, "A qualitative comparison method for rhetorical structures: identifying different discourse structures in multilingual corpora" (Lang Resources & Evaluation. June 2015, Volume 49, Issue 2).

For example, the meaning of rst-background ^ rst-enablement= (S increases the ability of R to comprehend an element in N) ^ (R comprehending S increases the ability of R to perform the action in N) = increase-VB the-DT ability-NN of-IN R-NN to-IN.

Also, since the rst-background ^ rst-enablement relationship varies, the RST relationship part is empty. Expressions that are verbal definitions of each RST relation are then generalized. For example, for each word, or for a placeholder for a word such as agent, this word (with its POS) is retained if the word is the same in each input phrase, and excluded if the word is different between these phrases. The resulting representation can be interpreted as a common meaning between two different RST relation definitions that are formally obtained.

The two arcs between the question and answer shown in FIG. 14 indicate generalized instances based on the RST relation "RST-contrast". For example, "I just had a baby" relates to the RST-contrast to "it does not look like me", which in turn relates to the RST-contrast to "have the basic legal and financial commitments", "husband to avoid contact". As can be seen from the above, the answer need not resemble the verb phrase of the question, but the rhetorical structure of the question and answer are similar. Not every phrase in the answer necessarily matches the phrase in the question. For example, unmatched phrases have a particular rhetorical relationship with phrases in the answer that are related to phrases in the question.

Building a Communicative Discourse Tree FIG. 15 illustrates an exemplary process for building a communicative discourse tree, according to one aspect. Application 102 may implement process 1500 . As noted above, the discourse tree for communication enables improved search engine results.

At block 1501, process 1500 includes accessing a sentence containing a fragment. At least one fragment includes a verb and multiple words. Each word contains multiple word roles within the fragment. Each fragment is a basic discourse unit. For example, application 102 accesses a sentence such as "Rebels, the self-proclaimed Donetsk People's Republic, deny that they controlled the territory from which the missile was allegedly fired," as described in connection with FIG.

Continuing with the example above, application 102 determines that the sentence contains several fragments. For example, the first fragment is "Rebels, ..., deny". The second fragment is "that they controlled the territory". The third fragment is "from which the missile was allegedly fired". Each fragment contains a verb (eg, "deny" for the first fragment and "controlled" for the second fragment). However, fragments need not contain verbs.

At block 1502, process 1500 includes generating a discourse tree representing rhetorical relationships between sentence fragments. A discourse tree contains multiple nodes. Each non-terminal node represents a rhetorical relationship between two of the sentence fragments, and each terminal node of the nodes of the discourse tree is associated with one of the sentence fragments.

Continuing with the example above, application 102 generates a discourse tree as shown in FIG. For example, the third fragment, "from which the missile was allegedly fired," details "that they controlled the territory." Both the second and third fragments relate to the attributes of what happened (ie, the attack could not have been an insurgent because the insurgent did not take over the area).

At block 1503, process 1500 includes accessing multiple verb signatures. For example, application 102 accesses a list of verbs, eg from VerbNet. Each verb matches or is related to the verb of the fragment. For example, for the first fragment the verb is "deny". Accordingly, application 102 accesses a list of verb signatures related to the verb "deny."

As noted above, each verb signature includes one or more of the fragment's verbs and thematic roles. For example, the signature includes one or more of noun phrases (NP), nouns (N), communicative behaviors (V), verb phrases (VP) or adverbs (ADV). Thematic roles describe the relationship between verbs and related words. For example, "The teacher amused the children" has a different signature than "small children amuse quickly". For the first fragment, the verb "deny", application 102 accesses the list of frames or verb signatures for verbs that match "deny". The list is "NP V NP to be NP", "NP V that S", and "NP V NP".

Each verb signature contains a thematic role. Thematic role refers to the role of the verb in the sentence fragment. Application 102 determines the thematic role in each verb signature. Exemplary subject roles are "actor", "agent", "asset", "attribute", "beneficiary", "cause", "location destination source", "destination", "source", "location", "experiencer", "extent", "instrument". ), "material and product", "material", "product", "patient", "predicate", "recipient", "stimulus", "theme", "time" or "topic".

At block 1504, the process 1500 includes, for each verb signature of the verb signatures, determining the number of thematic roles of each signature that match the word roles in the fragment. For the first fragment, the rhetorical classification application 102 determines that the verb "deny" has only three roles: "agent", "verb" and "theme".

At block 1505, process 1500 includes selecting a particular verb signature from the verb signatures based on the particular verb signature with the highest number of matches. For example, referring back to FIG. 13, "deny" in the first fragment "the rebels deny that they controlled the territory" matches the verb signature deny "NP V NP" and "control" matches control(rebels, territory). The verb signatures are nested, resulting in a nested signature of "deny(rebel, control(rebel, territory))".

Request-Response Representation Request-response pairs can be analyzed singly or as pairs. In one example, request-response pairs can be strung together. Upon splicing, rhetorical agreement is expected to be preserved not only between consecutive members, but also between tri-members and members of a 4-tuple. A discourse tree can be constructed for text that represents a sequence of request-response pairs. For example, in the domain of customer complaints, the request and response are in the same text from the point of view of the complaining customer. The customer complaint text can be split into request and response text parts, which can then form pairs of positive and negative data sets. In one example, all texts about supporters and all texts about opponents are combined. The first sentence of each paragraph below will form the request portion (which will contain three sentences) and the second sentence of each paragraph will form the response portion (which in this example will contain three sentences).

FIG. 16 shows discourse trees and scenario graphs, according to one aspect. FIG. 16 shows discourse tree 1601 and scenario graph 1602 . Discourse tree 1601 corresponds to the following three sentences.

(1) I explained that my check (written after I made the deposit) had been dishonored. A customer service representative accepted that it usually takes some time to process the deposit.

(2) I recall being unfairly charged an overdraft fee in a similar situation a month ago. They denied that it was unfair because the overdraft fee was disclosed in my account information.

(3) I disagreed with their charge and wanted this charge returned to my account. They explained that nothing can be done at this point and that I disagreed with their fee and wanted this fee deposited back to my account. They explained that nothing can be done at this point and that I need to look into the account rules closer.

As can be seen from the discourse tree in FIG. 16, it can be difficult to determine if the text represents dialogue or explanation. Therefore, implicit similarities between texts can be discovered by analyzing the arcs of communicative behavior of Perth crossovers. For example, in general conditions,
(1) One communicative action with a subject from the first tree against another communicative action with a subject from the second tree (communication action arcs are not used).

(2) one pair of communicative behaviors with themes from the first tree (the arc of the communicative behavior is used) against another pair of communicative behaviors from the second tree;

For example, in the above example, the generalization of cheating(husband, wife, another lady) ^ avoid(husband, contact(husband, another lady)) gives us communicative_action(husband, *), which introduces a constraint on A of the form that if a given agent (=husband) is mentioned as the subject of a CA in Q, he (she) should also be the subject of a (possibly another) CA in A.

A word can be applied to a vector model, such as the "word2vector" model, to handle the meaning of words that represent the subject of CA. More specifically, the following rules can be used to compute cross-subject generalizations of communication behavior: If subject1=subject2 then subject1^subject2 = <subject1, POS(subject1), 1>. In this case the subject remains and the score is 1. Otherwise, subject1^subject2 = <*, POS(subject1), word2vecDistance(subject1^subject2)> if the subjects have the same part-of-speech (POS). Here the '*' indicates that the lemma is a placeholder and the score is the word2vec distance between these words. If the POS are different, the generalization is an empty tuple and may not generalize further.

Classification Settings for Request-Response Pairs In traditional searches, as a baseline, match between request-response pairs can be measured in terms of keyword statistics such as TF*IDF, which stands for term frequency-inverse document frequency. To improve search relevance, this score is augmented by item popularity, item location or taxonomy-based scores (Galitsky (2015)). Search can also be formulated as a path re-ranking problem in a machine learning framework. A feature space contains request-response pairs as elements, and a separating hyperplane divides this feature space into exact and imprecise pairs. Thus, the search problem can be formulated in a local way as the similarity between Req and Resp or in a global learning way as the similarity between request-response pairs.

Other methods for determining matches between requests and responses are also possible. In a first example, application 102 extracts features for Req and Resp, compares these features as counts, and introduces a scoring function so that scores can indicate class (low score for incorrect pairs, high score for correct pairs).

In a second example, application 102 compares expressions for Req and Resp against each other and assigns a score for the comparison result. Similarly, the score would indicate class.

In a third example, application 102 constructs expressions for pairs Req and Resp<Req, Resp> as elements of the training set. Application 102 then performs learning in the feature space of all such elements <Req, Resp>.

FIG. 17 illustrates forming a request-response pair, according to one aspect. FIG. 17 shows request-response pairs 1701 , request trees (or objects) 1702 and response trees 1703 . To form the object of <Req, Resp>, application 102 combines the discourse tree for the request and the discourse tree for the response with the root RR into a single tree. Application 102 then classifies the objects into exact (high match) and inexact (low match) categories.

Nearest Neighbor Graph Based Classification Once the CDT is constructed, to identify arguments in the text, the application 102 computes the similarity compared to the CDT for the positive class and verifies that it is low enough to reach the CDT set for the negative class. Similarity between CDTs is defined by the largest common sub-CDT.

In one example,

An ordered set G of CDT(V,E) is constructed with vertex and edge labels from . The labeled CDT Γ from G is a pair of the multipaired form ((V, l), (E, b)). where V is a set of vertices, E is a set of edges,

is a function that assigns labels to vertices, and b:E→Λ _E is a function that assigns labels to edges. Isomorphic trees with identical labels are not identified.

The order is defined as follows. For two CDTs, from G we have Γ ₁ :=((V ₁ ,l ₁ ),(E ₁ ,b ₁ )) and Γ ₂ :=((V ₂ ,l ₂ ),((E ₂ ,b ₂ )) , and that Γ ₁ dominates Γ ₂ or Γ ₂ ≦Γ ₁ if there exists a one-to-one mapping φ:V ₂ →V ₁ (or , Γ ₂ is a sub-CDT of Γ ₁ ), so this is

(2)

will be compatible under
This definition allows for the computation of label similarity (“weakening”) between matched vertices when going from a “larger” CDT G ₁ to a “smaller” CDT G ₂ .

where the similarity CDT Z of a pair of CDT X and CDT Y denoted by X^Y=Z is the set of all containing-maximum common sub-CDTs of X and Y, each of which satisfies the following additional conditions for matching: Specifically, (1) two vertices from CDT X and CDT Y must exhibit the same RST relation. (2) Each common sub-CDT from Z contains at least one communication behavior with the same VerbNet signature as in X and Y;

This definition is easily extended so that some graph generalizations can be found. The order of subsumption for pairs of graphsets X and Y is naturally defined as XμY:=X*Y=X.

FIG. 18 shows a discourse tree for maximal common subcommunications, according to one aspect. Note that the tree is inverted and the arc labels are generalized. The communication behavior site() is generalized with the communication behavior say(). The CA Commission's first (agent) argument for the former is generalized in the CA Dutch's first argument for the latter. The same behavior applies to the second evidence for this pair of CAs: investigator^evidence.

CDT U belongs to the positive class, so that (1) U is similar (with non-empty common sub-CDT) in positive examples R ⁺ and (2) for any negative example R ⁻ , if U

, then U*R ⁻ μU*R ⁺ .
This condition introduces a similarity criterion, stating that in order to be assigned to a class, the similarity between an unknown CDT U and the closest CDT from the positive class should be higher than the similarity between U and each negative example. Condition 2 implies that if there is a positive example R ⁺ and therefore there is no R ⁻ to do, then we have U*R ⁺ μR ⁻ , i.e. there is no counterexample to this generalization of positive examples.

Interlace Kernel Learning for CDT Tree kernel learning for strings, partrees and parse intersections is a recently established research area. The partree kernel counts the number of common subtrees as a measure of discourse similarity between two instances. The tree kernel is the "Discriminative Reranking of Discourse Parses Using Tree Kernels" (Proceedings of EMNLP; 2014) specified for DT by Joty, Shafiq and A. Moschitti. See also Wang, W., Su, J., & Tan, CL (2010), "Kernel Based Discourse Relation Recognition with Temporal Ordering Information" (In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics). (We are using a special form of tree kernel for discursive relation recognition). A crossover kernel is defined for the CDT by augmenting the DT kernel with information about communication behavior.

A CDT can be represented by a vector V of integer counts of each subtree type (without taking into account its ancestors):
V(T)=(# of type 1 subtree, . . . , # of type I subtree, . . . , # of type n subtree). This results in very high dimensionality, since the number of different subtrees is exponential in size. So the feature vector

directly using is computationally infeasible. To solve the computational problem, we introduce a tree kernel function to efficiently compute the dot product between the above high-dimensional vectors. Assuming there are two tree segments CDT1 and CDT2, the tree kernel function is defined as follows:
K(CDT1, CDT2)=<V(CDT1), V(CDT2)>=ΣiV(CDT1)[i], V(CDT2)[i]=Σn1Σn2 Σi Ii(n1) ^* Ii(n2)
In this case n1εN1 and n2εN2, where N1 and N2 are the set of all nodes in CDT1 and CDT2, respectively.

Ii(n) is the indicator function.
Ii(n)={1 if a subtree of type i occurs at the root at the node, 0 otherwise}. K(CDT1, CDT2) is an instance of the convolution kernel for the tree structure (Collins and Duffy; 2002) and can be computed by a recursive definition.

Δ(n1,n2)=ΣIIi(n1) ^* Ii(n2)
where Δ(n1,n2)=0 if n1 and n2 are assigned the same POS tag or their children are different subtrees.

Otherwise, if n1 and n2 are both POS tags (pre-terminal nodes), then Δ(n1, n2)=1×λ.
Otherwise:

In this case, ch(n,j) is the jth child of node n, nc(n ₁ ) is the number of children of n ₁ , and λ(0<λ<1) is a corrosion factor to keep the kernel value from varying too much with subtree size. Additionally, the recursive rule (3) is retained. This is because, given two nodes of the same children, these children can be used to build a common subtree, as well as a common subtree of further descendants. The partree kernel counts the number of common subtrees as a syntactic similarity measure between two instances.

FIG. 19 shows trees in a kernel training format for discourse trees for communication, according to one aspect.

Terms on communicative behavior as labels are transformed into a tree that is added to each node on RST relations. For the text of EDUs as labels for terminal nodes, only the phrase structure is preserved. Terminal nodes are labeled with a sequence of phrase types rather than parse tree fragments.

If there is a rhetorical relational arc from node X to the terminal EDU node Y, labeled A(B,C(D)), then the subtree AB→(CD) is attached to X.

Implementation of the Rhetorical Match Classifier The rhetorical match classifier 120 can determine complementarity between two sentences, such as questions and answers, by using a communicative discourse tree. FIG. 20 illustrates an exemplary process used to implement a rhetorical match classifier, according to one aspect. FIG. 20 shows a process 2000 that may be implemented by application 102 . As described above, rhetorical matching classifier 120 is trained with training data 125 .

A rhetorical match classifier 120 determines a communicative discourse tree for both questions and answers. For example, rhetorical match classifier 120 builds a discourse tree for question communication from a question, such as question 171 or input text 130, and a discourse tree for answer communication from candidate answers.

At block 2001, process 2000 includes determining, for a question sentence, a discourse tree for question communication that includes a question root node. A question sentence can be a specific question, request or comment. Application 102 creates discourse tree 110 for question communication from input text 130 . Using the example described in connection with Figures 13 and 15, an exemplary question sentence is "are rebels responsible for the downing of the flight". Application 102 may use process 1500 described in connection with FIG. The exemplary question has a root node of "elaborate".

At block 2002, process 2000 includes determining a second communicative discourse tree for the answer sentence. In this case, the discourse tree for reply communication includes the reply root node. Continuing with the example above, application 102 creates a communication discourse tree 111 that also has a root node "explain", as shown in FIG.

At block 2003, process 2000 includes associating communication discourse trees by identifying that the question root node and the answer root node are the same. Application 102 determines that the discourse tree for question communication and the discourse tree for answer communication have the same root node. The resulting associated discourse tree for communication is shown in Figure 17 and can be labeled as "request-response pairs".

At block 2004, process 2000 includes computing the level of complementarity between the discourse trees for question communication and the discourse trees for answer communication by applying a predictive model to the merged discourse trees.

The rhetorical match classifier uses machine learning techniques. In one aspect, application 102 trains and uses rhetorical match classifier 120 . For example, application 102 defines a positive class and a negative class of request-response pairs. The positive class contains rhetorically exact request-response pairs, and the negative class contains related but rhetorically disparate request-response pairs.

For each request-response pair, application 102 constructs a CDT by parsing each sentence and obtaining verb signatures for sentence fragments.

Application 102 provides associated communicative discourse tree pairs to rhetorical match classifier 120 . The rhetorical match classifier 120 outputs the level of complementarity.

At block 2005, process 2000 includes identifying the question and answer sentences as complementary in response to determining that the level of complementarity exceeds the threshold. Application 102 can use the threshold level of complementarity to determine whether question-answer pairs are sufficiently complementary. For example, application 102 may output the answer as answer 172 or answer 150 if the classification score is greater than the threshold. Alternatively, application 102 may discard the answer, access answer database 105 or another public database for candidate answers, and repeat process 2000 as necessary.

In one aspect, application 102 obtains a co-reference. In a further aspect, application 102 obtains entities and subentities or hyponym links. Narrower terms are words that have a more specific meaning than the general terms or broader terms applicable to the word. For example, "spoon" is a narrower term for "cutlery."

In another aspect, application 102 applies interlaced kernel learning to the representation. Interlaced kernel learning may be performed in block 2004 instead of classification-based learning, for example. Application 102 builds parse interlaced pairs for a partree of request-response pairs. Application 102 parses the discourse to obtain discourse tree pairs for request-response pairs. Application 102 aligns the basic discourse units of discourse-tree request-response and partree request-response. The application 102 merges the basic discourse units of the discourse-tree request-response and the basic discourse units of the partree request-response.

In one aspect, the application 102 improves text similarity assessment with the word2vector model.

In a further aspect, application 102 sends sentences corresponding to discourse trees for question communication or sentences corresponding to discourse trees for answer communication to a device, such as mobile device 170 . Output from application 102 can be used as input to search queries, database lookups, or other systems. In this manner, application 102 can be integrated with a search engine system.

FIG. 21 illustrates a chatbot commenting on posts, according to one aspect. FIG. 21 shows chat 2100, user messages 2101-2104 and agent responses 2105. FIG. Agent response 2105 can be implemented by application 102 . As shown, agent response 2105 identified an appropriate answer to the thread of user messages 2101-2104.

FIG. 22 illustrates a chatbot commenting on posts, according to one aspect. FIG. 22 shows chat 2200, user messages 2201-2205 and agent responses 2206. FIG. FIG. 22 shows three messages from User 1, specifically 2201, 2203 and 2205, and two messages from User 2, specifically 2202 and 2204. FIG. Agent response 2206 can be implemented by application 102 . As shown, agent response 2106 identified an appropriate answer for the thread of messages 2201-2204.

The features shown in FIGS. 21 and 22 can be implemented by computing device 101 or by an apparatus that provides input text 130 to computing device 101 and receives answers 150 from computing device 101.

Additional Rules for RR Matches and RR Irrationality The following are examples of structural rules that introduce constraints to enforce RR matches:
1. Both Req and Resp have the same emotional polarity (if the request is positive, the response should be positive as well, and vice versa).

2. Both Req and Resp have a logical argument.
Under reasonable reasoning, the request and response would match perfectly. A rational agent will present an answer that will be both relevant and consistent with the rhetorical expression of the question. In the real world, however, not all responses are sufficiently rational. Cognitive bias research groups examine people's propensity to think in certain ways that may systematically deviate from their standards of rationality or good judgment.

Correspondence bias is the tendency of people to overemphasize personality-based explanations of observed behavior of others other than themselves when responding to questions. See Baumeister, R. F. & Bushman, B. J., Social psychology and human nature (International Edition: 2010). At the same time, people responding to queries underemphasize the role and power of contextual influence on the same behavior.

Confirmation bias is the tendency to seek or interpret information in a way that confirms the preconceived notions of those answering the question. They may distrust information that does not support their views. Confirmation bias is related to the concept of cognitive dissonance. This may allow individuals to reduce conflicts by seeking out information that reaffirms their views.

Anchoring leads to relying too heavily on or "anchoring" one property or piece of information in making a decision.

The availability heuristic causes us to overestimate the likelihood of an event occurring within a memory with a greater 'availability', which can be affected by how recent the memory is or how it can be abnormally or emotionally altered.

According to the bandwagon effect, people answer questions by believing that many other people will do (or believe) the same thing.

Belief bias is the effect in which someone's assessment of the logical strength of an argument is biased by the credibility of the conclusion.

A bias blind spot is the tendency to see oneself less biased toward oneself than toward other people, or the tendency to be able to identify others with a higher cognitive bias than oneself.

Evaluation The primary domain for test data is Yahoo! Derived from Question-Answer Pairs from Answers, which is a set of questions and answers paired on a wide range of topics. Of the set of questions of 4.4 million users, 20,000 questions containing more than 2 sentences were selected. No filtering was applied to the answers, as the answers to the majority of the questions were properly detailed. There are multiple answers per question and the best answer is marked. We consider question-best-answer pairs as members of the positive training set and question-other-answers as members of the negative training set. To derive the negative set, we randomly selected answers to different but somewhat related questions, or formed queries from the questions and obtained answers from web search results.

Our second dataset includes social media. We extracted request-response pairs primarily from Facebook posts. We also used smaller portions of conversations with LinkedIn.com and vk.com related to employment. In the social domain, the standard of writing is quite low. Text cohesiveness is very limited and often lacks logical structure and relevance. Over the years, the author has formed a training set from his own account and public Facebook accounts available via the API (at the time of writing writing the Facebook API to get the messages is not possible). Additionally, we used 860 email threads from the Enron dataset. In addition, we collected data on manual responses to posts by agents that automatically generate posts on behalf of human user hosts. See Galitsky B., Dmitri Ilvovsky, Nina Lebedeva and Daniel Usikov, Improving Trust in Automation of Social Promotion: AAAI Spring Symposium on The Intersection of Robust Intelligence and Trust in Autonomous Systems; Stanford, CA 2014 (“Galitsky 2014”). Formed 4000 pairs from various social network sources.

A third domain is customer complaints. For typical complaints, dissatisfied customers describe their problems with products and services, as well as the process by which the customer attempted to communicate these problems to the company and how the company responded. Complaints are often written with prejudices such as exaggerating product failures and portraying the other party's behavior as unfair and inappropriate. At the same time, complainants attempt to state their complaints in a persuasive, coherent, and logically consistent manner (“Galitsky 2014”). Complaints thus serve as a domain with a high rate of matching between requests and responses. For the purpose of assessing the match between user complaints and company responses (according to how the user describes them), we collected 670 complaints from planetfeedback.com over ten years.

A fourth domain is interviews by journalists. Professional journalists usually write interviews in such a way that the level of agreement between questions and answers is very high. We also collected over 1200 contributions by professional and citizen journalists from sources such as datran.com, allvoices.com, huffingtonpost.com and others.

We designed a crawler to facilitate data collection. A crawler searches a particular set of sites, downloads web pages, extracts candidate text, and verifies that this candidate text conforms to a response format for a question or request. Each pair of texts is then formed. This search is powered by the Bing Azure Search Engine API in the web and news domains.

Recognition of Valid and Invalid Responses Table 1 shows the accuracy of the response classification. Each row represents a particular method. Each class of methods is shown in the gray area.

It can be seen that the highest accuracy is achieved in the journalism and community response domains and the lowest accuracy is achieved in customer complaints and social social networks. We can conclude that the higher the accuracy with which the method is established, the higher the level of agreement between Req and Resp and, accordingly, the higher the competence of the responder.

The deterministic family of approaches (middle two rows, local RR-similarity-based classification) perform about 9% lower than SVM TK. This indicates that the similarity between Req and Res and p is substantially less important than the specific structure of the RR pair that indicates the RR match. This means that the match between Req and Resp cannot be assessed on an individual basis. If we require DT(Req) to be very similar to DT(Resp), the accuracy we get is reasonable but the recall will be very poor. Even progressing from DT to CDT is only 1% to 2% useful. This is because communicative behavior does not play a major role in constructing requests or shaping responses.

Regarding the statistical family of approaches (bottom five rows, tree kernel), the richest source of discourse data (SVM TK for RR-DT) gives the highest classification accuracy, which is almost the same as RR similarity-based classification. SVM TK for RST and CA (sufficient partrees) contained more linguistic data, but some parts of it (most likely syntactic) were redundant, leading to poorer results on the restricted training set. Using additional features under TK, such as emotion and argument, does not help. Most likely, these features are derived from RR-CDT features and do not contribute to classification accuracy by themselves.

Employing the TK family of CDT-based approaches yields an accuracy comparable to that achieved in classifying DTs as correct and incorrect. In rhetorical parsing tasks, state-of-the-art systems have come under intense competition over the past few years, yielding accuracies in excess of 80%.

The performance of the direct analytical approach in the deterministic family is rather weak. This means that more and more complex feature structures are needed, and simply counting and taking into account types of rhetorical relations is insufficient to determine how well the RRs match each other. If two RR pairs have the same type and count of rhetorical relations and even communicative behavior, they can still belong to opposite RR match classes in the vast majority of cases.

Nearest-neighbor pair learning for CDT is less accurate than SVM TK for CDT, but interesting examples of subtrees given by the former are typical of the discussion and shared among analogous data. The number of former groups of CDT subtrees will naturally be much higher. Unfortunately, the SVM TK approach does not help explain how exactly the RR matching problem is solved, it merely shows the final scoring and class labels. Logical arguments can also be expressed in responses without communicative behavior, but are rare (this observation is supported by our data).

Measuring RR Matches in the Evaluation Domain In terms of recognition accuracy evaluation, we have obtained the optimal method in the previous subsection. Now, having finalized this method, we will measure RR agreement in our evaluation domain. In addition, we will show how the sum of general matches provided by the optimal method correlates with individual match criteria such as sentiment, reasoning, topic and keyword relevance. When we use our optimal approach to label the training set (SVM TK for RR-CDT), its size can grow dramatically and we can explore interesting properties of RR matches in various domains. We will find some intuitive feature contributions to the RR agreement on larger datasets than previous evaluations.

In this subsection, we intend to demonstrate that the RR pair validity recognition framework can serve as a basis for matching between arbitrary requests and responses. Furthermore, this recognition framework can assess how strongly various features are correlated with RR pair effectiveness.

From the evaluation of recognition accuracy, we obtained the optimal method for recognizing whether RR pairs are valid or not. Now, having established this recognition method, we will measure the RR match in our evaluation domain and also infer how the sum of general matches provided by the best method correlates with individual match criteria such as sentiment, reasoning, topic and keyword relevance. When we use our optimal approach to label the training set (SVM TK for RR-CDT), its size can grow dramatically and we can explore interesting properties of RR matches in various domains. We will find the contribution of some intuitive features of the RR matches on a larger dataset than previous evaluations. We will measure this agreement, feature by feature, only on the positive training dataset of the above assessment as recognition accuracy (%, Table 2). Note that recall and negative data sets are not required for match assessment.

For example, we rate the accuracy of the observation that RR pairs determined by matching by topic as computed by the bag-of-words approach are valid RR pairs in the domain of customer complaints according to SVM TK for RR-CDT classification at 64.3%.

Emotional matches indicate the appropriate emotional match contribution in the RR pair. The sentiment rule specifically includes that if the RR polarity is the same, the response should confirm what the request says. Conversely, if the polarity is reversed, the response should attack what the request wants. A match by logical argument requires an appropriate communicative discourse in which the response does not match the assertion being requested.

This data helps elucidate the nature of the linguistic match between what the advocate is saying and how the other party is responding. Not all matching features need to be present for a valid dialog discourse. However, if most of these characteristics are not met, the given answer should be considered invalid and inappropriate and another answer should be selected. Table 2 shows us which features have to be used to support dialogs and to what extent in various domains. Therefore, the proposed technique can serve as an automated means of writing quality and customer support quality assessment.

Chatbot Applications A Conversational Agent for Social Promotion (CASP) is an agent provided as a simulated persona acting on behalf of its human host to facilitate and manage communications about him or her. Galitsky B., Dmitri Ilvovsky, Nina Lebedeva and Daniel Usikov, "Improving Trust in Automation of Social Promotion" (AAAI Spring Symposium on The Intersection of Robust Intelligence and Trust in Autonomous Systems; Stanford, CA 2014). CASP relieves its host from the routine of minor activities on social networks, such as sharing news and commenting on other people's messages, blogs, forums, images and videos. Conversational agents for social promotion are developing, but they can also lose credibility. The overall performance of CASP focused on RR pair matching filtering responses mined from the web is evaluated.

People have, on average, 200-300 friends or contacts on social networking systems such as Facebook and LinkedIn. Maintaining positive relationships with this large number of friends requires several hours a week to read what these friends post and comment on. In reality, people only maintain relationships with their 10-20 closest friends, family members and co-workers, and contact the rest of their friends very rarely. These not-so-close friends feel abandoned social network relationships. However, maintaining positive relationships with all members of a social network is beneficial to many aspects of life, from work-related to personal affairs. Users of social networks are expected to let their friends know that they are interested in them and care about them and therefore respond to the messages they post in response to events in their lives. Therefore, users of social networks must, and often do not have, a significant amount of time to maintain relationships on social networks. For close friends and family, the user will still socialize using manual input. For the rest of the network, they will use the proposed CASP for social promotion.

CASP tracks user chats, user posts on blogs and forums, and comments on shopping sites to suggest web documents and their pieces of information relevant to purchasing decisions. This requires quoting portions of the text, generating a search engine query, and having a search engine API such as Bing execute it to filter out search results determined to be irrelevant to the seed message. The last step is critical to the robust functioning of CASP, and poor relevance in the rhetorical space will lead to a loss of trust in that space. Therefore, accurate assessment of RR match is critical to the successful use of CASP.

A CASP is provided as a simulated persona that acts on behalf of a person's host to facilitate and manage one's communications for that person (FIGS. 21-22). Agents are designed to rescue their hosts from routine, non-trivial activities on social networks such as sharing news and commenting on other people's messages, blogs, forums, images and videos. Unlike most application domains for simulated personalities, their social partners are not necessarily aware that they exchange news, opinion and updates with automated agents. We experimented with CASP's rhetorical matching and reasoning about the mental states of its peers within several Facebook accounts. We evaluate its performance and accuracy for reasoning about mental states related to human users with whom it communicates. For conversational systems, users need to feel that the system responds appropriately to their actions and that what the system replies makes sense. To achieve this in the horizontal domain, we need to make full use of linguistic information so that messages can be exchanged in a meaningful way.

CASP inputs seeds (posts written by humans) and outputs messages formed by the seeds. The message is formed by tailoring content found on the web to be relevant to the entered post. This relevance is based on relevance in terms of content and relevance in terms of RR match or state of mind match (e.g., it answers by asking a question, answering a recommendation post asking for more questions, etc.).

Figures 21 and 22 illustrate chatbots commenting on posts.
We assess how much human users lose trust in CASPs and their hosts when there is no relevance in both content and state of mind. Instead of evaluating rhetorical relevance, which is an intermediate parameter in terms of system utility, we assess how much users lose trust in CASP when they are plagued by rhetorically irrelevant and inappropriate posts.

In Table 3 we show the results of user tolerance to CASP failures. After several failures, friends lose trust, file complaints, remove from friend lists, share negative information about their loss of trust with others, and even encourage other friends to remove CASP-enabled friends from their friend lists. The values in the cells indicate the average number of rhetorically disrelevant posts when each trust loss event occurred. These disrelevant postings occurred within one month of performing this assessment, and we do not obtain values for the relative frequency of occurrence of these postings. For each user, on average, 100 posts were responded to (1-4 per seed post).

It can be seen that different domains have different scenarios in which a user loses trust in a CASP. For domains such as travel and shopping where the information is less important, the tolerance for loss of relevance is relatively high.

Conversely, in domains that are taken more seriously, such as work-related, personal preferences, etc., users are more sensitive to CASP failures and their various forms of trust are eroded more quickly.

For all domains, acceptance slowly declines as posts become more complex. Users' perceptions, regardless of content or their expectations, are worse for longer texts compared to posting shorter, single sentences or phrases by CASP.

The Domain of Natural Language Descriptions of Algorithms The ability to map natural language to a formal query or command language is important for the development of more user-friendly interfaces to many computing systems such as databases. However, relatively little research has addressed the problem of learning such semantic parsers from corpora of sentences paired with their formal language counterparts. There is "Learning to transform natural to formal languages" by Kate, Rohit, YW Wong and R. Mooney (AAAI, 2005). Moreover, to our knowledge, no such investigation was conducted at the discourse level. Learning to convert natural language (NL) into a fully formal language makes it easier to develop NL interfaces to hybrid computing and AI systems.

Dijkstra, a Dutch computer scientist who coined the concept of "structured programming" more than 40 years ago, said: "I wonder if machines that are supposed to be programmed in our native language (whether it's Dutch, English, American, French, German, or Swahili) have become as extravagantly arcane as what they would use." The seer was clearly right. That is, the specialization and precision of programming languages have also enabled tremendous advances in computing and computing. Dijkstra compares the invention of programming languages to the invention of mathematical symbolism. Dijkstra also states: "Rather than viewing the obligation to use formal codes as a burden, we should regard the convenience of using formal codes as a privilege. These formal codes enable children to learn early on what only a genius could achieve." But forty-odd years later, we are still stuck with the amount of code occupied in typical industrial applications (tens and hundreds of millions of lines of code that are terrifying to support and develop). The phrase "code itself is the best description" has become a kind of bad joke.

Natural language description of programs is an area where text rhetoric is special and where agreement between descriptions is essential. We will focus on common rhetorical and domain-specific expressions that map algorithmic descriptions to software code.

FIG. 23 shows a discourse tree for an algorithmic text, according to one aspect. We have the following text and its DT (Fig. 23).

1) Find any pixel p1.
2) Find the convex area a_off to which this pixel p1 belongs such that all pixels are less than 128;

3) Verify that the boundaries of the selected area have all pixels greater than 128;

4) If the above verification is successful, stop with a positive result. Otherwise, add all pixels that are less than 128 to a_off.

5) Check that the size of a_off is less than the threshold. Then go to 2. Otherwise, stop with a negative result.

We now show how to convert a given sentence to a logical form and then to a software code representation. A number of rhetorical relations serve to combine the statements resulting from the translation of individual sentences.

Verify that the boundaries of the selected area have all pixels greater than 128.

FIG. 24 shows an annotated sentence according to one aspect. See FIG. 24 for the pseudocode annotated deconstruction of 1-1 through 1-3.

By converting all constants to variables, we try to minimize the number of free variables and not overconstrain the representation at the same time. Connected (edge-linked) arrows indicate that the same constant value (pixel) maps to an equal variable (pixel) according to the rules of logic programming. To achieve this, we add a predicate that requires constraining the free variables (unary).

1-4) Add predicates that suppress free variables.
epistemic_action(verify) & border(Area) & border(Pixel) & above(Pixel, 128) & area(Area)
Now we need to construct an articulation to quantify everything, but in this particular case we use a loop structure so it will not be used.

FIG. 25 shows an annotated sentence according to one aspect. See FIG. 25 for an annotated deconstruction of the pseudocode of 1-5 through 2-3.

Ultimately, we have:
2-3) Resulting code fragment

Related Work Discourse analysis has a limited number of applications in answering questions and summarizing and generalizing texts, but we have not found an application for automatically constructed discourse trees. We list research related to the application of discourse analysis to two areas: dialog management and dialog games. These areas may apply to the same problems for which this proposal is intended. Both of these proposals have not only a series of logic-based approaches, but also analytical- and machine-learning-based approaches.

Managing Dialogs and Answering Questions When questions and answers are logically related, matching their rhetorical structure becomes less important.

De Boni proposed a method for judging the adequacy of answers to questions based on evidence of logical relevance rather than logical evidence of truth. See "Using logical relevance for question answering" by De Boni and Marco, Journal of Applied Logic, Volume 5, Issue 1, March 2007, Pages 92-103. We define logical relevance as the idea that an answer should not be considered absolutely true or false with respect to the question, but should be more flexibly true with sliding relevance. In addition, it is possible to make rigorous inferences about the appropriateness of an answer even if the source of the answer is incomplete, inconsistent, or contains errors. The authors show how logical relevance can be achieved in a relaxed form by using measured simplifications to seek logical evidence that an answer is in fact the answer to a particular question.

Our model of CDT attempts to combine general rhetorical and speech action information in a single structure. While speech actions provide a useful characterization of certain practical powers, more recent work, especially in building dialog systems, has significantly extended this core concept to model a wider variety of conversational functions that can be performed by utterances. The resulting enriched behavior is called a dialog behavior. See Jurafsky, Daniel and Martin, James H. (2000), "Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition," Upper Saddle River, NJ, Prentice Hall. In their multi-level approach to dialogue behavior, Traum and Hinkelman identify four levels of dialogue behavior necessary to ensure both consistency and content of conversation. See Traum, David R. and James F. Allen (1994), "Discourse obligations in dialogue processing" (In Proceedings of the 32nd annual meeting on Association for Computational Linguistics (ACL '94). Association for Computational Linguistics, Stroudsburg, PA, USA, 1-8). The four levels of speech actions are turning actions, reasoning actions, core speaking actions and discussion actions.

Decades of research into the logical and philosophical underpinnings of Q/A have focused on small, restricted domains and systems and have proven to be of limited use in industrial settings. The idea of the logical evidence of "being an answer to", developed in linguistics and mathematical logic, has been presented as having limited applicability within practical systems. Current applied research aimed at creating general-purpose (“open domain”) working systems is based on relatively simple architectures that combine information extraction and retrieval, as demonstrated by systems provided in the standard evaluation framework given by the Text Retrieval Conference (TREC) Q/A track.

Sperber and Wilson (1986) judged the relevance of an answer according to the amount of effort required to "prove" that a particular answer is relevant to the question. This rule can be formulated with rhetorical terms as a criterion of relevance. The less hypothetical the rhetorical relation required to prove that the answer matches the question, the more relevant the answer will be. The effort required can be measured in terms of the amount of prior knowledge required, inferences from the text or assumptions. In order to provide a more manageable criterion, we propose to simplify the problem by focusing on how constraints or rhetorical relations can be left out of the way questions are formulated. In other words, we evaluate how the question can be simplified to prove the answer. The resulting rule is formulated as follows. The relevance of an answer is determined by how many rhetorical constraints must be left out of the question about which answer to prove. The fewer rhetorical constraints that must be left out, the more relevant the answer.

The corpus of research on how finding rhetorical relationships can help in Q/A is very limited. The system introduced by Kontos made it possible to exploit the rhetorical relationship between the "basic" text, which proposes a model of a biomedical system, and the abstract portion of the paper, which presents the experimental findings that support this model. See "Question Answering and Rhetoric Analysis of Biomedical Texts in the AROMA System" by Kontos, John, Ioanna Malagardi, and John Peros (2016) (unpublished manuscript).

Adjacent pairs are defined as pairs of adjacent utterances produced by separate speakers and ordered as first and second parts. A particular type of first portion requires a particular type of second portion. Some of these constraints may also be removed to cover more cases of dependencies between utterances. See Popescu-Belis, Andrei, "Dialogue Acts: One or More Dimensions?" (Tech Report ISSCO Working paper n.62. 2005).

Adjacent pairs are inherently related, but can be decomposed into labels (“first part”, “second part”, “none”), possibly augmented with pointers to other members of the pair. Observed types of adjacent pairs that are frequently encountered include: request / offer / invite → accept / refuse; assessment → agree / disagree; blame → denial / admission; question → answer; apology → downplay; thank → welcome; greeting → greeting. See Levinson, Stephen C (2000), "Presumptive Meanings; The Theory of Generalized Conversational Implicature" (Cambridge, MA: The MIT Press).

A rhetorical relation is a relational concept related to the relation between utterances and non-utterances in isolation, similar to adjacency pairs. However, if an utterance is a satellite to a nucleus in only one relationship, it becomes possible to assign a relationship label to the utterance. This makes a strong need for a deep analysis of the dialog structure. The number of rhetorical relations in the RST ranges from the "dominates" and "prioritizes gratification" classes used by (Grosz and Sidner (1986)) to over 100 types. Consistency relations are an alternative way of representing rhetorical structures within a text. See "Categories of coherence relations in discourse annotation" by Scholman, Merel, Jacqueline Evers-Vermeul, and Ted Sanders (Dialogue &Discourse; Vol 7, No 2 (2016)).

There are many classes of NLP applications that are expected to exploit textual information structures. DT can be a very useful text summary. Information about the characteristics of the text segments must be considered to form an accurate and consistent overview based on the structure of the nuclear-satellite relationships and inter-segment relationships proposed by Sparkk-Jones in 1995. See "Summarising: analytic framework, key component, experimental method, in Summarizing Text for Intelligent Communication" by Sparck Jones, K. (Ed. B. Endres-Niggemeyer, J. Hobbs and K. Sparck Jones; Dagstuhl Seminar Report 79 (1995)). Combining the most important segments of the detailing relationships starting from the root node produces the most informative summary. DT has been used for multi-document overviews. See Radev, Dragomir R., Hongyan Jing and Malgorzata Budzikowska (2000), "Centroid-based summarization of multiple documents: sentence extraction, utility-based evaluation, and user studies" (In Proceedings of the 2000 NAACL-ANLP Workshop on Automatic summarization - Volume 4).

In natural language generation problems where the main challenge is consistency, one can rely on the information structure of the text to organize the extracted fragments of the text in a consistent way. Methods for measuring textual coherence can be used in the automatic assessment of essays. Because DT can capture textual coherence, the essay's flexible discourse structure can be used to assess essay style and quality. Burstein described a semi-automatic method for essay assessment that assessed textual coherence. See Burstein, Jill C., Lisa Braden-Harder, Martin S. Chodoro, Bruce A. Kaplan, Karen Kukich, Chi Lu, Donald A. Rock and Susanne Wolff (2002).

A neural network language model proposed in engio in 2003 attempts to predict the next word using the concatenation of several previous word vectors to form the input of the neural network. See "A neural probabilistic language model" by Bengio, Yoshua, Rejean Ducharme, Pascal Vincent and Christian Janvin (2003), J. Mach. Learn. Res. 3 (March 2003), 1137-1155. The results are as follows. Specifically, after the model is trained, the word vectors are mapped into a vector space such that semantically similar words in distributed representations of sentences and documents have similar vector representations. This kind of model can potentially work in discourse, but it is difficult to supply as rich linguistic information as we supply for tree kernel learning. A corpus of research exists that extends the word2vec model beyond the word level to achieve phrase-level or sentence-level representation. For example, the naive approach uses a weighted average of all the words in the document (the weighted average of the word vectors), which impairs word order similar to what the bag-of-words approach does. A more sophisticated approach uses matrix-vector operations to combine word vectors in the order given by the parse tree of the sentence. See R. Socher, C. D. Manning and A. Y. Ng (2010), "Learning continuous phrase representations and syntactic parsing with recursive neural networks" (In Proceedings of the NIPS-2010 Deep Learning and Unsupervised Feature Learning Workshop). Combining word vectors using a partree has been shown to work only on sentences, since it relies on parsing.

Many early approaches to policy learning for dialog systems used small state spaces and action sets and focused only on limited policy learning experiments (eg, confirmation or initiative types). The communicator dataset (Walker et al., 2001) is the largest available corpus of human-machine dialogs, further annotated with dialog context. This corpus has been used extensively to train and test dialog managers, but for a limited number of attributes such as destination city, it is restricted to information requiring dialogs in the air travel domain. At the same time, in this work we relied on an extensive corpus of request-response pairs of varying nature.

Reichman (1985) provides a formal description and an Augmented Transition Network (ATN) model of speech movement for conventional methods for recognizing the speech behavior of an utterance. The authors used an analysis of linguistic markers similar to those currently used for rhetorical syntactic parsing of preverb ``please'', modal verbs, prosody, references, clue phrases, etc. match); "Because..." (support), as well as indicators in other utterances). See "Getting computers to talk like you and me: discourse context, focus and semantics (an ATN model)" by Reichman R. (1985), Cambridge, Mass. London: MIT Press.

Given a DT for text as candidate answers to complex queries, a rule system is proposed for the occurrence of validity and invalidity of query keywords in this DT. See Galisky (2015). To be a valid answer to a query, the keyword must occur in the sequence of basic discourse units of this answer such that the sequence of basic discourse units are sufficiently ordered and linked by the nuclear-satellite relationship. An answer may be invalid if the keywords of the query occur only in the satellite discourse units of the answer.

Dialog Games In any conversation, a question is typically followed by an answer, or some explicit statement of inability or refusal to answer. We have the following model for the intentional space of conversation. When a question is posed by agent B, agent A recognizes agent B's purpose in order to find an answer and employs the purpose of communicating the answer to B in order to collaborate. A then plans to generate an answer by achieving an objective. It provides a neat explanation in a simple case, but requires strong assumptions for cooperativeness. Agent A must adopt Agent B's objectives as her own. As a result, this does not explain why A is speaking when she does not know the answer or when she is not ready to accept B's purpose.

Litman and Allen introduced intentional analysis at the discourse level in addition to the domain level to assume a set of conventional multi-agent behaviors at the discourse level. See Litman, D. L. and Allen, J. F. (1987), "A plan recognition model for subdialogues in conversation," Cognitive Science, 11:163-2. Others have attempted to explain this kind of behavior using social intentional structures such as solidarity intentions. See Cohen P. R. and Levesque, H. J. (1990), "Intention is choice with commitment," Artificial Intelligence, 42:213-261. See also "Attentions, Intentions and the Structure of Discourse" by Grosz, Barbara J. and Sidner, Candace L. (1986), Computational Linguistics, 12(3), 175-204. While these explanations have helped explain some discourse phenomena more fully, they still require strong cooperation to account for dialogue coherence, but do not provide a simple explanation for why agents behave when they do not support high-level mutual objectives.

Imagine a stranger approaching a person and asking, "Do you have spare coins?" Since they have never met before, it is unlikely that there is any intention of solidarity or plans to be shared. From a purely strategic point of view, agents may not be interested in whether the stranger's objectives are met. However, typically the agent will respond in such situations. Therefore, Q/A explanations should go beyond recognizing speaker intent. Questions do more than provide evidence of the speaker's purpose, and something more than adopts the interlocutor's purpose is involved in formulating the response to the question.

Mann proposed a library of discourse-level behaviors, sometimes called dialogue games, that encode common communicative dialogues. See "Rhetorical structure theory: Towards a functional theory of text organization" by Mann, William and Sandra Thompson (1988), Text-Interdisciplinary Journal for the Study of Discourse, 8(3):243-281. To be cooperative, an agent must always participate in one of these games. So when a question is asked, only a certain number of activities (ie those introduced by the question) will result in a collaborative response. Games provide an excellent illustration of coherence, but still require agents to be aware of each other's intentions in order to perform dialogue games. As a result, this work can be viewed as a special case of intentional observation. Because of this separation, agents need not assume cooperation for the task each agent is performing, but must be aware of intent and cooperation at the conversational level. It remains unexplained what purpose motivates conversational cooperativeness.

Coulthard and Brazil suggested that multiple responses can play a dual role for both response and new initiation: Initiation ^ (Re-Initiation) ^ Response ^ (Follow-up). See "Exchange structure" by Coulthard RM and Brazil D. (1979) (Discourse analysis monographs no. 5.: Birmingham: The University of Birmingham, English Language Research). An exchange may consist of 2-4 utterances. Additionally, follow-ups themselves can be followed up. An opening move often marks the beginning of an exchange, but this does not limit the type of next move. Finally, there may be closing moves that do not require follow-up. When these observations are added to their formula, we end up with:
(Open) ^ Initiation ^ (Re-Initiation) ^ Response ^ (Feedback) ^(Follow-up) ^ (Close)
This can now deal with anything that comes from 2 to 7 or more exchanges.

FIG. 26 illustrates discourse behavior of a dialog according to one aspect. Tsui (1994) features discourse actions according to three-part transactions. Her system for initiation, response and follow-up selection is shown in FIG. 26, corresponding to the upper, middle and lower portions.

FIG. 27 shows discourse actions of a dialogue according to one aspect.
The problem of classifying invalid versus valid RR pairs is also applicable to the task of complete dialog generation beyond answering questions and automated dialog support. Popescu presented the building blocks of a logic-based rhetorical structure of a natural language generator for human-computer dialogue. See Popescu, Vladimir, Jean Caelen, Corneliu Burileanu, "Logic-Based Rhetorical Structuring for Natural Language Generation in Human-Computer Dialogue" (Lecture Notes in Computer Science Volume 4629, pp 309-317, 2007). Practical and contextual aspects are taken into account when interacting with task controllers that provide domain and application dependent information and are structured in a fully formalized task ontology. To achieve the goals of computational feasibility and generality, a discourse ontology has been constructed and several axioms have been proposed that introduce constraints on rhetorical relations.

For example, the axioms specifying the semantics of topic (α) are given as follows.

where K(α) is a clause that logically represents the semantics of utterance α.
Concepts about the topic of an utterance are defined here in terms of a set of objects in the domain ontology and are referenced in a defined manner within the utterance. Therefore, topical relationships between utterances are computed using the task/domain ontology addressed by the task controller.

As instances of such rules, the following can be considered.

In this case, t+ is "future and 'new'".
Rhetorical Relationships and Arguments Often the primary means of linking questions and answers is logical argument. There is a clear connection between the RST relationships and the argumentative relationships that this study attempts to learn. There are four types of relationships: directed relationships, support, attack, detail and undirected sequence relationships. The relationship between support and attack is a controversial one and is known from related work. See "From Argument Diagrams to Argumentation Mining in Texts" by Peldszus, A. and Stede, M. (2013), A Survey. Int. J of Cognitive Informatics and Natural Intelligence 7(1), 1-31). The latter two correspond to the discourse relations used in RST. The discussion sequence relation corresponds to "Sequence" in RST, and the discussion detail relation roughly corresponds to "background" and "detail".

The discussion detail relationship is important because in scientific publications often some background matter (eg definitions of terms) is important to the understanding of the overall discussion. A support relationship between an argumentation component Resp and another argumentation component Req indicates that Resp supports (deduces, proves) Req. Similarly, the attack relationship between Resp and Req is annotated if Resp is attacking (restricting, refuting) Req. Where Resp is a detail of Req and defines what is stated in Req without giving more information or inference by discussion, then detail relations are used. Finally, we link two argumentative constructs (in Req or Resp) with a sequence relation if the constructs belong to each other and only make sense when combined, i.e., if the constructs form a multi-sentence argumentative construct.

We find that SVM TK can be used to distinguish between a wide range of text styles (Galitsky, 2015), including text styles without arguments and text styles with various forms of arguments. Each text style and genre has its own rhetorical structure that is leveraged and automatically learned. Because of the somewhat low correlation between text style and text words, conventional classification approaches that only consider keyword statistics can lack accuracy in complex cases. We also performed text classification into somewhat abstract classes such as those belonging to linguistic objects and metalanguages within the literature domain, as well as style-based document classification into proprietary design documents. Galitsky, B; Ilvovsky, D.; and Kuznetsov SO, "Rhetoric Map of an Answer to Compound Queries" (Knowledge Trail Inc. ACL 2015, 681-686). Evaluation of textual integrity in the domain of invalid vs. valid customer complaints (including disjointed discussion streams indicative of the complainer's sullenness) shows a stronger contribution of rhetorical structure information compared to sentiment profile information. While the discourse structure obtained by the RST parser is sufficient to perform textual integrity assessments, the affective profile-based approach shows much weaker results and is not a strong complement to rhetorical structure.

Although an extensive corpus of research is dedicated to RST parsers, investigations into how to exploit RST parsing results on real NLP problems are limited to content generation, summarization and retrieval (Jansen et al., 2014). The DTs obtained by these parsers cannot be used directly in rule bases to filter or construct text. Learning is therefore required to exploit the implicit properties of DT. This work is, to our knowledge, the pioneering work that employs discourse trees and their extensions for general open-domain question answering, chatbots, dialog management and text construction.

A dialog chatbot system needs to be able to understand and match the user's communication intentions, reason with these intentions, construct the user's own respective communication intentions, and populate these intentions with the actual language to be communicated to the user. Discourse trees by themselves do not provide an expression for the intent of these communications. In this work, we introduced communicative discourse trees built on conventional discourse trees, which, on the one hand, can be generated on a large scale these days, and on the other hand, can constitute descriptive utterance-level models of dialogue. Processing dialogs by machine learning of communicative discourse trees has allowed us to model a wide range of dialog types of cooperation modes and interaction types (planning, execution, and interleaved planning and execution).

Statistical computational learning approaches offer several major potential advantages for dialog system development over manual, rule-based, hand-coding approaches.

A data-driven development cycle;
the optimal policy of conduct in some cases;
A more accurate model for response selection;
- Possibility of generalization to invisible states;
• Reduced development and deployment costs for industry.

Comparing inductive learning results with kernel-based statistical learning, reliance on the same information allowed us to perform more concise feature engineering than either approach does.

The extensive corpus of literature on RST parsers does not address the question of how the resulting DT can be employed in practical NLP systems. RST parsers are mostly evaluated on their match with a human-annotated test set rather than on their expressiveness of the features of interest. In this work, we focused on the interpretation of DTs and explored ways to express them in a form that indicates agreement or disagreement rather than fact-neutral enumeration.

We used the CDT to provide matching criteria for how the next message follows a given message within a dialog. The CDT now contains labels for communication actions in the form of substituted VerbNet frames. We investigated discourse features that indicated incorrect request-response and question-answer pairs versus correct request-response and question-answer pairs. We used two learning frameworks to recognize exact pairs. Two learning frameworks are deterministic nearest-neighbor learning for CDT as graphs and tree-kernel learning for CDT. In this case, the feature space of all CDT subtrees is subject to SVM training.

A positive training set was constructed from accurate pairs obtained from Yahoo Answers, social networks, company conversations including Enron emails, customer complaints, and interviews with journalists. A corresponding negative training set was created by appending responses to a variety of arbitrary requests and questions that contained relevant keywords such that the relevance similarity between requests and responses was high. According to the evaluation, valid pairs can be recognized in 68%-79% of cases in domains with weak request-response matching and 80%-82% of cases in domains with strong matching. These accuracies are essential to support automated conversations. These accuracies are comparable to benchmark tasks that classify the discourse tree itself as valid or invalid, as well as analogous question-answer systems.

We believe this work is the first to leverage automatically constructed discourse trees for question-answering support. Previous research has used specific customer discourse models and features that are systematically collected, explainably learned, reverse engineered, and difficult to compare with each other. We conclude that learning rhetorical structures in the form of CDTs is a major data source for supporting complex question answering, chatbots and dialog management.

Argument Detection Using a Communicative Discourse Tree Aspects described herein use a communicative discourse tree to determine whether a text contains an argument. Such an approach can be useful, for example, for chatbots that can determine whether users are arguing or not. Several argument patterns may be employed when a user attempts to provide an argument for something. An argument can be the key point of any communication, persuasive essay, or speech.

The communicative discourse tree for a given text reflects the arguments present in the text. For example, the basic points of an argument are reflected in the rhetorical structure of the text in which the argument is presented. A text without argument has a different rhetorical structure. See Marie-Francine, Erik Boiy, Raquel Mochales Palau, and Chris Reed. 2007. Automatic detection of arguments in legal texts. (Proceedings of the 11th International Conference on Artificial Intelligence and Law, ICAIL 07, pp. 225-230, Stanford, California, USA). Additionally, arguments may differ between domains. For example, for product recommendations, text with positive sentiment is used to encourage potential purchasers to purchase. In the political domain, the logic of emotion versus argument versus agency is much more complex.

Machine learning can be used in conjunction with a communicative discourse tree to determine arguments. Determining arguments can be approached as a binary classification task in which a communicative discourse tree representing a particular block of text is provided to a classification model. A classification model returns a prediction of whether the communicative discourse tree is a positive or negative class. The positive class corresponds to texts with arguments and the negative class corresponds to texts without arguments. Aspects described herein can perform classification based on different syntactic and discourse features associated with logical arguments. In one example, for text classified as argument-containing text, the text is similar to the elements of the first class assigned to this class. Two types of learning can be used to assess the contribution of our information sources: nearest neighbor learning and statistical learning techniques.

Nearest Neighbor (kNN) learning uses explicit engineering of graph descriptions. The similarity measure is the overlap between the graph of the given text and the graph of the given element of the training set. In statistical learning, various aspects learn structures with implicit features.

In general, machine learning methods estimate the contribution of each feature type and the learning method described above to the problem of argumentation identification involving the presence of opposing arguments (Stab and Gurevych, 2016). More specifically, various aspects make use of rhetorical relationships and how discourse and semantic relationships cooperate in an argument detection task.

Sentiment analysis is necessary for a wide range of industrial applications, but its accuracy remains rather low. Recognizing the existence of arguments, if done reliably, could potentially replace some opinion mining tasks when trying to distinguish strongly opinionated content from neutral content. The demonstrative recognition results can then serve as features for sentiment analysis classifiers, distinguishing cases of high emotional polarity from neutral and low polarity cases.

Examples of Using Communicative Discourse Trees to Analyze Arguments The following examples are introduced to illustrate the value of using communicative discourse trees to determine the presence of arguments in a text. The first example discusses Theranos, a healthcare company that hoped to revolutionize blood testing. Several sources, including The Wall Street Journal, claimed the company's conduct was fraudulent. These allegations were based on accusations by employees who left Theranos. At some point, the FDA got involved. In 2016, some believed Theranos' view that the issue was initiated by Theranos' competitors who were jealous of the efficiency of Theranos' promised blood-testing technology. However, using argument analysis, embodiments described herein show that the Theranos argument pattern mined on Theranos' website was incomplete. In fact, cases of cheating were brought to the fore, leading to a flood of cheating reports. According to the Securities and Exchange Commission, Theranos CEO Elizabeth Holmes has raised more than $700 million from investors "through years of scrupulous fraud" in which she exaggerated or misrepresented statements about the company's technology and finances.

Given content about Theranos, if a user is biased toward Theranos and not biased against its opponents, the discussion detection system will attempt to provide answers that favor Theranos' views. In this case, the good arguments of the proponents, or the bad arguments of their opponents, would also be useful. Table 4 shows flags for various combinations of agency, sentiment, and discussion to tailor search results for a given user with particular preferences for Entity A versus Entity B. The right gray side of the column has opposite flags for the second and third rows. For row 4, only cases with generally accepted opinion-sharing merit are flagged for display.

The chatbot can use the information in Table 4 to personalize responses or tailor search results or opinionated data to user expectations. For example, a chatbot can consider political perspectives when providing news to users. Additionally, personalizing responses is useful for product recommendations. For example, a particular user may prefer skiing to snowboarding, as evidenced by one user sharing stories of people who do not like snowboarding. As such, aspects described herein enable chatbots to exhibit empathy and behave like peers by ensuring that users are not irritated by a lack of common opinion with the chatbot.

Continuing with Theranos' example, an RST representation of the argument is constructed, and various aspects can be observed to see if the discourse tree can indicate whether a paragraph conveys both the assertion and the arguments that support it. Additional information is attached to the discourse tree so that it can be determined whether or not it expresses a discourse pattern. According to The Wall Street Journal, here's what actually happened: "Since October [2015], The Wall Street Journal has published a series of accusations by anonymous sources that portray Theranos inaccurately. Now, in its most recent article ("U.S. Investigates Theranos Complaints," December 20), The Wall Street Journal again relied on anonymous sources, this time to the Centers for Medicare and Medicaid. (Carreyrou, 2016) reported two undisclosed and unverified complaints allegedly filed by the U.S. Food and Drug Administration (FDA) and the U.S. Food and Drug Administration (CMS).

FIG. 28 illustrates an exemplary communicative discourse tree according to one aspect. FIG. 28 shows discourse tree 2800, communication behavior 2801, and communication behavior 2802. FIG. More specifically, discourse tree 2800 represents the following paragraph: "But Theranos has struggled behind the scenes to turn the excitement over its technology into reality. At the end of 2014, the lab instrument developed as the linchpin of its strategy handled just a small fraction of the tests then sold to consumers, according to four former employees." (At the end of 2014, the lab equipment developed as the cornerstone of that strategy processed only a small portion of the tests and then sold them to consumers.) As can be seen, when arbitrary communicative behaviors are attached to the discourse tree 2800 as terminal arc labels, it becomes clear that authors are trying to make people understand what they mean, rather than just sharing facts. As shown, communication behavior 2801 is "struggle" and communication behavior 2802 is "develop."

FIG. 29 illustrates an exemplary communicative discourse tree according to one aspect. Figure 29 shows a discourse tree 2900 representing the following text: "Theranos remains actively engaged with its regulators, including CMS and the FDA, and no one, including the Wall Street Journal, has provided a copy of the alleged complaints to those agencies. os has not seen these alleged complaints, it has no basis on which to evaluate the purported complaints." or "not see" can be relied upon to represent the fact that the author is actually arguing with the opponent.

FIG. 30 illustrates an exemplary communicative discourse tree according to one aspect. FIG. 30 shows a discourse tree 3000 representing the following text, in which Theranos attempts to save the day: "It is not unusual for disgruntled and terminated employees in the heavily regulated health care industry to file complaints in an effort to retaliate against employers for termination of employment. Regulatory agencies have a process for evaluating complaints, many of which are not substantiated. Theranos trusts its regulators to properly investigate any complaints."
Thus, discourse relations are necessary but insufficient to show the structure of arguments, and speech behaviors (communicative behaviors) are likewise necessary but insufficient. For the paragraphs associated with FIG. 30, it is necessary to know the discourse structure of interactions between agents and what kinds of interactions they are. Specifically, it is necessary to distinguish between neutral elaboration (which does not involve communicative behavior) and elaboration relationships that involve communicative behavior with emotions such as "not provide" that are correlated with arguments. Note that the domain of the interactions (eg, healthcare) is not necessary, nor is it necessary what the subject (company, magazine, institution) or entity of these interactions is. However, mental domain-independent relationships between these entities are useful.

FIG. 31 illustrates an exemplary communicative discourse tree according to one aspect. FIG. 31 shows a discourse tree 3100 representing the following text about Theranos' argument that the opponent's argument is incomplete: "By continually relying on mostly anonymous sources, while dismissing concrete facts, documents, and expert scientists and engineers in the field provided by Theranos, the Journal denies its readers the ability to scrutinize and weigh the sources' identities, motives, and the veracity of their statements." By continuing to rely on documents and largely anonymous sources while shunning expert scientists and engineers, The Wall Street Journal denies its readers the ability to scrutinize and consider the origins and motives of the sources and the authenticity of the journal's statements.)
From the point of view of common sense reasoning, Theranos has two options for confirming its argument that its trials are valid: (1) conduct independent research, compare their results with peers, publish the data, and confirm their analytical results, or (2) defeat arguments by opponents that their trial results are invalid and support the argument that they are wrong. Clearly, the former argument is much stronger, and the latter argument is usually chosen when the agent thinks the former argument is too difficult to enforce. On the one hand, readers may agree with Theranos that the WSJ should have provided more evidence for its charges against Theranos. On the other hand, the reader probably dislikes Theranos' choice of the latter argument type (2) above, thus weakening Theranos' position considerably. One reason Theranos' argument is weak is that Theranos seeks to refute opponents' allegations of client complaints about Theranos' services. Theranos' demand for evidence by forcing the WSJ to disclose the source and nature of the complaint is weak. One argument is that a third party (an independent investigative agent) is more rational and conclusive. However, some readers may find Theranos' argument (burden of proof avoidance) logical and valid. Note that the argument evaluator cannot identify rhetorical relationships within the text by relying solely on the text. Rather, the context of the situation is useful for grasping the intention of the speaker.

In the second example, the author's aim is to attack allegations that the Syrian government used chemical weapons in the spring of 2018. FIG. 32 illustrates an exemplary communicative discourse tree according to one aspect. FIG. 32 shows a communicative discourse tree 3200 for this second example.

Given this example, acceptable evidence would be to share certain observations that, from the perspective of peers, would be associated with the absence of chemical weapons attacks. For example, if it could be demonstrated that the timing of the alleged chemical weapons attack coincided with a period of very heavy rains, that would be a persuasive way of attacking this claim. However, since no such observations were confirmed, the source Russia Today resorted to plotting complex mental states as to how the allegations were conveyed, in which case most accounts of the mental states of those involved are difficult to verify. [Whatever the Douma residents ,][who had first-hand experience of the shooting of the water][dousing after chemical attack video ,][have to say ,][their words simply do not fit into the narrative][allowed in the West ,][analysts told RT .] [Footage of screaming bewildered civilians and children][being dosed with water ,][presumably to defuse] contaminate them ,][was a key part in convincing Western audiences][that a chemical attack happened in Douma .] [Russia brought the people][seen in the video][to Brussels ,][where they told anyone][interested in listening][that the scene was staged .] [Their testimonies , however , were swiftly branded as bizarre and underwhelming and even an obscene masquerade][staged by Russians .] [They refuse to see this as evidence,][obviously pending][what the OPCW team is going to come up with in Douma], [Middle East expert Ammar Waqqaf said in an interview with RT .] [The alleged chemical incident,][without any investigation, has already become a solid fact in the West,][which the US, Britain and France based their retaliatory strike on.] Whatever they say, their words simply don't fit the narrative accepted in the West, analysts told RT. Images of bewildered screaming civilians and children, presumably being doused with water for decontamination, were a key part of convincing Western viewers that a chemical weapons attack had taken place in Douma. Russia took the people in the video to Brussels, where they told those who would listen that the scene was a hoax. But their testimony was quickly branded as a bizarre, overwhelming, and even despicable fiction engineered by the Russians. Ammar Waqqaf, an expert on Middle East affairs, said in an interview with RT that they refused to see this as evidence and that the OPCW team was apparently pending on what to come up with in Douma. The alleged chemical weapons attack, without any investigation, has already become a solid fact in the West on which the US, UK and France base their retaliatory attacks.)

Note that the above text finds no refutation of the chemical weapons attack claims it seeks to refute. Instead, the above text states that dissenters are not interested in observing this rebuttal. The main statement of this article is that certain agents "disallow" certain types of evidence that attack the main allegations, rather than provide and support this evidence. Instead of debunking the chemical weapons attack claim, the article builds on a complex mental conflict between residents, the Russian agents who brought them to Brussels, and Western and Middle Eastern experts.

FIG. 33 illustrates an exemplary communicative discourse tree according to one aspect. FIG. 33 shows a communication discourse tree 3300 for another controversial article, Trump-Russia link acquisition (BBC 2018). For a long time, the BBC has been unable to confirm this claim, so this article has been repeated over and over to keep readers hopeful that it will one day be corroborated. The existence of dossiers is neither confirmed nor denied, and the author's goal is to convince the viewer that such dossiers exist without misrepresenting events. To achieve this goal, authors can attach several hypothetical statements about existing dossiers to various states of mind in order to impress readers of the authenticity and relevance of the topic.

As depicted in FIGS. 32 and 33, many rhetorical relationships are associated with mental states. Mental states are complex enough to make it difficult for humans to verify the accuracy of the main claims. The communicative discourse tree shows that the author tries to replace the logic chains that would support the argument with complex mental states. A simple look at the CDT shown in Figures 32 and 33, without reading the associated text, is sufficient to see that the line of argument is incomplete.

Handling Heated Arguments FIG. 34 illustrates an exemplary communicative discourse tree according to one aspect. FIG. 34 shows a communication discourse tree 3400 for an example heated argument. Specifically, the following text, represented by communication discourse tree 3400, provides a CDT example for a heated discussion of a customer who was badly treated by the credit card company American Express (Amex) in 2007. Communicative discourse tree 3400 shows the emotional profile. The affective profile is the emotional value attached to the instructions of the proposer (here "I") and the opponent (here "Amex"). As you can see, proponents almost always positively and opponents negatively confirm the line of argument for this complaint. A fluctuating sentiment value would indicate a problem with how the author presents the arguments.

The text is divided into logical chunks as follows: [I 'm another one of the many][that has been carelessly mistreated by American Express .] [I have had my card since 2004 and never late .] [In 2008][they reduced my credit limit from $16,600 to $6,000][citing several false excuses .] [Only one of their excuses was true - other credit card balances .] [They also increased. my interest rate by 3 %][at the same time .] [I have never been so insulted by a credit card company .] [I used to have a credit score of 830 , not anymore , thanks to their unfair credit practices .] [They screwed my credit score .] [In these bad economic times you'd think][they would appreciate consistent paying customers like us][but I guess][they are just so full of themselves .] [I just read today][that their CEO stated][that they will be hurt less than their competitors][because 80 percent of their revenues][are generated from fees. ] (I am one of many who have been blatantly wronged by American Express. I have had the card since 2004 and have never missed a payment. In 2008, they lowered my credit limit from $16,600 to $6,000, citing some false excuses. Only one of their excuses - another credit card balance - was true. At the same time, my interest rate also increased by 3%. I have never been so insulted by a credit card company. I used to have an 830 credit score, but now I don't because of their unfair credit practices. They ruined my credit score. In these bad economic times, you'd think they would appreciate consistent paying customers like us, but I think they're just thinking about themselves. I just read today that their CEO says that 80% of their revenue comes from fees, making them less vulnerable than their competitors. It explains their ruthless, arrogant and unacceptable credit practices. It looks like they need to exploit every cardholder they can before the new law comes into force. Come on America, learn from our horrible experience, stop using our American Express credit cards, and take revenge! ).

FIG. 35 illustrates an exemplary communicative discourse tree according to one aspect. FIG. 35 shows a communicative discourse tree 3500 representing text advising how to conduct behavior that conveys an argument: "When a person is in the middle of an argument, it can be easy to get caught up in the heat of the moment and say something that makes the situation even worse. Nothing can make someone more frenzied and hysterical than telling them to calm down. It causes the other person to feel as if one is putting the blame for the elevation of the situation on them. Rather than actually helping them calm down." , it comes off as patronizing and will most likely make them even angrier." (which will probably make the person even more angry). FIG. 35 is an example of a meta-argument. A meta-argument is an argument about how to conduct a heated argument and can be represented by the same rhetorical relation.

Using Machine Learning Models to Determine Arguments As discussed, the application 102 can detect arguments within text. FIG. 36 illustrates an exemplary process for determining arguments using machine learning, according to one aspect.

At block 3601, the process 3600 includes accessing text containing fragments. Application 102 can receive input text 130 or text from different sources such as chat, Twitter, and other Internet-based sources. Text can consist of fragments, sentences, paragraphs, or longer amounts.

At block 3602, the process 3600 includes creating a discourse tree from the text, the discourse tree containing nodes, each non-terminal node representing a rhetorical relationship between two of the fragments, and each terminal node of the nodes of the discourse tree being associated with one of the fragments. Application 102 creates a discourse in substantially the same manner as described in block 1502 of process 1500 .

At block 3603, the process 3600 involves matching each fragment with a verb with a verb signature, thereby creating a communicative discourse tree. Application 102 creates a discourse in substantially the same manner as described in steps 1503-1505 of process 1500. FIG.

At block 3604, the process 3600 includes determining whether the communication discourse tree contains arguments by applying a classification model trained to detect arguments to the communication discourse tree. Classification models can use different learning techniques. For example, a classification model can use support vector machines with tree kernel learning. In addition, the classification model can use nearest neighbor learning of maximal common subtrees.

As an example, application 102 can use machine learning to determine the similarity between the communicative discourse tree identified in block 3603 and one or more communicative discourse trees from the training set of communicative discourse trees. Application 102 may select additional communicative discourse trees from a training set containing multiple communicative discourse trees. Training can be based on the communicative discourse tree that has the highest number of similarities with the additional communicative discourse trees. Application 102 identifies whether additional communicative discourse trees are from the positive set or the negative set. A positive set is associated with texts containing arguments, and a negative set is associated with texts without arguments. Application 102 determines whether the text contains discussion or does not contain discussion based on this identification.

Evaluating Logical Argument Detection To evaluate argument detection, a positive dataset is made heterogeneous and created from several sources to pick together different styles, genres, and argument types. First, we used some of the more frequently debated data, such as opinionated data from newspapers such as the New York Times (1400 articles), Boston Globe (1150 articles), Los Angeles Times (2140) and others (1200). Also use the customer complaint text. In addition, we use text style and genre recognition datasets (Lee, 2001). This dataset has a specific dimension associated with the argument (section [ted] "Emotional speech on a political topic with an attempt to sound convincing"). And finally, we add some texts from standard argumentation mining datasets where the presence of arguments is established by the annotators: the "Fact and Feeling" dataset (Oraby et al., 2015), 680 articles and the dataset "Argument annotated essays v.2" (Stab and Gurevych, 2016), 430 articles.

For negative datasets: Wikipedia (3500 articles), factual news sources (Reuters-supplied with 3400 articles), and datasets that also include corpus sections such as (Lee, 2001) [tells] (450 articles), "Instructions for how to use software" (320 articles) [tele]; "Instructions for how to use hardware" (175 articles); [news], "A presentation of a news article in an objective, independent manner" (220 articles), and other mixed datasets without arguments (735 articles) can be used.

Both the positive and negative datasets contain 8800 texts. The average text size was 400 words (always >200 words and <1000 words). Using Amazon Mechanical Turk, we confirmed that the positive dataset contained arguments in common sense view according to the workers employed. Twelve workers with previous acceptance scores above 85% were assigned the labeling task. For manual confirmation of the presence or absence of arguments, we randomly selected representatives from each set (approximately 10%) and confirmed that they belonged properly to a class with greater than 95% confidence. We avoided sources with less than 95% confidence as such. For the first portion of text that received manual labeling, we performed an assessment of inter-annotator agreement and found it to exceed 90%. Therefore, we relied on one operator per text for the remaining annotations. For evaluation, we split the dataset into training and testing parts in a 4:1 ratio.

Specific Argument Patterns Dataset The purpose of this argumentation dataset is to collect texts of complaints that use a variety of argumentation tools to prove that the document creator is a business victim. A customer complaint is an emotional text containing a description of a problem a customer has experienced with a particular business. Raw complaints are collected from PlanetFeedback.com for a number of banks submitted from 2006-2010. The 400 complaints are manually tagged with respect to the following parameters relevant to the discussion:
the validity of the perceived complaint;
・Relevance of the discussion,
• Presence of specific argument patterns and • Detectable misspellings.

Judging by the complaints, most complainers are in a real quandary due to strong deviations from what they expected from the service, what they received and the manner in which it was communicated. Most complaint writers report incompetence, flawed policies, neglect, indifference to customer needs, and misrepresentations from customer service personnel.

Authors often exhaust the channels available to them, become confused, seek recommendations from other users, and advise others about avoiding certain financial services. Complaints focus on proving that the proponent is right and the opponent is wrong, the proposed resolution, and the desired outcome.

Several argument patterns are used in complaints:
• Most often, according to common sense, deviations of what actually happened from what was expected. This pattern covers both valid and invalid arguments (valid pattern).
• A second popular argument pattern cites the difference between what was promised (advertised, communicated) and what was received or actually happened. This pattern also mentions that the opponent does not act according to the rules (valid).
A number of complaints explicitly state that bank representatives are lying. Lying includes inconsistencies between information provided by different bank agents, misrepresentation of facts, and careless promises (valid).
• Another reason for complaints is due to the rudeness of bank agents and customer service personnel. Customers cite rudeness in both cases and then the objector points are valid or not (and complaint and argument validity are tagged accordingly). Even if there is no financial loss or inconvenience, complainants will not agree to anything a given bank does if they are treated disrespectfully. (invalid pattern).
• Complainants cite their needs as a reason for banks to behave in a certain way. A popular argument is that the administration bailed out banks through the taxpayers and now banks should benefit their customers (null).

This dataset contains more emotionally heated complaints compared to other argumentation mining datasets. For a given topic such as insufficient funding fees, this dataset offers many different ways of arguing that this fee is unfair. Our dataset therefore allows for a systematic search for clusters of topic-independent argument patterns and observes links between argument types and overall complaint relevance. Other argumentation datasets, including legal argumentation, student papers (Stab and Gurevych 2017), Internet argumentation corpora (Abbot et al., 2016), fact-sensory datasets (Oraby et al., 2016), and political debates, have strong topical variation and are therefore more difficult to track the spectrum of possible argumentation patterns per topic. Unlike professional writing in legal and political spheres, genuine writing of complaining users has a simple motivational structure, transparency of their purpose, and occurs in a given domain and context. In the dataset used in this study, argumentation plays an important role for the well-being of authors who are subject to unjustified large financial claims or evicted from their homes. Therefore, authors attempt to give as strong an argument as possible to back up their claims and strengthen the case.

If the complaint is untrue, it is usually invalid: the customer either complains or seeks compensation. However, even if the complaint is true, it can easily be invalidated, especially if the argument is flawed. If a false complaint has a valid argument pattern, it is difficult for the annotator to properly assign it as valid or invalid. Three annotators worked on this dataset and the agreement among the annotators is over 80%.

Evaluation Settings and Results For nearest neighbor classification, we used the maximal common subgraph for the DT method alongside the maximal common subgraph for the CA method based on scenario graphs constructed on CA extracted from the text (Table 5). For SVM TK classification, we employ the parse-crossover approach tree-kernel learning, where each paragraph is represented by a parse-crossover containing exhaustive syntactic and discourse information. We also use SVM TK for DT and CA information is not considered.

Our family of pre-baseline methods is based on keywords and keyword statistics. For the naive Bayes approach, we relied on the WEKA framework (Hall et al., 2009). Since most lexical and length-based features are reliable for finding poorly supported arguments (Stab and Gurevych 2017), we used non-NER as a feature along with the number of tokens in phrases that potentially express arguments. Also, the NER count was used assuming that it correlates with the strength of the argument. Even when these features are strongly correlated with arguments, they do not help us understand the nature of how arguments are structured and communicated in language, as represented by the CDT.

A naive approach simply relies on keywords to capture the presence of arguments. Usually several communicative behaviors, at least one of which has a negative emotional polarity (related to the opponent), are sufficient to deduce that a logical argument exists. This simple approach is 29% worse than the best performing CDT approach. The naive Bayes classifier provides only 2% improvement.

For nearest neighbor learning, it can be observed that DT and CA actually complement each other, yielding a CDT accuracy of over 26% for the former and over 30% for the latter. CA alone produced worse results than independent DT (Table 6). As can be seen, CDT SVM TK outperforms RST+CA and full syntactic feature SVM TK (SVM TK baseline) by 5%. This is due to feature engineering and reliance on less data, but more relevant than the baseline.

Nearest neighbor learning on CDT achieves slightly less accuracy than SVM TK on CDT, but the former provides an interesting example of subtrees typical of the discussion and shared among factual data. The number of former groups of CDT subtrees is of course significantly higher. Unfortunately, the SVM TK approach does not help explain how exactly the argument identification problem is solved. It just gives the final scoring and class label. Expressing logical arguments without CA is possible, but rare. This observation is reinforced by our data.

It is worth mentioning that our evaluation setup is close to the SVM-based ranking of RST parsers. The problem is formulated as classifying the DT into a set of correct and incorrect trees that are close to the manually annotated trees. Our settings are slightly different as they are better tuned to smaller datasets. Note that the improvement in argument detection going from DT to CDT demonstrates the validity of our extension of RST with speech action related information.

Table 7 shows the SVM TK argumentation detection results by source. As a positive set we take only individual sources here. A negative set is formed from the same source, but reduced in size to match the size of the smaller positive set. Cross-validation settings are similar to our evaluation of the entire positive set.

We found no correlation between the specificity of a particular domain and the contribution of discourse-level information to argumentative detection accuracy. At the same time, all four domains show monotonic improvement when we progress from Keywords and Naive Bayes to SVM TK. Since all four sources show improvement in argumentation detection rate with CDT, we conclude that the same is likely for other sources of argument-related information.

Table 8 shows the pattern-specific argument detection results. We calculate the classification accuracy as specific patterns versus other patterns and lack of arguments. The first and second types of argumentation are more difficult to recognize (7-10% lower than typical argumentation), and the third and fourth types are easier to detect (3% above general argumentation accuracy).

These argument recognition accuracies are comparable to state-of-the-art argument mining techniques. One study analyzed a text containing 128 premise-conclusion pairs and obtained an F scale of 63-67% to determine the directionality of inferential connections in an argument. See Lawrence, John and Chris Reed. Mining Argumentative Structure from Natural Language text using Automatically Generated Premise-Conclusion Topic Models. Proceedings of the 4th Workshop on Argument Mining, pages 39-48. 2017. Bar-Haim et al. show that automatic expansion of the initial vocabulary can significantly improve both the accuracy and coverage of argumentative stance recognition (what supports and what refutes) up to the 69% F scale. See Bar-Haim, Roy Lilach Edelstein, Charles Jochim and Noam Slonim. Improving Claim Stance Classification with Lexical Knowledge Expansion and Context Utilization. Proceedings of the 4th Workshop on Argument Mining, pages 32-38. 2017. Aker et al. provide a comparative analysis of the performance of different supervised machine learning methods and feature sets on the argument mining task, achieving an F-measure of 81% for detecting argument sentences and 59% for the argument structure prediction task. See Aker, Ahmet, Alfred Sliwa, Yuan Ma, Ruishen Liu Niravkumar Borad, Seyedeh Fatemeh Ziyaei, Mina Ghbadi What works and what does not: Classifier and feature analysis for argument mining. Proceedings of the 4th Workshop on Argument Mining, pages 91-96. Regarding the argumentative segmentation of argumentative texts into argumentative units and their non-argumentative counterparts, Ajjour et al achieve 88% with Bi-LSTM for essays and 84% for edits. See Ajjour, Yamen, Wei-Fan Chen, Johannes Kiesel, Henning Wachsmuth and Benno Stein. Unit Segmentation of Argumentative Texts. Proceedings of the 4th Workshop on Argument Mining, pages 118-128, 2017. Taking into account the complexity of the argumentation mining task, these classification accuracies are comparable to current studies, but lack the exploration of arguments' causes via discourse-level analysis. This work therefore proposes a much more direct feature engineering of the general discussion and its specific patterns.

CDT Construction Although the partitioning into EDUs works reasonably well, the assignment of RST relations is noisy and in some domains its accuracy can be as low as 50%. However, if the RST relational label is random, it does not significantly degrade the performance of our argument detection system, because random discourse trees are not very similar to the positive or negative training set elements and most likely do not participate in the positive or negative decision. To overcome the noisy input problem, a broader training dataset is needed so that the number of reliable and valid discourse trees is large enough to cover the cases to be classified. As long as this number is large enough, the contribution of noisy and poorly constructed discourse trees is low.

There are some systematic deviations from the correct intuitive discourse tree obtained by the discourse parser. In this section, we assess whether there is a correlation between CDT deviations and our training set. We allow the possibility that the CDT deviations for texts with arguments are stronger than those for texts without arguments.

For each source, we calculated the number of CDTs that deviated significantly. For the purposes of this assessment, we considered the CDT to be deviated if more than 20% of rhetorical relationships were judged inappropriate. We do not distinguish between specific RST relationships associated with arguments such as attributes and contrasts. The strain assessment dataset is significantly smaller than the detection dataset because it requires substantial manual effort and the task cannot be submitted to Amazon Mechanical Turk workers.

It can be observed that there is no clear correlation between recognition class and CDT distortion rate (Table 9). Therefore, we conclude that the noisy CDT training set can be adequately evaluated for argument detection. As can be seen, there is a strong correlation between these noisy CDTs and the presence of logical arguments.

Emotion Reliable emotion detection in any domain is difficult, so we focus on specific emotion-related features such as logical arguments with specific polarities. Logical argument detection can help improve the detection performance of emotion detection. We formulate the emotion detection problem at the paragraph level. Detects emotional polarity only.

Categorizing emotions based on individual words can be misleading because it is possible to modify (weaken, strengthen, or reverse) atomic emotion carriers based on lexical, discourse, or contextual factors. Words interact with each other to produce expression level polarities. For example, the meaning of a compound expression is a function of the meanings of its parts and the syntactic rules with which they are combined. Considering more linguistic structures than required by RST is therefore what motivates our combination of these insights from various discourse analysis models. Our hypothesis is that it is possible to very accurately compute the polarity values of the larger syntactic elements of a text as a function of the polarity of their subcomponents, in a manner similar to the ``composition principle'' in formal semantics. In other words, if the meaning of a sentence is a function of the meanings of its parts, then the overall polarity of a sentence is a function of the polarities of its parts. For example, we can ascribe a negative trait to the verb "reduce", but a positive polarity in "reduce risk" even though "risk" is itself negative (see negative polarity in "reduce productivity"). This polarity reversal is only captured when extending the analysis beyond the sentence level to compute the overall polarity of the text as a whole. Therefore, any polarity conflicts are resolved as a function of the overall meaning of the text, based on textual and contextual factors. Polarity weights are not properties of the individual elements of the text, but are functions of properties that operate at the level of cohesion and are implicit coherence relations at the syntactic, discourse and pragmatic levels of discourse analysis.

Many studies have shown that discourse-related information can successfully improve the performance of sentiment analysis, for example, the importance of EDUs can be reweighted based on their association type or depth in DT (Hogenboom et al, 2015a). Some methods prune the discourse tree at a certain threshold to result in a tree of fixed depth between two and four levels. Other approaches train machine learning classifiers based on relationship types as input features (Hogenboom et al, 2015b). Most studies on emotion in RDSTs attempt to map DT structures to mathematically simpler representations, as it is virtually impossible to encode unstructured data of arbitrary complexity in fixed-length vectors (Markle-Hus et al., 2017).

FIG. 37 is a fragment of a discourse tree according to one aspect. FIG. 37 shows a discourse tree 3700 representing the following text. To show that the nuclear-satellite relationship is important in determining the sentiment of an entity, we use the following two sentences: [Although the camera worked well ,] [I could not use it because of the viewfinder], which expresses negative feelings about the camera; and [The camera worked well ], [although the viewfinder was inconvenient], which expresses positive feelings about the camera.

For the emotion detection evaluation, we used a dataset of positive and negative, real and fake traveler reviews of hotels in the Chicago area. See Ott, C. Cardie, and J.T. Hancock. 2013. Negative Deceptive Opinion Spam. In Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Authors compile datasets with the goal of distinguishing between genuine and fake reviews. It turns out that the fact that a review is fake does not correlate strongly with the presence of a logical argument. Fake reviews created by Mechanical Turn workers support authors' opinions in the same way real travelers do. The test corpus contains four groups of 400 reviews of 1-3 paragraphs each. 1) 400 true positive reviews from TripAdvisor; 2) 400 false positive reviews from Mechanical Turk; 3) 400 true negative reviews from Expedia, Hotels.com, Orbitz, Priceline, TripAdvisor, and 4) 400 false negative reviews from Mechanical Turk.

As a baseline method, we use Stanford NLP Sentiment. Take the sentence-level polarity and aggregate it to the paragraph level. Usually, when the opinion is positive, the author simply lists what he likes. However, if the opinion is negative, the author will often try to back it up and carry out comparisons, explanations, and arguments as to why he is correct and why his assessment is appropriate.

Therefore, the rules for the integration of default emotion detectors and argument-based emotion detectors are as follows (Table 10). This rule is directed at consumer review data and will need modification to better handle other text genres.

The following case is a borderline positive review that could easily be flipped to negative: "Like all hotels in Chicago, this hotel caters to wealthy and/or business clients with very high parking price. However, if you are aware of that prior to arrival, it's not a big deal. It makes sense to find a different place to park the car and bring your own snacks for the room. It would be nice though if hotels such as the Swissotel had a fridge in the room for guest use. Staff was very helpful. Overall, if I can get a good rate again, I'll stay at the Swissotel the next time I am in Chicago. All in all, if I can get a good rate again, I will stay at the Swissotel next time I'm in Chicago." This text looks like an overall negative review from DT's point of view. Most reviews with similar DT are negative.

FIG. 38 illustrates a discourse tree for boundary review, according to one aspect. FIG. 38 shows a discourse tree 3800 for boundary review. The Boundary Review is negative from the point of view of discourse and neutral from the point of view of the reader.

Extending Constructive Semantics toward Discourse We examine how sentiment in the first sentence is assessed by the semantic constructive model. See R. Socher, A. Perelygin, J. Wu, J. Chuang, C. Manning, A. Ng and C. Potts. Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank. Conference on Empirical Methods in Natural Language Processing (EMNLP 2013). Here, it is difficult to understand that "high price" has a negative sentiment value, judging by the individual words and their composition. In a movie database for training, 'high' is likely to be assigned a positive sentiment and 'high price' is not tagged as negative. Even if "high price" were perceived to be negative, it would be difficult to determine how the rest of the tree would affect it, such as the phrase "wealthy and/or business clients". Note that neither the words in this phrase are assigned the correct emotion, even in the realm of the movie.

Judging the emotional polarity of this sentence in isolation is rather difficult given its words and phrasing. Instead, the discourse of consecutive sentences can be considered to determine the overall paragraph sentiment and the sentiment of a given sentence with greater accuracy.

FIG. 39 shows a discourse tree of sentences illustrating a constructive semantic approach to sentiment analysis according to one aspect. FIG. 39 shows discourse tree 3900 .

We argue that a sentiment analysis that benefits from 'constructive semantics' insights would correctly assign polar sentiments in the above example if it captures more than just the word 'high' (assigned a negative sentiment polarity), the phrase 'high price' (with a negative sentiment polarity), or the sentence-level construct 'Like all....price' (sentiment polarity is difficult to determine because one needs to read the entire text for global sentiment polarity attributes). state. Sentiment analysis is calculated based on the global polarity and not on the individual elements of the sentence, but more interestingly on the discourse-level structure (macrostructure). For example, "high reliability" is neutral in the sense of "I want a car with high reliability" because it is a positive characteristic but does not refer to a particular car.

Results Because our evaluation of emotion detection is domain-independent, the baseline system (Socher et al., 2013) is trained on a different domain than the test domain.

Results of sentiment analysis achieved by hybrid constructive semantics and discourse analysis are shown in Table 11. The first row shows the accuracy of the baseline system for our data. The second gray row shows the improvement due to the hybrid system. This improvement is achieved by discovering overall negative sentiment at the paragraph level when the presence of arguments is recognized. In some of these cases, negative sentiment is implicit and can only be detected indirectly from discourse structures in which individual words do not exhibit negative sentiment.

We investigate a standalone SVM TK emotion recognition system with various representations (rows 3-5). The CDT representation outperforms the parse intersection and DT representations. Emotion recognition accuracy is much lower with simpler representations that do not consider discourse-level information at all (not shown).

We also investigated whether fake opinion texts have a different rhetorical structure than the real ones. See Jindal and Liu, Opinion Spam and Analysis, Department of Computer Science, University of Illinois at Chicago, 2008. Jindal and Liu addressed the problem of troll spam: detection of obvious instances that are easily identified by human readers, such as advertisements, questions, and other irrelevant or opinionless text. (Ott et al. investigated potentially more insidious types of opinion spam, such as rant spam, intentionally written to sound like the real thing to deceive readers. See M. Ott, Y. Choi, C. Cardie, and J.T. Hancock. 2011. Finding Deceptive Opinion Spam by Any Stretch of the Imagination. Proceedings of the 49th Annual Meeting of the Society for Computational Linguistics: Human Language Technology. , written by workers at Amazon Mechanical Turk.The instructions assume that they are hired by a hotel's marketing department and are asked to pretend they are being asked to write a fake review (as if they were a customer) posted on a travel review website;

Although our SVM TK system did not achieve 90% performance, the task of detecting fake review text was performed by a general-purpose text classification system (76-77% accuracy, two bottom gray rows), which extracts arguments and assesses sentiment polarity.

Verification of the Argument Aspects of this disclosure verify the argument. To be persuasive, the text or speech contains valid arguments. Application 102 extracts the argument structure from the body of the text and presents the argument via a communicative discourse tree (CDT). The application 102 can then verify that the assertion in the text or the target assertion is valid, i.e., not logically attacked by other assertions, and consistent with external truths, i.e. rules. Domain knowledge can be used to verify the validity of claims. However, in some cases, domain knowledge may not be available and other domain independent information such as writing style and writing logic is used.

Some aspects enable applications such as customer relationship management (CRM). CRM deals with handling customer complaints (Galitsky and de la Rosa 2011). In customer complaints, the writer is upset about the product or service they received and how the problem was communicated by customer support. Complainants often write their complaints in very strong emotional language, which can skew the logic of the argument and thus make it difficult to judge the validity of the complaint. Both emotional and logical arguments are heavily used.

To facilitate improved autonomous agents, some aspects employ linguistic-based argument mining and logic-based logical verification of arguments. The concept of automatically identifying argumentation schemes was first discussed in (Walton et al., 2008). Ghosh et al. (2014) investigate the discussion discourse structure of specific types of communication-online dialogue threads. Identifying arguments in text leads to problems identifying truth, misinformation, and disinformation on the web (Pendyala and Figueira, 2015, Galitsky 2015, Pisarevskaya et al 2015). (Lawrence and Reed, 2015) combine three types of argumentative structure identification: linguistic features, topic change and machine learning. As further described herein, some aspects employ Defeasible Logic Programming (DeLP) (Garcia and Simari, 2004; Alsinet et al., 2008) in conjunction with discourse trees for communication.

FIG. 40 illustrates an example process 4000 for validating an argument according to one aspect. Application 102 may execute process 4000 .

At block 4001, process 4000 includes accessing text containing fragments. At block 4001 , process 4000 performs steps substantially similar to those described at block 3601 of process 3600 . The text may include input text 130, which may be paragraphs, sentences, utterances, or other text.

At block 4002, the process 4000 includes identifying the presence of arguments in a subset of the texts by creating a communicative discourse tree from the text and applying a classification model trained to detect arguments to the communicative discourse tree. At block 4002 , process 4000 performs steps substantially similar to those described in blocks 3602 - 3604 of process 3600 . Other discussion detection methods can be used.

At block 4003, process 4000 includes evaluating arguments by using a logic system. Applications 102 can evaluate arguments using different types of logic systems. For example, Retractable Logic Programming (DeLP) can be used. FIG. 42 illustrates example operations by which block 4003 may be implemented. For illustrative purposes, process 4000 is described with respect to FIG.

FIG. 41 illustrates an exemplary communicative discourse tree for an argument according to one aspect. FIG. 41 includes discourse tree 4101 for communication. Communicative discourse tree 4101 includes node 4120 and other nodes, some of which are labeled communicative behaviors 4110-4117.

In one example, a judge hears an eviction case and wants to determine whether rent was verifiably paid (deposited) or not (denoted as rent_receipt). The input is text in which the defendant expresses his point. CDT4101 represents the following text: "The landlord contacted me, the tenant, and the rent was requested.However, I refused the rent since I demanded repair to be done.I reminded the landlord about necessary repairs, but the landlord issued the three-day notice confirming that the rent was overdue.Regretfully, the property still stayed unrepaired."
FIG. 42 illustrates an exemplary method for validating arguments using reversible logic programming, according to one aspect. Retractable Logic Programming (DeLP) is a set of facts, absolute rules Π of the form (A:-B), and sets of reversible rules Δ of the form A-<B, whose intended meaning is "if B applies, usually A also applies". Let P=(Π,Δ) be a DeLP program and L be a base literal. Absolute rules, even if based on opinion, cannot be changed. In contrast, revocable rules can be false in some cases.

In the example above, the underlined words form clauses in DeLP, and other expressions can form facts. An example of a fact is "rent_refused", ie the landlord refused the rent. An example of an absolute rule is "the earth is flat". An example of a revocable rule is "rent_receipt -< rent_deposit_transaction", which usually means "rent_receipt" if "rent_deposit_transaction". However, the reversibility rule may not always be true if, for example, the rent is credited to the wrong account or there is an error with the bank.

Application 102 can use the results from the communicative discourse tree created in block 4002 as input to DeLP. A communicative discourse tree shows valuable information such as how facts are interconnected by revocable rules. CDT basic discourse units of rhetorical relation type "contrast" and communicative behaviors of type "disagree" indicate retractable rules.

At block 4201, method 4200 includes creating a fixed portion of the logical system. A fixed part of a logical system includes one or more claim clauses and one or more domain definition clauses. A domain definition clause is associated with a domain of text and can include legal, scientific terms, and common sense knowledge in a particular domain. A scientific example is that if an object is moving with acceleration, it experiences a physical force. In the area of landlord-tenant law, an example of a standard definition is "if repairs are made -> house is habitable and appliances are in working order".

Continuing the above example, the text contains the target claim "rent_receipt" to be evaluated, ie, "Has the rent been received?" Application 102 also extracts the following clause "repair_is_done -< rent_refused" from the text "refused the rent since I demanded repair to be done."

At block 4202, method 4200 includes creating a variable portion of the logical system by determining a set of revocable rules and a set of facts. Application 102 determines a set of retractable rules from the communicative discourse tree by extracting from the communicative discourse tree one or more of (i) basic discourse units that are rhetorical relation type contrasts and (ii) communicative behaviors that are class type disagreeable. Class disagreements include actions such as "deny", "disagree", "disbelieve", "refusing to believe", "contradict", "deviate", "deviate", "contrary", "differ", "significant", "dissimilar". Other examples are possible.

Application 102 determines the following revocable rules.

In addition, application 102 determines additional facts from communication behavior that is of type "disagree." Continuing with this example and referring back to FIG. 41, application 102 determines the following facts from the CDT's communication behavior objects: contact_tenant (communication behavior 4111), rent_is_requested (communication behavior 4112), rent_refused (communication behavior 4113), stay_unrepaired (communication behavior 4114), remind_about_repair (communication behavior 4115), three_days_notice_is. _issued (communication behavior 4116), and rent_is_overdue (communication behavior 4117).

At block 4203, the method 4200 includes determining, from the set of revocable rules, revocable derivations that include a set of non-conflicting revocable rules. A revocable derivation _of L from P consists of _{a finite} sequence L _{1 ,} L _{2 ,} . , _{L j} ( _j < _{i )} . Let h be a literal and P=(Π,Δ) be a DeLP program. If <A,h> is an argument for h and A is the set of reversible rules of Δ, then
1. There exists a reversible derivation for h from =(Π∪A),
2. The set (Π∪A) is not contradictory and
3. A is the smallest and there is no suitable subset A ₀ of A that satisfies conditions (1) and (2).

Thus, the argument <A,h> is the minimal consistent set of reversible rules resulting from reversible derivation for a given literal h associated with program P. As mentioned above, minimal subset means that there is no subset that satisfies conditions 1 and 2.

At block 4204, the method 4200 includes creating one or more rebuttals from the set of facts. Counter-arguments are possible arguments in order that can be attacked by other arguments, as in human dialogue. An argument line is a sequence of arguments in which each element in the sequence overrules its predecessor. For DeLP, there are some admissibility requirements for argument lines to avoid fallacies (such as circular reasoning by repeating the same argument twice).

A counter-argument can be formulated as follows. For example, an argument _{<A 1 ,h 1 >} attacks _{<A 2 ,h 2 >} if and only if there is a subargument <A,h> of _{<A 2 ,h 2 >} (A⊆A ₁ ) and h and _hi contradict (i.e., Π∪{h, _hi } derives complementary literals). We would say that _{<A 1 ,h 1 >} overrules _{<A 2 ,h 2 >} if _{<A 1 ,h 1 >} attacks _{<A 2 ,h 2 >} with subarguments <A,h> such that _{<A 1 ,h 1 >} is strictly preferred (or not comparable) to <A,h>. In the first case, _{<A 1 , h 1 >} is called a proper defeater, and in the second case a blocking defeater.

At block 4205, the method 4200 includes building a dialectical tree including root nodes representing arguments and leaf nodes representing rebuttals from the retractable derivations. A target claim can be thought of as a DeLP query resolved with respect to a dialectical tree containing all possible lines of argument for a given query. The definition of a dialectical tree provides us with an algorithmic view for discovering implicit self-aggressive relationships in user assertions. Let _{<A 0 , h 0 >} be an argument from program P (target assertion). For purposes of explanation, block 4205 is described with respect to FIG.

FIG. 43 is a diagram illustrating an exemplary dialectical tree according to one aspect. Figure 43 depicts the dialectical tree for the text developed above. FIG. 43 includes a dialectical tree 4300 that includes a root node 4301 and nodes 4302-4307. The dialectical tree 4300 is based on _{<A 0 , h 0 >} , which is defined as follows:
1. The root of the tree (root node 4301) is labeled with _{<A 0 ,h 0 >} .
2. Let N be a non- _root vertex labeled _{<A n , h n >} , and let Λ=[ _{<A 0 , h 0 >} , _{<A 1 , h 1 >} , . Let all [<B ₀ , q ₀ >, <B ₁ , q ₁ >, . . . <B _k , q _k >] attack _{<A n , h n >} . For each attacker <B _i , q _i > with an acceptable line of argument [Λ, <B _i , q _i >], we have an arc between N and its child N ₁ .

Labeling for the dialectical tree can then be done as follows:
1. All leaves (nodes 4302-4307) are labeled as U nodes (unrefuted nodes).
2. Any interior node is labeled as a U node whenever all of its associated child nodes are labeled as D nodes.
3. Any interior node is labeled as a D node whenever at least one of its associated child nodes is labeled as a U node.

At block 4206, the method 4200 includes evaluating the dialectical tree by recursively evaluating the objections.

In the DeLP example, the literal rent_receipt is supported by <A, rent_receipt> = <{ (rent_receipt -< rent_deposit_transaction), (rent_deposit_transaction -< tenant_short_on_money)}, rent_receipt>, for which there are three objections, with three respective lines of argument:

(1) and (2) are proper rebuttals and the last one is a blocking rebuttal. The first argumentation structure has a counterargument, <{rent_deposit_transaction -<tenant_short_on_money}, rent_deposit_transaction), which is not a counterargument, because the former is more specific. Therefore, there is no counter-argument and the argument line ends there.

_B3 above is the blocking objection <{(rent_deposit_transaction -<tenant_short_on_money)}, rent_deposit_transaction>, which is a dissenting subargument of <A, rent_receipt> and cannot be introduced because it creates an unacceptable line of argument. _B2 has two objections that can be introduced:

D ₁ and C ₂ have blocking objections, but cannot be introduced as they render the argument line unacceptable. Therefore, the state rent_receipt cannot be reached, since no arguments are guaranteed to support the literal rent_receipt.

At block 4207, method 4200 entails identifying as valid the argument supported by the argument in response to determining that none of the counterarguments are inconsistent with the reversible derivation. A determination that there are no conflicting arguments indicates that the assertion is valid, and a determination that there are conflicting arguments indicates that the assertion is invalid. The rhetorical classifier 102 can then take action based on the verification, such as providing different answers to the user device based on the validity of the assertion.

Argumentary Verification Results Argumentative verification is evaluated based on argumentation detection (by linguistic means) and then verification (logical means). A dataset of 623 legal cases scraped from Landlord vs Tenant (2018) is formed. Each year, this website provides summaries of over 700 recent landlord-tenant court cases and agency decisions. Landlord v. Tenant covers over a dozen courts and agencies, including the New York Civil Court, the New York City Department of Housing and Community Redevelopment (DHCR), the New York City Environmental Board of Trustees, and others. This website allows users to gain access to their dynamic database of cases dating back to 1993, as well as the New York Landlord v. Tenant Newsletter Archive, and to perform searches of specified case summaries. The full judgment and comments are also available from this source.

A typical case summary reads: "Tenants complained of a reduction in building-wide services. They said that the building super didn't make needed repairs as requested and that landlord had refused to perform repairs in their apartment. They also complained about building accessibility issues. Among other things, the building side door walkway was reconstructed and made narrower. This made it hard to navigate a wheelchair through that doorway. The DRA ruled against tenants, who appealed and lost." Tenants complained of reduced services.The tenants stated that the building manager did not make the necessary repairs as requested, and that the landlord refused to make repairs in their apartments.The residents also complained of accessibility problems to the building.In particular, the side door passages of the building were modified and narrowed.This made it difficult to pass a wheelchair through the doorway.The DRA dismissed the resident's complaint, and the resident appealed but lost.Firstly, We extract sentences containing arguments and then try to find the assertions conveyed from the DeLP ontology. The assertion to be verified in the above example is that repair_is_done. We then test this claim. We get claim validity values from tags on the web page assigned by the judge who heard the case, such as rent_reduction_denied. Table 12 below shows the evaluation results conveyed along with the discussion in the landlord-tenant case text.

For the argumentation detection task, we use this landlord vs. tenant as the positive training set. As negative data sets, we use various text sources that should not contain argumentative or opinion data. We also used components of the dataset (Lee, 2001), including Wikipedia, factual news sources, and sections of the corpus such as ['tells'], instructions on how to use software; ['tele'], instructions on how to use hardware, and [news], presentation of news articles in an objective and independent manner. Further details on the negative, non-controversial data set are available in (Galitsky et al 2018 and Chapter 10).

The baseline argument detection approach relies on keywords and syntactic features to detect arguments (Table 13.8). Often a coordinated pair of communicative behaviors (thus at least one having a negative emotional polarity associated with the opponent) is a hint that a logical argument exists. This naive approach is 29% worse than the top performing TK-learned CDT approach. The SVM TK of CDT outperforms the SVM TK of RST+CA and RST+full partree (Galitsky, 2017) by about 5% due to noisy syntactic data that is frequently redundant for argument detection.

The SVM TK approach provides an acceptable F-measure, but does not help explain how the emotional argumentation identification problem is solved exactly, giving only the final scoring and class labels. The nearest neighbor maximum common subgraph algorithm is much more informative in this regard (Galitsky et al., 2015). Comparing the bottom two rows, we observe that it is possible, but rare, to express emotional arguments without CA.

Evaluating the logical arguments extracted from the text, we were concerned with cases where one author provided an invalid, inconsistent, and self-contradictory case. This is important for chatbots as a front-end to CRM systems that focus on customer retention and facilitate customer communication (Galitsky et al., 2009). A domain of residential real estate complaints was selected and a DeLP thesaurus was constructed for this domain. Automated grievance systems can be essential in decision support procedures, for example for property management companies (Constantinos et al., 2003).

In our plausibility assessment, we focus on target features regarding how a given complaint needs to be addressed, such as compensation_required, proceed_with_eviction, rent_receipt.

Validity evaluation results are shown in Table 13. The first and second rows show the results of the simplest complaint with a single CA indicating a single rhetorical relation such as contrast and an extracted argument-aggressive relation, respectively. In the third and fourth rows, correspondingly, we show the verification results for legal cases with two non-default rhetorical relations and two CAs of disagreement type. Row 5 evaluates complaints of average complexity, and the bottom row evaluates the most complex and longer complaints for their CDT. The third column shows the detection accuracy of invalid arguments in complaints in the stand-alone argumentation verification system. Finally, the fourth column shows the accuracy of the integrated argument extraction and verification system.

In our plausibility assessment, we focus on target features (claims) about what kind of eviction needs to be issued, such as compensation_required, proceed_with_eviction, rent_receipt. A system decision is determined by whether an identified claim is plausible, and if plausible, the ruling upholds this claim, and if not plausible, decides against it.

In these results, the reproducibility is low because in the majority of cases the invalidity of the claims is due to factors other than self-destruction. Accuracy is relatively high because if a logical flaw in the argument is established, other factors besides the argument (such as false facts) also contribute, so the whole claim is likely invalid. As the complexity of the complaint and its discourse tree increases, F1 first improves because more logical terms are available and then degrades because there is a higher chance of inference error due to noisier input.

It is important to maintain a low false positive rate in decision support systems. Missing invalid complaints is acceptable, but the confidence should be fairly high for invalid complaints that are detected. If a human agent is encouraged to view a given complaint as invalid, the agent's expectations should be met in most cases. Although the F1 value of the overall argumentation detection and verification system is low compared to modern recognition systems, it is still considered usable as a component of a CRM decision support system.

Syntax Generation Some of the techniques discussed herein, including the improved discourse parser, can use syntax generalization. Performing syntactic generalization of two sentences includes identifying words in each of the sentences and/or identical parts of speech (POS) in each of the sentences. Headwords refer to words that do not have associated part-of-speech information. If two words have different lemmas but the part of speech of each word is the same, then the part of speech is part of the generalization result. If the lemma is the same but the part of speech is different, it is the heading part of the generalized result.

To illustrate this concept, consider two examples of natural language expressions. The meaning of each expression is represented by a logical expression. Integrations and anti-integrations of these formulas are constructed. Some words (entities) map to predicates, some to their arguments, and some others do not occur explicitly in the logical form representation, but indicate the above instantiations of predicates with arguments.

Consider the following two sentences "camera with digital zoom" and "camera with zoom for beginners". The following logical predicates are used to express meaning.

camera(name_of_feature, type_of_users) and
zoom(type_of_zoom).

Note that this is a simplified example and thus may have a reduced number of arguments compared to more typical examples. Continuing with the example, the above expression can be expressed as:
camera(zoom(digital), AnyUser),
camera(zoom(AnyZoom), beginner)
According to the notation, variables (dematerialized values not specified in the NL expression) are capitalized. Given the above pair of equations, the integration computes their most general specialization camera(zoom(digital), beginner) and the anti-integration computes their most specific generalization camera(zoom(AnyZoom), AnyUser).

At the syntactic level, these expressions receive two noun phrase generalizations (‘^’) as {NN-camera, PRP-with, [digital], NN-zoom [for beginners]}. Expressions in square brackets are removed because they occur in one expression but not in the other. The result is obtain{NN-camera, PRP-with, NN-zoom]}, the syntactic analog of the semantic generalization.

The goal of abstract generalization is to find commonalities between parts of text at various semantic levels. Generalization operations are performed at one or more levels. Examples of levels are paragraph level, sentence level, phrase level and word level.

At each level (except word level), at an individual word, the result of the generalization of the two expressions is a set of expressions. In such a set, for each pair of expressions in which one is less common than the other, the latter is eliminated. A generalization of two sets of representations is a set of sets that is the result of a pairwise generalization of these representations.

There is only one generalization for a pair of words: if the words are the same in the same form, the result is the node that has this word in this form. To involve the word2vec model, the following rules are used to compute two different word generalizations. If subject1=subject2 then subject1^subject2 = <subject1, POS(subject1), 1>. otherwise subject1^subject2 =<*,POS(subject1), word2vecDistance(subject1^subject2)> if they have the same part of speech. If the parts of speech are different, the generalization is an empty tuple. It cannot be generalized further.

For a pair of phrases, the generalization includes all the maximal ordered sets of generalized nodes for the words in the phrase such that the order of the words is preserved. In the example below,
"To buy digital camera today, on Monday."
"Digital camera was a good buy today, first Monday of the month."
The generalization is {<JJ-digital, NN-camera>,<NN-today,ADV,Monday>}, where noun phrase generalizations are followed by adverbial phrase generalizations. The verb buy is excluded from both generalizations because it occurs in a different order in the above phrases. Buy-digital-camera is not a generalized phrase because buy occurs in a different sequence with other generalized nodes.

Improved Discourse Parser Certain aspects relate to an improved discourse parser. Predicting the rhetorical relationship between two sentences is a goal of discourse analysis along with text segmentation (dividing a sentence into basic discourse units). A document can be analyzed as a sequence of hierarchical discourse structures, but the issue in discourse coherence is how rhetorical relationships are signaled by the source text (and can be identified by the parser). For example, rhetorical relations are often signaled by discourse markers such as and, because, however and while, and when relations contain such markers they are sometimes classified as explicit relations. Discourse markers are reliable coherence relation signals.

However, existing discourse parsers use machine learning techniques and datasets, which can be difficult and time consuming to scale. The task of discourse prediction is already complex, compounded by the time-consuming nature of extending these annotated datasets. Therefore, many available discourse parsers assign elaboration and join relations where other more descriptive rhetorical relations are more suitable. Therefore, the reproducibility of establishing other more specific rhetorical relationships can be relatively low.

However, existing discourse parsers can be improved by performing additional analysis of the discourse tree, eg semantic analysis, and adjusting the discourse tree accordingly. In the examples discussed below, the discourse parser is applied to the text and any resulting elaboration or joint rhetorical relation can be replaced with a better rhetorical relation obtained by using semantic analysis (e.g., abstract semantic representation) patterns, if available. In general, this approach applies to intra-sentence rhetorical relations.

FIG. 44 shows a discourse tree and a semantic tree, according to one aspect. FIG. 44 shows discourse tree 4400 and semantic tree 4410 .

Discourse tree 4400 and semantic tree 4410 each represent the following text: "It was a question of life or death for me: I had scarcely enough drinking water to last a week."

The discourse tree 4410 is represented in text-based form as follows (indentation refers to the level of nesting in the tree):
detailed
TEXT: It was a question of life or death for me:
detailed
TEXT: I had scarcely enough drinking water
TEXT: to last a week.
As can be seen from the discourse tree 4410, the second elaboration relation 4412 generated by the discourse parser and relating to "I had scarcely enough drinking water" and "to last a week" is not as accurate as possible for the text, because "to last a week" is simply more than the elaboration of "I had scarcely enough drinking water". This is improved by leveraging AMR relationships in semantic tree 4420 . The semantic tree 4410 is also presented below in text-based form.

As seen in semantic tree 4420 , the semantic relation “purpose” identified as relation 4422 has a semantic role associated with the verb drink identified as role 4424 . Since discourse tree 4410 and semantic tree 4420 have the common entity “drink”, we can identify the kernel EDU (“I had scarcely enough drinking water”) that has drink in discourse tree 4400 .

The higher the number of common entities between the discourse tree and the semantic tree template, the better the match for improving rhetorical relations. Continuing the example, a satellite EDU ("to last a week") is identified and linked with the rhetorical relation "Detailed". Finally, 'Detailed' is replaced with 'Purpose' for a more accurate discourse tree. This link is shown as link 4430 .

Exploiting semantic information improves the discourse tree in certain situations, such as when there is a lack of discourse markers, when discourse markers are ambiguous or misleading, or when deeper semantic representations of sentences imply specific rhetorical relationships, such as AMR. Once the syntactic similarities between the text being parsed and the AMR patterns are established, the semantic roles from the AMR verbs can be interpreted as respective rhetorical relations at the discourse level. This mapping between semantic relationships and specific rhetorical relationships in AMR is established regardless of how connected EDUs, nuclear EDUs and satellite EDUs are connected.

As a result of manual generalization of the available AMR annotations, a mapping between AMR semantic and rhetorical relations was developed and is shown in Table 14 below. Table 14 shows examples of semantic roles and corresponding rhetorical relationships. In Table 14, the first column lists the rhetorical relation to be detected. The second column represents the mapping of AMR semantic relations to rhetorical relations. The third column provides exemplary sentences that are rematched to the sentence being parsed rhetorically. The fourth column shows the AMR parsing of the template.

A list of rhetorical relations is considered to create this mapping of rhetorical relations to semantic roles, which can be performed offline (eg, before runtime). For each rhetorical relation, a set of AMR annotations for a particular semantic relation is determined. Once systematic correlations are identified, corresponding mappings, represented by entries in Table 14, are created. Table 14 shows the rhetorical relationships elaborated by the AMR examples.

Table 15 below provides an example of a fragmented discourse tree whose elaboration is turned into specific relationships. Build and subdivide templates. The template shows the detected rhetorical-related aspects in bold. The second example shows the actual refinement when applying the template from the second row from the bottom and turning the specification into a concession. A syntactic generalization between this template and the sentence is also shown.

Syntactic similarities are established between the elaboration core and satellite basic discourse units and templates to replace the rhetorical relationships of the elaboration with those obtained by manual tagging of AMR. If such similarity is high (patterns from the AMR dataset have been parsed), the elaboration can be overwritten with high confidence. The higher the syntactic similarity score, the higher the confidence that the semantic role derived from the pattern accurately describes the rhetorical relationship. In the absence of extensive mappings to sufficient AMR pattern data and rhetorical relation data, formal learning of such mappings is difficult. Therefore, this similarity score threshold is used.

Table 16 shows co-occurrence values and percentages for lexical, syntactic and semantic correlations with rhetorical relations. This data helps improve the scoring of "and" and "as" (which are usually ignored due to syntactic generalization), as well as while, however, because, because they usually have very low scores.

Another example of using semantic relationships and roles to improve rhetorical relationships is shown in FIG.
FIG. 45 shows a discourse tree and a semantic tree, according to one aspect. FIG. 45 shows discourse tree 4510 and semantic tree 4520 . Discourse tree 4510 represents the text "I ate the most wonderful hamburger that she had ever bought for me." Semantic tree 4520 does not represent the same text as discourse tree 4510 . Rather, semantic tree 4520 represents the text of a template that is a good match with the text of discourse tree 4510 and that can be used to improve discourse tree 4510 .

The discourse tree 4510 is represented in text form as follows:
detailed
TEXT: I ate the most wonderful hamburger
TEXT: that she had ever bought for me.
As can be seen from discourse tree 4510, the two basic discourse units "I ate the most wonderful hamburger" and "that she had ever bought for me." are connected by the rhetorical relation "explain". Therefore, discourse tree 4510 is a good candidate for improvement, as "elaborate" may not be the most correct rhetorical relation.

The AMR semantic role of compared-to is mapped to the rhetorical relation of comparison. Default discourse parsers provide renditions that can be turned into more accurate rhetorical relationships when EDU pairs with default rhetorical relationships are semantically similar to templates with specific semantic relationships that can be mapped to rhetorical relationships. To establish the correct rhetorical relationship between EDUs in a sentence, a match is attempted against templates found in a set of semantic templates (eg, AMR repository). The template matched is for the sentence "It was the most magnificent and stately planet that he had ever seen."

The EDU and template are aligned and generalized to match the EDU pair being parsed with the template. In this case, the syntactic generalization between EDU pairs and templates becomes [VB-* DT-the RBS-most JJ-(wonderful ^magnificent) IN-that PRP-she VB-had RB-ever VB-*] and there is significant evidence that the sentences and patterns being parsed share a common syntactic structure. For example, wonderful ^magnificent produces an abstract adjective with the meaning of what is common among these adjectives. Connection 4530 indicates the correspondence between the adjective magnificent in the AMR representation and the adjective wonderful in the original DT.

Therefore, the elaboration of discourse tree 4500 is replaced with a rhetorical relation of type "compare". The corrected discourse tree is as follows:
comparison
TEXT: I ate the most wonderful hamburger
TEXT: that she had ever bought for me.
Based on the example above, the process of improving the discourse tree is further described.

FIG. 46 is a flowchart of an exemplary process 4600 for generating an improved discourse tree, according to one aspect. It should be appreciated that in some cases one or more actions in process 4600 may not be performed. Process 4600 may be performed by application 102 .

At block 4602, process 4600 includes creating a discourse tree from the text by identifying basic discourse units within the text. At block 4602 , process 4600 includes operations substantially similar to block 1502 of process 1500 . The determined discourse tree contains nodes. Each non-terminal node of the node represents a rhetorical relationship between two basic discourse units, and each terminal node of the nodes of the discourse tree is associated with a basic discourse unit.

At block 4604, the process 4600 includes identifying rhetorical relationships of type elaboration or joints in the discourse tree. A rhetorical relation relates two base discourse units, e.g., a first base discourse unit and a second base discourse unit (rather than associating two other rhetorical relations or one rhetorical relation and one base discourse unit).

The first elementary discourse unit and the second elementary discourse unit form a reference sentence. For example, referring back to Figure 45, the first EDU is "I ate the most wonderful hamburger", the second EDU is "that she had ever bought for me", and the rhetorical relation (before the update) is "detail".

At block 4606, process 4600 includes determining a syntactic generalization score for each candidate sentence in the set of candidate sentences. As described above in Tables 14 and 15, each candidate sentence has a corresponding semantic relationship (eg, AMR expressions). In a simplified example, the syntactic generalization score is the number of common entities between the reference sentence and the candidate sentence. Each common entity shares a common part of speech between the candidate sentence and the reference sentence. However, the syntactic generalization score can be calculated differently in other aspects, as described below.

The goal of abstract generalization is to find commonalities between parts of text at various semantic levels. Generalization can be done at the paragraph, sentence, EDU, phrase, and individual word levels. The result of the generalization of the two expressions, except at the word level, is a set of expressions. In such a set, for each pair of expressions where one is less common than the other, the latter is eliminated. A generalization of two sets of representations is a set of representation sets that is the result of a pairwise generalization of these representations. For illustrative purposes, FIG. 46 will be discussed with respect to FIG. 47 showing generalization and FIG. 48 showing alignment.

FIG. 47 illustrates generalization of sentences and templates with known semantic relationships, according to one aspect. FIG. 47 shows a generalization of sentence 4710 "If you read a book at night, your knowledge will improve" and template 4720 "If one gets lost in the night, such knowledge is valuable" from Table 14. The resulting generalization 4730 is as follows:
[IN-If PRP-* VB-* ... NN-night ... NN-knowledge ]
In this template, IN-If PRP-* VB-* is the signature of the semantic relation of :condition() and also the discourse relation of the condition, but there happen to be more general words that may or may not be used to establish an affinity between the sentence and the template, such as "NN-night ... NN-knowledge".

To determine how to calculate an appropriate generalization score, we performed a computational study to determine the part-of-speech weights that yielded the most accurate similarity measures between sentences. The problem was formulated as finding the optimal weights for nouns, adjectives, verbs, and their forms (such as gerund and past tense) to maximize the resulting search relevance. Search relevance was measured as the deviation in search result order from the best for a given query, and the current search order was determined based on the generalization score for a given set of POS weights (with other generalization parameters fixed). This optimization performed results in W _NN =1.0, W _JJ =0.32, W _RB =0.71, W _CD =0.64, W _VB =0.83, W _PRP =0.35, except for common frequent verbs like get, take, set, and put where W _VBcommon =0.57. W _<POS,*> is set to 0.2 (different word but same POS) and W _<*,word> = 0.3 (same word but occurs as different POS in two sentences). W _{{and,as,but,while,however,because}} is computed as a default value of 1 normalized to the values in the second column of Table 16. Note that the default syntactic generalization ignores discourse directives in most cases.

The generalization score between the reference sentence (ref_sentence) and the candidate template (Template) can then be expressed as the sum over the phrases of the weighted sum over the words Word _{ref_sentence} and word _template .

The greatest generalization can then be defined as the one with the highest score.
At the phrase level, generalization begins with finding alignments between two phrases (as many word correspondences as possible between two phrases). Alignment operations are performed such that phrase integrity is preserved. For example, two phrases can be aligned only if correspondence between their initial nouns is established. There are similar integrity constraints for aligning verbs, prepositions and other types of phrases.

FIG. 48 shows alignment between two sentences, according to one aspect. 48 shows the alignment between the sentence 4810 that reads "use the screw driver from this tool for fixing heaters," and the sentence 4820 that reads "get short screw driver holder for electric heaters."The resulting alignment 4830 is as follows:
VB-* JJ-* NN-zoom NN-* IN-for NN-*
In an aspect, separate kernel and satellite generalizations can be used to generate an improved discourse tree. For example, a discourse tree is created as in block 4602 of process 4600 . From the discourse tree, rhetorical relations are identified. Preferred rhetorical relations are identified. Examples of suitable rhetorical relationships include elaboration and joint innermost relationships, including nested relationships (details above another elaboration [above another elaboration]).

Nuclear EDUs and satellite EDUs are identified. These EDUs can be reduced in size and/or complexity if they are too complex or long. Nuclear EDU is generalized with each template (eg, Table 14 and/or Table 15). Similar to block 4608, the candidate sentence with the highest generalization score is selected. If the score is above the threshold, the satellite EDU corresponding to the rhetorical relation is generalized with the template. If the satellite EDU's generalization score is above a threshold, then the rhetorical relation is used to replace the rhetorical relation in the reference sentence. Examples of generalized thresholds are 2.0 (for nuclei) and 3.3 (for satellites).

Returning to Figure 46, at block 4608, the process 4600 includes selecting the candidate sentence with the highest syntactic generalization score of the syntactic generalization scores.

In some embodiments, no match is found. For example, application 102 searches an Abstract Semantic Representation (AMR) dataset (e.g., Table 14 and/or Table 15) to identify that the identified semantic relationship is not in the AMR dataset, and then replaces the rhetorical relationship in the discourse tree with the additional semantic relationship that is in the AMR dataset.

At block 4610, process 4600 includes identifying semantic relationships corresponding to the candidate sentences. Semantic relationships correspond to words in a candidate sentence and define roles in the candidate sentence. For example, semantic relationships in candidate sentences are identified in Table 14 and/or Table 15.

At block 4612, the process 4600 includes replacing the rhetorical relations in the discourse tree with updated rhetorical relations corresponding to the semantic relations, thereby creating an updated discourse tree. Rhetorical relationships that match the semantic relationships identified at block 4610 are identified. The identified rhetorical relation is inserted into the discourse tree in place of the rhetorical relation identified at block 4604 .

Levels of generalization: from syntactic to semantic to discourse To demonstrate how syntactic generalization rises from the syntactic level to the semantic level, we can follow Mill's method of direct correspondence (induction) applied to linguistic structures. The English philosopher JS Mills wrote in his 1843 book A System of Logic, "If two or more instances of the phenomenon under investigation have only one situation in common, then only the situation in which all the cases agree is the cause (or effect) of the given phenomenon."

Consider the linguistic property A of a phrase f. For A to be a prerequisite for some effect E, A must always be present in multiple phrases addressing E. In the realm of linguistics, A is the linguistic structure and E is its meaning. Therefore, the presence or absence of linguistic features considered as "possible prerequisites" in the sentence is verified. Clearly, any linguistic property A that is absent if the meaning E is present cannot be a prerequisite for this meaning E of a phrase.

For example, the matching method can be represented as the phrase f1 where the words {ABCD} occur with the meaning formally represented as <wxyz>. Consider also another phrase f2 in which the word {AEFG} occurs with the same meaning <wtuv> as in phrase f1. Now we obtain {A} by applying a generalization to the words {ABCD} and {AEFG} (here ignoring the syntactic structures of f1 and f2 for the sake of example). Thus, we now know that word A is the cause of w (has meaning w). Throughout this book, we consider the linguistic structures covering ABCD, in addition to this list itself, and apply the matching method.

Thus, (inductive) semantics can be generated that apply syntactic generalizations. Considering only the syntactic information of a sample, we cannot get the semantics, but generalizing two or more phrases (samples), we get not only the syntactic structure, but also the (inductive) semantic structure. Viewing syntactic generalization as an inductive cognitive procedure, we can formally define the transition from the syntactic level to the semantic level. In this work, we do not mix syntactic and semantic features for learning classes, instead we derive semantic features from syntax according to the inductive framework above.

Evaluating Improved Parsers False positives and false negatives are identified in the refinement of rhetorical relations in detail. To do so, we analyze the functionality of downstream applications of discourse parsing such as summarization, dialog management, and argumentation analysis.

If the specification is the correct relationship but is changed to a more specific relationship, the search results obtained may not match the query and some sentences may not occur in the result summary as the rules are adjusted for the specification. In order to make the search and summarization system less sensitive to the false positives obtained by our refinement system, the matching and selection rules need to be updated to take into account that causes, concessions, conditions are partial cases of recitation, and discourse trees with these node labels should be matched accordingly.

If the specification is not overridden with a more specific relationship, the accuracy of downstream systems is adversely affected. If a particular idiosyncratic relationship in the question is identified, it must be addressed in the answer, so selection of a particular answer that matches the question in style will fail if it remains detailed. Less relevant sentences or phrases could be included in the summary, or the number of candidates for such inclusion would be reduced. We conclude that false negatives are worse than false positives.

Another consideration concerns the genre of the training dataset for discourse parsing. The RST Discourse Treebank (RST-DT, Carlson et al., 2001) contains news articles and is not a good source for modeling textual structure in other genres such as fiction, scientific texts, engineering systems descriptions and legal documents. The application of discourse parsing in these areas is important. Therefore, even if elaboration is sufficient in news presentations, more specific structures are needed to model authors' reasoning in other genres and domains, such as professional texts.

Four problems are used to evaluate the developed discourse parser refinement (Table 16):
1) Search complex long questions (Section 2), implementing coordination between questions and answers. using the Yahoo!Answers dataset;
2) Find utterances that form a dialog from a text or document. We use the dataset and techniques developed in Chapter 1 of Volume 2. This technique relies heavily on discourse parsing;
3) Generate a dialog from the initial query (Volume 2, Chapter 1). Rhetorical relations are considered to filter unsuitable utterances (Galitsky and Ilvovsky 2016);
4) Evaluate the veracity of a document by its discourse structure (Volume 2, Chapter 6). The structure of the discourse tree, how authors present information, along with which rhetorical relationships, is important for detecting lies or fake news.

It can be observed that refinement of discourse parser results achieves over 4% improvement in utterance and text classification. However, in search problems, it is less sensitive to proper rhetorical relations, and we achieved improvements close to 3%. We now turn to an evaluation of how the individual rhetorical relational refinements work (Table 17). We use one AMR corpus for training and another AMR corpus for testing.

In lines 3 to 8, the relationship detection results for individual relationship types such as contrast are analyzed. The baseline parser recognizes contrasts along with causes and conditions and does not recognize the remaining relationships. This is due to the fact that these unique rare relationships are poorly represented in the Discourse TreeBank. Overall, only for these rhetorical relations, the performance of baseline classifiers weighted by the amount of these relations is rather poor. Learning from AMR, it is possible to achieve a performance of 77.7 on this data. Therefore, we get an improvement of 36.1 for the refinement of these 6 relations. This is important for tasks where the emphasis is on specific relation types rather than on the discourse tree structure.

Overall, since these six relations are rare, parser improvements result in an average improvement of up to 6% for all rhetorical relations. This is still valuable for discourse parser applications.

Related Work The six rhetorical relations that are the focus of discourse parser improvements are very important for retrieval, summarization, and some applications of text classification to abstract classes. For retrieval purposes, if a question involves any of these six relations, a good comprehensive answer cannot contain only the default rhetorical relations, but instead properly communicates with its author these specific relations in question. Such an answer cannot simply be a factual answer.
Contrasts need to be conveyed by presenting both the thesis and antithesis, possibly using argumentation patterns.
Causes must be communicated through one form or another of explanation, providing reasons and setting a framework for reasoning.
Conditions and comparisons must be addressed by communicating and presenting factual data in both (If and Then, or Item A vs. Item B) parts.

Therefore, if these six relationships in a user question are not properly recognized, it is likely that the wrong answers will be presented in the delivery manner, and the user will be frustrated with the search, even if it is relevant in terms of the queried entities and their attributes.

The Penn Discourse Treebank (PDTB, Prasad et al 2017)) is the primary resource for training discourse parsers. Version 3.0 is the third release in the Penn Discourse Treebank project, which aims to annotate the Wall Street Journal section of Treebank-2 using discourse relations. The Penn Discourse Treebank version 3 contains over 53,600 annotated relationship tokens. We normalized some pairwise annotations, impregnated them with new meanings, and performed consistency checks on the corpus. Further details regarding the development of PDTB are available (PDBT 2019). Since PDBR only includes the news genre, the performance of discourse parsers trained on it in other genres is limited. In this study, we attempted to resolve these limitations by using a source of semantic relations from which discourse relations can be deduced.

The PDTB project was inspired by the observation that discourse relations are based on an identifiable set of explicit words or phrases (discourse junctions, discourse directives) or simply the adjacency of two sentences. PTDB is used by many researchers in the field of natural language processing and, more recently, by researchers in psycholinguistics. It has also stimulated the development of similar resources in other languages and domains. There is a strong discrepancy between the complexity with which PDTB 3.0 is built and the complexity a discourse parser can handle. Discourse parsers trained on previous models perform very poorly in recognizing the very specific relationships used in PDTB 3.0. Therefore, the improvements described in this chapter are essential for downstream applications of discourse parsing.

Exemplary Computing System FIG. 49 is a simplified diagram illustrating a distributed system 4900 for implementing one of the aspects described above. In the illustrated aspect, distributed system 4900 includes one or more client computing devices 4902, 4904, 4906 and 4908 configured to run and operate client applications such as web browsers, proprietary clients (e.g., Oracle forms), etc., via one or more networks 4910. Server 4912 may be communicatively coupled to remote client computing devices 4902 , 4904 , 4906 and 4908 via network 4910 .

In various aspects, server 4912 may be adapted to run one or more services or software applications provided by one or more of the components of the system. A service or software application may include non-virtual and virtual environments. Virtual environments, whether two-dimensional or three-dimensional (3D) representations, page-based logical environments, etc., can include those used for virtual events, trade shows, simulators, classrooms, purchasing merchandise, and corporate activities. In some aspects, these services may be provided to users of client computing devices 4902, 4904, 4906 and/or 4908 as web-based or cloud services or under a Software as a Service (SaaS) model. Users operating client computing devices 4902, 4904, 4906 and/or 4908 may then utilize one or more client applications to interact with server 4912 and utilize the services provided by these components.

In the illustrated configuration, software components 4918 , 4920 and 4922 of system 4900 are shown implemented on server 4912 . Also, in other aspects, one or more of the components of distributed system 4900 and/or services provided by those components may be implemented by one or more of client computing devices 4902, 4904, 4906 and/or 4908. A user operating a client computing device may then utilize one or more client applications to use the services provided by these components. These components may be implemented in hardware, firmware, software, or a combination thereof. It should be appreciated that a variety of different system configurations are possible that may differ from distributed system 4900 . Accordingly, the aspects shown in the figures are an example of distributed systems for implementing systems of aspects and are not intended to be limiting.

Client computing devices 4902, 4904, 4906 and/or 4908 may be handheld mobile devices (e.g., iPhones, cell phones, iPads, computing tablets, personal digital assistants (PDAs)) or wearable devices (e.g., Google Glass head-mounted displays) that run software such as Microsoft Windows Mobile and/or iOS. , Windows Phone, Android, BlackBerry 10, Palm OS, Internet, email, short message service (SMS), BlackBerry®, or any other communication protocol available. The client computing device may be a general purpose personal computer, which includes, by way of example, personal computers and/or laptop computers running various versions of Microsoft Windows®, Apple Macintosh® and/or Linux® operating systems. The client computing device may be a workstation computer running any of a variety of commercially available UNIX or UNIX-like operating systems including, but not limited to, various GNU/Linux operating systems such as Google Chrome OS. Alternatively or additionally, client computing devices 4902, 4904, 4906, and 4908 may be thin client computers, Internet-enabled gaming systems (e.g., Microsoft X-Box game consoles with or without Kinect® gesture input devices), and/or other electronic devices such as personal messaging devices capable of communicating over network 4910.

Although exemplary distributed system 4900 is shown with four client computing devices, any number of client computing devices may be supported. Other devices, such as devices with sensors, may interact with server 4912 .

Network 4910 in distributed system 4900 may be any type of network familiar to those skilled in the art capable of supporting data communications using any of a variety of commercially available protocols, including but not limited to TCP/IP (Transmission Control Protocol/Internet Protocol), SNA (Systems Network Architecture), IPX (Internet Packet Exchange), AppleTalk, and the like. Solely by way of example, network 4910 may be a local area network (LAN), such as those based on Ethernet, Token Ring, or the like. Network 4910 may be a wide area network and the Internet. Network 4910 may include a virtual network, which may include a virtual private network (VPN), an intranet, an extranet, a public switched telephone network (PSTN), an infrared network, a wireless network (eg, a network operating under any of the Institute of Electrical and Electronics (IEEE) 802.28 suite of protocols, Bluetooth and/or other wireless protocols), and/or these. and/or other networks.

Server 4912 may consist of one or more general-purpose computers, dedicated server computers (including, by way of example, personal computer (PC) servers, UNIX servers, midrange servers, mainframe computers, rack-mounted servers, etc.), server farms, server clusters, or other suitable configurations and/or combinations. Server 4912 may include one or more virtual machines running virtual operating systems, or other computing architectures that include virtualization. One or more flexible pools of logical storage can be virtualized to maintain virtual storage devices for servers. The virtual network can be controlled by server 4912 using software defined networking. In various aspects, server 4912 may be adapted to run one or more services or software applications described in the above disclosures. For example, server 4912 may correspond to a server for performing the above-described processes according to aspects of this disclosure.

Server 4912 may run an operating system, including any of those listed above, as well as any commercially available server operating system. Server 4912 may also run any of a variety of additional server and/or middle-tier applications, including hypertext transport protocol (HTTP) servers, file transfer protocol (FTP) servers, common gateway interface (CGI) servers, JAVA servers, database servers, and the like. Exemplary database servers include, but are not limited to, those commercially available from Oracle, Microsoft, Sybase, International Business Machines, and the like.

In some implementations, server 4912 may include one or more applications for analyzing and aggregating data feeds and/or event updates received from users of client computing devices 4902 , 4904 , 4906 and 4908 . By way of example, data feeds and/or event updates may include, but are not limited to, Twitter feeds, Facebook updates, or real-time updates received from one or more third-party sources and continuous data streams, and may include real-time events associated with sensor data applications, financial tickers, network performance measurement tools (e.g., network monitoring and traffic management applications), clickstream analysis tools, automotive traffic monitoring, and the like. Server 4912 may also include one or more applications for displaying data feeds and/or real-time events via one or more displays of client computing devices 4902 , 4904 , 4906 and 4908 .

Distributed system 4900 may also include one or more databases 4914 and 4916 . Databases 4914 and 4916 may reside in various locations. As an example, one or more of databases 4914 and 4916 may reside on non-transitory storage media local to (and/or residing on) server 4912 . Alternatively, databases 4914 and 4916 may be remote from server 4912 and communicate with server 4912 via network-based or dedicated connections. In one set of aspects, databases 4914 and 4916 may reside in a storage-area network (SAN). Similarly, any necessary files to perform functions attributed to server 4912 may be stored locally and/or remotely on server 4912 as appropriate. In one set of aspects, databases 4914 and 4916 may comprise relational databases, such as those provided by Oracle Corporation, adapted to store, update and retrieve data in response to SQL formatted commands.

FIG. 50 is a simplified block diagram of one or more components of a system environment 5000 in which services provided by one or more components of one aspect system can be provided as cloud services, according to one aspect of the present disclosure. In the depicted aspect, the system environment 5000 includes one or more client computing devices 5004, 5006 and 5008 that can be used by users to interact with a cloud infrastructure system 5002 that provides cloud services. A client computing device may be configured to run a client application, such as a web browser, proprietary client application (e.g., Oracle Forms), or other application that may be used by a user of the client computing device to interact with the cloud infrastructure system 5002 to use the services provided by the cloud infrastructure system 5002.

It should be appreciated that the illustrated cloud infrastructure system 5002 may have other components than those shown. Additionally, the aspects shown in the figures are only one example of a cloud infrastructure system that may incorporate aspects of the present invention. In some other aspects, the cloud infrastructure system 5002 may have more or fewer components than shown, may combine two or more components, or may have different configurations or arrangements of components.

Client computing devices 5004 , 5006 and 5008 may be devices similar to those described above for 4902 , 4904 , 4906 and 4908 .

Although the exemplary system environment 5000 is shown with three client computing devices, any number of client computing devices may be supported. Other devices, such as devices having sensors and the like, may interact with cloud infrastructure system 5002 .

Network 5010 may facilitate communication and exchange of data between clients 5004 , 5006 and 5008 and cloud infrastructure system 5002 . Each network may be any type of network familiar to those skilled in the art capable of supporting data communication using any of a variety of commercially available protocols, including those described above for network 4910.

Cloud infrastructure system 5002 may comprise one or more computers and/or servers, which may include those described above for server 4912 .

In certain aspects, the services provided by the cloud infrastructure system may include a number of services available on demand to users of the cloud infrastructure system, such as online data storage and backup solutions, web-based email services, hosted office suites and document collaboration services, database processing, administrative technical support services, and the like. Services provided by a cloud infrastructure system can be dynamically scaled to meet the needs of its users. A concrete instantiation of a service provided by a cloud infrastructure system is referred to herein as a "service instance." Generally, any service available to users from a cloud service provider's system over a communication network such as the Internet is referred to as a "cloud service." Typically, in public cloud environments, the servers and systems that make up a cloud service provider's system are different from the customer's own on-premise servers and systems. For example, a cloud service provider's system may host the application, and users may order and use the application on demand over a communication network such as the Internet.

In some examples, services in a computer network cloud infrastructure may include storage, hosted databases, hosted web servers, secure computer network access to software applications, or other services provided to users by cloud vendors or otherwise known in the art. For example, a service may include password-protected access to remote storage on the cloud over the Internet. As another example, a service may include a web services-based hosted relational database and scripting language middleware engine for private use by networked developers. As another example, a service may include access to an email software application hosted on a cloud vendor's website.

In certain aspects, cloud infrastructure system 5002 may include a set of application, middleware, and database service offerings delivered to customers in a self-service, subscription-based, elastically scalable, reliable, highly available, and secure manner. An example of such a cloud infrastructure system is the Oracle Public Cloud provided by the Assignee.

Large amounts of data, sometimes referred to as big data, can be hosted and/or manipulated by infrastructure systems at multiple levels and at different scales. Such data may contain data sets that are so large and complex that they can be difficult to process using typical database management tools or conventional data processing applications. For example, terabytes of data may be difficult to store, retrieve, and process using personal computers or their rack-based counterparts. Data of this size can be difficult to work with modern relational database management systems and desktop statistical and visualization packages. They can go beyond the fabric of commonly used software tools and require massively parallel processing software running thousands of server computers to capture, curate, manage and process data within an acceptable elapsed time.

In order to visualize, detect trends, and/or interact with large amounts of data, analysts and researchers can store and process extremely large data sets. Dozens, hundreds or thousands of processors linked in parallel can operate on such data, thereby displaying such data or simulating external forces on the data or what it represents. These datasets may require structured data such as data organized in databases or according to a structured model and/or unstructured data (e.g. emails, images, data blobs (binary large objects), web pages, complex event processing). By enhancing an aspect's ability to focus more (or fewer) computing resources on a target relatively quickly, cloud infrastructure systems are better available to perform tasks on large data sets based on demand from businesses, government agencies, research organizations, private individuals, groups of like-minded individuals or organizations, or other entities.

In various aspects, cloud infrastructure system 5002 may be adapted to automatically provision, manage and track customer subscriptions to services provided by cloud infrastructure system 5002 . Cloud infrastructure system 5002 may provide cloud services via various deployment models. For example, cloud infrastructure system 5002 may be owned by an organization that sells cloud services (e.g., owned by Oracle Corporation), and serviced under a public cloud model in which the service is available to the general public or various industrial enterprises. As another example, cloud infrastructure system 5002 may be operated solely for a single organization and service may be provided under a private cloud model, which may serve one or more entities within that organization. Cloud services may also be provided under a community cloud model in which the cloud infrastructure system 5002 and the services provided by the cloud infrastructure system 5002 are shared by several organizations within an associated community. Cloud services may also be provided under a hybrid cloud model, which is a combination of two or more different models.

In some aspects, the services provided by cloud infrastructure system 5002 may include one or more services provided under a Software as a Service (SaaS) category, a Platform as a Service (PaaS) category, an Infrastructure as a Service (IaaS) category, or other categories of services, including hybrid services. A customer may order one or more services provided by the cloud infrastructure system 5002 through a subscription order. The cloud infrastructure system 5002 then performs processing to provide the service on the customer's subscription order.

In some aspects, services provided by cloud infrastructure system 5002 may include, but are not limited to, application services, platform services, and infrastructure services. In some examples, application services may be provided by a cloud infrastructure system via a SaaS platform. A SaaS platform may be configured to provide cloud services that fall into the SaaS category. For example, a SaaS platform may provide functionality for building and delivering a suite of on-demand applications on a unified development and deployment platform. A SaaS platform may manage and control the underlying software and infrastructure for providing SaaS services. By using the services provided by the SaaS platform, customers can take advantage of applications running on cloud infrastructure systems. Customers can acquire application services without the need for customers to purchase separate licenses and support. A variety of different SaaS services may be offered. Examples include, but are not limited to, services that provide solutions for sales performance management, enterprise consolidation and business flexibility for large organizations.

In some aspects, platform services may be provided by a cloud infrastructure system via a PaaS platform. A PaaS platform may be configured to provide cloud services that fall into the PaaS category. Examples of platform services include, but are not limited to, services that enable an organization (such as Oracle) to integrate existing applications on a shared common architecture, and the ability to build new applications that leverage shared services provided by the platform. A PaaS platform may manage and control the underlying software and infrastructure for providing PaaS services. Customers can acquire PaaS services provided by cloud infrastructure systems without the need for customers to purchase separate licenses and support. Examples of platform services include, but are not limited to, Oracle Java Cloud Service (JCS), Oracle Database Cloud Service (DBCS), and the like.

By using the services provided by the PaaS platform, customers can take advantage of the programming languages and tools supported by the cloud infrastructure system and also control the deployed services. In some aspects, platform services provided by the cloud infrastructure system may include database cloud services, middleware cloud services (eg, Oracle Fusion middleware services) and Java cloud services. In one aspect, the database cloud service may support a shared service deployment model that allows an organization to pool database resources and provide database-as-a-service to customers in the form of a database cloud. The middleware cloud service may provide a platform for customers to develop and deploy various business applications on the cloud infrastructure system, and the Java cloud service may provide a platform for customers for deploying Java applications on the cloud infrastructure system.

Various different infrastructure services may be provided by the IaaS platform in the cloud infrastructure system. Infrastructure services facilitate the management and control of basic computing resources such as storage, networks, and other basic computing resources for customers who utilize services provided by SaaS and PaaS platforms.

In certain aspects, cloud infrastructure system 5002 may also include infrastructure resources 5030 for providing resources used to provide various services to customers of the cloud infrastructure system. In one aspect, infrastructure resources 5030 may include a pre-integrated optimal combination of hardware such as servers, storage and networking resources for running the PaaS platform and services provided by the SaaS platform.

In some aspects, resources in the cloud infrastructure system 5002 may be shared by multiple users and dynamically reallocated on a per demand basis. Also, resources may be allocated to users at different times. For example, the cloud infrastructure system 5030 may allow a first set of users in a first time zone to utilize resources of the cloud infrastructure system for a defined period of time, and may allow reallocation of the same resources to another set of users located in different time zones, thereby maximizing resource utilization.

In certain aspects, a number of internal shared services 5032 shared by various components or modules of cloud infrastructure system 5002 and services provided by cloud infrastructure system 5002 may be provided. These internal shared services may include, but are not limited to, security and identity services, integration services, enterprise repository services, enterprise manager services, virus scanning and whitelisting services, high availability and backup and recovery services, services to enable cloud support, email services, notification services, file transfer services, and the like.

In certain aspects, cloud infrastructure system 5002 may provide comprehensive management of cloud services (eg, SaaS, PaaS and IaaS services) in the cloud infrastructure system. In one aspect, cloud management functions may include functions for provisioning, managing, tracking, etc. customer subscriptions received by cloud infrastructure system 5002 .

In one aspect, as shown, cloud management functionality can be provided by one or more modules such as order management module 5020, order orchestration module 5022, order provisioning module 5024, order management and monitoring module 5026, and identity management module 5028. These modules may include or be provided with one or more computers and/or servers, which may be general purpose computers, dedicated server computers, server farms, server clusters, or other suitable configurations and/or combinations.

At exemplary operation 5034, a customer using a client device, such as client device 5004, 5006, or 5008, may interact with cloud infrastructure system 5002 by requesting one or more services provided by cloud infrastructure system 5002 and placing an order for subscriptions to one or more services provided by cloud infrastructure system 5002. In certain aspects, a customer may access a cloud User Interface (UI), namely cloud UI 5012, cloud UI 5014 and/or cloud UI 5016, and place a subscription order through these UIs. Order information received by cloud infrastructure system 5002 in response to a customer placing an order may include information identifying the customer and one or more services provided by cloud infrastructure system 5002 to which the customer is to subscribe.

After the order is placed by the customer, the order information is received via cloud UI 5050, 5014 and/or 5016.

At operation 5036 the order is stored in order database 5018 . Orders database 5018 may be one of several databases operated by cloud infrastructure system 5018 and operated in conjunction with other system elements.

At operation 5038 the order information is transferred to order management module 5020 . In some examples, order management module 5020 may be configured to perform order-related billing and accounting functions, such as order confirmation and order reservation upon confirmation.

At operation 5040 information about the order is communicated to order orchestration module 5022 . Order orchestration module 5022 may utilize order information to orchestrate the provisioning of services and resources for orders placed by customers. In some examples, order orchestration module 5022 may orchestrate the provisioning of resources to support services subscribed to using the services of order provisioning module 5024 .

In certain aspects, order orchestration module 5022 enables management of business processes associated with each order and applies business logic to determine whether an order should proceed to provisioning. At operation 5042, upon receiving an order for a new subscription, order orchestration module 5022 sends a request to order provisioning module 5024 to allocate resources and configure those resources needed to fulfill the subscription order. Order provisioning module 5024 enables allocation of resources for services ordered by customers. Order provisioning module 5024 provides a level of abstraction between the cloud services provided by cloud infrastructure system 5000 and the physical implementation layer used to provision the resources to provide the requested service. Thus, the order orchestration module 5022 can be decoupled from implementation details such as whether services and resources are actually provisioned on the fly, or pre-provisioned and only allocated/assigned on demand.

At operation 5049 , once the services and resources have been provisioned, a notification of the services provided may be sent by the order provisioning module 5024 of the cloud infrastructure system 5002 to the customer on the client computing device 5004 , 5006 and/or 5008 .

At operation 5051 , the customer's subscription orders may be managed and tracked by order management and monitoring module 5026 . In some examples, the order management and monitoring module 5026 may be configured to collect usage statistics about the services in the subscription order, such as the amount of storage used, amount of data transferred, number of users, and amount of system uptime and system downtime.

In certain aspects, cloud infrastructure system 5000 may include identity management module 5028 . Identity management module 5028 may be configured to provide identity services such as access management and authorization services in cloud infrastructure system 5000 . In some aspects, identity management module 5028 may control information about customers who want to use services provided by cloud infrastructure system 5002 . Such information may include information authenticating the identity of such customers and information describing what actions those customers are authorized to perform on various system resources (e.g., files, directories, applications, communication ports, memory segments, etc.). Identity management module 5028 may also include management of descriptive information about each customer and how and by whom this descriptive information may be accessed and changed.

FIG. 51 illustrates an exemplary computer system 5100 upon which various aspects of the invention can be implemented. Computer system 5100 may be used to implement any of the computer systems described above. As shown, computer system 5100 includes a processing unit 5104 that communicates with several peripheral subsystems via bus subsystem 5102 . These peripheral subsystems may include processing acceleration unit 5106 , I/O subsystem 5108 , storage subsystem 5118 and communication subsystem 5124 . Storage subsystem 5118 includes tangible computer-readable storage media 5122 and system memory 5110 .

Bus subsystem 5102 provides a mechanism for causing the various components and subsystems of computer system 5100 to communicate with each other as intended. Although bus subsystem 5102 is shown schematically as a single bus, alternate aspects of the bus subsystem may utilize multiple buses. Bus subsystem 5102 may be any of several types of bus structures including memory buses or memory controllers, peripheral buses, and local buses using any of a variety of bus architectures. For example, such architectures include Industry Standard Architecture (ISA) buses, Micro Channel Architecture (MCA) buses, Enhanced ISA (EISA) buses, Video Electronics Standards Association (VESA) local buses and Peripheral Component Interconnect (PCI) buses, which can be implemented as mezzanine buses manufactured to the IEEE P3086.1 standard. can include

A processing unit 5104 , which may be implemented as one or more integrated circuits (eg, a conventional microprocessor or microcontroller), controls the operation of computer system 5100 . Processing unit 5104 may include one or more processors. These processors may include single-core or multi-core processors. In particular aspects, processing unit 5104 may be implemented as one or more independent processing units 5132 and/or 5134, with single-core or multi-core processors included in each processing unit. Also, in other aspects, processing unit 5104 may be implemented as a quad-core processing unit formed by incorporating two dual-core processors into a single chip.

In various aspects, the processing unit 5104 may execute various programs in response to program code and may maintain multiple programs or processes executing simultaneously. At any given moment, some or all of the program code to be executed may reside in processing unit 5104 and/or storage subsystem 5118 . Through suitable programming, processing unit 5104 may provide the various functions described above. Computer system 5100 may also additionally include a processing acceleration unit 5106, which may include a digital signal processor (DSP), special purpose processor, or the like.

The I/O subsystem 5108 may include user interface input devices and user interface output devices. User interface input devices may include keyboards, pointing devices such as mice or trackballs, touch pads or touch screens, which are incorporated into displays, scroll wheels, click wheels, dials, buttons, switches, keypads, audio input devices, as well as voice command recognition systems, microphones and other types of input devices. User interface input devices may include, for example, motion sensing and/or gesture recognition devices such as the Microsoft Kinect® motion sensor that allow a user to control and interact with an input device such as the Microsoft Xbox® 360 game controller through a natural user interface using gestures and spoken commands. The user interface input device may also include an eye gesture recognizer, such as the Google Glass® Blink Detector, which detects eye movements from the user (e.g., "blinking" while taking pictures and/or menu selections) and converts the eye gestures as input to the input device. A user interface input device may also include a voice recognition sensing device that allows a user to interact with a voice recognition system (eg, Siri® navigator) via voice commands.

User interface input devices may also include, but are not limited to, audio/visual devices such as three-dimensional (3D) mice, joysticks or pointing sticks, gamepads and graphics tablets, and speakers, digital cameras, digital camcorders, portable media players, webcams, image scanners, fingerprint scanners, bar code readers 3D scanners, 3D printers, laser range finders, and gaze detection devices. User interface input devices may also include, for example, medical imaging input devices such as computed tomography, magnetic resonance imaging, position emission tomography, medical ultrasound, and the like. User interface input devices may also include audio input devices such as MIDI keyboards, digital musical instruments, and the like.

User interface output devices may include display subsystems, indicator lights, non-visual displays such as audio output devices, and the like. The display subsystem may be a flat panel display, such as one that uses a cathode ray tube (CRT), liquid crystal display (LCD) or plasma display, a projection device, a touch screen, or the like. In general, use of the term "output device" is intended to include all possible types of devices and mechanisms for outputting information from computer system 5100 to a user or other computer. For example, user interface output devices can include, but are not limited to, various display devices that visually convey text, graphics and audio/video information, such as monitors, printers, speakers, headphones, automotive navigation systems, plotters, audio output devices and modems.

Computer system 5100 may comprise a storage subsystem 5118 comprising software elements currently shown residing within system memory 5110 . System memory 5110 may store program instructions loadable and executable on processing unit 5104 and data generated during execution of these programs.

Depending on the configuration and type of computer system 5100, system memory 5110 may be volatile (such as random access memory (RAM)) and/or non-volatile (such as read-only memory (ROM), flash memory, etc.). RAM typically contains data and/or program modules that are immediately accessible to processing unit 5104 and/or are currently being operated on and executed by processing unit 5104 . In some implementations, system memory 5110 may include multiple different types of memory, such as static random access memory (SRAM) or dynamic random access memory (DRAM). In some implementations, a basic input/output system (BIOS), containing the basic routines that help to transfer information between elements within computer system 5100, such as during start-up, can be typically stored in ROM. By way of example and not limitation, system memory 5110 also illustrates application programs 5112, program data 5114, and an operating system 5116, which may include client applications, web browsers, middle-tier applications, relational database management systems (RDBMS), and the like. By way of example, operating system 5116 may include various versions of Microsoft Windows, Apple Macintosh and/or Linux operating systems, various commercially available UNIX or UNIX-like operating systems (including, but not limited to, various GNU/Linux operating systems, Google Chrome OS, etc.), and/or iOS, Windows Phone, Android. ® OS, BlackBerry® 10 OS and Palm® OS operating systems.

Storage subsystem 5118 may also provide tangible computer-readable storage media for storing the basic programming and data structures that provide the functionality of some aspects. Software (programs, code modules, instructions) that provide the functions described above when executed by the processor may be stored in storage subsystem 5118 . These software modules or instructions may be executed by processing unit 5104 . Storage subsystem 5118 may also provide a repository for storing data used in accordance with the present invention.

Storage subsystem 5118 may also include a computer readable storage media reader 5120 that is further connectable to computer readable storage media 5122 . Together and optionally in combination with system memory 5110, computer-readable storage media 5122 may collectively represent remote, local, fixed and/or removable storage devices, as well as storage media for temporarily and/or permanently containing, storing, transmitting, and retrieving computer-readable information.

A computer-readable storage medium 5122 containing code or portions of code may include any suitable medium known or used in the art. Such media include, but are not limited to, storage media and communication media such as volatile and nonvolatile, removable and non-removable media implemented in any method or technology for the storage and/or transmission of information. This may include tangible temporary computer-readable storage media such as RAM, ROM, electronically erasable programmable ROM (EEPROM), flash memory or other memory technology, CD-ROM, digital versatile disk (DVD), or other optical storage device, magnetic cassette, magnetic tape, magnetic disk storage device or other magnetic storage device, or other tangible computer-readable medium. It may also include intangible, transitory computer-readable media such as data signals, data transmissions, or any other medium that can be used to transmit desired information and that is accessible by computer system 5100.

By way of example, the computer readable storage medium 5122 may include a hard disk drive that reads from or writes to non-removable, nonvolatile magnetic media, a magnetic disk drive that reads from or writes to removable nonvolatile magnetic disks, and an optical disk drive that reads from or writes to removable nonvolatile optical disks such as CD ROM, DVD and Blu-ray disks or other optical media. Computer readable storage media 5122 can include, but are not limited to, Zip drives, flash memory cards, universal serial bus (USB) flash drives, secure digital (SD) cards, DVD discs, digital video tapes, and the like. The computer-readable storage medium 5122 may also include flash memory-based SSDs, enterprise flash drives, solid-state drives (SSDs) based on non-volatile memory such as solid-state ROM, SSDs based on volatile memory such as solid-state RAM, dynamic RAM, static RAM, DRAM-based SSDs, magnetoresistive RAM (MRAM) SSDs, and hybrid SSDs that use a combination of DRAM and flash memory-based SSDs. The disk drives and their associated computer-readable media can provide computer readable instructions, data structures, program modules and other data for computer system 5100 .

Communications subsystem 5124 provides an interface with other computer systems and networks. Communications subsystem 5124 serves as an interface for receiving data from other systems and for transmitting data from computer system 5100 to other systems. For example, communications subsystem 5124 may allow computer system 5100 to connect to one or more devices over the Internet. In some aspects, the communication subsystem 5124 includes a radio frequency (RF) transceiver component for accessing wireless voice and/or data networks (e.g., using cellular technologies, advanced data network technologies such as 3G, 4G or enhanced data rates for global evolution (EDGE)), WiFi (IEEE 802.28 family standards or other mobile communication technologies or any combination thereof), a global positioning system (GPS) receiver component, and/or may include other components. In some aspects, communication subsystem 5124 may provide a wired network connection (eg, Ethernet) in addition to or instead of a wireless interface.

Also, in some aspects, communications subsystem 5124 may receive input communications in the form of structured and/or unstructured data feeds 5126, event streams 5128, event updates 5151, etc., on behalf of one or more users who may be using computer system 5100.

As an example, the communication subsystem 5124 may be configured to receive data feeds 5126, such as web feeds such as Twitter feeds, Facebook updates, Rich Site Summary (RSS) feeds, etc., in real-time from users of social media networks and/or other communication services, and/or receive real-time updates from one or more third-party sources.

Additionally, communication subsystem 5124 may be configured to receive data in the form of a continuous data stream. Such data may include an event stream 5128 and/or event updates 5151 of real-time events that may be continuous or essentially without distinct edges and without boundaries. Examples of applications that generate continuous data may include, for example, sensor data applications, financial tickers, network performance measurement tools (eg, network monitoring and traffic management applications), clickstream analysis tools, automotive traffic monitoring, and the like.

Communication subsystem 5124 may also be configured to output structured and/or unstructured data feeds 5126, event streams 5128, event updates 5151, etc. to one or more databases that may communicate with one or more streaming data source computers coupled to computer system 5100.

Computer system 5100 may be one of a variety of types, including handheld portable devices (e.g., iPhone mobile phones, iPad computing tablets, PDAs), wearable devices (e.g., Google Glass head-mounted displays), PCs, workstations, mainframes, kiosks, server racks, or other data processing systems.

Due to the ever-changing nature of computers and networks, the depicted description of computer system 5100 is intended only as a specific example. Many other configurations are possible, having more or fewer components than the system shown in the figures. Customized hardware may be used and/or particular elements may be implemented, for example, in hardware, firmware, software (including applets), or a combination. Additionally, connections to other computing devices, such as network input/output devices, may be utilized. Based on the disclosure and teachings provided herein, a person of ordinary skill in the art will appreciate other means and/or methods for implementing various aspects.

In the foregoing specification, aspects of the invention have been described with reference to specific aspects thereof, but those skilled in the art will appreciate that the invention is not so limited. Various features and aspects of the above-described invention may be used individually or together. Moreover, aspects may be utilized in numerous environments and applications beyond those described herein without departing from the broader spirit and scope of the specification. The specification and drawings are, accordingly, to be regarded in an illustrative rather than a restrictive sense.

Claims (20)

A method for improving the accuracy of a discourse tree, comprising:
creating a discourse tree from the text by identifying base discourse units in the text, the discourse tree comprising a plurality of nodes, each non-terminal node of the nodes in the discourse tree representing a rhetorical relationship between two base discourse units, each terminal node of the nodes in the discourse tree being associated with a base discourse unit, the method further comprising:
identifying a rhetorical relation of type elaboration or joint in said discourse tree, said rhetorical relation relating a first elementary discourse unit and a second elementary discourse unit, said first elementary discourse unit and said second elementary discourse unit forming a reference sentence, said method further comprising:
determining a syntactic generalization score for each candidate sentence of a set of candidate sentences, each candidate sentence having a corresponding semantic relationship, said determining comprising:
identifying one or more common entities between the candidate sentence and the reference sentence;
calculating a syntactic generalization score equal to the number of the identified one or more common entities, the method further comprising:
selecting the candidate sentence with the highest syntactic generalization score among the syntactic generalization scores;
identifying semantic relationships corresponding to the candidate sentences, the semantic relationships corresponding to words in the candidate sentences and defining roles in the candidate sentences, the method further comprising:
creating an updated discourse tree by replacing the rhetorical relations in the discourse tree with updated rhetorical relations corresponding to the semantic relations. Creating the discourse tree from the text includes:
providing the text to a classification model;
2. The method of claim 1, comprising using the classification model to identify the first basic discourse unit, the second basic discourse unit, and the rhetorical relation. 2. The method of claim 1, wherein the updated rhetorical relation is one of ends, means, causes, or temporal sequences. forming a response from the updated discourse tree;
2. The method of claim 1, further comprising outputting the response to an external device. The method further comprises:
forming a first syntactic parse tree from each candidate sentence;
forming a second syntactic parse tree from the reference sentence, and wherein identifying the one or more common entities between the candidate sentence and the reference sentence comprises, for each common entity, identifying the common entity in the first syntactic parse tree and the second syntactic parse tree. The method further comprises:
forming a communicative discourse tree from the updated discourse tree by matching each fragment with a verb to a verb signature;
identifying that the text contains arguments by applying a classification model trained to detect arguments to the communicative discourse tree;
forming a response from the text and outputting the response to an external device. The method further comprises:
forming a communicative discourse tree from the updated discourse tree by matching each fragment with a verb to a verb signature;
identifying that the text contains arguments corresponding to assertions by applying a classification model trained to detect arguments to the communication discourse tree;
evaluating the consistency of the argument with respect to itself and with respect to domain definition clauses associated with domains of the text by solving a logic system, the logic system comprising:
(a) a fixed portion comprising said claim clause and said domain definition clause;
(b) a set of revocable rules from said communicative discourse tree and a variable portion comprising facts from communicative behaviors of said communicative discourse tree, said method further comprising:
2. The method of claim 1, comprising forming a text response from the text and outputting the text response to an external device in response to determining that the evaluated consistency is greater than a threshold. a system,
a non-transitory computer-readable medium storing computer-executable program instructions;
a processing device communicatively coupled to the non-transitory computer-readable medium for executing the computer-executable program instructions, wherein executing the computer-executable program instructions configures the processing device to perform an action, the action comprising:
creating a discourse tree from the text by identifying a base discourse unit in the text, the discourse tree comprising a plurality of nodes, each non-terminal node of the node in the discourse tree representing a rhetorical relationship between two base discourse units, each terminal node of the node in the discourse tree being associated with a base discourse unit;
identifying a rhetorical relation of type elaboration or joint in said discourse tree, said rhetorical relation relating first and second elementary discourse units, said first and second elementary discourse units forming a reference sentence, said action further comprising:
determining a syntactic generalization score for each candidate sentence of a set of candidate sentences, each candidate sentence having a corresponding semantic relationship, said determining comprising:
identifying one or more common entities between the candidate sentence and the reference sentence;
calculating a syntactic generalization score equal to the number of the identified one or more common entities, the operation further comprising:
selecting the candidate sentence with the highest syntactic generalization score among the syntactic generalization scores;
identifying semantic relationships corresponding to the candidate sentences, the semantic relationships corresponding to words in the candidate sentences and defining roles in the candidate sentences, the acts further comprising:
creating an updated discourse tree by replacing the rhetorical relations in the discourse tree with updated rhetorical relations corresponding to the semantic relations. Creating the discourse tree from the text includes:
providing the text to a classification model;
9. The system of claim 8, comprising using the classification model to identify the first basic discourse unit, the second basic discourse unit, and the rhetorical relation. 9. The system of claim 8, wherein the updated rhetorical relation is one of ends, means, causes, or temporal sequences. 9. The system of claim 8, wherein each of the one or more common entities shares an utterance intersection between the candidate sentence and the reference sentence. Said operation further comprises:
forming a first syntactic parse tree from each candidate sentence;
forming a second syntactic parse tree from the reference sentence, and wherein identifying the one or more common entities between the candidate sentence and the reference sentence comprises, for each common entity, identifying the common entity in the first syntactic parse tree and the second syntactic parse tree. Said operation further comprises:
forming a communicative discourse tree from the updated discourse tree by matching each fragment with a verb to a verb signature;
identifying that the text contains arguments by applying a classification model trained to detect arguments to the communicative discourse tree;
forming a response from the text and outputting the response to an external device. Said operation further comprises:
forming a communicative discourse tree from the updated discourse tree by matching each fragment with a verb to a verb signature;
identifying that the text contains arguments corresponding to assertions by applying a classification model trained to detect arguments to the communication discourse tree;
evaluating the consistency of the argument with respect to itself and with respect to domain definition clauses associated with domains of the text by solving a logic system, the logic system comprising:
(a) a fixed portion comprising said claim clause and said domain definition clause;
(b) a set of revocable rules from said communicative discourse tree and a variable portion comprising facts from communicative behaviors of said communicative discourse tree, said action further comprising:
9. The system of claim 8, comprising forming a text response from the text and outputting the text response to an external device in response to determining that the evaluated consistency is greater than a threshold. A non-transitory computer-readable storage medium storing computer-executable program instructions that, when executed by a processing device, cause the processing device to perform an action, the action comprising:
creating a discourse tree from the text by identifying a base discourse unit in the text, the discourse tree comprising a plurality of nodes, each non-terminal node of the node in the discourse tree representing a rhetorical relationship between two base discourse units, each terminal node of the node in the discourse tree being associated with a base discourse unit;
identifying a rhetorical relation of type elaboration or joint in said discourse tree, said rhetorical relation relating first and second elementary discourse units, said first and second elementary discourse units forming a reference sentence, said action further comprising:
determining a syntactic generalization score for each candidate sentence of a set of candidate sentences, each candidate sentence having a corresponding semantic relationship, said determining comprising:
identifying one or more common entities between the candidate sentence and the reference sentence;
calculating a syntactic generalization score equal to the number of the identified one or more common entities, the operation further comprising:
selecting the candidate sentence with the highest syntactic generalization score among the syntactic generalization scores;
identifying semantic relationships corresponding to the candidate sentences, the semantic relationships corresponding to words in the candidate sentences and defining roles in the candidate sentences, the acts further comprising:
creating an updated discourse tree by replacing the rhetorical relations in the discourse tree with updated rhetorical relations corresponding to the semantic relations. Creating the discourse tree from the text includes:
providing the text to a classification model;
16. The non-transitory computer-readable storage medium of claim 15, comprising using the classification model to identify the first basic discourse unit, the second basic discourse unit, and the rhetorical relation. 16. The non-transitory computer-readable storage medium of claim 15, wherein the updated rhetorical relation is one of ends, means, causes, or temporal sequences. 16. The non-transitory computer-readable storage medium of claim 15, wherein each of the one or more common entities shares a common portion of utterance between the candidate sentence and the reference sentence. When executed by a processing device, the computer-executable program instructions cause the processing device to perform an action, the action comprising:
forming a first syntactic parse tree from each candidate sentence;
16. The non-transitory computer-readable storage medium of claim 15, wherein identifying the one or more common entities between the candidate sentence and the reference sentence comprises identifying the common entity in the first syntactic parsetree and the second syntactic partree for each common entity. When executed by a processing device, the computer-executable program instructions cause the processing device to perform an action, the action comprising:
forming a communicative discourse tree from the updated discourse tree by matching each fragment with a verb to a verb signature;
identifying that the text contains arguments by applying a classification model trained to detect arguments to the communicative discourse tree;
forming a response from the text and outputting the response to an external device.

Images