US20070112555A1 - Dynamic Natural Language Understanding - Google Patents
Dynamic Natural Language Understanding Download PDFInfo
- Publication number
- US20070112555A1 US20070112555A1 US11/562,132 US56213206A US2007112555A1 US 20070112555 A1 US20070112555 A1 US 20070112555A1 US 56213206 A US56213206 A US 56213206A US 2007112555 A1 US2007112555 A1 US 2007112555A1
- Authority
- US
- United States
- Prior art keywords
- text
- semantic category
- natural language
- value
- category
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Abandoned
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
- G06F40/284—Lexical analysis, e.g. tokenisation or collocates
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
Definitions
- This invention relates to natural language understanding.
- Grammar based natural language understanding systems and methods typically use a parser to parse a text into a tree, i.e. a hierarchical (“depth”) structure. Elements of the tree are processed in a hierarchical manner, either bottom up or top down. In order to achieve successful understanding of the text, the sentence structure/grammar generally needs to conform to rules, thereby placing constraints on the freedom of expression of the submitter of the text.
- depth a hierarchical
- Statistically based natural language understanding systems and methods typically use many statistical methods including classification to understand a text. Freedom of expression by the submitter of the text is therefore enhanced.
- U.S. Pat. No. 5,680,511 to Baker, et al. in one aspect, provides word recognition systems that operate to recognize an unrecognized or ambiguous word that occurs within a passage of words.
- the system can offer several words as choice words for inserting into the passage to replace the unrecognized word.
- the system can select the best choice word by using the choice word to extract from a reference source, sample passages of text that relate to the choice word. For example, the system can select the dictionary passage that defines the choice word.
- the system compares the selected passage to the current passage, and generates a score that indicates the likelihood that the choice word would occur within that passage of text.
- the system can select the choice word with the best score to substitute into the passage.
- the passage of words being analyzed can be any word sequence including an utterance, a portion of handwritten text, a portion of typewritten text or other such sequence of words, numbers and characters.
- Alternative embodiments of the present invention are disclosed which function to retrieve documents from a library as a function of context.
- U.S. Pat. No. 5,642,519 to Martin provide a unified grammar for a speech interpreter capable of real-time speech understanding for user applications running on a general purpose microprocessor-based computer.
- the speech interpreter includes a unified grammar (UG) compiler, a speech recognizer and a natural language (NL) processor.
- the UG compiler receives a common UG lexicon and unified grammar description, and generates harmonized speech recognition (SR) and NL grammars for the speech recognizer and natural language processor, respectively.
- the lexicon includes a plurality of UG word entries having predefined characteristics, i.e., features, while the UG description includes a plurality of complex UG rules which define grammatically allowable word sequences.
- the UG compiler converts the complex UG rules (complex UG rules include augmentations for constraining the UG rules) into permissible SR word sequences and SR simple rules (simple rules do not include any augmentation) for the SR grammar.
- the SR grammar is a compact representation of the SR word entries corresponding to the UG word entries, permissible SR word sequences and simple SR rules corresponding to the augmentations of the complex UG rules.
- the NL grammar provides the NL processor with NL patterns enabling the NL processor to extract the meaning of the validated word sequences passed from the speech recognizer.
- Speech recognition systems consistent with the present invention include a lexicon database with words and associated lexical properties.
- the systems receive exemplary clauses containing permissible word combinations for speech recognition, and identify additional lexical properties for selected words in the lexicon database corresponding to words in the received exemplary clauses using lexical property tests of a grammar database.
- Certain lexical property tests are switchable to a disabled state.
- the exemplary clauses are parsed with the switchable lexical property tests disabled to produce an index of the lexical properties corresponding to the exemplary clauses.
- the lexicon database is updated with the identified additional lexical properties by assigning the lexical properties to the corresponding words of the lexicon database.
- the grammar database is compiled with the lexical property tests enabled and the lexicon database with the assigned lexical properties to produce a grammar that embodies constraints of the lexical property tests and the lexical properties.
- U.S. Pat. No. 5,918,222 to Fukui, et al. teaches a data storage means for storing data in a predetermined information form.
- An information retrieval means retrieves the data stored in the data storage means.
- a reception means receives an information disclosure demand from a demander, a response rule storage means for storing general knowledge for generating a response responding to the demander and a personal relationship information associated with a unique personal relationship between a user having the data on an information provider side and a user on an information demander side.
- a response plan formation means responsive to the demand received by the reception means, plans a response for exhibiting, to the information demander, data obtained by causing the retrieval means to retrieve the data stored in the data storage means on the basis of the knowledge and the personal relationship information stored in the response rule storage means.
- a response generation means generates the response to the information demander in accordance with the plan formed by the response plan formation means.
- U.S. Pat. No. 5,987,404 to Della Pietra, et. al. proposes using statistical methods to do natural language understanding.
- the key notion is that there are “strings” of words in the natural language, that correspond to a single semantic concept.
- U.S. Pat. No. 5,576,954 to Driscoll teaches a procedure for determining text relevancy and can be used to enhance the retrieval of text documents by search queries.
- This system helps a user intelligently and rapidly locate information found in large textual databases.
- a first embodiment determines the common meanings between each word in the query and each word in the document. Then an adjustment is made for words in the query that are not in the documents. Further, weights are calculated for both the semantic components in the query and the semantic components in the documents. These weights are multiplied together, and their products are subsequently added to one another to determine a real value number (similarity coefficient) for each document. Finally, the documents are sorted in sequential order according to their real value number from largest to smallest value.
- Another, embodiment is for routing documents to topics/headings (sometimes referred to as filtering).
- topics/headings sometimes referred to as filtering.
- the importance of each word in both topics and documents are calculated.
- the real value number (similarity coefficient) for each document is determined.
- each document is routed one at a time according to their respective real value numbers to one or more topics.
- the documents can be sorted.
- This system can be used to search and route all kinds of document collections, such as collections of legal documents, medical documents, news stories, and patents.
- U.S. Pat. No. 5,642,502 also to Driscoll teaches a system and method for retrieving relevant documents from a text data base collection comprised of patents, medical and legal documents, journals, news stories and the like. Each small piece of text within the documents such as a sentence, phrase and semantic unit in the data base is treated as a document.
- Natural language queries are used to search for relevant documents from the data base. A first search query creates a selected group of documents. Each word in both the search query and in the documents are given weighted values. Combining the weighted values creates similarity values for each document which are then ranked according to their relevant importance to the search query. A user reading and passing through this ranked list checks off which documents are relevant or not.
- the system automatically causes the original search query to be updated into a second search query which can include the same words, less words or different words than the first search query. Words in the second search query can have the same or different weights compared to the first search query.
- the system automatically searches the text data base and creates a second group of documents, which as a minimum does not include at least one of the documents found in the first group.
- the second group can also be comprised of additional documents not found in the first group.
- the ranking of documents in the second group is different than the first ranking such that the more relevant documents are found closer to the top of the list.
- U.S. Pat. No. 5,893,092 also to Driscoll teaches a search system and method for retrieving relevant documents from a text data base collection comprised of patents, medical and legal documents, journals, news stories and the like. Each small piece of text within the documents such as a sentence, phrase and semantic unit in the data base is treated as a document.
- Natural language queries are used to search for relevant documents from the data base.
- a first search query creates a selected group of documents.
- Each word in both the search query and in the documents are given weighted values. Combining the weighted values creates similarity values for each document which are then ranked according to their relevant importance to the search query.
- a user reading and passing through this ranked list checks off which documents are relevant or not.
- the system automatically causes the original search query to be updated into a second search query which can include the same words, less words or different words than the first search query. Words in the second search query can have the same or different weights compared to the first search query.
- the system automatically searches the text data base and creates a second group of documents, which as a minimum does not include at least one of the documents found in the first group.
- the second group can also be comprised of additional documents not found in the first group.
- the ranking of documents in the second group is different than the first ranking such that the more relevant documents are found closer to the top of the list.
- U.S. Pat. No. 6,088,692 also to Driscoll teaches a natural language search system and method for retrieving relevant documents from a text data base collection comprised of patents, medical and legal documents, journals, news stories and the like. Each small piece of text within the documents such as a sentence, phrase and semantic unit in the data base is treated as a document.
- Natural language queries are used to search for relevant documents from the data base. A first search query creates a selected group of documents. Each word in both the search query and in the documents are given weighted values. Combining the weighted values creates similarity values for each document which are then ranked according to their relevant importance to the search query. A user reading and passing through this ranked list checks off which document are relevant or not.
- the system automatically causes the original search query to be updated into a second search query which can include the same words, less words or different words than the first search query. Words in the second search query can have the same or different weights compared to the first search query.
- the system automatically searches the text data base and creates a second group of documents, which as a minimum does not include at least one of the documents found in the first group.
- the second group can also be comprised of additional documents not found in the first group.
- the ranking of documents in the second group is different than the first ranking such that the more relevant documents are found closer to the top of the list.
- U.S. Pat. No. 5,694,592 also to Driscoll teaches a procedure for determining text relevancy that can be used to enhance the retrieval of text documents by search queries.
- This system helps a user intelligently and rapidly locate information found in large textual databases.
- a first embodiment determines the common meanings between each word in the query and each word in the document. Then an adjustment is made for words in the query that are not in the documents. Further, weights are calculated for both the semantic components in the query and the semantic components in the documents. These weights are multiplied together, and their products are subsequently added to one another to determine a real value number(similarity coefficient) for each document. Finally, the documents are sorted in sequential order according to their real value number from largest to smallest value.
- Another, embodiment is for routing documents to topics/headings (sometimes referred to as faltering).
- topics/headings sometimes referred to as faltering
- the importance of each word in both topics and documents are calculated.
- the real value number (similarity coefficient) for each document is determined.
- each document is routed one at a time according to their respective real value numbers to one or more topics.
- the documents can be sorted.
- This system can be used to search and route all kinds of document collections, such as collections of legal documents, medical documents, news stories, and patents.
- U.S. Pat. No. 6,138,085 to Richardson, et al. teaches a facility for determining, for a semantic relation that does not occur in a lexical knowledge base, whether this semantic relation should be inferred despite its absence from the lexical knowledge base.
- This semantic relation to be inferred is preferably made up of a first word, a second word, and a relation type relating the meanings of the first and second words.
- the facility identifies a salient semantic relation having the relation type of the semantic relation to be inferred and relating the first word to an intermediate word other than the second word. The facility then generates a quantitative measure of the similarity in meaning between the intermediate word and the second word.
- the facility further generates a confidence weight for the semantic relation to be inferred based upon the generated measure of similarity in meaning between the intermediate word and the second word.
- the facility may also generate a confidence weight for the semantic relation to be inferred based upon the weights of one or more paths connecting the first and second words
- U.S. Pat. No. 5,675,710 to Lewis teaches a method and apparatus for training a text classifier.
- a supervised learning system and an annotation system are operated cooperatively to produce a classification vector which can be used to classify documents with respect to a defined class.
- the annotation system automatically annotates documents with a degree of relevance annotation to produce machine annotated data.
- the degree of relevance annotation represents the degree to which the document belongs to the defined class.
- This machine annotated data is used as input to the supervised learning system.
- the supervised learning system can also receive manually annotated data and/or a user request.
- the machine annotated data, along with the manually annotated data and/or the user request, are used by the supervised learning system to produce a classification vector.
- the supervised learning system comprises a relevance feedback mechanism.
- the relevance feedback mechanism is operated cooperatively with the annotation system for multiple iterations until a classification vector of acceptable accuracy is produced.
- the classification vector produced by the invention is the result of a combination of supervised and unsupervised learning
- U.S. Pat. No. 6,311,152 to Bai, et. al teaches a system ( 100 , 200 ) for tokenization and named entity recognition of ideographic language.
- a word lattice is generated for a string of ideographic characters using finite state grammars ( 150 ) and a system lexicon ( 240 ).
- Segmented text is generated by determining word boundaries in the string of ideographic characters using the word lattice dependent upon a contextual language model ( 152 A) and one or more entity language models ( 152 B).
- One or more named entities is recognized in the string of ideographic characters using the word lattice dependent upon the contextual language model ( 152 A) and the one or more entity language models ( 152 B).
- the contextual language model ( 152 A) and the one or more entity language models ( 152 B) are each class-based language models.
- the lexicon ( 240 ) includes single ideographic characters, words, and predetermined features of the characters and words.
- What is needed in the art is a method and system for understanding natural language that includes inter alia statistical steps and elements which also take advantage of hierarchical-structure. What is also needed in the art is a system and method where the extraction of one part of a text which belongs to one semantic category assists in the extraction of another part which belongs to a semantic category of a different hierarchical level. In addition, what is needed in the art is a method and system for understanding natural language where later steps of the process are affected based on the results of earlier steps, thereby introducing a dynamic aspect to the method and system.
- a method for use in a method for understanding a natural language text comprising performing the following selectively in a statistical manner: attempting to extract at least one value belonging to a semantic category from a natural language text or a form thereof; and if a result of the attempting complies with a predetermined criterion, attempting to extract, based on the result, at least one value belonging to another semantic category of a different hierarchical level than the semantic category, else performing at least one action from a group of actions including: asking a submitter of the text a question whose content depends on the result and giving up on understanding the natural language text.
- the predetermined criterion is at least one from a group including: at least one value for the semantic category was extracted, only one value for the semantic category was extracted, one of the at least one value extracted for the semantic category is selected based on a grade thereof, a correct number of values for the semantic category were extracted, a correct number of values for the semantic category are selected based on grades thereof from among the at least one value extracted for the semantic category, at least some values belonging to other previously extracted at least one semantic category are appropriate for at least one value extracted for the semantic category, values belonging to other previously extracted at least one semantic category are appropriate for only one value extracted for the semantic category, the semantic category is a particular semantic category where an unlimited number of extracted values is allowed, it is desired to process in parallel more than one extracted value for the semantic category, there is a default value corresponding to each required value for the semantic category which was not extracted, there is only one possible value for the semantic category, and there is only a correct number of possible values for the semantic category.
- a method for understanding a natural language text comprising: receiving a natural language text; processing each at least two semantic categories, the each on a different hierarchical level, by performing the following selectively in a statistical manner: (i) attempting to determine at least one value belonging to the each semantic category through extraction, wherein if the each semantic category is not a first processed of the at least two semantic categories, then the attempting is based on results of previously processed semantic categories, and(ii) if the each semantic category is not a last processed of the at least two semantic categories and a result of the attempting does not comply with a predetermined criterion, dialoging with a submitter of the text and receiving at least one answer from the submitter, wherein at least one value determined from the at least one answer augments the result so as to comply with the predetermined criterion and allow extraction attempts for other of the at least two semantic categories to be subsequently processed; and evaluating values determined for the at least two semantic categories with respect to one another to determine whether the
- a method for training at least two classifiers to understand a natural language text comprising: introducing entries into a database, the entries belonging to at least two semantic categories of different hierarchical levels; defining examples of natural language texts, wherein at least some of the examples include embedded syntactic tokens based on the entries; and training at least two classifiers for the at least two semantic categories using the examples or a form thereof.
- a module for use in a system for natural language understanding comprising: at least one classifier or pseudo classifier configured to extract values belonging to a semantic category from a natural language text or a form thereof; and an action resolver configured if a result of extracting values of the semantic category complies with a predetermined criterion to employ based on the result at least one classifier or pseudo classifier to extract values belonging to another semantic category of a different hierarchical level, and configured if the result does not comply with a predetermined criterion to perform at least one action from a group of actions including: employing based on the result a dialog management module and giving up on understanding the natural language text.
- a system for natural language understanding comprising: at least two classifiers or pseudo classifiers configured to extract values belonging to at least two semantic categories on different hierarchical levels from a natural language text or a form thereof; a dialog management module configured to dialog with a submitter of the natural language text; at least one evaluation module configured to evaluate values belonging to the at least two semantic categories; and an action resolver configured to cause the text to be understood by (i) employing, if a result of extracting values of a semantic category complies with a predetermined criterion and the semantic category is not a last to be processed semantic category, a classifier or pseudo classifier based on the result to extract values belonging to another semantic category, by (ii) employing, if the result does not comply with a predetermined criterion and the semantic category is not a last to be processed semantic category, a dialog management module and then employing, based on the result as augmented by at least one answer received from the submitter by the dialog management module,
- a system for training classifiers for natural language understanding comprising: a real time database including entries related to semantic categories on at least two different hierarchical levels; classifiers for the semantic categories; and a knowledge work tool configured to develop syntactic tokens from the entries, embed the tokens in examples and train the classifiers at least partially on the examples.
- a method for understanding a natural language text comprising performing the following in a selectively statistical manner: receiving a natural language text; extracting at least one parameter value from the text or a form thereof; identifying at least one parameter type related to each extracted parameter value; providing at least one restatement of the received text, each at least one restatement having embedded within, at least one of the identified parameter types; extracting at least one overall category value from the at least one restatement or a form thereof,; selecting a subcategory extractor corresponding to one of the extracted at least one overall category, and using the selected subcategory extractor to extract at least one subcategory value; choosing one of the at least one extracted subcategory values; evaluating the at least one identified parameter type in relation to the chosen subcategory value; and concluding that the natural language text is understood.
- system for understanding a natural language text comprising: one classifier configured to extract an overall category value from a natural language text or a form thereof; a different classifier corresponding to each overall category value configured to extract subcategory values from a natural language text or a form thereof; one classifier configured to extract parameter values from a natural language text or a form thereof; a dialog management module configured to dialog with a submitter of the natural language text; at least one evaluation component configured to evaluate extracted values; and an action resolver configured to employ different parts of the system in turn in order to understand the natural language text, including employing the one classifier for parameter values before the one overall category classifier and employing the overall category classifier before the corresponding subcategory classifier.
- a program storage device readable by machine, tangibly embodying a program of instructions executable by the machine to perform method steps for use in a method for understanding a natural language text, comprising performing the following selectively in a statistical manner: attempting to extract at least one value belonging to a semantic category from a natural language text or a form thereof; and if a result of the attempting complies with a predetermined criterion, attempting to extract, based on the result, at least one value belonging to another semantic category of a different hierarchical level than the semantic category, else performing at least one action from a group of actions including: asking a submitter of the text a question whose content depends on the result and giving up on understanding the natural language text.
- a computer program product comprising a computer useable medium having computer readable program code embodied therein for use in a computer program product comprising: computer readable program code for causing the computer to perform the following selectively in a statistical manner: computer readable program code for causing the computer to attempt to extract at least one value belonging to a semantic category from a natural language text or a form thereof; and computer readable program code for causing the computer if a result of the attempting complies with a predetermined criterion to attempt to extract, based on the result, at least one value belonging to another semantic category of a different hierarchical level than the semantic category, else performing at least one action from a group of actions including: asking a submitter of the text a question whose content depends on the result and giving up on understanding the natural language text.
- a program storage device readable by machine, tangibly embodying a program of instructions executable by the machine to perform method steps for understanding a natural language text, comprising: receiving a natural language text; processing each at least two semantic categories, the each on a different hierarchical level, by performing the following selectively in a statistical manner: (i) attempting to determine at least one value belonging to the each semantic category through extraction, wherein if the each semantic category is not a first processed of the at least two semantic categories, then the attempting is based on results of previously processed semantic categories, and (ii) if the each semantic category is not a last processed of the at least two semantic categories and a result of the attempting does not comply with a predetermined criterion, dialoging with a submitter of the text and receiving at least one answer from the submitter, wherein at least one value determined from the at least one answer augments the result so as to comply with the predetermined criterion and allow extraction attempts for other of the at least two semantic categories to be
- a computer program product comprising a computer useable medium having computer readable program code embodied therein for understanding a natural language text
- the computer program product comprising: computer readable program code for causing the computer to receive a natural language text; computer readable program code for causing the computer to process each at least two semantic categories, the each on a different hierarchical level, by performing the following selectively in a statistical manner: computer readable program code for causing the computer to (i) attempt to determine at least one value belonging to the each semantic category through extraction, wherein if the each semantic category is not a first processed of the at least two semantic categories, then the attempting is based on results of previously processed semantic categories, and computer readable program code for causing the computer to (ii) if the each semantic category is not a last processed of the at least two semantic categories, and a result of the attempting does not comply with a predetermined criterion, dialog with a submitter of the text and receive at least one answer from the submitter, where
- a program storage device readable by machine, tangibly embodying a program of instructions executable by the machine to perform method steps for training at least two classifiers to understand a natural language text, comprising: introducing entries into a database, the entries belonging to at least two semantic categories of different hierarchical levels; defining examples of natural language texts, wherein at least some of the examples include embedded syntactic tokens based on the entries; and training at least two classifiers for the at least two semantic categories using the examples or a form thereof.
- a computer program product comprising a computer useable medium having computer readable program code embodied therein for training at least two classifiers to understand a natural language text
- the computer program product comprising: computer readable program code for causing the computer to introduce entries into a database, the entries belonging to at least two semantic categories of different hierarchical levels; computer readable program code for causing the computer to define examples of natural language texts, wherein at least some of the examples include embedded syntactic tokens based on the entries; and computer readable program code for causing the computer to train at least two classifiers for the at least two semantic categories using the examples or a form thereof.
- a program storage device readable by machine, tangibly embodying a program of instructions executable by the machine to perform method steps for understanding a natural language text, comprising performing the following in a selectively statistical manner: receiving a natural language text; extracting at least one parameter value from the text or a form thereof; identifying at least one parameter type related to each extracted parameter value; providing at least one restatement of the received text, each at least one restatement having embedded within, at least one of the identified parameter types; extracting at least one overall category value from the at least one restatement or a form thereof; selecting a subcategory extractor corresponding to one of the extracted at least one overall category, and using the selected subcategory extractor to extract at least one subcategory value; choosing one of the at least one extracted subcategory values; evaluating the at least one identified parameter type in relation to the chosen subcategory value; and concluding that the natural language text is understood.
- a computer program product comprising a computer useable medium having computer readable program code embodied therein for understanding a natural language text
- the computer program product comprising: computer readable program code for causing the computer to perform the following in a selectively statistical manner: computer readable program code for causing the computer to receive a natural language text; computer readable program code for causing the computer to extract at least one parameter value from the text or a form thereof; computer readable program code for causing the computer to identify at least one parameter type related to each extracted parameter value; computer readable program code for causing the computer to provide at least one restatement of the received text, each at least one restatement having embedded within, at least one of the identified parameter types; computer readable program code for causing the computer to extract at least one overall category value from the at least one restatement or a form thereof; computer readable program code for causing the computer to select a subcategory extractor corresponding to one of the extracted at least one overall category, and use the selected subcatego
- FIG. 1 is a block diagram of a system for understanding a natural language input and optionally executing a request included therein, according to a preferred embodiment of the present invention
- FIG. 2 is a sample screen of a knowledge base work tool, according to a preferred embodiment of the present invention.
- FIG. 3 is a sample screen showing the processing of an active browsing script according to a preferred embodiment of the present invention.
- FIG. 4 is a block diagram of a module for understanding a natural language text, according to a preferred embodiment of the present invention.
- FIG. 5 is a flow chart of a method for understanding a natural language text, according to a preferred embodiment of the present invention
- FIG. 6 is a flow chart of a method for evaluating extraction results, according to a preferred embodiment of the present invention.
- FIGS. 7A and 7B show a sequence for employing different modules of the natural language module, according to a preferred embodiment of the present invention.
- FIG. 8 is a flow chart for preparing a text for extraction, according to a preferred embodiment of the present invention.
- FIG. 9 is a flow chart for selecting a classifier or pseudo classifier based on previous extraction results, according to a preferred embodiment of the present invention.
- FIG. 10 is a flow chart for interaction with the submitter of a natural language text, according to a preferred embodiment of the present invention.
- FIG. 11 is an entity-relationship (ER) diagram of a real time database, according to a preferred embodiment of the present invention.
- FIG. 12 is a flow chart of a method for training a natural language module, according to a preferred embodiment of the present invention.
- the preferred embodiment relates to a system and method for understanding natural language.
- the natural language to be understood includes a request.
- the invention is not limited to understanding requests and also applies to understanding natural language inputs which do not include a request.
- FIG. 1 is an example of a block diagram of a system 100 for understanding natural language and if the natural language includes a request, optionally executing the request, according to a preferred embodiment of the present invention.
- System 100 assumes remote access through a network, such as the Internet, but it should be evident that local access is within the scope of the invention.
- Client 110 is shown here as a Laptop Computer however it should be evident that client 110 can be any input device, for example computers, PDAs (Personal Digital Assistants), phones, cellular phones, cellular phones with SMS or WAP capabilities, fax, scanner, etc.
- pre-handling of the input may be necessary, for example speech to text conversion, optical character recognition etc. These pre-handling techniques are well known in the art.
- An optional network profiler 120 manages remote networks (not shown), controlling network properties as well as the inflow and outflow of requests to and from the remote systems.
- An optional control server 130 manages the connection between client 110 and the rest of system 100 , and also the internal connections within the rest of system 100 .
- a natural language understanding (NLU) server 140 includes the software for understanding the natural language.
- NLU server 140 In order for the software on NLU server 140 to be able to comprehend the natural language, a preparation stage is necessary which includes for example the implementation of training activity. The preparation stage can be conducted, for example, using a knowledge builder work tool 150 .
- a sample screen of work tool 150 is shown in FIG. 2 .
- a request implementation server 160 allows the implementation of the understood request.
- request implementation server 160 is an active browsing server, implementation can occur through control of Internet sites automatically during runtime.
- every request supported by system 100 is represented by a script code of several lines that defines how and where the request should be carried out.
- the active browsing script can be prepared, for example, using an active browsing studio work tool 170 .
- FIG. 3 shows a sample screen of the processing of an active browsing script so as to carry out the request. (Browsing work tool 170 in some preferred embodiments may also be used in building a real time database 445 as explained below with reference to FIG. 11 ).
- NLU server 140 and request implementation server 160 in certain preferred embodiments are replaced by natural language understanding (NLU) module 140 and request implementation module 160 , respectively.
- NLU module or server 140 and request implementation server or module 160 in certain preferred embodiments are separately implemented so that for example, NLU module/server 140 can be used with another means of request implementation or even with no request implementation means. For example, even if the natural language is assumed to include a request, the request may be understood without being implemented. In preferred embodiments where the user input is not a request, request implementation means may be unnecessary.
- NLU module/server 140 in a system with request implementation module/server 160 .
- NLU module/server 140 and implementation module/server 160 can be integrated into one unit or separated into more than two units.
- module as in NLU module and request implementation module is assumed to refer to both modules and servers, which may form part of network or non-network systems.
- FIG. 4 shows an example of NLU module 140 , according to a preferred embodiment of the present invention. It should be evident that the modules shown in FIG. 4 as forming part of NLU module 140 and discussed below can be integrated or divided into a smaller or larger number of modules. The actual separation of the functions of NLU module 140 into the modules shown in FIG. 4 is for ease of understanding only. In one preferred embodiment of the invention, the modules shown on the bottom of FIG. 4 are associated with the online (i.e. using) stage and the modules shown on the top of FIG. 4 are associated with the offline (training) stage. FIG. 4 is discussed below in conjunction with flow charts illustrating the methods for using and training module 140 . The order of the steps in one or more of the methods illustrated in the flow charts may be varied in other preferred embodiments. In other preferred embodiments, some steps in one or more of the methods in the flow charts may be omitted and/or additional steps may be added.
- FIG. 5 An example of the overall method for understanding the submission from a user is shown in FIG. 5 , according to a preferred embodiment of the present invention.
- the method shown in FIG. 5 enables understanding of the text, without compelling the text to comply to a pre-defined grammar.
- User input is received (step 510 ) from a submitter, for example through client 110 .
- input can optionally include a request. It is assumed that any necessary pre-handling of the input has already been performed as explained above so that the input is received by module 140 in a format compatible with module 140 .
- the format is ASCII.
- the input is also referred to as “text”.
- the next step is preparing the text, if necessary, for processing (step 512 ) using a text pre-preprocessing module 435 .
- An action resolver module 410 decides which classifier module (also sometimes termed extractor) 420 or pseudo-classifier static component 425 to employ (step 515 ). Each classifier or group of classifiers 420 extracts values belonging to a different semantic category.
- One or more pseudo-classifier static components 425 extract phrases that can not be learned or do not need to be learned (as will be explained further below) belonging to one or more semantic categories.
- Selected classifier 420 or pseudo classifier static component 425 is employed on the text (step 520 ), and the results of the extraction are evaluated by action resolver 410 (step 525 ), as will be described below.
- the results of selected classifier 420 are sufficient to understand the text (i.e. it is concluded that the text is understood) (step 535 )
- the results are optionally output (for example to the submitter, or to control server 130 and from there to request implementation module 160 ) in step 530 .
- Outputting the results is one possible way of indicating that the text is understood.
- other indications of understanding can be used, for example an indication that a request optionally included in the input was implemented.
- a further extraction can be performed by a classifier 420 or pseudo classifier 425 (step 540 ), action resolver 410 prepares the text, if necessary, for further extraction (step 512 ) and action resolver 410 chooses the next classifier 420 or the next pseudo classifier 425 for the next semantic category to extract (step 515 ).
- a further extraction may not be able to be performed and the results are insufficient to understand the text. This situation may occur, for example, if one or more of classifiers 420 could not classify the text into any one of the possible semantic meanings that classifier 420 knows.
- NLU module 140 may be considered to have failed to understand the text and may stop any further processing (step 560 ).
- the results of the failure can be optionally communicated to the submitter.
- action resolver module 410 may dialog with the submitter in step 545 , and receive submitter response in step 550 . Once the response is received a further extraction may be performed on the submitter response in step 520 using the classifier 420 or pseudo classifier 425 for the same semantic category, or a classifier 420 /pseudo classifier 425 for another previously extracted semantic category.
- the method described above with reference to FIG. 5 includes some steps performed in a selectively statistical manner.
- the step is a statistically based step
- the step is typically non-statistically based.
- Classifiers are well known in the art.
- An example of a public domain algorithm which can be used by classifiers 420 of this invention includes Naive-bayes text-classification developed by Carnegie Mellon University and available on the world wide web at www.cs.cmu.edu/afs/cs/project/theo-11/www/naive-bayes.html. This public domain algorithm is based on “Machine Learning” by Tom Mitchell, Chapter 6, McGraw Hill 1997 , ISBN 0070428077.
- a pseudo classifier is a type of static component 425 . Pseudo classifiers do not need to be trained.
- Non limiting examples of pseudo classifiers 425 include time extractors and currency extractors, which detect and extract time phrases and currency phrases respectively.
- a data structure for time is used which keeps values for seconds, minutes, days, months, years, etc. Time extractors can therefore translate phrases such as “next Tuesday” into next Tuesday's date, or translate phrases like “for three days”, “every Sunday” etc.
- the data structure for money is in the form of #x.yy, where x is an integer and yy is any number between 00 and 99 and #can be replaced by any currency type.
- Pseudo classifiers such as time and currency extractors may in some cases be based on limited strict grammars, suitable only for specific time or money expressions, and may use well known parsing methods for detecting and parsing the temporal expressions (time-phrases) or currency phrases.
- the parsing may be partial and include a part of the given text that could be parsed based on the limited grammars and later transformed to a data structure that can hold the time or currency expressions.
- grammar based temporal expression parsing (which as well known in the art can be easily modified for currency phrases) can be found at odur.let.rug.nl/ ⁇ tilde over ( ) ⁇ vannoord/papers/yearbook/node2.html as a part of a work called “Grammatical Analysis in a Spoken Dialogue System” by Gosse Bouma, Rob Koeling, Mark-Jan Nederhof and Gertjan Van Noord.
- a semantic category should be understood for the purpose of the description below to be a grouping of values including at least one common property which distinguishes those values from values in other semantic categories that are defined for a particular embodiment.
- the semantic categories there is a hierarchical structure among the semantic categories which is exploited when understanding the text.
- overall category highest level
- subcategory medium level
- parameter values lowest level
- one or more classifier 420 or pseudo classifiers 425 may extract value(s) belonging to the overall category.
- the overall category value in this example can be considered the domain or topic of interest of the text.
- one or more classifier 420 or pseudo classifiers 425 may extract value(s) belonging to the subcategory of the overall category, such as operations related to the overall topic of interest.
- one or more classifier 420 or pseudo classifiers 425 may extract parameter value(s).
- the extracted parameter values are of the parameter types required by the extracted subcategory value.
- the subcategory values share the common property of being subcategory values of overall category values, whereas overall category values share the common property of having subcategory values.
- parameter values share the common property of having corresponding parameter types which can be accepted by subcategory values, whereas subcategory values share the common property of typically accepting parameter values of particular parameter types (It should be noted that in some cases no parameter types are defined for a particular subcategory value).
- the text may include more than one subtext (for example more than one request) with each subtext represented by a separate hierarchical structure that can be processed either in parallel or sequentially.
- each subtext represented by a separate hierarchical structure that can be processed either in parallel or sequentially.
- the text can be represented by a single hierarchical structure.
- FIG. 6 illustrates a preferred embodiment of a method for evaluating results of the extraction by classifier 420 (corresponding to step 525 of FIG. 5 ).
- classifier 420 for a given semantic category searches a knowledge base 430 using the text prepared for extraction. (Details on how knowledge base 430 is developed will be explained below).
- Classifier 420 returns one or more possible matches (step 610 ) and a grade for each match (step 620 ). Grading is preferably performed by classifier 420 based on the training undergone by classifier 420 during the preparation of knowledge-base 430 .
- the process iterates in step 630 until all matches and grades of the matches are output.
- not all matches are output but only matches that meet certain criteria, for example the highest graded matches, the most clustered matches, etc.
- certain criteria for example the highest graded matches, the most clustered matches, etc.
- clustered matches are to be output, matches whose grades are within a small range are identified and considered relevant.
- the two clustered groups of matches correspond to ⁇ 8.3 to 7.8 ⁇ and ⁇ 6.2 to 6 ⁇ .
- the outputted matches may in this example be those which correspond to the higher cluster ⁇ 8.3 to 7.8 ⁇ , perhaps together with other outputted matches, for example the match corresponding to the highest grade 9 .
- the matches are sorted by grade in step 640 .
- all matches are stored until no longer needed, i.e. until subsequent processing, for example further extractions, dialoging with submitters, etc. renders the match unlikely to be the correct match.
- the matches may be stored until the final results are output in step 530 .
- step 642 If the results are for the first semantic category extracted and more than one semantic category exists, then the results are considered insufficient to understand the text in step 650 (i.e. a ‘no’ answer to step 535 ). If there is only one existing semantic category (step 642 ) then the results are considered sufficient in step 670 (i.e. a ‘yes’ answer to step 535 ). If the results are for subsequent extracted semantic categories, but the results can not be evaluated in conjunction with results from previous semantic category extractions then the results are considered insufficient in step 650 (corresponding to a ‘no’ answer to step 535 ). An example of a situation where the results can not be evaluated in conjunction with previous results is if the current extracted semantic category value(s) is not directly related to any of the previously extracted category value(s).
- the overall category value extracted is “financial operation” and the only two possible subcategory values for this overall category value are “get a stock quote” and “buy stocks”. If neither of these subcategory values are extracted, but instead the subcategory value “get a horoscope forecast” is extracted, then the results are considered insufficient because the extracted values for the overall category and subcategory are not related to each other.
- the results can be evaluated in conjunction with results from previous semantic category extractions, then the results are evaluated in conjunction with the results from the previous sets (step 658 ). For example, the evaluation can determine whether the results for the current semantic category correspond to the results from previous semantic category extractions (see below FIG. 7 for more details on a possible evaluation process). If no weighted grade is to be calculated, then if the results are sufficient to understand the text, the method proceeds directly to step 670 (corresponding to a ‘yes’ answer to step 535 ). Otherwise, if no weighted grade is to be calculated and the results are insufficient, the method proceeds directly to step 650 (corresponding to a ‘no’ answer to step 535 ). The results may be considered sufficient to understand the text, for example, if all required values for each semantic category are known and the values for the different semantic categories correspond with one another.
- step 645 if the results are not for the last extracted semantic category, then the results are considered insufficient in step 650 . In these embodiments, only when the results are for the last extracted semantic category does the method continue with step 655 where a decision is made on whether evaluation in conjunction with results from previous semantic category extractions is feasible.
- the meaning of the text is guessed at prior to extracting all semantic categories and/or a final conclusion on a value of a semantic category is reached prior to completing an evaluation of results of that semantic category in conjunction with previously extracted semantic categories (In other words, in these embodiments the threshold of “sufficiency” is lower).
- an optional weighted grade may be calculated in step 660 as a final test of the combination of results from the different semantic categories.
- the weighted grade is derived using a formula which takes into account the grades achieved by the current results and the previous semantic category results.
- the formula could be an average with either equal weights for each semantic category or different weights for each semantic category.
- one possible formula might assign the overall category a weight of 2 and the subcategory a weight of 1. If the weighted grade is high then the results are considered sufficient in step 670 (corresponding to a ‘yes’ answer to step 535 ).
- step 658 If the weighted grade is not high enough then if further evaluation i.e. of other combinations of results from the different semantic categories is possible, more evaluations are performed in step 658 . If no further evaluation of other combinations are possible than the results are considered insufficient in step 650 (corresponding to a ‘no’ answer to step 535 ).
- pseudo classifier 425 results are also graded.
- the time extractor may in some preferred embodiments return the results in a sequence of decreasing probability.
- the results may be evaluated in conjunction with other results as described in the method of FIG. 6 .
- pseudo classifiers return either a Boolean true if a match was extracted, or a Boolean false if no match was extracted, without any grading.
- FIG. 7 illustrates in more detail a preferred embodiment of a possible sequence followed by action resolver 410 in turning to different modules of natural language module 140 in order to implement the method of FIG. 5 .
- action resolver 410 is for example a state automaton.
- the sequence shown in FIG. 7 processes the semantic categories in an order which takes advantage of the hierarchy of the semantic categories, so that results of a semantic category of a certain hierarchical level helps in the processing of a semantic category of another hierarchical level.
- parameter values belonging to the lowest level semantic category, are extracted first. At least some of the results of the parameter values extraction are used to embed tokens into the text for overall category extraction, i.e. the highest level semantic category, and for subcategory extraction, i.e. the medium level semantic category. The results of the overall category extraction is used to select a subcategory classifier, i.e. the medium level semantic category.
- the hierarchical structure of the semantic categories is therefore advantageous to the overall processing.
- text preprocessing module 435 is employed (step 702 corresponding to step 512 ).
- parameter values classifier 420 is employed (step 705 corresponding to step 520 ).
- Real time database 445 is used to identify the one or more possible corresponding parameter type(s) for each extracted parameter value.
- parameter values pseudo classifier(s) 425 is employed (step 710 corresponding to step 520 ) and corresponding parameter types are identified.
- text pre-processing module 435 is again employed in preparation for overall classifier 420 (step 712 corresponding to step 512 ).
- extracted overall category values can be evaluated in relation to extracted parameter values, for example by employing a first evaluation static component 425 to check if the parameter types of the extracted parameter values are in sync with subcategory values associated with the extracted overall category values.
- dialog management module 440 is employed (step 720 corresponding to step 545 ). If the overall category value is ambiguous (i.e. more than one possibility), then in some preferred embodiments, dialog management module 440 is employed, but in other preferred embodiments, a corresponding subcategory classifier 420 is employed for each of the possible overall category values. If dialog management module 440 has been employed, overall category classifier 420 is again employed to check the response received from the submitter (step 721 corresponding to step 520 ). Alternatively, if the response is obvious and does not need to be understood (for example the response is a selection of a multiple choice option) then overall classifier 420 does not need to be employed to check the response and step 721 may be skipped.
- Examples of situations when the method can proceed directly from step 718 to step 722 (without dialoging) include inter-alia: when only one overall category value is extracted, when there is a default overall category value for the particular embodiment, when there is only one overall category value for a particular embodiment, when more than one overall category value was extracted but the parameter types of the extracted parameter values point to one of the extracted overall category values or to subcategory values associated with one of the extracted overall category values, when more than one overall category value was extracted but one overall category value can be selected based on the grades of the extracted overall category values, and when it is decided to process more than one extracted overall category value in parallel and employ a corresponding subcategory classifier for more than one overall category values.
- Corresponding subcategory classifier 420 is employed (step 722 corresponding to step 520 ). Evaluations are then performed (corresponding to step 658 of FIG. 6 ) using evaluation static specific components 425 .
- First specific static component 425 for evaluation is employed which evaluates all the results of the previous extractions by extractors 420 and pseudo extractors 425 (step 725 ).
- First specific evaluation static component 425 checks extracted parameter values against extracted subcategory value(s) to see if the parameter values (for example based on the identified parameter types) are suitable for the extracted subcategory value(s). For example for each extracted subcategory value, first evaluation component 425 may match the parameter type(s) identified for each extracted parameter value with the parameter types expected for the extracted subcategory value as predefined in real time database 445 . The matching in this example, may result in some expected parameter types (as predefined) not matched with any extracted parameter values, matched with exactly one parameter value, or matched with more than one extracted parameter value.
- dialog management module 440 is employed (step 727 corresponding to step 545 ).
- sub-category classifier 420 is again employed to check the response received from the submitter (step 728 corresponding to step 520 ).
- the response is obvious and does not need to be understood (for example the response is a selection of a multiple choice option) then sub-category classifier 420 does not need to be employed to check the response and step 728 may be skipped.
- First specific evaluation static component 425 is again employed in step 729 .
- Examples of situation when the method can continue directly from step 726 to step 735 (without dialoging) include inter-alia: when only one subcategory value is extracted, when no subcategory value is extracted but there is a default subcategory value corresponding to the overall category value, when there is only one pre-defined subcategory value corresponding to the overall category value, when there is more than one extracted subcategory value but the parameter types of the extracted parameter values point to one of the extracted subcategory values, and when there is more than one extracted subcategory value but one subcategory value can be selected based on the grades of the extracted subcategory values.
- a second specific static evaluation (relational) component 425 is employed. Relationship evaluation component 425 evaluates the correspondence between the at least one extracted parameter value and the more than one parameter value defined for the subcategory value(step 740 ). For example if two names of cities were extracted for a ticket purchase, the second static evaluation component 425 recognizes which is a destination and which is a source.
- relational component 425 may search real time database 445 for a predefined grammar line or utterance for example in the form String/s ⁇ ParameterType X> String/s ⁇ [Arg A]; String/s ⁇ ParameterType X> String/s ⁇ [Arg b]” which means that when a parameter value of type “ParameterType X” is extracted, the parameter value will be matched with the arguments A and B required by the subcategory value according to the String/s in its context.
- a third specific static evaluation component 425 is employed in step 742 .
- This third evaluation component 425 checks if parameter values corresponding to all parameter types defined for the subcategory value were extracted (step 745 ).
- third evaluation component 425 can use a check list against the parameter types predefined for the subcategory value in real time database 445 Continuing with the example, if no parameter values were extracted for certain parameter types defined for the subcategory value, third component 425 can check if there are default parameter values which can be assigned or if the parameter types with missing parameter values are optional. Still continuing with the example, if no parameter value or if more than one parameter value was extracted for a mandatory parameter type (as predefined) which requires one parameter value and has no default value, then dialoging occurs.
- dialog management module 440 is employed (step 750 corresponding to step 545 ).
- Parameter values classifier 420 and/or one or more parameter values pseudo classifier(s) 425 is again employed to evaluate the response (step 752 corresponding to step 520 ) Alternatively, if the response is obvious and does not need to be understood (for example the response is a selection of a multiple choice option) then parameter values classifier 420 does not need to be employed to check the response and step 752 may be skipped.
- Third static evaluation component 425 is again employed in step 742
- Examples of when the dialoging of step 750 need not occur include inter-alia, when the correct number of parameter values for the parameter types required by the subcategory value were extracted, when the subcategory values requires no parameter types, and when the parameter types required by the subcategory value have default values.
- a weighted grade can first be evaluated (step 680 ) as a final test that the results are sufficient.
- One of the distinct advantages of the preferred embodiments of the present invention are the one or more dynamic aspects of the natural language understanding. Later steps of the process are adaptable based on the results of earlier steps.
- the methods illustrated in FIGS. 8, 9 , and 10 each include steps which are influenced by the results of the previous steps.
- Each of the dynamic aspects illustrated in FIGS. 8, 9 , and 10 can be separately implemented, and one or more of the dynamic aspects constitutes a separate preferred embodiment.
- the results of a previous extraction may be used to develop tokens that may be embedded in the text used as an input for the next extraction(s). The tokens become part of the input for the next extraction(s) and are therefore termed syntactic tokens.
- FIG. 8 the results of a previous extraction may be used to develop tokens that may be embedded in the text used as an input for the next extraction(s). The tokens become part of the input for the next extraction(s) and are therefore termed syntactic tokens.
- FIG. 8 the results of a previous extraction may be used to
- more than one extractor 420 or more than one pseudo extractor 425 is available for the same semantic category and the selection of extractor 420 or pseudo extractor 425 depends on the results of previous extractions.
- the dialog with a submitter can vary based on the results (including unsuccessful or no results) of previous extractions.
- FIG. 8 illustrates a preferred embodiment of a method for preparing the natural language text for extraction (step 512 ).
- the first step is to determine if the results of one or more previous extractions (by either classifiers 420 or pseudo classifiers 425 ) can be used to develop one or more syntactic tokens (step 810 ). This step is only performed during certain subsequent extractions and not for the first extraction.
- the tokens are embedded into one or more restatements of the text (step 815 ), thereby allowing results of previous extractions to directly influence subsequent extractions.
- the embedded tokens can either replace the text parts which serve as sources for the tokens, or the tokens can supplement those text parts.
- action resolver 410 is responsible for embedding the tokens, but in other preferred embodiments, other modules such as text pre-processing module 435 embeds the tokens.
- N-grams are constructed for the text or the restatements of the text (including embedded tokens) in step 820 , if required.
- n-grams are required for classifiers and some pseudo classifiers but not for all pseudo classifiers.
- text pre-processing module 435 constructs the n-grams.
- N-grams are well known in the art.
- a non-limiting definition for an n-gram based partially on “Text retrieval from Document Images based on N-gram Algorithm”, Chew Lim Tan, Sam Yuan Sung, Zhaohui Yu, and Yi Xu available at http:/citeseer.nj.nec.com/400555.html is:
- An N-Gram is a sequence of N consecutive items of a stream obtained by sliding an N-item wide window over the text one item forward at a time. Every possible N-Gram is given a number, so called the hash key. How the N-Grams are numbered is not important, as long as each instance of a certain N-Gram is always given the same number, and that distinct numbers are assigned to different N-Grams.
- the n-grams are sparse n-grams (i.e. also reflect the distance between words in the text).
- the use of sparse n-grams in some preferred embodiments of the present invention is advantageous. Sparse n-grams improve the probability of correct natural language understanding because sparse n-grams takes into account the specific order of words in a sentence.
- sparse n-grams are also trained upon (see below discussion with regard to FIG. 12 )
- the True/False indicator can be seen as an “Adjacency” indicator. If the words, doublets or triplets are composed of words which are adjacent in the text, a “True” indicator is indicated in the sparse n-gram. It should be noted that in this example in order to gain flexibility if the doublet or triplet is composed of words which are adjacent in the text, two sparse n-grams are created, one with a “true” indicator and one with a “false” indicator.
- the added flexibility enables a match between the text after n-grams are generated and the n-grams created for the trained sentence (see below FIG. 12 ), both if in the trained sentence these words were adjacent or were not adjacent.
- a sparse n-gram with a “false” indicator is created.
- the constructed n-grams are used as the input to selected classifier 420 or pseudo classifier 425 (step 825 )
- n-grams may be skipped (i.e. skip step 820 ) and the selected classifier 420 and/or pseudo classifier 425 may extract based on other techniques for example by using word spotting.
- FIG. 9 illustrates a preferred embodiment of how the results of previous extractions can influence the selection of the next classifier 420 /pseudo classifier 425 .
- the method of FIG. 9 may be included in step 515 .
- the sequence of semantic category extractions is preprogrammed.
- the method of FIG. 9 allows the selection of appropriate classifier 420 or pseudo classifier 425 .
- step 905 a decision is made on whether more than one classifier 420 or pseudo classifier 425 is available for the semantic category to be extracted, for example by checking the structure of real time database 435 If no, the one available is employed (i.e. proceed directly to step 520 ). If yes, the method continues with step 910 where a further decision is made whether a selection of less than all available classifier 420 or less than all available pseudo classifier 425 for the semantic category can be made based on previous extractions. If a selection is possible, the selection is made in step 915 . Otherwise, results of the previous extractions are clarified in step 920 , for example by dialoging with the submitter. In other cases, in step 920 , all available classifiers or all available pseudo classifiers for the semantic category are used, for example looking back at FIG. 7 in step 710 all available parameter pseudo classifiers are initially employed.
- the semantic categories include an overall category and a subcategory and also assume that there is more than one subcategory classifier 420 , with a different subcategory classifier 420 used depending on the overall category value. If the results of the overall category extraction are clear, then in step 915 the subcategory classifier 420 corresponding to the extracted overall category value is employed. If the results of the overall category extraction are unclear, then clarification is received in step 920 .
- the (four) semantic categories include an overall category, a pre-subcategory, a subcategory and parameter values.
- a different pre-subcategory classifier 420 is selected depending on the overall category value and a different subcategory classifier 420 is selected depending on the pre-subcategory value.
- Further added hierarchical levels can be processed in a complementary manner.
- step 745 it is found that a certain parameter type, for example a money expression, required by the extracted subcategory value is missing.
- a question is asked and once the answer from the submitter is received, only the currency pseudo classifier (corresponding to the missing parameter type) out of all pseudo classifiers 425 would be selected and employed on the answer.
- step 745 if it is found in step 745 that a money expression and a time expression are missing, then currency pseudo classifier 425 and time pseudo classifier 425 out of all the available pseudo classifiers 425 would be employed on the answer.
- FIG. 10 illustrates a preferred embodiment of a method for dialoguing with the submitter of the natural language (step 545 ). Additional user information may be required for example, to resolve an ambiguity, provide a missing piece of information, or restate the submission.
- the answers received from the submitter via the dialog augments previous extraction results so as to aid in understanding the natural language text.
- the term augments is used to include one or more of the following: clarifies, supplements, pinpoints, expands, narrows, etc., i.e. the answers from the dialog allows the text to be better understood than had the dialog not taken place and only the previous extraction results were available.
- a dialog management module 440 is called by action resolver 410 (step 1030 ).
- Ambiguity may result from more than one correct interpretation of a submission (for example, a request for the flight schedules leaving New York City can be interpreted as leaving Kennedy or La Guardia airport and more information from the submitter would be required to resolve the ambiguity). Further processing may also not be possible if information is missing (for example parameter values of all parameter types required by the subcategory value were not extracted).
- dialog management module 440 searches real time data base 445 for pre-defined questions associated with one or more categories. In other preferred embodiments, dialog management module 440 does not pose predefined questions, but instead formulates questions. In either case open questions and multiple choice questions may be used according to the type of missing information In one preferred embodiment, ambiguity problems result in a close-ended question, whereas missing information (for example values of missing categories) result in open-ended questions. In step 1050 , the question is posed to the submitter.
- dialog management module 440 based on this system logic and the current step in the understanding process of FIG. 5 .
- the system logic should be understood to mean logic inputted into real time database 445 to aid in the natural language understanding and which as a bonus also aids in question formulation/question string retrieval.
- the logic for a subcategory value can include the parameter types related to that subcategory value, characteristics of these parameter types independently and in relation to one another (such as when the types are mandatory), relative importance of each of these parameter types, etc.
- the question posed to the submitter is varied based on previous extraction results (where results in this context can also include non-results i.e. unsuccessful extraction).
- results in this context can also include non-results i.e. unsuccessful extraction).
- the previous extraction results can be compared to the logic in order to formulate an appropriate question.
- a first question formulated and posed to the submitter may relate only to the more important type.
- a pre-defined question may include all required parameter types for a given subcategory according to the logic. However the question strings retrieved from database 445 and used in the question posed to the submitter will relate only to those parameter types with no previously extracted parameter values.
- a pre-defined question may include all possible airports in New York State, but if the previous extractions extracted New York City, the multiple-choice question posed to the submitter will be modified so as to offer as possible responses only airports in New York City.
- Real time database 445 includes entries for the semantic categories. As mentioned above, real time database 445 in some preferred embodiments also includes dialog questions and/or suggested answers.
- FIG. 11 showing an entity-relationship (ER) diagram of database 445 .
- the entries in database 445 are divided into four types.
- the first type includes entries related to the semantic category “overall category” 1105 .
- the second type includes entries related to the semantic category “subcategory” 1115 .
- Each overall category entry 1105 has a number of subcategory entries 1115 .
- Each subcategory entry 1115 requires or accepts certain parameter type entries 1125 .
- Each parameter type entry 1125 is able to take on one or more parameter value entries 1135 .
- the parameter value entries 1135 are related to the semantic category “parameter values”.
- the invention is not bound by the illustrated ER structure or contents.
- parameter value entries 1135 are not stored under subcategory entries 1115 in the illustrated example of database 445 , parameter values under a common conceptual view would be considered to be of a lower hierarchical level as explained above.
- parameter value entries 1135 in database 445 are not stored under subcategory entries 1115 so that the same parameter value entry 1135 can relate to more than one subcategory entry 1115 .
- not all parameter values related to parameter type entries 1125 are stored as parameter value entries 1135 in database 445 .
- a particular subcategory entry 1115 may be associated with a parameter type entry 1125 (for example date) whose parameter values are extracted by pseudo classifier 420 and therefore are not stored as entries in database 445 .
- the definition of the hierarchical structure of database 445 and the characterization of the entries into the different hierarchical levels is typically performed by a system analyst with knowledge of the requirements of a particular implementation and is therefore beyond the scope of this invention.
- entries for one or more categories are manually entered in database 445 .
- entries for one or more categories can be at least partially automatically gathered from the Internet, preferably using active browsing studio work tool 170 . In preferred embodiments including request implementation through control of internet sites, this approach implies that at least part of the information used in building data base 445 originates from the medium where request implementation takes place.
- FIG. 12 shows a method for training real time database 445 so as to generate knowledge base 430 (used by classifiers 420 ), according to a preferred embodiment, which assumes the same semantic categories as in FIG. 11 .
- the first step 1205 is the defining of natural language examples.
- the second step 1210 if required for some examples, is the embedding of syntactic tokens based on entries in real time database 445 within the natural language examples. Tokens for example can include overall category entries 1105 , subcategory entries 1115 , parameter type entries 1125 and/or parameter values entries 1135 .
- Train database 450 preferably includes some examples with embedded tokens and some examples without embedded tokens (step 1212 ), so that classifiers 420 are trained to understand both text which includes proper nouns (for example, Intel) and/or common nouns (for example, stocks).
- the next step 1215 is the transformation of the examples into n-grams, preferably sparse, if required (for example if it is expected that in operation the extraction will be performed on n-grams). In some preferred embodiments, the transformation is performed by pre-processing module 435 .
- the examples, in the form of n-grams if required, are input into classifiers in 420 step 1220 .
- examples with embedded tokens corresponding to parameter type entries 1125 are used for training classifiers 420 for the overall category and subcategory.
- examples with embedded tokens corresponding to parameter value entries 1135 are used for training classifiers 420 for parameter values.
- the classifiers are trained in step 1225 . It should be evident that the same algorithm referenced above with respect to classifiers 420 can be used in training step 1225 .
- Knowledge-base 430 is obtained from the training in the final step 1230 .
- knowledge-base 430 is a data structure that is saved to a disk so knowledgebase 430 can be used later.
- knowledge work tool 150 assists in the training process. For example, once an operator of the training has chosen an entry from database 445 and the placement of a token in an example, tool 150 can develop a token from the entry and embed the token in the correct place. As an additional example, work tool 150 can employ pre-processing module 435 for developing n-grams. In certain preferred embodiments, work tool 150 also operates the training of classifiers 420 and saves the results in knowledgebase 430 .
- the training phase of knowledgebase 430 is completely separated from the usage phase—i.e. all training is performed prior to use of knowledgebase 430 .
- knowledgebase 430 continues to be expanded during the usage phase, by learning from the texts received from submitters.
- n grams are sparse and comply with the exemplary form defined above with reference to FIG. 8 :
- Action resolver 410 selects one classifier 420 to employ in step 515 .
- classifiers 420 there are three types of classifiers 420 : one overall category classifier (the domain of interest); a subcategory classifier corresponding to each overall category value (requested operation for the domain of interest); and one parameter value classifier (items required by operation).
- Parameter values classifier 420 and all available parameter value pseudo classifier(s) 425 are employed in step 520 .
- the parameter values pseudo-classifier 425 which is a time phrase extractor extracts one item: Feb. 6, 2001 i.e. the date yesterday.
- the parameter type of Feb. 6, 2001 is identified as date.
- the parameter values classifier 420 extracts items from the word “Columbia” along with the grades of the items (steps 610 and 620 ).
- Real time database 445 is used to identify the parameter-types i.e. item-types of the different items extracted from the word Columbia. Assume that five items are extracted corresponding to Columbia as a country, a university, a hospital, and twice as a stock. Due to the ambiguity, more than one item and the corresponding item type are saved.
- the text is prepared for the next extraction in step 512 , using all possible item-type matches.
- the syntactic tokens (in this example, the item types) are embedded into the text in step 815 .
- Sparse n-grams are constructed in step 820 for each of the four item-type matches (which now include the embedded tokens). It should be noted that when constructing the sparse n-grams the embedded tokens are treated as if the tokens are words and an integral part of the text.
- Action resolver 410 employs domain extractor 420 on the new sparse n-grams in step 825 . Results are evaluated in step 525 (see method of FIG. 6 ). It is assumed that two possible domains are outputted in step 610 . The first domain is hospital policies (i.e. prices for a stay at Columbia Hospital) and the second domain is Nasdaq (the market where the stocks for Columbia Records and Columbia Hospital are listed).
- step 1030 which in this example searches real time database 445 for a question (step 1040 ).
- the question and answers are “Please clarify the topic of interest a) the hospital policies of Columbia Hospital, b) The stock results of Columbia Hospital or, c) The stock results of Columbia records.
- the question is posed to the submitter in step 1050 .
- the user response is assumed to be “hospital stock” implying the stock results of Columbia Hospital (step 550 ), which is inputted into domain classifier 420 (step 520 ) to extract Nasdaq
- step 915 the operation extractor 420 related to Nasdaq is selected in step 915 and employed in step 520 .
- classifiers 420 had been previously trained. As an example, assume that the following examples were defined in step 1205 :
- parameter-type tokens may have been transformed into n-grams in step 1215 and used to train domain classifier 420 for the Nasdaq domain and/or operation classifier 420 for the operation “stock quote” in step 1225 .
- “Columbia Medical” may have been used to train parameter values classifier 420 for Columbia Hospital corresponding to both hospital parameter type and stock parameter type.
- parameter values of parameter type “stock” include Intel, Yahoo, Microsoft, AT&T, etc.
- Parameter values of parameter type “location” as in Avis dealership location include Los Angeles airport, Los Angeles downtown, San Francisco, Calif, etc. No specific time parameter values are specified for the “time” parameter type.
- Parameter values of parameter type “car group” as in rental car group include compact, sub compact, sports, 2-door, etc.
- ParameterType ParameterValues Stock Intel Yahoo, Microsoft, AT&T, . . . Location LA Airport, LA Downtown, (Avis Dealership) San Francisco, Sacramento, . . . Time No specific Items Car Group (rental) Compact, Sub Compact, Spats, 2-Door, . . .
- step 610 to 640 the parameter values extracted by parameter value classifier 420 and/or parameter values pseudo classifier 425 are output. In this example, only one parameter value “Intel” is extracted. In step 645 , as this is the first semantic category extracted, the results are insufficient.
- the overall category classifier 420 is then called and applied to n-grams created from a restatement of the original text which includes a token based on the result of the parameter value extraction, i.e. “Get a quote for ⁇ ParameterType: Stock>” (in this example the token replaces “Intel”).
- the outputted results of overall category classifier 420 are the two possible domains, with financial Information receiving a high grade and car rentals a low grade.
- the results are sorted by grade in step 640 and in step 655 , the results are evaluated in conjunction with the parameter value results. As the subcategory value is still unknown, the results are considered insufficient.
- the subcategory classifier 420 corresponding to overall category value “financial information” is called in steps 610 to 658 .
- the results include the operation with the highest grade, assumed to be “Get Stock Quote”.
- the results are checked for compliance with previous results.
- the evaluation shows that the highest graded operation is a member of the found domain and that the found parameter value is of a type accepted by the found operation as an argument.
- a weighted grade corresponding to the highest graded operation is calculated by a simple formula giving equal weights to each semantic category and the weighted grade is checked to see whether the weighted grade is above a given threshold. If the weighted grade is below the threshold, in step 675 evaluation can be attempted for other sets of results with lower grades (for example including a lower graded operation), and it can be checked whether the resulting weighted grade is higher than the given threshold.
- step 702 the text is preprocessed into n-grams because in this example it is assumed that n-grams are inputted to classifiers 420 and/or pseudo classifiers 425 .
- the n-grams are of the sparse form described above with reference to FIG. 8 .
- parameter values classifier 420 and pseudo classifiers 425 are initially called.
- the extracted parameter values include several values: LA Airport, Feb. 8, 2001 08:00 (Tomorrow's date), Mar. 13, 2001 12:00, and Sacramento.
- the text is restated so as to include tokens based on the found parameter values, namely: “rent a car ⁇ ParameterType: Time> in ⁇ ParameterType AvisDealershipLocation> until ⁇ ParameterType: Time> return to ⁇ ParameterType AvisDealershipLocation>.
- New n-Grams are created from the restated text again using the sparse n-gram form described above with the embedded tokens treated as words.
- Overall category classifier 420 is called and extracts the car rentals domain.
- step 718 because the overall category was unambiguously found the method proceeds with step 722 .
- dialoging with the user could take the form of posing a closed multiple choice question to the submitter which includes the two possible domains as choices.
- subcategory classifier 420 is called.
- First evaluation static component 425 is then called in order to try to find a match between the parameter types of the found parameter values and the expected arguments of the highest graded extracted operation.
- Static evaluation component 425 matches the parameter types “Time” and “AvisDealershipLocation” corresponding to the extracted parameter values with the corresponding arguments of the “Make Car Rental Reservation” operation
- step 735 If there had been ambiguity, a typical multiple choice question could display as choices all available operations for the found domain or all operations for the found domain which received a high grade from subcategory classifier 420 )
- Second static evaluation component 425 In steps 735 to 740 because there are parameter types which are acceptable for more than one argument of the found operation, there is a need to call second static evaluation component (Relational Static Component) 425 .
- Relational Static Component 425 identifies which values belong to which arguments by checking the context of the values. The time value Mar. 13, 2001 12:00 is recognized as the return time due to the preceding word “until”, and the value Sacramento is recognized as the return location by the preceding words “return to”. Once these values are assigned correctly to the arguments of the operation the other time and AvisDealershiplocation values follow naturally.
- third static component 425 is called to check if all required arguments have been assigned suitable values.
- third static component 425 finds that four out of the five arguments have values assigned.
- the car group argument is as yet unassigned. Therefore in step 750 in a dialog with the submitter either an open question is posed to prompt the submitter to enter the car group or a closed question is posed including as choice all possible car groups (as predefined). Once the answer is received, the last required parameter is known and results can be output.
- Step 810 checks if there are any previous results that can be developed into tokens.
- tokens for parameter types Time and AvisDealershipLocation which correspond to the extracted parameter values can be developed.
- a token is developed for each text part that had been used as a source for extraction of a parameter value.
- step 815 the developed tokens are embedded in the text in place of the source texts that were used to extract the parameter values.
- the restatement of the original request “rent a car tomorrow morning in LA airport until March 13th at noon, return to Sacramento” is restated as “rent a car ⁇ ParameterType: Time> in ⁇ ParameterType: AvisDealershipLocation> until ⁇ ParameterType: Time>return to ⁇ ParameterType: AvisDealershipLocation>”.
- steps 820 to 825 from the restatement new n-grams are constructed in the sparse n gram form described above with reference to FIG. 8 .
- Some of the n-Grams include tokens, which are dealt with as regular words.
- step 905 the answer to the question is yes when deciding whether there is more than one possible subcategory classifier 420 which can be called.
- a selection needs to be made from among the two possible subcategory classifiers 420 , one that classifies operations for the Financial Information domain and one that classifies operations for the Car Rentals domain.
- step 910 to 920 because the domain “car rentals” is assumed to have already been found, the car rentals subcategory classifier 420 is used. (If after using the overall category classifier there is still ambiguity with regard to the correct domain, dialoging in step 920 would be attempted to clarify the correct domain)
- step 1020 redundant interactions with the submitter are avoided by performing additional automatic processing to try to solve any problems without the help of the submitter.
- the parameter values LA airport, Sacramento, Feb. 8, 2001 08.00, and March 13, 12.00
- domain domain
- further processing can be performed using the first static evaluation component 420 in order to determine the desired operation by looking at the parameter types of the extracted parameter values and comparing these parameter types with the possible accepted arguments of the available operations, thereby avoiding dialoging.
- the car group value is still missing and so dialoging with the submitter is required to obtain the car group value.
- both the overall category and the subcategory can not be extracted based on the text alone.
- the overall category can be extracted from a restatement which includes a token based on an extracted parameter value, i.e. ( ⁇ parameterType:Stock>).
- This restatement implies that the desired operation to be found accepts the stock parameter type as an argument.
- only operations in the financial information domain (and not in the car rental domain) receive such values. Therefore the domain can be determined without dialoging.
- the operation is still ambiguous because all four operations in this domain accept stock as an argument. Therefore dialoging with the submitter is required to allow the submitter to select the correct operation.
- dialog management module 440 is called if no further processing is possible. Dialog management module 440 generates the correct interaction based on the current status of the handling of the request. If dialog management 440 is called while processing the first text to determine the car group value, dialog module 440 needs to create an interaction for determining the car group parameter value. Therefore dialog module 440 goes to real time database 445 and finds the string that was prepared as a question for this case specifically, i.e.—a question regarding the lack of value for this specific argument. If dialog module 440 is called for the second text in order to determine the operation, dialog module 440 needs to create an interaction that clarifies an ambiguity in the operation and presents the submitter with all possible options. Therefore, dialog module 440 goes to real time database 445 and finds the String that was prepared for this specific case, i.e.—operation ambiguity interaction. Once the question is formatted, the question is transferred to the submitter and the reply of the submitter is analyzed.
- the initial creation of knowledge base 430 includes the following steps.
- step 1205 natural language examples are defined for the supported domains, operations and parameter values.
- the following examples may be used, inter-alia for training:
- LA Airport “Los Angeles Airport” ⁇ Parameter value: LA Airport
- step 1210 tokens are embedded in some of the above examples. For example:
- the training examples are turned into n-Grams, and the classifiers are trained on the n-grams, with the results serialized into Knowledgebase 430 .
- the training process is classifier-specific allowing the examples in their n-gram representation to be associated with the categories and values which were trained on those n-grams.
- system may be a suitably programmed computer.
- the invention contemplates a computer program being readable by a computer for executing the method of the invention.
- the invention further contemplates a machine-readable memory tangibly embodying a program of instructions executable by the machine for executing the method of the invention.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Machine Translation (AREA)
Abstract
Methods and systems for dynamic natural language understanding. A hierarchical structure of semantic categories is exploited to assist in the natural language understanding. Optionally, the natural language to be understood includes a request.
Description
- This application is a continuation of application Ser. No. 10/097,537 filed Mar. 13, 2002, which, under 35 U.S.C. 119, claims the priority date of Mar. 13, 2001 of provisional application 60/275,598, all of which are incorporated herein by reference.
- This invention relates to natural language understanding.
- Natural language understanding systems and methods traditionally use strict grammar or statistics.
- Grammar based natural language understanding systems and methods typically use a parser to parse a text into a tree, i.e. a hierarchical (“depth”) structure. Elements of the tree are processed in a hierarchical manner, either bottom up or top down. In order to achieve successful understanding of the text, the sentence structure/grammar generally needs to conform to rules, thereby placing constraints on the freedom of expression of the submitter of the text.
- Statistically based natural language understanding systems and methods typically use many statistical methods including classification to understand a text. Freedom of expression by the submitter of the text is therefore enhanced.
- Systems of the related art include the following:
- U.S. Pat. No. 5,680,511 to Baker, et al. in one aspect, provides word recognition systems that operate to recognize an unrecognized or ambiguous word that occurs within a passage of words. The system can offer several words as choice words for inserting into the passage to replace the unrecognized word. The system can select the best choice word by using the choice word to extract from a reference source, sample passages of text that relate to the choice word. For example, the system can select the dictionary passage that defines the choice word. The system then compares the selected passage to the current passage, and generates a score that indicates the likelihood that the choice word would occur within that passage of text. The system can select the choice word with the best score to substitute into the passage. The passage of words being analyzed can be any word sequence including an utterance, a portion of handwritten text, a portion of typewritten text or other such sequence of words, numbers and characters. Alternative embodiments of the present invention are disclosed which function to retrieve documents from a library as a function of context.
- U.S. Pat. No. 5,642,519 to Martin provide a unified grammar for a speech interpreter capable of real-time speech understanding for user applications running on a general purpose microprocessor-based computer. The speech interpreter includes a unified grammar (UG) compiler, a speech recognizer and a natural language (NL) processor. The UG compiler receives a common UG lexicon and unified grammar description, and generates harmonized speech recognition (SR) and NL grammars for the speech recognizer and natural language processor, respectively. The lexicon includes a plurality of UG word entries having predefined characteristics, i.e., features, while the UG description includes a plurality of complex UG rules which define grammatically allowable word sequences. The UG compiler converts the complex UG rules (complex UG rules include augmentations for constraining the UG rules) into permissible SR word sequences and SR simple rules (simple rules do not include any augmentation) for the SR grammar. The SR grammar is a compact representation of the SR word entries corresponding to the UG word entries, permissible SR word sequences and simple SR rules corresponding to the augmentations of the complex UG rules. The NL grammar provides the NL processor with NL patterns enabling the NL processor to extract the meaning of the validated word sequences passed from the speech recognizer.
- U.S. Pat. No. 5,991,712 also to Martin teaches that improved word accuracy of speech recognition can be achieved by providing a scheme for automatically limiting the acceptable word sequences. Speech recognition systems consistent with the present invention include a lexicon database with words and associated lexical properties. The systems receive exemplary clauses containing permissible word combinations for speech recognition, and identify additional lexical properties for selected words in the lexicon database corresponding to words in the received exemplary clauses using lexical property tests of a grammar database. Certain lexical property tests are switchable to a disabled state. To identify the additional lexical properties, the exemplary clauses are parsed with the switchable lexical property tests disabled to produce an index of the lexical properties corresponding to the exemplary clauses. The lexicon database is updated with the identified additional lexical properties by assigning the lexical properties to the corresponding words of the lexicon database. The grammar database is compiled with the lexical property tests enabled and the lexicon database with the assigned lexical properties to produce a grammar that embodies constraints of the lexical property tests and the lexical properties.
- U.S. Pat. No. 5,918,222 to Fukui, et al. teaches a data storage means for storing data in a predetermined information form. An information retrieval means retrieves the data stored in the data storage means. A reception means receives an information disclosure demand from a demander, a response rule storage means for storing general knowledge for generating a response responding to the demander and a personal relationship information associated with a unique personal relationship between a user having the data on an information provider side and a user on an information demander side. A response plan formation means, responsive to the demand received by the reception means, plans a response for exhibiting, to the information demander, data obtained by causing the retrieval means to retrieve the data stored in the data storage means on the basis of the knowledge and the personal relationship information stored in the response rule storage means. A response generation means generates the response to the information demander in accordance with the plan formed by the response plan formation means.
- U.S. Pat. No. 5,987,404 to Della Pietra, et. al. proposes using statistical methods to do natural language understanding. The key notion is that there are “strings” of words in the natural language, that correspond to a single semantic concept. One can then define an alignment between an entire semantic meaning (consisting of a set of semantic concepts), and the English. This is modeled using P(E,A|S). One can model p(S) separately. This allows each parameter to be modeled using many different statistical models.
- U.S. Pat. No. 5,576,954 to Driscoll teaches a procedure for determining text relevancy and can be used to enhance the retrieval of text documents by search queries. This system helps a user intelligently and rapidly locate information found in large textual databases. A first embodiment determines the common meanings between each word in the query and each word in the document. Then an adjustment is made for words in the query that are not in the documents. Further, weights are calculated for both the semantic components in the query and the semantic components in the documents. These weights are multiplied together, and their products are subsequently added to one another to determine a real value number (similarity coefficient) for each document. Finally, the documents are sorted in sequential order according to their real value number from largest to smallest value. Another, embodiment is for routing documents to topics/headings (sometimes referred to as filtering). Here, the importance of each word in both topics and documents are calculated. Then, the real value number (similarity coefficient) for each document is determined. Then each document is routed one at a time according to their respective real value numbers to one or more topics. Finally, once the documents are located with their topics, the documents can be sorted. This system can be used to search and route all kinds of document collections, such as collections of legal documents, medical documents, news stories, and patents.
- U.S. Pat. No. 5,642,502 also to Driscoll teaches a system and method for retrieving relevant documents from a text data base collection comprised of patents, medical and legal documents, journals, news stories and the like. Each small piece of text within the documents such as a sentence, phrase and semantic unit in the data base is treated as a document. Natural language queries are used to search for relevant documents from the data base. A first search query creates a selected group of documents. Each word in both the search query and in the documents are given weighted values. Combining the weighted values creates similarity values for each document which are then ranked according to their relevant importance to the search query. A user reading and passing through this ranked list checks off which documents are relevant or not. Then the system automatically causes the original search query to be updated into a second search query which can include the same words, less words or different words than the first search query. Words in the second search query can have the same or different weights compared to the first search query. The system automatically searches the text data base and creates a second group of documents, which as a minimum does not include at least one of the documents found in the first group. The second group can also be comprised of additional documents not found in the first group. The ranking of documents in the second group is different than the first ranking such that the more relevant documents are found closer to the top of the list.
- U.S. Pat. No. 5,893,092 also to Driscoll teaches a search system and method for retrieving relevant documents from a text data base collection comprised of patents, medical and legal documents, journals, news stories and the like. Each small piece of text within the documents such as a sentence, phrase and semantic unit in the data base is treated as a document. Natural language queries are used to search for relevant documents from the data base. A first search query creates a selected group of documents. Each word in both the search query and in the documents are given weighted values. Combining the weighted values creates similarity values for each document which are then ranked according to their relevant importance to the search query. A user reading and passing through this ranked list checks off which documents are relevant or not. Then the system automatically causes the original search query to be updated into a second search query which can include the same words, less words or different words than the first search query. Words in the second search query can have the same or different weights compared to the first search query. The system automatically searches the text data base and creates a second group of documents, which as a minimum does not include at least one of the documents found in the first group. The second group can also be comprised of additional documents not found in the first group. The ranking of documents in the second group is different than the first ranking such that the more relevant documents are found closer to the top of the list.
- U.S. Pat. No. 6,088,692 also to Driscoll teaches a natural language search system and method for retrieving relevant documents from a text data base collection comprised of patents, medical and legal documents, journals, news stories and the like. Each small piece of text within the documents such as a sentence, phrase and semantic unit in the data base is treated as a document. Natural language queries are used to search for relevant documents from the data base. A first search query creates a selected group of documents. Each word in both the search query and in the documents are given weighted values. Combining the weighted values creates similarity values for each document which are then ranked according to their relevant importance to the search query. A user reading and passing through this ranked list checks off which document are relevant or not. Then the system automatically causes the original search query to be updated into a second search query which can include the same words, less words or different words than the first search query. Words in the second search query can have the same or different weights compared to the first search query. The system automatically searches the text data base and creates a second group of documents, which as a minimum does not include at least one of the documents found in the first group. The second group can also be comprised of additional documents not found in the first group. The ranking of documents in the second group is different than the first ranking such that the more relevant documents are found closer to the top of the list.
- U.S. Pat. No. 5,694,592 also to Driscoll teaches a procedure for determining text relevancy that can be used to enhance the retrieval of text documents by search queries. This system helps a user intelligently and rapidly locate information found in large textual databases. A first embodiment determines the common meanings between each word in the query and each word in the document. Then an adjustment is made for words in the query that are not in the documents. Further, weights are calculated for both the semantic components in the query and the semantic components in the documents. These weights are multiplied together, and their products are subsequently added to one another to determine a real value number(similarity coefficient) for each document. Finally, the documents are sorted in sequential order according to their real value number from largest to smallest value. Another, embodiment is for routing documents to topics/headings (sometimes referred to as faltering). Here, the importance of each word in both topics and documents are calculated. Then, the real value number (similarity coefficient) for each document is determined. Then each document is routed one at a time according to their respective real value numbers to one or more topics. Finally, once the documents are located with their topics, the documents can be sorted. This system can be used to search and route all kinds of document collections, such as collections of legal documents, medical documents, news stories, and patents.
- U.S. Pat. No. 6,138,085 to Richardson, et al. teaches a facility for determining, for a semantic relation that does not occur in a lexical knowledge base, whether this semantic relation should be inferred despite its absence from the lexical knowledge base. This semantic relation to be inferred is preferably made up of a first word, a second word, and a relation type relating the meanings of the first and second words. In a preferred embodiment, the facility identifies a salient semantic relation having the relation type of the semantic relation to be inferred and relating the first word to an intermediate word other than the second word. The facility then generates a quantitative measure of the similarity in meaning between the intermediate word and the second word. The facility further generates a confidence weight for the semantic relation to be inferred based upon the generated measure of similarity in meaning between the intermediate word and the second word. The facility may also generate a confidence weight for the semantic relation to be inferred based upon the weights of one or more paths connecting the first and second words
- U.S. Pat. No. 5,675,710 to Lewis teaches a method and apparatus for training a text classifier. A supervised learning system and an annotation system are operated cooperatively to produce a classification vector which can be used to classify documents with respect to a defined class. The annotation system automatically annotates documents with a degree of relevance annotation to produce machine annotated data. The degree of relevance annotation represents the degree to which the document belongs to the defined class. This machine annotated data is used as input to the supervised learning system. In addition to the machine annotated data, the supervised learning system can also receive manually annotated data and/or a user request. The machine annotated data, along with the manually annotated data and/or the user request, are used by the supervised learning system to produce a classification vector. In one embodiment, the supervised learning system comprises a relevance feedback mechanism. The relevance feedback mechanism is operated cooperatively with the annotation system for multiple iterations until a classification vector of acceptable accuracy is produced. The classification vector produced by the invention is the result of a combination of supervised and unsupervised learning
- U.S. Pat. No. 6,311,152 to Bai, et. al teaches a system (100, 200) for tokenization and named entity recognition of ideographic language. In the system, a word lattice is generated for a string of ideographic characters using finite state grammars (150) and a system lexicon (240). Segmented text is generated by determining word boundaries in the string of ideographic characters using the word lattice dependent upon a contextual language model (152A) and one or more entity language models (152B). One or more named entities is recognized in the string of ideographic characters using the word lattice dependent upon the contextual language model (152A) and the one or more entity language models (152B). The contextual language model (152A) and the one or more entity language models (152B) are each class-based language models. The lexicon (240) includes single ideographic characters, words, and predetermined features of the characters and words.
- What is needed in the art is a method and system for understanding natural language that includes inter alia statistical steps and elements which also take advantage of hierarchical-structure. What is also needed in the art is a system and method where the extraction of one part of a text which belongs to one semantic category assists in the extraction of another part which belongs to a semantic category of a different hierarchical level. In addition, what is needed in the art is a method and system for understanding natural language where later steps of the process are affected based on the results of earlier steps, thereby introducing a dynamic aspect to the method and system.
- According to the present invention, there is provided a method for use in a method for understanding a natural language text, comprising performing the following selectively in a statistical manner: attempting to extract at least one value belonging to a semantic category from a natural language text or a form thereof; and if a result of the attempting complies with a predetermined criterion, attempting to extract, based on the result, at least one value belonging to another semantic category of a different hierarchical level than the semantic category, else performing at least one action from a group of actions including: asking a submitter of the text a question whose content depends on the result and giving up on understanding the natural language text.
- In one embodiment, the predetermined criterion is at least one from a group including: at least one value for the semantic category was extracted, only one value for the semantic category was extracted, one of the at least one value extracted for the semantic category is selected based on a grade thereof, a correct number of values for the semantic category were extracted, a correct number of values for the semantic category are selected based on grades thereof from among the at least one value extracted for the semantic category, at least some values belonging to other previously extracted at least one semantic category are appropriate for at least one value extracted for the semantic category, values belonging to other previously extracted at least one semantic category are appropriate for only one value extracted for the semantic category, the semantic category is a particular semantic category where an unlimited number of extracted values is allowed, it is desired to process in parallel more than one extracted value for the semantic category, there is a default value corresponding to each required value for the semantic category which was not extracted, there is only one possible value for the semantic category, and there is only a correct number of possible values for the semantic category.
- According to the present invention, there is also provided, a method for understanding a natural language text, comprising: receiving a natural language text; processing each at least two semantic categories, the each on a different hierarchical level, by performing the following selectively in a statistical manner: (i) attempting to determine at least one value belonging to the each semantic category through extraction, wherein if the each semantic category is not a first processed of the at least two semantic categories, then the attempting is based on results of previously processed semantic categories, and(ii) if the each semantic category is not a last processed of the at least two semantic categories and a result of the attempting does not comply with a predetermined criterion, dialoging with a submitter of the text and receiving at least one answer from the submitter, wherein at least one value determined from the at least one answer augments the result so as to comply with the predetermined criterion and allow extraction attempts for other of the at least two semantic categories to be subsequently processed; and evaluating values determined for the at least two semantic categories with respect to one another to determine whether the values are sufficient to understand the text, and if the values are not sufficient: dialoging with the submitter, receiving at least one answer from the submitter, determining from the at least one answer at least one value belonging to at least one of the at least two semantic categories, the at least one value in conjunction with earlier determined values being sufficient to understand the text.
- According to the present invention there is further provided a method for training at least two classifiers to understand a natural language text, comprising: introducing entries into a database, the entries belonging to at least two semantic categories of different hierarchical levels; defining examples of natural language texts, wherein at least some of the examples include embedded syntactic tokens based on the entries; and training at least two classifiers for the at least two semantic categories using the examples or a form thereof.
- According to the present invention, there is provided a module for use in a system for natural language understanding, comprising: at least one classifier or pseudo classifier configured to extract values belonging to a semantic category from a natural language text or a form thereof; and an action resolver configured if a result of extracting values of the semantic category complies with a predetermined criterion to employ based on the result at least one classifier or pseudo classifier to extract values belonging to another semantic category of a different hierarchical level, and configured if the result does not comply with a predetermined criterion to perform at least one action from a group of actions including: employing based on the result a dialog management module and giving up on understanding the natural language text.
- According to the present invention, there is also provided: a system for natural language understanding, comprising: at least two classifiers or pseudo classifiers configured to extract values belonging to at least two semantic categories on different hierarchical levels from a natural language text or a form thereof; a dialog management module configured to dialog with a submitter of the natural language text; at least one evaluation module configured to evaluate values belonging to the at least two semantic categories; and an action resolver configured to cause the text to be understood by (i) employing, if a result of extracting values of a semantic category complies with a predetermined criterion and the semantic category is not a last to be processed semantic category, a classifier or pseudo classifier based on the result to extract values belonging to another semantic category, by (ii) employing, if the result does not comply with a predetermined criterion and the semantic category is not a last to be processed semantic category, a dialog management module and then employing, based on the result as augmented by at least one answer received from the submitter by the dialog management module, a classifier or pseudo classifier to extract values belonging to another semantic category, and by (iii) employing the evaluation module to evaluate the values of the at least two semantic categories in relation to one another in order to determine if the values are sufficient to understand the text and if the values are not sufficient employing the dialog management module to determine at least one value, the at least one value in conjunction with the values being sufficient to understand the text.
- According to the present invention, there is further provided a system for training classifiers for natural language understanding, comprising: a real time database including entries related to semantic categories on at least two different hierarchical levels; classifiers for the semantic categories; and a knowledge work tool configured to develop syntactic tokens from the entries, embed the tokens in examples and train the classifiers at least partially on the examples.
- According to the present invention there is still further provided a method for understanding a natural language text, comprising performing the following in a selectively statistical manner: receiving a natural language text; extracting at least one parameter value from the text or a form thereof; identifying at least one parameter type related to each extracted parameter value; providing at least one restatement of the received text, each at least one restatement having embedded within, at least one of the identified parameter types; extracting at least one overall category value from the at least one restatement or a form thereof,; selecting a subcategory extractor corresponding to one of the extracted at least one overall category, and using the selected subcategory extractor to extract at least one subcategory value; choosing one of the at least one extracted subcategory values; evaluating the at least one identified parameter type in relation to the chosen subcategory value; and concluding that the natural language text is understood.
- According to the present invention, there is yet further provided system for understanding a natural language text, comprising: one classifier configured to extract an overall category value from a natural language text or a form thereof; a different classifier corresponding to each overall category value configured to extract subcategory values from a natural language text or a form thereof; one classifier configured to extract parameter values from a natural language text or a form thereof; a dialog management module configured to dialog with a submitter of the natural language text; at least one evaluation component configured to evaluate extracted values; and an action resolver configured to employ different parts of the system in turn in order to understand the natural language text, including employing the one classifier for parameter values before the one overall category classifier and employing the overall category classifier before the corresponding subcategory classifier.
- According to the present invention, there is provided a program storage device readable by machine, tangibly embodying a program of instructions executable by the machine to perform method steps for use in a method for understanding a natural language text, comprising performing the following selectively in a statistical manner: attempting to extract at least one value belonging to a semantic category from a natural language text or a form thereof; and if a result of the attempting complies with a predetermined criterion, attempting to extract, based on the result, at least one value belonging to another semantic category of a different hierarchical level than the semantic category, else performing at least one action from a group of actions including: asking a submitter of the text a question whose content depends on the result and giving up on understanding the natural language text.
- According to the present invention, there is also provided a computer program product comprising a computer useable medium having computer readable program code embodied therein for use in a computer program product comprising: computer readable program code for causing the computer to perform the following selectively in a statistical manner: computer readable program code for causing the computer to attempt to extract at least one value belonging to a semantic category from a natural language text or a form thereof; and computer readable program code for causing the computer if a result of the attempting complies with a predetermined criterion to attempt to extract, based on the result, at least one value belonging to another semantic category of a different hierarchical level than the semantic category, else performing at least one action from a group of actions including: asking a submitter of the text a question whose content depends on the result and giving up on understanding the natural language text.
- According to the present invention, there is further provided a program storage device readable by machine, tangibly embodying a program of instructions executable by the machine to perform method steps for understanding a natural language text, comprising: receiving a natural language text; processing each at least two semantic categories, the each on a different hierarchical level, by performing the following selectively in a statistical manner: (i) attempting to determine at least one value belonging to the each semantic category through extraction, wherein if the each semantic category is not a first processed of the at least two semantic categories, then the attempting is based on results of previously processed semantic categories, and (ii) if the each semantic category is not a last processed of the at least two semantic categories and a result of the attempting does not comply with a predetermined criterion, dialoging with a submitter of the text and receiving at least one answer from the submitter, wherein at least one value determined from the at least one answer augments the result so as to comply with the predetermined criterion and allow extraction attempts for other of the at least two semantic categories to be subsequently processed; and evaluating values determined for the at least two semantic categories with respect to one another to determine whether the values are sufficient to understand the text, and if the values are not sufficient: dialoging with the submitter, receiving at least one answer from the submitter, determining from the at least one answer at least one value belonging to at least one of the at least two semantic categories, the at least one value in conjunction with earlier determined values being sufficient to understand the text.
- According to the present invention, there is still further provided a computer program product comprising a computer useable medium having computer readable program code embodied therein for understanding a natural language text, the computer program product comprising: computer readable program code for causing the computer to receive a natural language text; computer readable program code for causing the computer to process each at least two semantic categories, the each on a different hierarchical level, by performing the following selectively in a statistical manner: computer readable program code for causing the computer to (i) attempt to determine at least one value belonging to the each semantic category through extraction, wherein if the each semantic category is not a first processed of the at least two semantic categories, then the attempting is based on results of previously processed semantic categories, and computer readable program code for causing the computer to (ii) if the each semantic category is not a last processed of the at least two semantic categories, and a result of the attempting does not comply with a predetermined criterion, dialog with a submitter of the text and receive at least one answer from the submitter, wherein at least one value determined from the at least one answer augments the result so as to comply with the predetermined criterion and allow extraction attempts for other of the at least two semantic categories to be subsequently processed; and computer readable program code for causing the computer to: evaluate values determined for the at least two semantic categories with respect to one another to determine whether the values are sufficient to understand the text, and if the values are not sufficient: dialog with the submitter, receive at least one answer from the submitter, determine from the at least one answer at least one value belonging to at least one of the at least two semantic categories, the at least one value in conjunction with earlier determined values being sufficient to understand the text.
- According to the present invention, there is provided a program storage device readable by machine, tangibly embodying a program of instructions executable by the machine to perform method steps for training at least two classifiers to understand a natural language text, comprising: introducing entries into a database, the entries belonging to at least two semantic categories of different hierarchical levels; defining examples of natural language texts, wherein at least some of the examples include embedded syntactic tokens based on the entries; and training at least two classifiers for the at least two semantic categories using the examples or a form thereof.
- According to the present invention there is also provided a computer program product comprising a computer useable medium having computer readable program code embodied therein for training at least two classifiers to understand a natural language text, the computer program product comprising: computer readable program code for causing the computer to introduce entries into a database, the entries belonging to at least two semantic categories of different hierarchical levels; computer readable program code for causing the computer to define examples of natural language texts, wherein at least some of the examples include embedded syntactic tokens based on the entries; and computer readable program code for causing the computer to train at least two classifiers for the at least two semantic categories using the examples or a form thereof.
- According to the present invention, there is further provided a program storage device readable by machine, tangibly embodying a program of instructions executable by the machine to perform method steps for understanding a natural language text, comprising performing the following in a selectively statistical manner: receiving a natural language text; extracting at least one parameter value from the text or a form thereof; identifying at least one parameter type related to each extracted parameter value; providing at least one restatement of the received text, each at least one restatement having embedded within, at least one of the identified parameter types; extracting at least one overall category value from the at least one restatement or a form thereof; selecting a subcategory extractor corresponding to one of the extracted at least one overall category, and using the selected subcategory extractor to extract at least one subcategory value; choosing one of the at least one extracted subcategory values; evaluating the at least one identified parameter type in relation to the chosen subcategory value; and concluding that the natural language text is understood.
- According to the present invention there is yet further provided, a computer program product comprising a computer useable medium having computer readable program code embodied therein for understanding a natural language text, the computer program product comprising: computer readable program code for causing the computer to perform the following in a selectively statistical manner: computer readable program code for causing the computer to receive a natural language text; computer readable program code for causing the computer to extract at least one parameter value from the text or a form thereof; computer readable program code for causing the computer to identify at least one parameter type related to each extracted parameter value; computer readable program code for causing the computer to provide at least one restatement of the received text, each at least one restatement having embedded within, at least one of the identified parameter types; computer readable program code for causing the computer to extract at least one overall category value from the at least one restatement or a form thereof; computer readable program code for causing the computer to select a subcategory extractor corresponding to one of the extracted at least one overall category, and use the selected subcategory extractor to extract at least one subcategory value; computer readable program code for causing the computer to choose one of the at least one extracted subcategory values; computer readable program code for causing the computer to evaluate the at least one identified parameter type in relation to the chosen subcategory value; and computer readable program code for causing the computer to conclude that the natural language text is understood.
- In order to understand the invention and to see how it may be carried out in practice, a preferred embodiment will now be described, by way of non-limiting example only, with reference to the accompanying drawings, in which:
-
FIG. 1 is a block diagram of a system for understanding a natural language input and optionally executing a request included therein, according to a preferred embodiment of the present invention; -
FIG. 2 is a sample screen of a knowledge base work tool, according to a preferred embodiment of the present invention; -
FIG. 3 is a sample screen showing the processing of an active browsing script according to a preferred embodiment of the present invention; -
FIG. 4 is a block diagram of a module for understanding a natural language text, according to a preferred embodiment of the present invention; -
FIG. 5 is a flow chart of a method for understanding a natural language text, according to a preferred embodiment of the present invention -
FIG. 6 is a flow chart of a method for evaluating extraction results, according to a preferred embodiment of the present invention; -
FIGS. 7A and 7B show a sequence for employing different modules of the natural language module, according to a preferred embodiment of the present invention; -
FIG. 8 is a flow chart for preparing a text for extraction, according to a preferred embodiment of the present invention; -
FIG. 9 is a flow chart for selecting a classifier or pseudo classifier based on previous extraction results, according to a preferred embodiment of the present invention; -
FIG. 10 is a flow chart for interaction with the submitter of a natural language text, according to a preferred embodiment of the present invention; -
FIG. 11 is an entity-relationship (ER) diagram of a real time database, according to a preferred embodiment of the present invention; and -
FIG. 12 is a flow chart of a method for training a natural language module, according to a preferred embodiment of the present invention. - The preferred embodiment relates to a system and method for understanding natural language.
- In certain preferred embodiments of the present invention, the natural language to be understood includes a request. However, the invention is not limited to understanding requests and also applies to understanding natural language inputs which do not include a request.
- All examples given below are non-limiting illustrations of the invention described and defined herein.
-
FIG. 1 is an example of a block diagram of asystem 100 for understanding natural language and if the natural language includes a request, optionally executing the request, according to a preferred embodiment of the present invention.System 100 assumes remote access through a network, such as the Internet, but it should be evident that local access is within the scope of the invention. - It is assumed that a user inputs natural language through a
client 110.Client 110 is shown here as a Laptop Computer however it should be evident thatclient 110 can be any input device, for example computers, PDAs (Personal Digital Assistants), phones, cellular phones, cellular phones with SMS or WAP capabilities, fax, scanner, etc. Depending on the type ofclient 110 used, pre-handling of the input may be necessary, for example speech to text conversion, optical character recognition etc. These pre-handling techniques are well known in the art. Anoptional network profiler 120 manages remote networks (not shown), controlling network properties as well as the inflow and outflow of requests to and from the remote systems. Anoptional control server 130 manages the connection betweenclient 110 and the rest ofsystem 100, and also the internal connections within the rest ofsystem 100. - A natural language understanding (NLU)
server 140 includes the software for understanding the natural language. In order for the software onNLU server 140 to be able to comprehend the natural language, a preparation stage is necessary which includes for example the implementation of training activity. The preparation stage can be conducted, for example, using a knowledgebuilder work tool 150. A sample screen ofwork tool 150 is shown inFIG. 2 . - Assuming that the input is a natural language request, a
request implementation server 160, allows the implementation of the understood request. For example ifrequest implementation server 160 is an active browsing server, implementation can occur through control of Internet sites automatically during runtime. In one preferred embodiment, every request supported bysystem 100 is represented by a script code of several lines that defines how and where the request should be carried out. The active browsing script can be prepared, for example, using an active browsingstudio work tool 170.FIG. 3 shows a sample screen of the processing of an active browsing script so as to carry out the request. (Browsing work tool 170 in some preferred embodiments may also be used in building areal time database 445 as explained below with reference toFIG. 11 ). - In a non-network system,
NLU server 140 andrequest implementation server 160 in certain preferred embodiments are replaced by natural language understanding (NLU)module 140 andrequest implementation module 160, respectively. NLU module orserver 140 and request implementation server ormodule 160 in certain preferred embodiments are separately implemented so that for example, NLU module/server 140 can be used with another means of request implementation or even with no request implementation means. For example, even if the natural language is assumed to include a request, the request may be understood without being implemented. In preferred embodiments where the user input is not a request, request implementation means may be unnecessary. - Similarly, in certain preferred embodiments, other means of natural language understanding could be used instead of NLU module/
server 140 in a system with request implementation module/server 160. In other preferred embodiments, NLU module/server 140 and implementation module/server 160 can be integrated into one unit or separated into more than two units. - For the purposes of the description below the term “module” as in NLU module and request implementation module is assumed to refer to both modules and servers, which may form part of network or non-network systems.
-
FIG. 4 shows an example ofNLU module 140, according to a preferred embodiment of the present invention. It should be evident that the modules shown inFIG. 4 as forming part ofNLU module 140 and discussed below can be integrated or divided into a smaller or larger number of modules. The actual separation of the functions ofNLU module 140 into the modules shown inFIG. 4 is for ease of understanding only. In one preferred embodiment of the invention, the modules shown on the bottom ofFIG. 4 are associated with the online (i.e. using) stage and the modules shown on the top ofFIG. 4 are associated with the offline (training) stage.FIG. 4 is discussed below in conjunction with flow charts illustrating the methods for using andtraining module 140. The order of the steps in one or more of the methods illustrated in the flow charts may be varied in other preferred embodiments. In other preferred embodiments, some steps in one or more of the methods in the flow charts may be omitted and/or additional steps may be added. - An example of the overall method for understanding the submission from a user is shown in
FIG. 5 , according to a preferred embodiment of the present invention. The method shown inFIG. 5 enables understanding of the text, without compelling the text to comply to a pre-defined grammar. - User input is received (step 510) from a submitter, for example through
client 110. As mentioned above, input can optionally include a request. It is assumed that any necessary pre-handling of the input has already been performed as explained above so that the input is received bymodule 140 in a format compatible withmodule 140. In one preferred embodiment, the format is ASCII. Herein below, once the input has been pre-handled to a format compatible withmodule 140, the input is also referred to as “text”. - The next step is preparing the text, if necessary, for processing (step 512) using a
text pre-preprocessing module 435. Anaction resolver module 410 decides which classifier module (also sometimes termed extractor) 420 or pseudo-classifierstatic component 425 to employ (step 515). Each classifier or group ofclassifiers 420 extracts values belonging to a different semantic category. One or more pseudo-classifierstatic components 425 extract phrases that can not be learned or do not need to be learned (as will be explained further below) belonging to one or more semantic categories. Selectedclassifier 420 or pseudo classifierstatic component 425 is employed on the text (step 520), and the results of the extraction are evaluated by action resolver 410 (step 525), as will be described below. If the results of selectedclassifier 420 are sufficient to understand the text (i.e. it is concluded that the text is understood) (step 535), the results are optionally output (for example to the submitter, or to controlserver 130 and from there to request implementation module 160) instep 530. Outputting the results is one possible way of indicating that the text is understood. Alternatively, other indications of understanding can be used, for example an indication that a request optionally included in the input was implemented. - If the results are insufficient to understand the text at this stage, but based on the results, a further extraction can be performed by a
classifier 420 or pseudo classifier 425 (step 540),action resolver 410 prepares the text, if necessary, for further extraction (step 512) andaction resolver 410 chooses thenext classifier 420 or the nextpseudo classifier 425 for the next semantic category to extract (step 515). Alternatively, a further extraction may not be able to be performed and the results are insufficient to understand the text. This situation may occur, for example, if one or more ofclassifiers 420 could not classify the text into any one of the possible semantic meanings that classifier 420 knows. In this case,NLU module 140 may be considered to have failed to understand the text and may stop any further processing (step 560). The results of the failure can be optionally communicated to the submitter. Alternatively and more preferably,action resolver module 410 may dialog with the submitter instep 545, and receive submitter response instep 550. Once the response is received a further extraction may be performed on the submitter response instep 520 using theclassifier 420 orpseudo classifier 425 for the same semantic category, or aclassifier 420/pseudo classifier 425 for another previously extracted semantic category. - It should be noted that the method described above with reference to
FIG. 5 includes some steps performed in a selectively statistical manner. For example, when aclassifier 420 is used instep 520, the step is a statistically based step, whereas when apseudo classifier 425 is used instep 520, the step is typically non-statistically based. - Classifiers are well known in the art. An example of a public domain algorithm which can be used by
classifiers 420 of this invention includes Naive-bayes text-classification developed by Carnegie Mellon University and available on the world wide web at www.cs.cmu.edu/afs/cs/project/theo-11/www/naive-bayes.html. This public domain algorithm is based on “Machine Learning” by Tom Mitchell, Chapter 6, McGraw Hill 1997, ISBN 0070428077. - A pseudo classifier is a type of
static component 425. Pseudo classifiers do not need to be trained. Non limiting examples ofpseudo classifiers 425 include time extractors and currency extractors, which detect and extract time phrases and currency phrases respectively. In certain preferred embodiments, a data structure for time is used which keeps values for seconds, minutes, days, months, years, etc. Time extractors can therefore translate phrases such as “next Tuesday” into next Tuesday's date, or translate phrases like “for three days”, “every Sunday” etc. In certain preferred embodiments, the data structure for money is in the form of #x.yy, where x is an integer and yy is any number between 00 and 99 and #can be replaced by any currency type. Pseudo classifiers such as time and currency extractors may in some cases be based on limited strict grammars, suitable only for specific time or money expressions, and may use well known parsing methods for detecting and parsing the temporal expressions (time-phrases) or currency phrases. In these preferred embodiments, the parsing may be partial and include a part of the given text that could be parsed based on the limited grammars and later transformed to a data structure that can hold the time or currency expressions. An example of grammar based temporal expression parsing (which as well known in the art can be easily modified for currency phrases) can be found at odur.let.rug.nl/{tilde over ( )}vannoord/papers/yearbook/node2.html as a part of a work called “Grammatical Analysis in a Spoken Dialogue System” by Gosse Bouma, Rob Koeling, Mark-Jan Nederhof and Gertjan Van Noord. - A semantic category should be understood for the purpose of the description below to be a grouping of values including at least one common property which distinguishes those values from values in other semantic categories that are defined for a particular embodiment.
- In preferred embodiments of the present invention, there is a hierarchical structure among the semantic categories which is exploited when understanding the text. In certain preferred embodiments, there are three semantic categories: overall category (highest level), subcategory (medium level), and parameter values (lowest level). As an example one or
more classifier 420 orpseudo classifiers 425 may extract value(s) belonging to the overall category. The overall category value in this example can be considered the domain or topic of interest of the text. Continuing with the example, one ormore classifier 420 orpseudo classifiers 425 may extract value(s) belonging to the subcategory of the overall category, such as operations related to the overall topic of interest. Continuing with the same example, one ormore classifier 420 orpseudo classifiers 425 may extract parameter value(s). In successful extractions for certain preferred embodiments, the extracted parameter values are of the parameter types required by the extracted subcategory value. In this example, the subcategory values share the common property of being subcategory values of overall category values, whereas overall category values share the common property of having subcategory values. Also in this example, parameter values share the common property of having corresponding parameter types which can be accepted by subcategory values, whereas subcategory values share the common property of typically accepting parameter values of particular parameter types (It should be noted that in some cases no parameter types are defined for a particular subcategory value). - In other preferred embodiments, there may be fewer or more semantic categories in a hierarchical structure. For example, there may be an overall category, a pre-subcategory, a subcategory, and parameter values.
- In some preferred embodiments, the text may include more than one subtext (for example more than one request) with each subtext represented by a separate hierarchical structure that can be processed either in parallel or sequentially. For the sake of simplicity, it is assumed below that the text can be represented by a single hierarchical structure.
- It should be noted that the example given above of a possible hierarchical structure (i.e. overall category such as domain as the highest level, subcategory such as operation as the medium level, and parameter values as the lowest level) conforms with one conceptual view (as implemented typically by a system analyst-see below
FIG. 11 ). In other words, if the following words were presented: currency conversion, banking, depositing, checks, and dollars, a common conceptual view would classify banking as belonging to the highest level category (domain), currency conversion and depositing as belonging to the middle level category (operations), and checks and dollars as belonging to the lowest level category (parameter values). As another example, out of the following words: studying, teaching, school, books and tests, school would be considered by many to belong to the highest level category, studying and teaching to the middle level category and books and tests to the lowest level category. The hierarchical structure used in other preferred embodiments are by no means bound to this conceptual view and may embrace other conceptual views. -
FIG. 6 illustrates a preferred embodiment of a method for evaluating results of the extraction by classifier 420 (corresponding to step 525 ofFIG. 5 ). In theearlier step 520 ofFIG. 5 ,classifier 420 for a given semantic category searches aknowledge base 430 using the text prepared for extraction. (Details on howknowledge base 430 is developed will be explained below).Classifier 420 returns one or more possible matches (step 610) and a grade for each match (step 620). Grading is preferably performed byclassifier 420 based on the training undergone byclassifier 420 during the preparation of knowledge-base 430. The process iterates instep 630 until all matches and grades of the matches are output. In other preferred embodiments, not all matches are output but only matches that meet certain criteria, for example the highest graded matches, the most clustered matches, etc. For example, if clustered matches are to be output, matches whose grades are within a small range are identified and considered relevant. Continuing with the example, if there are ten results whose grades are {9, 8.3, 8.1, 7.9, 6.2, 6.1, 6, 4, and 1.2}, the two clustered groups of matches correspond to {8.3 to 7.8} and {6.2 to 6}. The outputted matches may in this example be those which correspond to the higher cluster {8.3 to 7.8}, perhaps together with other outputted matches, for example the match corresponding to the highest grade 9. - The matches are sorted by grade in
step 640. Generally, all matches are stored until no longer needed, i.e. until subsequent processing, for example further extractions, dialoging with submitters, etc. renders the match unlikely to be the correct match. In other preferred embodiments, the matches may be stored until the final results are output instep 530. - If the results are for the first semantic category extracted and more than one semantic category exists, then the results are considered insufficient to understand the text in step 650 (i.e. a ‘no’ answer to step 535). If there is only one existing semantic category (step 642) then the results are considered sufficient in step 670 (i.e. a ‘yes’ answer to step 535). If the results are for subsequent extracted semantic categories, but the results can not be evaluated in conjunction with results from previous semantic category extractions then the results are considered insufficient in step 650 (corresponding to a ‘no’ answer to step 535). An example of a situation where the results can not be evaluated in conjunction with previous results is if the current extracted semantic category value(s) is not directly related to any of the previously extracted category value(s). To illustrate the example, assume the overall category value extracted is “financial operation” and the only two possible subcategory values for this overall category value are “get a stock quote” and “buy stocks”. If neither of these subcategory values are extracted, but instead the subcategory value “get a horoscope forecast” is extracted, then the results are considered insufficient because the extracted values for the overall category and subcategory are not related to each other.
- If on the other hand, the results can be evaluated in conjunction with results from previous semantic category extractions, then the results are evaluated in conjunction with the results from the previous sets (step 658). For example, the evaluation can determine whether the results for the current semantic category correspond to the results from previous semantic category extractions (see below
FIG. 7 for more details on a possible evaluation process). If no weighted grade is to be calculated, then if the results are sufficient to understand the text, the method proceeds directly to step 670 (corresponding to a ‘yes’ answer to step 535). Otherwise, if no weighted grade is to be calculated and the results are insufficient, the method proceeds directly to step 650 (corresponding to a ‘no’ answer to step 535). The results may be considered sufficient to understand the text, for example, if all required values for each semantic category are known and the values for the different semantic categories correspond with one another. - In alternative preferred embodiments, in
step 645 if the results are not for the last extracted semantic category, then the results are considered insufficient instep 650. In these embodiments, only when the results are for the last extracted semantic category does the method continue withstep 655 where a decision is made on whether evaluation in conjunction with results from previous semantic category extractions is feasible. - In alternative preferred embodiments, the meaning of the text is guessed at prior to extracting all semantic categories and/or a final conclusion on a value of a semantic category is reached prior to completing an evaluation of results of that semantic category in conjunction with previously extracted semantic categories (In other words, in these embodiments the threshold of “sufficiency” is lower).
- Continuing with the illustrated preferred embodiment, once the results from all semantic categories are available, an optional weighted grade may be calculated in
step 660 as a final test of the combination of results from the different semantic categories. The weighted grade is derived using a formula which takes into account the grades achieved by the current results and the previous semantic category results. For example, the formula could be an average with either equal weights for each semantic category or different weights for each semantic category. Continuing with the example, in preferred embodiments which include an overall category and subcategory one possible formula might assign the overall category a weight of 2 and the subcategory a weight of 1. If the weighted grade is high then the results are considered sufficient in step 670 (corresponding to a ‘yes’ answer to step 535). If the weighted grade is not high enough then if further evaluation i.e. of other combinations of results from the different semantic categories is possible, more evaluations are performed instep 658. If no further evaluation of other combinations are possible than the results are considered insufficient in step 650 (corresponding to a ‘no’ answer to step 535). - In some preferred embodiments,
pseudo classifier 425 results are also graded. For example the time extractor may in some preferred embodiments return the results in a sequence of decreasing probability. In certain preferred embodiments with pseudo classifier graded results, the results may be evaluated in conjunction with other results as described in the method ofFIG. 6 . In other preferred embodiments, pseudo classifiers return either a Boolean true if a match was extracted, or a Boolean false if no match was extracted, without any grading. -
FIG. 7 illustrates in more detail a preferred embodiment of a possible sequence followed byaction resolver 410 in turning to different modules ofnatural language module 140 in order to implement the method ofFIG. 5 . In the preferred embodiment illustrated inFIG. 7 ,action resolver 410 is for example a state automaton. Assume for the sake of the example hierarchical semantic categories including overall category, subcategory and parameter values. The sequence shown inFIG. 7 , processes the semantic categories in an order which takes advantage of the hierarchy of the semantic categories, so that results of a semantic category of a certain hierarchical level helps in the processing of a semantic category of another hierarchical level. Specifically in this example there is assumed to be oneclassifier 420 for the overall category, aseparate subcategory classifier 420 associated with each overall category value, one parameter valuesclassifier 420, and one or more parameter valuespseudo classifiers 425. Continuing with the example, parameter values, belonging to the lowest level semantic category, are extracted first. At least some of the results of the parameter values extraction are used to embed tokens into the text for overall category extraction, i.e. the highest level semantic category, and for subcategory extraction, i.e. the medium level semantic category. The results of the overall category extraction is used to select a subcategory classifier, i.e. the medium level semantic category. The hierarchical structure of the semantic categories is therefore advantageous to the overall processing. - As the sequence (other than the dynamic features to be discussed below) is pre-programmed by the designer of natural
language understanding module 140, the sequence shown inFIG. 7 is one of many examples of possible sequences. - First,
text preprocessing module 435 is employed (step 702 corresponding to step 512). Next, parameter values classifier 420 is employed (step 705 corresponding to step 520).Real time database 445 is used to identify the one or more possible corresponding parameter type(s) for each extracted parameter value. Then, parameter values pseudo classifier(s) 425 is employed (step 710 corresponding to step 520) and corresponding parameter types are identified. Afterwards,text pre-processing module 435 is again employed in preparation for overall classifier 420 (step 712 corresponding to step 512). - Next,
overall category classifier 420 is employed (step 715 corresponding to step 520). Optionally (not shown), extracted overall category values can be evaluated in relation to extracted parameter values, for example by employing a first evaluationstatic component 425 to check if the parameter types of the extracted parameter values are in sync with subcategory values associated with the extracted overall category values. - If no overall category value is found,
dialog management module 440 is employed (step 720 corresponding to step 545). If the overall category value is ambiguous (i.e. more than one possibility), then in some preferred embodiments,dialog management module 440 is employed, but in other preferred embodiments, acorresponding subcategory classifier 420 is employed for each of the possible overall category values. Ifdialog management module 440 has been employed,overall category classifier 420 is again employed to check the response received from the submitter (step 721 corresponding to step 520). Alternatively, if the response is obvious and does not need to be understood (for example the response is a selection of a multiple choice option) thenoverall classifier 420 does not need to be employed to check the response and step 721 may be skipped. - Examples of situations when the method can proceed directly from
step 718 to step 722 (without dialoging) include inter-alia: when only one overall category value is extracted, when there is a default overall category value for the particular embodiment, when there is only one overall category value for a particular embodiment, when more than one overall category value was extracted but the parameter types of the extracted parameter values point to one of the extracted overall category values or to subcategory values associated with one of the extracted overall category values, when more than one overall category value was extracted but one overall category value can be selected based on the grades of the extracted overall category values, and when it is decided to process more than one extracted overall category value in parallel and employ a corresponding subcategory classifier for more than one overall category values. -
Corresponding subcategory classifier 420 is employed (step 722 corresponding to step 520). Evaluations are then performed (corresponding to step 658 ofFIG. 6 ) using evaluation staticspecific components 425. - First specific
static component 425 for evaluation is employed which evaluates all the results of the previous extractions byextractors 420 and pseudo extractors 425 (step 725). First specific evaluationstatic component 425 checks extracted parameter values against extracted subcategory value(s) to see if the parameter values (for example based on the identified parameter types) are suitable for the extracted subcategory value(s). For example for each extracted subcategory value,first evaluation component 425 may match the parameter type(s) identified for each extracted parameter value with the parameter types expected for the extracted subcategory value as predefined inreal time database 445. The matching in this example, may result in some expected parameter types (as predefined) not matched with any extracted parameter values, matched with exactly one parameter value, or matched with more than one extracted parameter value. - If based on this evaluation no suitable subcategory value is found
dialog management module 440 is employed (step 727 corresponding to step 545). Oncedialog management module 440 has been employed,sub-category classifier 420 is again employed to check the response received from the submitter (step 728 corresponding to step 520). Alternatively, if the response is obvious and does not need to be understood (for example the response is a selection of a multiple choice option) thensub-category classifier 420 does not need to be employed to check the response and step 728 may be skipped. First specific evaluationstatic component 425 is again employed instep 729. - Examples of situation when the method can continue directly from
step 726 to step 735 (without dialoging) include inter-alia: when only one subcategory value is extracted, when no subcategory value is extracted but there is a default subcategory value corresponding to the overall category value, when there is only one pre-defined subcategory value corresponding to the overall category value, when there is more than one extracted subcategory value but the parameter types of the extracted parameter values point to one of the extracted subcategory values, and when there is more than one extracted subcategory value but one subcategory value can be selected based on the grades of the extracted subcategory values. - In cases where more than one parameter value of the same parameter type are defined for the subcategory value and at least one parameter value of that same parameter type was extracted (step 735), a second specific static evaluation (relational)
component 425 is employed.Relationship evaluation component 425 evaluates the correspondence between the at least one extracted parameter value and the more than one parameter value defined for the subcategory value(step 740). For example if two names of cities were extracted for a ticket purchase, the secondstatic evaluation component 425 recognizes which is a destination and which is a source. Continuing with the example,relational component 425 may searchreal time database 445 for a predefined grammar line or utterance for example in the form String/s <ParameterType X> String/s→[Arg A]; String/s <ParameterType X> String/s→[Arg b]” which means that when a parameter value of type “ParameterType X” is extracted, the parameter value will be matched with the arguments A and B required by the subcategory value according to the String/s in its context. In this example: “From <ParameterType: City>→Arg: SourceCity; To <ParameterType: City>→Arg: DestCity”, allows the extracted city following the word “from” to be recognized as the source city and the extracted city following the word “to” to be recognized as the destination city. - A third specific
static evaluation component 425 is employed instep 742. Thisthird evaluation component 425 checks if parameter values corresponding to all parameter types defined for the subcategory value were extracted (step 745). For examplethird evaluation component 425 can use a check list against the parameter types predefined for the subcategory value inreal time database 445 Continuing with the example, if no parameter values were extracted for certain parameter types defined for the subcategory value,third component 425 can check if there are default parameter values which can be assigned or if the parameter types with missing parameter values are optional. Still continuing with the example, if no parameter value or if more than one parameter value was extracted for a mandatory parameter type (as predefined) which requires one parameter value and has no default value, then dialoging occurs. - If dialoging is required
dialog management module 440 is employed (step 750 corresponding to step 545). Parameter valuesclassifier 420 and/or one or more parameter values pseudo classifier(s) 425 is again employed to evaluate the response (step 752 corresponding to step 520) Alternatively, if the response is obvious and does not need to be understood (for example the response is a selection of a multiple choice option) then parameter values classifier 420 does not need to be employed to check the response and step 752 may be skipped. Thirdstatic evaluation component 425 is again employed instep 742 - Examples of when the dialoging of
step 750 need not occur include inter-alia, when the correct number of parameter values for the parameter types required by the subcategory value were extracted, when the subcategory values requires no parameter types, and when the parameter types required by the subcategory value have default values. - After
action resolver 410 finishes the sequence of employment of the various modules, for example as illustrated inFIG. 7 , there should be sufficient results to understand the text (corresponding to a ‘yes’ answer to step 535). Optionally, a weighted grade can first be evaluated (step 680) as a final test that the results are sufficient. - Note that the three specific evaluation
static components 425 mentioned here are separated in the description for ease of understanding and in alternative preferred embodiments may be combined into less or separated into more than three modules. - One of the distinct advantages of the preferred embodiments of the present invention are the one or more dynamic aspects of the natural language understanding. Later steps of the process are adaptable based on the results of earlier steps. The methods illustrated in
FIGS. 8, 9 , and 10 each include steps which are influenced by the results of the previous steps. Each of the dynamic aspects illustrated inFIGS. 8, 9 , and 10 can be separately implemented, and one or more of the dynamic aspects constitutes a separate preferred embodiment. InFIG. 8 , the results of a previous extraction may be used to develop tokens that may be embedded in the text used as an input for the next extraction(s). The tokens become part of the input for the next extraction(s) and are therefore termed syntactic tokens. InFIG. 9 , more than oneextractor 420 or more than onepseudo extractor 425 is available for the same semantic category and the selection ofextractor 420 orpseudo extractor 425 depends on the results of previous extractions. InFIG. 10 , the dialog with a submitter can vary based on the results (including unsuccessful or no results) of previous extractions. -
FIG. 8 illustrates a preferred embodiment of a method for preparing the natural language text for extraction (step 512). The first step is to determine if the results of one or more previous extractions (by eitherclassifiers 420 or pseudo classifiers 425) can be used to develop one or more syntactic tokens (step 810). This step is only performed during certain subsequent extractions and not for the first extraction. - If one or more syntactic tokens can be developed, the tokens are embedded into one or more restatements of the text (step 815), thereby allowing results of previous extractions to directly influence subsequent extractions. Depending on the embodiment, in the restatement of the text the embedded tokens can either replace the text parts which serve as sources for the tokens, or the tokens can supplement those text parts. In certain preferred embodiments,
action resolver 410 is responsible for embedding the tokens, but in other preferred embodiments, other modules such astext pre-processing module 435 embeds the tokens. - As an example of a token based on an extracted parameter value, assume the parameter value “crayon” was extracted. A syntactic token of the parameter types (for example “parameter type: writing utensil”) corresponding to the extracted parameter value “crayon” may be developed and embedded in the restatement of the text.
- N-grams are constructed for the text or the restatements of the text (including embedded tokens) in
step 820, if required. In certain preferred embodiments, n-grams are required for classifiers and some pseudo classifiers but not for all pseudo classifiers. In certain preferred embodiments,text pre-processing module 435 constructs the n-grams. - N-grams are well known in the art. A non-limiting definition for an n-gram based partially on “Text retrieval from Document Images based on N-gram Algorithm”, Chew Lim Tan, Sam Yuan Sung, Zhaohui Yu, and Yi Xu available at http:/citeseer.nj.nec.com/400555.html is: An N-Gram is a sequence of N consecutive items of a stream obtained by sliding an N-item wide window over the text one item forward at a time. Every possible N-Gram is given a number, so called the hash key. How the N-Grams are numbered is not important, as long as each instance of a certain N-Gram is always given the same number, and that distinct numbers are assigned to different N-Grams.
- Preferably the n-grams are sparse n-grams (i.e. also reflect the distance between words in the text). The use of sparse n-grams in some preferred embodiments of the present invention is advantageous. Sparse n-grams improve the probability of correct natural language understanding because sparse n-grams takes into account the specific order of words in a sentence. In preferred embodiments of the present invention, sparse n-grams are also trained upon (see below discussion with regard to
FIG. 12 ) - As an example, the sparse n-grams used are words, doublets, and triplets in the following form:
(wiwjwk,True) where i=j−1=k−2,i=1 . . . n−2
(wiwjwk,False) where k>j>i,i=1 . . . n−2
(wiwj,True) where i=j−1,i=1 . . . n−1
(wiwj, False) where j>i,i=1 . . . n−1
(wi,True) where i=1 . . . n - Note that in this example following the comma in each sparse n-gram there is a True or False indicator. The True/False indicator can be seen as an “Adjacency” indicator. If the words, doublets or triplets are composed of words which are adjacent in the text, a “True” indicator is indicated in the sparse n-gram. It should be noted that in this example in order to gain flexibility if the doublet or triplet is composed of words which are adjacent in the text, two sparse n-grams are created, one with a “true” indicator and one with a “false” indicator. The added flexibility enables a match between the text after n-grams are generated and the n-grams created for the trained sentence (see below
FIG. 12 ), both if in the trained sentence these words were adjacent or were not adjacent. In this example, if the doublet or triplet is composed of words which are not adjacent in the text, a sparse n-gram with a “false” indicator is created. - The constructed n-grams are used as the input to selected
classifier 420 or pseudo classifier 425(step 825) - In other embodiments of the invention, the construction of n-grams may be skipped (i.e. skip step 820) and the selected
classifier 420 and/orpseudo classifier 425 may extract based on other techniques for example by using word spotting. -
FIG. 9 illustrates a preferred embodiment of how the results of previous extractions can influence the selection of thenext classifier 420/pseudo classifier 425. The method ofFIG. 9 may be included instep 515. As mentioned above, in preferred embodiments of the present invention, the sequence of semantic category extractions is preprogrammed. However, in the cases where there is more than oneclassifier 420 or more than onepseudo classifier 425 for a semantic category, the method ofFIG. 9 allows the selection ofappropriate classifier 420 orpseudo classifier 425. Instep 905, a decision is made on whether more than oneclassifier 420 orpseudo classifier 425 is available for the semantic category to be extracted, for example by checking the structure ofreal time database 435 If no, the one available is employed (i.e. proceed directly to step 520). If yes, the method continues withstep 910 where a further decision is made whether a selection of less than allavailable classifier 420 or less than all availablepseudo classifier 425 for the semantic category can be made based on previous extractions. If a selection is possible, the selection is made instep 915. Otherwise, results of the previous extractions are clarified instep 920, for example by dialoging with the submitter. In other cases, instep 920, all available classifiers or all available pseudo classifiers for the semantic category are used, for example looking back atFIG. 7 instep 710 all available parameter pseudo classifiers are initially employed. - As an example, assume that the semantic categories include an overall category and a subcategory and also assume that there is more than one
subcategory classifier 420, with adifferent subcategory classifier 420 used depending on the overall category value. If the results of the overall category extraction are clear, then instep 915 thesubcategory classifier 420 corresponding to the extracted overall category value is employed. If the results of the overall category extraction are unclear, then clarification is received instep 920. - As another example, assume that there is an additional hierarchical level so that the (four) semantic categories include an overall category, a pre-subcategory, a subcategory and parameter values. In this example a different
pre-subcategory classifier 420 is selected depending on the overall category value and adifferent subcategory classifier 420 is selected depending on the pre-subcategory value. Further added hierarchical levels can be processed in a complementary manner. - As yet another example, assume that there are a plurality of parameter value
pseudo classifiers 425. Assume also that in step 745 (FIG. 7 ) it is found that a certain parameter type, for example a money expression, required by the extracted subcategory value is missing. In this case, a question is asked and once the answer from the submitter is received, only the currency pseudo classifier (corresponding to the missing parameter type) out of allpseudo classifiers 425 would be selected and employed on the answer. As another example, assuming more than two parameter valuepseudo classifiers 425, if it is found instep 745 that a money expression and a time expression are missing, then currencypseudo classifier 425 and timepseudo classifier 425 out of all the availablepseudo classifiers 425 would be employed on the answer. -
FIG. 10 illustrates a preferred embodiment of a method for dialoguing with the submitter of the natural language (step 545). Additional user information may be required for example, to resolve an ambiguity, provide a missing piece of information, or restate the submission. The answers received from the submitter via the dialog augments previous extraction results so as to aid in understanding the natural language text. The term augments is used to include one or more of the following: clarifies, supplements, pinpoints, expands, narrows, etc., i.e. the answers from the dialog allows the text to be better understood than had the dialog not taken place and only the previous extraction results were available. - If further processing can not be performed (
step 1020 corresponding to step 540 ofFIG. 5 ,step 718, step 726 or 745 ofFIG. 7 , or step 920 ofFIG. 9 ) adialog management module 440 is called by action resolver 410 (step 1030). Ambiguity may result from more than one correct interpretation of a submission (for example, a request for the flight schedules leaving New York City can be interpreted as leaving Kennedy or La Guardia airport and more information from the submitter would be required to resolve the ambiguity). Further processing may also not be possible if information is missing (for example parameter values of all parameter types required by the subcategory value were not extracted). In some preferred embodimentsdialog management module 440 searches realtime data base 445 for pre-defined questions associated with one or more categories. In other preferred embodiments,dialog management module 440 does not pose predefined questions, but instead formulates questions. In either case open questions and multiple choice questions may be used according to the type of missing information In one preferred embodiment, ambiguity problems result in a close-ended question, whereas missing information (for example values of missing categories) result in open-ended questions. Instep 1050, the question is posed to the submitter. - In preferred embodiments of the present invention, there is no need to design in advance a dialog tree which covers all possible questions for all possible missing information/ambiguities. Instead dialog is created on the fly or predefined question strings are retrieved from
real time database 445 based on system logic previously inserted in realtime data base 445. The decision as to what and when to ask is taken bydialog management module 440 based on this system logic and the current step in the understanding process ofFIG. 5 . The system logic should be understood to mean logic inputted intoreal time database 445 to aid in the natural language understanding and which as a bonus also aids in question formulation/question string retrieval. For example, the logic for a subcategory value can include the parameter types related to that subcategory value, characteristics of these parameter types independently and in relation to one another (such as when the types are mandatory), relative importance of each of these parameter types, etc. - In preferred embodiments of the present invention, the question posed to the submitter is varied based on previous extraction results (where results in this context can also include non-results i.e. unsuccessful extraction). For example when formulating a question, the previous extraction results can be compared to the logic in order to formulate an appropriate question. Continuing with the example of the previous paragraph if parameter values for two parameter types related to the subcategory value are missing but the logic dictates that one type is more important, then a first question formulated and posed to the submitter may relate only to the more important type. As another example, a pre-defined question may include all required parameter types for a given subcategory according to the logic. However the question strings retrieved from
database 445 and used in the question posed to the submitter will relate only to those parameter types with no previously extracted parameter values. - As another example, a pre-defined question may include all possible airports in New York State, but if the previous extractions extracted New York City, the multiple-choice question posed to the submitter will be modified so as to offer as possible responses only airports in New York City.
-
Real time database 445 includes entries for the semantic categories. As mentioned above,real time database 445 in some preferred embodiments also includes dialog questions and/or suggested answers. - There will now be explained a structure for
real time database 445, according to a preferred embodiment of the present invention, which assumes a hierarchical structure to semantic categories. Refer toFIG. 11 showing an entity-relationship (ER) diagram ofdatabase 445. The entries indatabase 445 are divided into four types. The first type includes entries related to the semantic category “overall category” 1105. The second type includes entries related to the semantic category “subcategory” 1115. Eachoverall category entry 1105 has a number ofsubcategory entries 1115. Eachsubcategory entry 1115 requires or accepts certainparameter type entries 1125. Eachparameter type entry 1125 is able to take on one or moreparameter value entries 1135. Theparameter value entries 1135 are related to the semantic category “parameter values”. The invention is not bound by the illustrated ER structure or contents. - It should be noted that although
parameter value entries 1135 are not stored undersubcategory entries 1115 in the illustrated example ofdatabase 445, parameter values under a common conceptual view would be considered to be of a lower hierarchical level as explained above. For convenience, in the illustrated example ofdatabase 445,parameter value entries 1135 indatabase 445 are not stored undersubcategory entries 1115 so that the sameparameter value entry 1135 can relate to more than onesubcategory entry 1115. It should also be noted that in many preferred embodiments, not all parameter values related toparameter type entries 1125 are stored asparameter value entries 1135 indatabase 445. For example aparticular subcategory entry 1115 may be associated with a parameter type entry 1125 (for example date) whose parameter values are extracted bypseudo classifier 420 and therefore are not stored as entries indatabase 445. - The definition of the hierarchical structure of
database 445 and the characterization of the entries into the different hierarchical levels is typically performed by a system analyst with knowledge of the requirements of a particular implementation and is therefore beyond the scope of this invention. - In some preferred embodiments, entries for one or more categories are manually entered in
database 445. In other preferred embodiments, entries for one or more categories can be at least partially automatically gathered from the Internet, preferably using active browsingstudio work tool 170. In preferred embodiments including request implementation through control of internet sites, this approach implies that at least part of the information used in buildingdata base 445 originates from the medium where request implementation takes place. -
FIG. 12 shows a method for trainingreal time database 445 so as to generate knowledge base 430 (used by classifiers 420), according to a preferred embodiment, which assumes the same semantic categories as inFIG. 11 . Thefirst step 1205 is the defining of natural language examples. Thesecond step 1210, if required for some examples, is the embedding of syntactic tokens based on entries inreal time database 445 within the natural language examples. Tokens for example can includeoverall category entries 1105,subcategory entries 1115,parameter type entries 1125 and/orparameter values entries 1135.Train database 450 preferably includes some examples with embedded tokens and some examples without embedded tokens (step 1212), so thatclassifiers 420 are trained to understand both text which includes proper nouns (for example, Intel) and/or common nouns (for example, stocks). Thenext step 1215 is the transformation of the examples into n-grams, preferably sparse, if required (for example if it is expected that in operation the extraction will be performed on n-grams). In some preferred embodiments, the transformation is performed bypre-processing module 435. The examples, in the form of n-grams if required, are input into classifiers in 420step 1220. In one preferred embodiment examples with embedded tokens corresponding toparameter type entries 1125 are used fortraining classifiers 420 for the overall category and subcategory. In one embodiment, examples with embedded tokens corresponding toparameter value entries 1135 are used fortraining classifiers 420 for parameter values. The classifiers are trained instep 1225. It should be evident that the same algorithm referenced above with respect toclassifiers 420 can be used intraining step 1225. Knowledge-base 430 is obtained from the training in thefinal step 1230. In one preferred embodiment, knowledge-base 430 is a data structure that is saved to a disk soknowledgebase 430 can be used later. - In certain preferred embodiments,
knowledge work tool 150 assists in the training process. For example, once an operator of the training has chosen an entry fromdatabase 445 and the placement of a token in an example,tool 150 can develop a token from the entry and embed the token in the correct place. As an additional example,work tool 150 can employpre-processing module 435 for developing n-grams. In certain preferred embodiments,work tool 150 also operates the training ofclassifiers 420 and saves the results inknowledgebase 430. - In some preferred embodiments, the training phase of
knowledgebase 430 is completely separated from the usage phase—i.e. all training is performed prior to use ofknowledgebase 430. In other preferred embodiments,knowledgebase 430 continues to be expanded during the usage phase, by learning from the texts received from submitters. - An example will now be given to further illustrate certain aspects of the overall process of
FIG. 5 and the sub-processes of FIGS. 6 to 10. Assume a user submission of “price for Columbia yesterday” received instep 510. The submission in the example includes a request. The text is prepared for extraction instep 512 by constructing the following n grams (step 815). In the example, it is assumed that the n grams are sparse and comply with the exemplary form defined above with reference toFIG. 8 : - (price for Columbia, True)
- (for Columbia yesterday, True)
- (price for Columbia, False)
- (for Columbia yesterday, False)
- (price for yesterday, False)
- (price Columbia yesterday, False)
- (price for, True)
- (for Columbia, True)
- (Columbia yesterday, True)
- (price for, False)
- (for Columbia, False)
- (Columbia yesterday, False)
- (price Columbia, False)
- (price yesterday, False)
- (for yesterday, False)
- (price, True)
- (for, True)
- (Columbia, True)
- (yesterday, True)
-
Action resolver 410 selects oneclassifier 420 to employ instep 515. In this case it is assumed that there are three types of classifiers 420: one overall category classifier (the domain of interest); a subcategory classifier corresponding to each overall category value (requested operation for the domain of interest); and one parameter value classifier (items required by operation). Parameter valuesclassifier 420 and all available parameter value pseudo classifier(s) 425 are employed instep 520. The parameter values pseudo-classifier 425 which is a time phrase extractor extracts one item: Feb. 6, 2001 i.e. the date yesterday. The parameter type of Feb. 6, 2001 is identified as date. The parameter valuesclassifier 420 extracts items from the word “Columbia” along with the grades of the items (steps 610 and 620).Real time database 445 is used to identify the parameter-types i.e. item-types of the different items extracted from the word Columbia. Assume that five items are extracted corresponding to Columbia as a country, a university, a hospital, and twice as a stock. Due to the ambiguity, more than one item and the corresponding item type are saved. - The text is prepared for the next extraction in
step 512, using all possible item-type matches. The syntactic tokens (in this example, the item types) are embedded into the text instep 815. In this example, it is assumed that an identical item-type is embedded in only one restatement of the text, even if more than one item was found of the same item-type (in this example, the tokens supplement “Columbia”). - price for Columbia <parameter type:hospital> yesterday <date>
- price for Columbia <parameter type:country> yesterday <date>
- price for Columbia <parameter type:university> yesterday <date>
- price for Columbia <parameter type:stock> yesterday <date>
- Sparse n-grams are constructed in
step 820 for each of the four item-type matches (which now include the embedded tokens). It should be noted that when constructing the sparse n-grams the embedded tokens are treated as if the tokens are words and an integral part of the text.Action resolver 410 employsdomain extractor 420 on the new sparse n-grams instep 825. Results are evaluated in step 525 (see method ofFIG. 6 ). It is assumed that two possible domains are outputted instep 610. The first domain is hospital policies (i.e. prices for a stay at Columbia Hospital) and the second domain is Nasdaq (the market where the stocks for Columbia Records and Columbia Hospital are listed). It is assumed that the second domain is outputted with a higher grade instep 620. However it is also assumed that the grades are close enough thataction resolver 410 decides to approach the submitter (step 545).Dialog management module 440 is called (step 1030) which in this example searchesreal time database 445 for a question (step 1040). In this example, the question and answers (adapted to the two possible extracted domains) are “Please clarify the topic of interest a) the hospital policies of Columbia Hospital, b) The stock results of Columbia Hospital or, c) The stock results of Columbia records. The question is posed to the submitter instep 1050. The user response is assumed to be “hospital stock” implying the stock results of Columbia Hospital (step 550), which is inputted into domain classifier 420 (step 520) to extract Nasdaq - Therefore the
operation extractor 420 related to Nasdaq is selected instep 915 and employed instep 520. The sparse n-grams earlier derived from: - price for Columbia <parameter type:stock> yesterday <date>are inputted into the
operation extractor 420 related to Nasdaq. The results of the operation are “get stock price”. The operation is evaluated in conjunction with previous results instep 658. The operation “get stock price” requires parameters of type stock and date. Both of these types have been extracted. As a final test a weighted grade is calculated which is assumed to be sufficiently high (steps 660 and 665). The request is therefore assumed to have been correctly understood (i.e. sufficient results-step 670) and the results are output instep 530. - In order for the request to have been correctly understood, it is assumed that
classifiers 420 had been previously trained. As an example, assume that the following examples were defined in step 1205: - “I want a price for <parameter type: stock>at <date:exact date>” This example in which parameter-type tokens were embedded in
step 1210 may have been transformed into n-grams instep 1215 and used to traindomain classifier 420 for the Nasdaq domain and/oroperation classifier 420 for the operation “stock quote” instep 1225. - “I want to trade with stocks” This example with no embedded tokens may have been transformed into n-grams in
step 1215 and used to traindomain classifier 420 for the Nasdaq domain (and possibly other domain stock markets). - “University Columbia of New York” may have been used to train parameter values classifier 420 for the item Columbia University of New York.
- “Columbia Medical” may have been used to train parameter values classifier 420 for Columbia Hospital corresponding to both hospital parameter type and stock parameter type.
- To further illustrate the flowcharts of
FIGS. 6, 7 , 8, 9, 10, and 12 another comprehensive example is presented. In the example, there is assumed to be two possible overall category values (here domains), “financial information” and “car rentals”. Subcategory values (here operations) for “financial information” are “get stock quote”, “get stock rate of change”, “get stock high value”, and “get stock low value”, each of which is associated with a parameter value of parameter type “stock”. There is assumed to be two subcategory values for “car rentals”, namely “get address of dealership” which is associated with a parameter value of parameter type location and “make a car rental reservation” which is associated with parameter values of parameter types “location”, “time”, and “car group, where two locations are required: pickup and return and two times are required: pickup time and return time. It is also assumed that parameter values of parameter type “stock” include Intel, Yahoo, Microsoft, AT&T, etc. Parameter values of parameter type “location” as in Avis dealership location include Los Angeles airport, Los Angeles downtown, San Francisco, Sacramento, etc. No specific time parameter values are specified for the “time” parameter type. Parameter values of parameter type “car group” as in rental car group include compact, sub compact, sports, 2-door, etc. - The table below summarizes the scope of the example:
Overall Category Parameter (domain) Subcategory (operation) Types (arguments) Financial Get Stock Quote Stock Information Get Stock Rate Of Change Stock Get Stock High Value Stock Gel Stock Low Value Stock Car Rentals Get address of dealership Location Mate Car Rental Reservation location (pickup) location (return) Time (pickup) Time (return) Car Group -
ParameterType ParameterValues Stock Intel Yahoo, Microsoft, AT&T, . . . Location LA Airport, LA Downtown, (Avis Dealership) San Francisco, Sacramento, . . . Time No specific Items Car Group (rental) Compact, Sub Compact, Spats, 2-Door, . . . - Referring to
FIG. 6 . assume that the text in this example is the request “get a quote for Intel”. Insteps 610 to 640, the parameter values extracted byparameter value classifier 420 and/or parameter valuespseudo classifier 425 are output. In this example, only one parameter value “Intel” is extracted. Instep 645, as this is the first semantic category extracted, the results are insufficient. - Assume that the
overall category classifier 420 is then called and applied to n-grams created from a restatement of the original text which includes a token based on the result of the parameter value extraction, i.e. “Get a quote for <ParameterType: Stock>” (in this example the token replaces “Intel”). Insteps 610 to 640, the outputted results ofoverall category classifier 420 are the two possible domains, with financial Information receiving a high grade and car rentals a low grade. The results are sorted by grade instep 640 and instep 655, the results are evaluated in conjunction with the parameter value results. As the subcategory value is still unknown, the results are considered insufficient. - Assume then that the
subcategory classifier 420 corresponding to overall category value “financial information” is called insteps 610 to 658. The results include the operation with the highest grade, assumed to be “Get Stock Quote”. The results are checked for compliance with previous results. The evaluation shows that the highest graded operation is a member of the found domain and that the found parameter value is of a type accepted by the found operation as an argument. In step 660 a weighted grade corresponding to the highest graded operation is calculated by a simple formula giving equal weights to each semantic category and the weighted grade is checked to see whether the weighted grade is above a given threshold. If the weighted grade is below the threshold, instep 675 evaluation can be attempted for other sets of results with lower grades (for example including a lower graded operation), and it can be checked whether the resulting weighted grade is higher than the given threshold. - Referring to
FIG. 7 , assume that the text in the example is instead the request “rent a car tomorrow morning in LA airport until March 13th at noon, return to Sacramento” - In
step 702 the text is preprocessed into n-grams because in this example it is assumed that n-grams are inputted toclassifiers 420 and/orpseudo classifiers 425. The n-grams are of the sparse form described above with reference toFIG. 8 . - In
steps 705 to 715parameter values classifier 420 andpseudo classifiers 425 are initially called. The extracted parameter values include several values: LA Airport, Feb. 8, 2001 08:00 (Tomorrow's date), Mar. 13, 2001 12:00, and Sacramento. The text is restated so as to include tokens based on the found parameter values, namely: “rent a car <ParameterType: Time> in <ParameterType AvisDealershipLocation> until <ParameterType: Time> return to <ParameterType AvisDealershipLocation>. New n-Grams are created from the restated text again using the sparse n-gram form described above with the embedded tokens treated as words.Overall category classifier 420 is called and extracts the car rentals domain. - In
step 718 because the overall category was unambiguously found the method proceeds withstep 722. (If there had been ambiguity with regard to the domain, dialoging with the user could take the form of posing a closed multiple choice question to the submitter which includes the two possible domains as choices.) - In
steps 722 to 725,subcategory classifier 420 is called. First evaluationstatic component 425 is then called in order to try to find a match between the parameter types of the found parameter values and the expected arguments of the highest graded extracted operation. In this example, because the request text is clear regarding the desired operation,subcategory classifier 420 returns only one operation.Static evaluation component 425 matches the parameter types “Time” and “AvisDealershipLocation” corresponding to the extracted parameter values with the corresponding arguments of the “Make Car Rental Reservation” operation - Because the subcategory value was unambiguously found, no dialoging is required and the method proceeds with step 735 (If there had been ambiguity, a typical multiple choice question could display as choices all available operations for the found domain or all operations for the found domain which received a high grade from subcategory classifier 420)
- In
steps 735 to 740 because there are parameter types which are acceptable for more than one argument of the found operation, there is a need to call second static evaluation component (Relational Static Component) 425. In this example both Time and AvisDealershipLocation are twice accepted as arguments by the operation “make car rental reservation”. Relationalstatic component 425 identifies which values belong to which arguments by checking the context of the values. The time value Mar. 13, 2001 12:00 is recognized as the return time due to the preceding word “until”, and the value Sacramento is recognized as the return location by the preceding words “return to”. Once these values are assigned correctly to the arguments of the operation the other time and AvisDealershiplocation values follow naturally. - In
steps 742 to 752 thirdstatic component 425 is called to check if all required arguments have been assigned suitable values. In this example thirdstatic component 425 finds that four out of the five arguments have values assigned. The car group argument is as yet unassigned. Therefore instep 750 in a dialog with the submitter either an open question is posed to prompt the submitter to enter the car group or a closed question is posed including as choice all possible car groups (as predefined). Once the answer is received, the last required parameter is known and results can be output. - Referring to
FIG. 8 it is assumed that the text is the same request as inFIG. 7 , namely “rent a car tomorrow morning in LA airport until March 13th at noon, return to Sacramento”. Step 810 checks if there are any previous results that can be developed into tokens. In this example tokens for parameter types Time and AvisDealershipLocation which correspond to the extracted parameter values can be developed. A token is developed for each text part that had been used as a source for extraction of a parameter value. - In
step 815 the developed tokens are embedded in the text in place of the source texts that were used to extract the parameter values. In this example the restatement of the original request “rent a car tomorrow morning in LA airport until March 13th at noon, return to Sacramento” is restated as “rent a car <ParameterType: Time> in <ParameterType: AvisDealershipLocation> until <ParameterType: Time>return to <ParameterType: AvisDealershipLocation>”. - In
steps 820 to 825 from the restatement, new n-grams are constructed in the sparse n gram form described above with reference toFIG. 8 . Some of the n-Grams include tokens, which are dealt with as regular words. - Refer now to
FIG. 9 . Again assume the text is the request “rent a car tomorrow morning in LA airport until March 13th at noon, return to Sacramento” Instep 905 the answer to the question is yes when deciding whether there is more than onepossible subcategory classifier 420 which can be called. In this example, a selection needs to be made from among the twopossible subcategory classifiers 420, one that classifies operations for the Financial Information domain and one that classifies operations for the Car Rentals domain. - In
steps 910 to 920 because the domain “car rentals” is assumed to have already been found, the car rentals subcategoryclassifier 420 is used. (If after using the overall category classifier there is still ambiguity with regard to the correct domain, dialoging instep 920 would be attempted to clarify the correct domain) - Refer to
FIG. 10 . Assume now that there are two texts received from the submitter the first text being “LA Airport to Sacramento, tomorrow morning until Mar. 13, 2001 at noon” and the second text being “Intel”. - In
step 1020 redundant interactions with the submitter are avoided by performing additional automatic processing to try to solve any problems without the help of the submitter. Assume that the parameter values (LA airport, Sacramento, Feb. 8, 2001 08.00, and March 13, 12.00) and domain (car rentals) have been extracted from the first text. Although the operation is not given in the first text, further processing can be performed using the firststatic evaluation component 420 in order to determine the desired operation by looking at the parameter types of the extracted parameter values and comparing these parameter types with the possible accepted arguments of the available operations, thereby avoiding dialoging. However, after calling second and thirdstatic components 420 the car group value is still missing and so dialoging with the submitter is required to obtain the car group value. Referring now at the second text, both the overall category and the subcategory can not be extracted based on the text alone. However, the overall category can be extracted from a restatement which includes a token based on an extracted parameter value, i.e. (<parameterType:Stock>). This restatement implies that the desired operation to be found accepts the stock parameter type as an argument. In this example, only operations in the financial information domain (and not in the car rental domain) receive such values. Therefore the domain can be determined without dialoging. However, after calling the subcategory classifier, the operation is still ambiguous because all four operations in this domain accept stock as an argument. Therefore dialoging with the submitter is required to allow the submitter to select the correct operation. - In
steps 1030 to 1050dialog management module 440 is called if no further processing is possible.Dialog management module 440 generates the correct interaction based on the current status of the handling of the request. Ifdialog management 440 is called while processing the first text to determine the car group value,dialog module 440 needs to create an interaction for determining the car group parameter value. Thereforedialog module 440 goes toreal time database 445 and finds the string that was prepared as a question for this case specifically, i.e.—a question regarding the lack of value for this specific argument. Ifdialog module 440 is called for the second text in order to determine the operation,dialog module 440 needs to create an interaction that clarifies an ambiguity in the operation and presents the submitter with all possible options. Therefore,dialog module 440 goes toreal time database 445 and finds the String that was prepared for this specific case, i.e.—operation ambiguity interaction. Once the question is formatted, the question is transferred to the submitter and the reply of the submitter is analyzed. - Refer to
FIG. 12 . In this example, the initial creation ofknowledge base 430 includes the following steps. Instep 1205, natural language examples are defined for the supported domains, operations and parameter values. For example, the following examples may be used, inter-alia for training: - “I want to receive financial information”→Domain: Financials
- “I want to get a stock quote”→Operation: Get Stock Quote
- “I would like to rent a car”→Domain: Car Rentals
- “I would like to rent a car”→Operation: Make Car Rental Reservation
- “Intel”→Parameter value: Intel
- “Los Angeles Airport”→Parameter value: LA Airport
- In
step 1210 tokens are embedded in some of the above examples. For example: - “I want to get a stock quote for <ParameterType: Stock>”→Operation: Get Stock Quote
- “I would like to rent a car <ParameterType: Time> in <ParameterType AvisDealershipLocation>”→Domain: Car Rentals
- In
steps 1212 to 1230 the training examples are turned into n-Grams, and the classifiers are trained on the n-grams, with the results serialized intoKnowledgebase 430. Typically, the training process is classifier-specific allowing the examples in their n-gram representation to be associated with the categories and values which were trained on those n-grams. - It will also be understood that the system according to the invention may be a suitably programmed computer. Likewise, the invention contemplates a computer program being readable by a computer for executing the method of the invention. The invention further contemplates a machine-readable memory tangibly embodying a program of instructions executable by the machine for executing the method of the invention.
- While the invention has been described with respect to a limited number of embodiments, it will be appreciated that many variations, modifications and other applications of the invention may be made.
Claims (13)
1. A method for use in a method for understanding a natural language text, comprising performing the following selectively in a statistical manner:
attempting to extract at least one value belonging to a semantic category from a natural language text or a form thereof; and
if a result of said attempting complies with a predetermined criterion, attempting to extract, based on said result, at least one value belonging to another semantic category of a different hierarchical level than said semantic category, else performing at least one action from a group of actions including: asking a submitter of said text a question whose content depends on said result and giving up on understanding said natural language text.
2. The method of claim 1 , wherein said attempting to extract at least one value belonging to said another semantic category includes:
selecting at least one classifier for said another semantic category from among more than said at least one classifier for said another semantic category, wherein said selecting is based on at least one extracted value belonging to said semantic category; and
employing said at least one classifier in an attempt to extract at least one value belonging to said another semantic category.
3. The method of claim 2 , wherein said another semantic category is a hierarchically lower level semantic category than said semantic category.
4. The method of claim 3 , wherein said at least one value belonging to said another semantic category is at least one operation and said at least one value belonging to said semantic category is at least one domain.
5. The method of claim 1 , wherein said form are n grams constructed from said text or from a restatement of said text which includes at least one embedded token.
6. The method of claim 5 , wherein said n-grams are sparse n-grams.
7. The method of claim 1 , wherein said asking a question includes: formulating said content of said question on the fly based on said result.
8. The method of claim 1 , wherein said asking a question includes: modifying a predefined question based on said result.
9. The method of claim 1 , further comprising: if said question is asked, attempting to extract at least one value belonging to a previously extracted semantic category from said answer.
10. The method of claim 9 , wherein said previously extracted semantic category is said semantic category.
11. The method of claim 1 , wherein said at least one value belonging to a semantic category and said at least one value belonging to another semantic category are at least one from a group including: at least one domain, at least one operation, and at least one parameter value.
12. A program storage device readable by machine, tangibly embodying a program of instructions executable by the machine to perform method steps for use in a method for understanding a natural language text, comprising performing the following selectively in a statistical manner:
attempting to extract at least one value belonging to a semantic category from a natural language text or a form thereof; and
if a result of said attempting complies with a predetermined criterion, attempting to extract, based on said result, at least one value belonging to another semantic category of a different hierarchical level than said semantic category, else performing at least one action from a group of actions including: asking a submitter of said text a question whose content depends on said result and giving up on understanding said natural language text.
13. A computer program product comprising a computer useable medium having computer readable program code embodied therein for use in a computer program product comprising:
computer readable program code for causing the computer to perform the following selectively in a statistical manner:
computer readable program code for causing the computer to attempt to extract at least one value belonging to a semantic category from a natural language text or a form thereof; and
computer readable program code for causing the computer, if a result of said attempting complies with a predetermined criterion, to attempt to extract, based on said result, at least one value belonging to another semantic category of a different hierarchical level than said semantic category, else to perform at least one action from a group of actions including: asking a submitter of said text a question whose content depends on said result and giving up on understanding said natural language text.
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US11/562,132 US20070112555A1 (en) | 2001-03-13 | 2006-11-21 | Dynamic Natural Language Understanding |
US12/047,003 US20080154581A1 (en) | 2001-03-13 | 2008-03-12 | Dynamic natural language understanding |
Applications Claiming Priority (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US27559801P | 2001-03-13 | 2001-03-13 | |
US10/097,537 US7216073B2 (en) | 2001-03-13 | 2002-03-13 | Dynamic natural language understanding |
US11/562,132 US20070112555A1 (en) | 2001-03-13 | 2006-11-21 | Dynamic Natural Language Understanding |
Related Parent Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US10/097,537 Continuation US7216073B2 (en) | 2001-03-13 | 2002-03-13 | Dynamic natural language understanding |
Related Child Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US12/047,003 Continuation US20080154581A1 (en) | 2001-03-13 | 2008-03-12 | Dynamic natural language understanding |
Publications (1)
Publication Number | Publication Date |
---|---|
US20070112555A1 true US20070112555A1 (en) | 2007-05-17 |
Family
ID=23053035
Family Applications (4)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US10/097,537 Expired - Fee Related US7216073B2 (en) | 2001-03-13 | 2002-03-13 | Dynamic natural language understanding |
US11/562,132 Abandoned US20070112555A1 (en) | 2001-03-13 | 2006-11-21 | Dynamic Natural Language Understanding |
US11/562,142 Expired - Fee Related US7840400B2 (en) | 2001-03-13 | 2006-11-21 | Dynamic natural language understanding |
US12/047,003 Abandoned US20080154581A1 (en) | 2001-03-13 | 2008-03-12 | Dynamic natural language understanding |
Family Applications Before (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US10/097,537 Expired - Fee Related US7216073B2 (en) | 2001-03-13 | 2002-03-13 | Dynamic natural language understanding |
Family Applications After (2)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US11/562,142 Expired - Fee Related US7840400B2 (en) | 2001-03-13 | 2006-11-21 | Dynamic natural language understanding |
US12/047,003 Abandoned US20080154581A1 (en) | 2001-03-13 | 2008-03-12 | Dynamic natural language understanding |
Country Status (4)
Country | Link |
---|---|
US (4) | US7216073B2 (en) |
EP (1) | EP1490790A2 (en) |
AU (1) | AU2002237495A1 (en) |
WO (1) | WO2002073451A2 (en) |
Cited By (42)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20030023423A1 (en) * | 2001-07-03 | 2003-01-30 | Kenji Yamada | Syntax-based statistical translation model |
US20050228643A1 (en) * | 2004-03-23 | 2005-10-13 | Munteanu Dragos S | Discovery of parallel text portions in comparable collections of corpora and training using comparable texts |
US20060142995A1 (en) * | 2004-10-12 | 2006-06-29 | Kevin Knight | Training for a text-to-text application which uses string to tree conversion for training and decoding |
US20080109209A1 (en) * | 2006-11-02 | 2008-05-08 | University Of Southern California | Semi-supervised training for statistical word alignment |
US20090326927A1 (en) * | 2008-06-27 | 2009-12-31 | Microsoft Corporation | Adaptive generation of out-of-dictionary personalized long words |
US20100017293A1 (en) * | 2008-07-17 | 2010-01-21 | Language Weaver, Inc. | System, method, and computer program for providing multilingual text advertisments |
US8234106B2 (en) | 2002-03-26 | 2012-07-31 | University Of Southern California | Building a translation lexicon from comparable, non-parallel corpora |
US8380486B2 (en) | 2009-10-01 | 2013-02-19 | Language Weaver, Inc. | Providing machine-generated translations and corresponding trust levels |
US20130054226A1 (en) * | 2011-08-31 | 2013-02-28 | International Business Machines Corporation | Recognizing chemical names in a chinese document |
US8468149B1 (en) | 2007-01-26 | 2013-06-18 | Language Weaver, Inc. | Multi-lingual online community |
US8548794B2 (en) | 2003-07-02 | 2013-10-01 | University Of Southern California | Statistical noun phrase translation |
US8615389B1 (en) | 2007-03-16 | 2013-12-24 | Language Weaver, Inc. | Generation and exploitation of an approximate language model |
US8676563B2 (en) | 2009-10-01 | 2014-03-18 | Language Weaver, Inc. | Providing human-generated and machine-generated trusted translations |
US8694303B2 (en) | 2011-06-15 | 2014-04-08 | Language Weaver, Inc. | Systems and methods for tuning parameters in statistical machine translation |
US8825466B1 (en) | 2007-06-08 | 2014-09-02 | Language Weaver, Inc. | Modification of annotated bilingual segment pairs in syntax-based machine translation |
US8831928B2 (en) | 2007-04-04 | 2014-09-09 | Language Weaver, Inc. | Customizable machine translation service |
US8886517B2 (en) | 2005-06-17 | 2014-11-11 | Language Weaver, Inc. | Trust scoring for language translation systems |
US8886515B2 (en) | 2011-10-19 | 2014-11-11 | Language Weaver, Inc. | Systems and methods for enhancing machine translation post edit review processes |
US8886518B1 (en) | 2006-08-07 | 2014-11-11 | Language Weaver, Inc. | System and method for capitalizing machine translated text |
US8943080B2 (en) | 2006-04-07 | 2015-01-27 | University Of Southern California | Systems and methods for identifying parallel documents and sentence fragments in multilingual document collections |
US8942973B2 (en) | 2012-03-09 | 2015-01-27 | Language Weaver, Inc. | Content page URL translation |
US8977536B2 (en) | 2004-04-16 | 2015-03-10 | University Of Southern California | Method and system for translating information with a higher probability of a correct translation |
US8990064B2 (en) | 2009-07-28 | 2015-03-24 | Language Weaver, Inc. | Translating documents based on content |
US9122674B1 (en) | 2006-12-15 | 2015-09-01 | Language Weaver, Inc. | Use of annotations in statistical machine translation |
US9152622B2 (en) | 2012-11-26 | 2015-10-06 | Language Weaver, Inc. | Personalized machine translation via online adaptation |
US9213694B2 (en) | 2013-10-10 | 2015-12-15 | Language Weaver, Inc. | Efficient online domain adaptation |
WO2016044321A1 (en) * | 2014-09-16 | 2016-03-24 | Min Tang | Integration of domain information into state transitions of a finite state transducer for natural language processing |
US9305548B2 (en) | 2008-05-27 | 2016-04-05 | Voicebox Technologies Corporation | System and method for an integrated, multi-modal, multi-device natural language voice services environment |
US20160140187A1 (en) * | 2014-11-19 | 2016-05-19 | Electronics And Telecommunications Research Institute | System and method for answering natural language question |
US9406078B2 (en) | 2007-02-06 | 2016-08-02 | Voicebox Technologies Corporation | System and method for delivering targeted advertisements and/or providing natural language processing based on advertisements |
US9570070B2 (en) | 2009-02-20 | 2017-02-14 | Voicebox Technologies Corporation | System and method for processing multi-modal device interactions in a natural language voice services environment |
US9620113B2 (en) | 2007-12-11 | 2017-04-11 | Voicebox Technologies Corporation | System and method for providing a natural language voice user interface |
US9626703B2 (en) | 2014-09-16 | 2017-04-18 | Voicebox Technologies Corporation | Voice commerce |
US9747896B2 (en) | 2014-10-15 | 2017-08-29 | Voicebox Technologies Corporation | System and method for providing follow-up responses to prior natural language inputs of a user |
US10261994B2 (en) | 2012-05-25 | 2019-04-16 | Sdl Inc. | Method and system for automatic management of reputation of translators |
US10297249B2 (en) | 2006-10-16 | 2019-05-21 | Vb Assets, Llc | System and method for a cooperative conversational voice user interface |
US10319252B2 (en) | 2005-11-09 | 2019-06-11 | Sdl Inc. | Language capability assessment and training apparatus and techniques |
US10331784B2 (en) | 2016-07-29 | 2019-06-25 | Voicebox Technologies Corporation | System and method of disambiguating natural language processing requests |
US10417646B2 (en) | 2010-03-09 | 2019-09-17 | Sdl Inc. | Predicting the cost associated with translating textual content |
US10431214B2 (en) | 2014-11-26 | 2019-10-01 | Voicebox Technologies Corporation | System and method of determining a domain and/or an action related to a natural language input |
US10614799B2 (en) | 2014-11-26 | 2020-04-07 | Voicebox Technologies Corporation | System and method of providing intent predictions for an utterance prior to a system detection of an end of the utterance |
US11003838B2 (en) | 2011-04-18 | 2021-05-11 | Sdl Inc. | Systems and methods for monitoring post translation editing |
Families Citing this family (330)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2001013255A2 (en) | 1999-08-13 | 2001-02-22 | Pixo, Inc. | Displaying and traversing links in character array |
US8645137B2 (en) | 2000-03-16 | 2014-02-04 | Apple Inc. | Fast, language-independent method for user authentication by voice |
US6714939B2 (en) * | 2001-01-08 | 2004-03-30 | Softface, Inc. | Creation of structured data from plain text |
WO2002073451A2 (en) * | 2001-03-13 | 2002-09-19 | Intelligate Ltd. | Dynamic natural language understanding |
US20020010715A1 (en) * | 2001-07-26 | 2002-01-24 | Garry Chinn | System and method for browsing using a limited display device |
ITFI20010199A1 (en) | 2001-10-22 | 2003-04-22 | Riccardo Vieri | SYSTEM AND METHOD TO TRANSFORM TEXTUAL COMMUNICATIONS INTO VOICE AND SEND THEM WITH AN INTERNET CONNECTION TO ANY TELEPHONE SYSTEM |
US7415445B2 (en) * | 2002-09-24 | 2008-08-19 | Hewlett-Packard Development Company, L.P. | Feature selection for two-class classification systems |
US8065151B1 (en) * | 2002-12-18 | 2011-11-22 | At&T Intellectual Property Ii, L.P. | System and method of automatically building dialog services by exploiting the content and structure of websites |
US7373300B1 (en) | 2002-12-18 | 2008-05-13 | At&T Corp. | System and method of providing a spoken dialog interface to a website |
US20040148170A1 (en) * | 2003-01-23 | 2004-07-29 | Alejandro Acero | Statistical classifiers for spoken language understanding and command/control scenarios |
US8335683B2 (en) * | 2003-01-23 | 2012-12-18 | Microsoft Corporation | System for using statistical classifiers for spoken language understanding |
US7720781B2 (en) * | 2003-01-29 | 2010-05-18 | Hewlett-Packard Development Company, L.P. | Feature selection method and apparatus |
US7669134B1 (en) | 2003-05-02 | 2010-02-23 | Apple Inc. | Method and apparatus for displaying information during an instant messaging session |
US20050055216A1 (en) * | 2003-09-04 | 2005-03-10 | Sbc Knowledge Ventures, L.P. | System and method for the automated collection of data for grammar creation |
US20130304453A9 (en) * | 2004-08-20 | 2013-11-14 | Juergen Fritsch | Automated Extraction of Semantic Content and Generation of a Structured Document from Speech |
US7584103B2 (en) * | 2004-08-20 | 2009-09-01 | Multimodal Technologies, Inc. | Automated extraction of semantic content and generation of a structured document from speech |
US7769579B2 (en) * | 2005-05-31 | 2010-08-03 | Google Inc. | Learning facts from semi-structured text |
EP1844464B1 (en) * | 2005-02-03 | 2013-06-26 | Nuance Communications, Inc. | Methods and apparatus for automatically extending the voice-recognizer vocabulary of mobile communications devices |
US9208229B2 (en) * | 2005-03-31 | 2015-12-08 | Google Inc. | Anchor text summarization for corroboration |
US8682913B1 (en) | 2005-03-31 | 2014-03-25 | Google Inc. | Corroborating facts extracted from multiple sources |
US7587387B2 (en) | 2005-03-31 | 2009-09-08 | Google Inc. | User interface for facts query engine with snippets from information sources that include query terms and answer terms |
US8996470B1 (en) | 2005-05-31 | 2015-03-31 | Google Inc. | System for ensuring the internal consistency of a fact repository |
US7831545B1 (en) | 2005-05-31 | 2010-11-09 | Google Inc. | Identifying the unifying subject of a set of facts |
US8677377B2 (en) | 2005-09-08 | 2014-03-18 | Apple Inc. | Method and apparatus for building an intelligent automated assistant |
US7633076B2 (en) | 2005-09-30 | 2009-12-15 | Apple Inc. | Automated response to and sensing of user activity in portable devices |
US20070162544A1 (en) * | 2005-10-03 | 2007-07-12 | Steven Rosenbaum | Method of and a system for accepting user-created content on a computer network site |
US20070143310A1 (en) * | 2005-12-16 | 2007-06-21 | Vigen Eric A | System and method for analyzing communications using multi-dimensional hierarchical structures |
US20070143329A1 (en) * | 2005-12-16 | 2007-06-21 | Eric Arno Vigen | System and method for analyzing communications using multi-dimensional hierarchical structures |
US7835911B2 (en) * | 2005-12-30 | 2010-11-16 | Nuance Communications, Inc. | Method and system for automatically building natural language understanding models |
US8260785B2 (en) | 2006-02-17 | 2012-09-04 | Google Inc. | Automatic object reference identification and linking in a browseable fact repository |
US7831423B2 (en) * | 2006-05-25 | 2010-11-09 | Multimodal Technologies, Inc. | Replacing text representing a concept with an alternate written form of the concept |
JP2009541800A (en) * | 2006-06-22 | 2009-11-26 | マルチモダル テクノロジーズ,インク. | Speech recognition method |
US7925507B2 (en) * | 2006-07-07 | 2011-04-12 | Robert Bosch Corporation | Method and apparatus for recognizing large list of proper names in spoken dialog systems |
US9318108B2 (en) | 2010-01-18 | 2016-04-19 | Apple Inc. | Intelligent automated assistant |
US8122026B1 (en) | 2006-10-20 | 2012-02-21 | Google Inc. | Finding and disambiguating references to entities on web pages |
US8204738B2 (en) * | 2006-11-03 | 2012-06-19 | Nuance Communications, Inc. | Removing bias from features containing overlapping embedded grammars in a natural language understanding system |
US9082406B2 (en) | 2006-11-30 | 2015-07-14 | Robert Bosch Llc | Method and system for extending dialog systems to process complex activities for applications |
JP4451435B2 (en) * | 2006-12-06 | 2010-04-14 | 本田技研工業株式会社 | Language understanding device, language understanding method, and computer program |
US8117022B2 (en) * | 2006-12-07 | 2012-02-14 | Linker Sheldon O | Method and system for machine understanding, knowledge, and conversation |
US8380511B2 (en) * | 2007-02-20 | 2013-02-19 | Intervoice Limited Partnership | System and method for semantic categorization |
US8347202B1 (en) | 2007-03-14 | 2013-01-01 | Google Inc. | Determining geographic locations for place names in a fact repository |
US8977255B2 (en) | 2007-04-03 | 2015-03-10 | Apple Inc. | Method and system for operating a multi-function portable electronic device using voice-activation |
DE102007026841A1 (en) * | 2007-06-06 | 2008-12-11 | Satisloh Ag | Polishing disc for a tool for fine machining of optically effective surfaces on in particular spectacle lenses and method for its production |
US8200644B2 (en) * | 2007-06-15 | 2012-06-12 | Bryte Computer Technologies, Inc. | Methods, systems, and computer program products for search result driven charitable donations |
US9015279B2 (en) * | 2007-06-15 | 2015-04-21 | Bryte Computer Technologies | Methods, systems, and computer program products for tokenized domain name resolution |
US8521511B2 (en) | 2007-06-18 | 2013-08-27 | International Business Machines Corporation | Information extraction in a natural language understanding system |
US9342588B2 (en) * | 2007-06-18 | 2016-05-17 | International Business Machines Corporation | Reclassification of training data to improve classifier accuracy |
US9058319B2 (en) * | 2007-06-18 | 2015-06-16 | International Business Machines Corporation | Sub-model generation to improve classification accuracy |
US7970766B1 (en) | 2007-07-23 | 2011-06-28 | Google Inc. | Entity type assignment |
ITFI20070177A1 (en) | 2007-07-26 | 2009-01-27 | Riccardo Vieri | SYSTEM FOR THE CREATION AND SETTING OF AN ADVERTISING CAMPAIGN DERIVING FROM THE INSERTION OF ADVERTISING MESSAGES WITHIN AN EXCHANGE OF MESSAGES AND METHOD FOR ITS FUNCTIONING. |
US8260619B1 (en) | 2008-08-22 | 2012-09-04 | Convergys Cmg Utah, Inc. | Method and system for creating natural language understanding grammars |
US8868479B2 (en) | 2007-09-28 | 2014-10-21 | Telogis, Inc. | Natural language parsers to normalize addresses for geocoding |
US9053089B2 (en) | 2007-10-02 | 2015-06-09 | Apple Inc. | Part-of-speech tagging using latent analogy |
US8595642B1 (en) | 2007-10-04 | 2013-11-26 | Great Northern Research, LLC | Multiple shell multi faceted graphical user interface |
US8165886B1 (en) | 2007-10-04 | 2012-04-24 | Great Northern Research LLC | Speech interface system and method for control and interaction with applications on a computing system |
US8364694B2 (en) | 2007-10-26 | 2013-01-29 | Apple Inc. | Search assistant for digital media assets |
US8812435B1 (en) | 2007-11-16 | 2014-08-19 | Google Inc. | Learning objects and facts from documents |
US8620662B2 (en) | 2007-11-20 | 2013-12-31 | Apple Inc. | Context-aware unit selection |
US10002189B2 (en) | 2007-12-20 | 2018-06-19 | Apple Inc. | Method and apparatus for searching using an active ontology |
US8738486B2 (en) * | 2007-12-31 | 2014-05-27 | Mastercard International Incorporated | Methods and apparatus for implementing an ensemble merchant prediction system |
US7925652B2 (en) * | 2007-12-31 | 2011-04-12 | Mastercard International Incorporated | Methods and systems for implementing approximate string matching within a database |
US8666976B2 (en) | 2007-12-31 | 2014-03-04 | Mastercard International Incorporated | Methods and systems for implementing approximate string matching within a database |
US9330720B2 (en) | 2008-01-03 | 2016-05-03 | Apple Inc. | Methods and apparatus for altering audio output signals |
US8327272B2 (en) | 2008-01-06 | 2012-12-04 | Apple Inc. | Portable multifunction device, method, and graphical user interface for viewing and managing electronic calendars |
US8065143B2 (en) | 2008-02-22 | 2011-11-22 | Apple Inc. | Providing text input using speech data and non-speech data |
US8289283B2 (en) | 2008-03-04 | 2012-10-16 | Apple Inc. | Language input interface on a device |
US20090249182A1 (en) * | 2008-03-31 | 2009-10-01 | Iti Scotland Limited | Named entity recognition methods and apparatus |
US8996376B2 (en) | 2008-04-05 | 2015-03-31 | Apple Inc. | Intelligent text-to-speech conversion |
US10496753B2 (en) | 2010-01-18 | 2019-12-03 | Apple Inc. | Automatically adapting user interfaces for hands-free interaction |
US8275803B2 (en) | 2008-05-14 | 2012-09-25 | International Business Machines Corporation | System and method for providing answers to questions |
US8332394B2 (en) | 2008-05-23 | 2012-12-11 | International Business Machines Corporation | System and method for providing question and answers with deferred type evaluation |
US8464150B2 (en) | 2008-06-07 | 2013-06-11 | Apple Inc. | Automatic language identification for dynamic text processing |
US20100030549A1 (en) | 2008-07-31 | 2010-02-04 | Lee Michael M | Mobile device having human language translation capability with positional feedback |
US8768702B2 (en) | 2008-09-05 | 2014-07-01 | Apple Inc. | Multi-tiered voice feedback in an electronic device |
US8898568B2 (en) | 2008-09-09 | 2014-11-25 | Apple Inc. | Audio user interface |
US8380687B1 (en) * | 2008-09-19 | 2013-02-19 | Symantec Corporation | Method and apparatus for providing a message trail of conversationally related messages |
US8352272B2 (en) | 2008-09-29 | 2013-01-08 | Apple Inc. | Systems and methods for text to speech synthesis |
US20100082328A1 (en) * | 2008-09-29 | 2010-04-01 | Apple Inc. | Systems and methods for speech preprocessing in text to speech synthesis |
US8352268B2 (en) | 2008-09-29 | 2013-01-08 | Apple Inc. | Systems and methods for selective rate of speech and speech preferences for text to speech synthesis |
US8583418B2 (en) | 2008-09-29 | 2013-11-12 | Apple Inc. | Systems and methods of detecting language and natural language strings for text to speech synthesis |
US8355919B2 (en) | 2008-09-29 | 2013-01-15 | Apple Inc. | Systems and methods for text normalization for text to speech synthesis |
US8396714B2 (en) | 2008-09-29 | 2013-03-12 | Apple Inc. | Systems and methods for concatenation of words in text to speech synthesis |
US8712776B2 (en) | 2008-09-29 | 2014-04-29 | Apple Inc. | Systems and methods for selective text to speech synthesis |
US8676904B2 (en) | 2008-10-02 | 2014-03-18 | Apple Inc. | Electronic devices with voice command and contextual data processing capabilities |
US8326809B2 (en) * | 2008-10-27 | 2012-12-04 | Sas Institute Inc. | Systems and methods for defining and processing text segmentation rules |
WO2010067118A1 (en) | 2008-12-11 | 2010-06-17 | Novauris Technologies Limited | Speech recognition involving a mobile device |
US8862252B2 (en) | 2009-01-30 | 2014-10-14 | Apple Inc. | Audio user interface for displayless electronic device |
US8380507B2 (en) | 2009-03-09 | 2013-02-19 | Apple Inc. | Systems and methods for determining the language to use for speech generated by a text to speech engine |
US8290273B2 (en) * | 2009-03-27 | 2012-10-16 | Raytheon Bbn Technologies Corp. | Multi-frame videotext recognition |
US10241752B2 (en) | 2011-09-30 | 2019-03-26 | Apple Inc. | Interface for a virtual digital assistant |
US10540976B2 (en) | 2009-06-05 | 2020-01-21 | Apple Inc. | Contextual voice commands |
US10241644B2 (en) | 2011-06-03 | 2019-03-26 | Apple Inc. | Actionable reminder entries |
US10255566B2 (en) | 2011-06-03 | 2019-04-09 | Apple Inc. | Generating and processing task items that represent tasks to perform |
US9858925B2 (en) | 2009-06-05 | 2018-01-02 | Apple Inc. | Using context information to facilitate processing of commands in a virtual assistant |
US20130219333A1 (en) * | 2009-06-12 | 2013-08-22 | Adobe Systems Incorporated | Extensible Framework for Facilitating Interaction with Devices |
US9431006B2 (en) | 2009-07-02 | 2016-08-30 | Apple Inc. | Methods and apparatuses for automatic speech recognition |
KR20110006004A (en) * | 2009-07-13 | 2011-01-20 | 삼성전자주식회사 | Apparatus and method for optimizing concatenate recognition unit |
US8380485B1 (en) * | 2009-08-13 | 2013-02-19 | The United States Of America As Represented By The Director, National Security Agency | Device for and method of language processing |
US20110078192A1 (en) * | 2009-09-30 | 2011-03-31 | International Business Machines Corporation | Inferring lexical answer types of questions from context |
US8682649B2 (en) | 2009-11-12 | 2014-03-25 | Apple Inc. | Sentiment prediction from textual data |
US8775160B1 (en) | 2009-12-17 | 2014-07-08 | Shopzilla, Inc. | Usage based query response |
US8428933B1 (en) | 2009-12-17 | 2013-04-23 | Shopzilla, Inc. | Usage based query response |
US8600743B2 (en) | 2010-01-06 | 2013-12-03 | Apple Inc. | Noise profile determination for voice-related feature |
US8311838B2 (en) | 2010-01-13 | 2012-11-13 | Apple Inc. | Devices and methods for identifying a prompt corresponding to a voice input in a sequence of prompts |
US8381107B2 (en) | 2010-01-13 | 2013-02-19 | Apple Inc. | Adaptive audio feedback system and method |
US10553209B2 (en) | 2010-01-18 | 2020-02-04 | Apple Inc. | Systems and methods for hands-free notification summaries |
US10276170B2 (en) | 2010-01-18 | 2019-04-30 | Apple Inc. | Intelligent automated assistant |
US10705794B2 (en) | 2010-01-18 | 2020-07-07 | Apple Inc. | Automatically adapting user interfaces for hands-free interaction |
US10679605B2 (en) | 2010-01-18 | 2020-06-09 | Apple Inc. | Hands-free list-reading by intelligent automated assistant |
DE202011111062U1 (en) | 2010-01-25 | 2019-02-19 | Newvaluexchange Ltd. | Device and system for a digital conversation management platform |
US8682667B2 (en) | 2010-02-25 | 2014-03-25 | Apple Inc. | User profiling for selecting user specific voice input processing information |
US8639516B2 (en) | 2010-06-04 | 2014-01-28 | Apple Inc. | User-specific noise suppression for voice quality improvements |
US8380719B2 (en) * | 2010-06-18 | 2013-02-19 | Microsoft Corporation | Semantic content searching |
US8713021B2 (en) | 2010-07-07 | 2014-04-29 | Apple Inc. | Unsupervised document clustering using latent semantic density analysis |
US9104670B2 (en) | 2010-07-21 | 2015-08-11 | Apple Inc. | Customized search or acquisition of digital media assets |
US8719006B2 (en) | 2010-08-27 | 2014-05-06 | Apple Inc. | Combined statistical and rule-based part-of-speech tagging for text-to-speech synthesis |
EP2616927A4 (en) | 2010-09-24 | 2017-02-22 | International Business Machines Corporation | Using ontological information in open domain type coercion |
US8892550B2 (en) | 2010-09-24 | 2014-11-18 | International Business Machines Corporation | Source expansion for information retrieval and information extraction |
US8943051B2 (en) | 2010-09-24 | 2015-01-27 | International Business Machines Corporation | Lexical answer type confidence estimation and application |
CN103221952B (en) | 2010-09-24 | 2016-01-20 | 国际商业机器公司 | The method and system of morphology answer type reliability estimating and application |
CN103250129A (en) | 2010-09-24 | 2013-08-14 | 国际商业机器公司 | Providing question and answers with deferred type evaluation using text with limited structure |
US20120078062A1 (en) | 2010-09-24 | 2012-03-29 | International Business Machines Corporation | Decision-support application and system for medical differential-diagnosis and treatment using a question-answering system |
US8719014B2 (en) | 2010-09-27 | 2014-05-06 | Apple Inc. | Electronic device with text error correction based on voice recognition data |
WO2012047530A1 (en) | 2010-09-28 | 2012-04-12 | International Business Machines Corporation | Providing answers to questions using logical synthesis of candidate answers |
EP2622592A4 (en) * | 2010-09-28 | 2017-04-05 | International Business Machines Corporation | Providing answers to questions using multiple models to score candidate answers |
WO2012047532A1 (en) | 2010-09-28 | 2012-04-12 | International Business Machines Corporation | Providing answers to questions using hypothesis pruning |
US8959102B2 (en) | 2010-10-08 | 2015-02-17 | Mmodal Ip Llc | Structured searching of dynamic structured document corpuses |
US10762293B2 (en) | 2010-12-22 | 2020-09-01 | Apple Inc. | Using parts-of-speech tagging and named entity recognition for spelling correction |
US10515147B2 (en) | 2010-12-22 | 2019-12-24 | Apple Inc. | Using statistical language models for contextual lookup |
US8781836B2 (en) | 2011-02-22 | 2014-07-15 | Apple Inc. | Hearing assistance system for providing consistent human speech |
US9262612B2 (en) | 2011-03-21 | 2016-02-16 | Apple Inc. | Device access using voice authentication |
US9940672B2 (en) * | 2011-03-22 | 2018-04-10 | Isentium, Llc | System for generating data from social media messages for the real-time evaluation of publicly traded assets |
US10057736B2 (en) | 2011-06-03 | 2018-08-21 | Apple Inc. | Active transport based notifications |
US10672399B2 (en) | 2011-06-03 | 2020-06-02 | Apple Inc. | Switching between text data and audio data based on a mapping |
US8812294B2 (en) | 2011-06-21 | 2014-08-19 | Apple Inc. | Translating phrases from one language into another using an order-based set of declarative rules |
US9367526B1 (en) * | 2011-07-26 | 2016-06-14 | Nuance Communications, Inc. | Word classing for language modeling |
US8706472B2 (en) | 2011-08-11 | 2014-04-22 | Apple Inc. | Method for disambiguating multiple readings in language conversion |
US8994660B2 (en) | 2011-08-29 | 2015-03-31 | Apple Inc. | Text correction processing |
US8762156B2 (en) | 2011-09-28 | 2014-06-24 | Apple Inc. | Speech recognition repair using contextual information |
US20130275116A1 (en) * | 2011-12-31 | 2013-10-17 | Electionear, Inc. | Interactive, live-connection, specifically targetable, database-supported, dynamic dialogue management engine |
US10134385B2 (en) | 2012-03-02 | 2018-11-20 | Apple Inc. | Systems and methods for name pronunciation |
US9483461B2 (en) | 2012-03-06 | 2016-11-01 | Apple Inc. | Handling speech synthesis of content for multiple languages |
US20130311362A1 (en) | 2012-04-26 | 2013-11-21 | Mastercard International Incorporated | Systems and methods for verifying payee information in electronic payments |
US9280610B2 (en) | 2012-05-14 | 2016-03-08 | Apple Inc. | Crowd sourcing information to fulfill user requests |
US8775442B2 (en) | 2012-05-15 | 2014-07-08 | Apple Inc. | Semantic search using a single-source semantic model |
US10417037B2 (en) | 2012-05-15 | 2019-09-17 | Apple Inc. | Systems and methods for integrating third party services with a digital assistant |
WO2013185109A2 (en) | 2012-06-08 | 2013-12-12 | Apple Inc. | Systems and methods for recognizing textual identifiers within a plurality of words |
US9721563B2 (en) | 2012-06-08 | 2017-08-01 | Apple Inc. | Name recognition system |
US9495129B2 (en) | 2012-06-29 | 2016-11-15 | Apple Inc. | Device, method, and user interface for voice-activated navigation and browsing of a document |
US9576574B2 (en) | 2012-09-10 | 2017-02-21 | Apple Inc. | Context-sensitive handling of interruptions by intelligent digital assistant |
US10614725B2 (en) | 2012-09-11 | 2020-04-07 | International Business Machines Corporation | Generating secondary questions in an introspective question answering system |
US9547647B2 (en) | 2012-09-19 | 2017-01-17 | Apple Inc. | Voice-based media searching |
US8935167B2 (en) | 2012-09-25 | 2015-01-13 | Apple Inc. | Exemplar-based latent perceptual modeling for automatic speech recognition |
DE112014000709B4 (en) | 2013-02-07 | 2021-12-30 | Apple Inc. | METHOD AND DEVICE FOR OPERATING A VOICE TRIGGER FOR A DIGITAL ASSISTANT |
US10652394B2 (en) | 2013-03-14 | 2020-05-12 | Apple Inc. | System and method for processing voicemail |
US10572476B2 (en) | 2013-03-14 | 2020-02-25 | Apple Inc. | Refining a search based on schedule items |
US9733821B2 (en) | 2013-03-14 | 2017-08-15 | Apple Inc. | Voice control to diagnose inadvertent activation of accessibility features |
US10642574B2 (en) | 2013-03-14 | 2020-05-05 | Apple Inc. | Device, method, and graphical user interface for outputting captions |
US9368114B2 (en) | 2013-03-14 | 2016-06-14 | Apple Inc. | Context-sensitive handling of interruptions |
US9977779B2 (en) | 2013-03-14 | 2018-05-22 | Apple Inc. | Automatic supplementation of word correction dictionaries |
CN105190607B (en) | 2013-03-15 | 2018-11-30 | 苹果公司 | Pass through the user training of intelligent digital assistant |
US10748529B1 (en) | 2013-03-15 | 2020-08-18 | Apple Inc. | Voice activated device for use with a voice-based digital assistant |
WO2014144579A1 (en) | 2013-03-15 | 2014-09-18 | Apple Inc. | System and method for updating an adaptive speech recognition model |
US9922642B2 (en) | 2013-03-15 | 2018-03-20 | Apple Inc. | Training an at least partial voice command system |
KR101904293B1 (en) | 2013-03-15 | 2018-10-05 | 애플 인크. | Context-sensitive handling of interruptions |
US9262938B2 (en) | 2013-03-15 | 2016-02-16 | International Business Machines Corporation | Combining different type coercion components for deferred type evaluation |
WO2014197336A1 (en) | 2013-06-07 | 2014-12-11 | Apple Inc. | System and method for detecting errors in interactions with a voice-based digital assistant |
WO2014197334A2 (en) | 2013-06-07 | 2014-12-11 | Apple Inc. | System and method for user-specified pronunciation of words for speech synthesis and recognition |
US9582608B2 (en) | 2013-06-07 | 2017-02-28 | Apple Inc. | Unified ranking with entropy-weighted information for phrase-based semantic auto-completion |
WO2014197335A1 (en) | 2013-06-08 | 2014-12-11 | Apple Inc. | Interpreting and acting upon commands that involve sharing information with remote devices |
DE112014002747T5 (en) | 2013-06-09 | 2016-03-03 | Apple Inc. | Apparatus, method and graphical user interface for enabling conversation persistence over two or more instances of a digital assistant |
US10176167B2 (en) | 2013-06-09 | 2019-01-08 | Apple Inc. | System and method for inferring user intent from speech inputs |
AU2014278595B2 (en) | 2013-06-13 | 2017-04-06 | Apple Inc. | System and method for emergency calls initiated by voice command |
IN2013MU02217A (en) * | 2013-07-01 | 2015-06-12 | Tata Consultancy Services Ltd | |
US10229106B2 (en) * | 2013-07-26 | 2019-03-12 | Nuance Communications, Inc. | Initializing a workspace for building a natural language understanding system |
JP6163266B2 (en) | 2013-08-06 | 2017-07-12 | アップル インコーポレイテッド | Automatic activation of smart responses based on activation from remote devices |
US10521866B2 (en) | 2013-10-15 | 2019-12-31 | Mastercard International Incorporated | Systems and methods for associating related merchants |
US10296160B2 (en) | 2013-12-06 | 2019-05-21 | Apple Inc. | Method for extracting salient dialog usage from live data |
US9817813B2 (en) * | 2014-01-08 | 2017-11-14 | Genesys Telecommunications Laboratories, Inc. | Generalized phrases in automatic speech recognition systems |
US10839432B1 (en) | 2014-03-07 | 2020-11-17 | Genesys Telecommunications Laboratories, Inc. | Systems and methods for automating customer interactions with enterprises |
US9620105B2 (en) | 2014-05-15 | 2017-04-11 | Apple Inc. | Analyzing audio input for efficient speech and music recognition |
US10592095B2 (en) | 2014-05-23 | 2020-03-17 | Apple Inc. | Instantaneous speaking of content on touch devices |
US9502031B2 (en) | 2014-05-27 | 2016-11-22 | Apple Inc. | Method for supporting dynamic grammars in WFST-based ASR |
US9430463B2 (en) | 2014-05-30 | 2016-08-30 | Apple Inc. | Exemplar-based natural language processing |
US10170123B2 (en) | 2014-05-30 | 2019-01-01 | Apple Inc. | Intelligent assistant for home automation |
US10289433B2 (en) | 2014-05-30 | 2019-05-14 | Apple Inc. | Domain specific language for encoding assistant dialog |
US10078631B2 (en) | 2014-05-30 | 2018-09-18 | Apple Inc. | Entropy-guided text prediction using combined word and character n-gram language models |
US9734193B2 (en) | 2014-05-30 | 2017-08-15 | Apple Inc. | Determining domain salience ranking from ambiguous words in natural speech |
US9842101B2 (en) | 2014-05-30 | 2017-12-12 | Apple Inc. | Predictive conversion of language input |
US9715875B2 (en) | 2014-05-30 | 2017-07-25 | Apple Inc. | Reducing the need for manual start/end-pointing and trigger phrases |
EP3480811A1 (en) | 2014-05-30 | 2019-05-08 | Apple Inc. | Multi-command single utterance input method |
US9760559B2 (en) | 2014-05-30 | 2017-09-12 | Apple Inc. | Predictive text input |
US9785630B2 (en) | 2014-05-30 | 2017-10-10 | Apple Inc. | Text prediction using combined word N-gram and unigram language models |
US9633004B2 (en) | 2014-05-30 | 2017-04-25 | Apple Inc. | Better resolution when referencing to concepts |
US10659851B2 (en) | 2014-06-30 | 2020-05-19 | Apple Inc. | Real-time digital assistant knowledge updates |
US9338493B2 (en) | 2014-06-30 | 2016-05-10 | Apple Inc. | Intelligent automated assistant for TV user interactions |
US10446141B2 (en) | 2014-08-28 | 2019-10-15 | Apple Inc. | Automatic speech recognition based on user feedback |
US9818400B2 (en) | 2014-09-11 | 2017-11-14 | Apple Inc. | Method and apparatus for discovering trending terms in speech requests |
US10789041B2 (en) | 2014-09-12 | 2020-09-29 | Apple Inc. | Dynamic thresholds for always listening speech trigger |
US10127911B2 (en) | 2014-09-30 | 2018-11-13 | Apple Inc. | Speaker identification and unsupervised speaker adaptation techniques |
US10074360B2 (en) | 2014-09-30 | 2018-09-11 | Apple Inc. | Providing an indication of the suitability of speech recognition |
US9646609B2 (en) | 2014-09-30 | 2017-05-09 | Apple Inc. | Caching apparatus for serving phonetic pronunciations |
US9886432B2 (en) | 2014-09-30 | 2018-02-06 | Apple Inc. | Parsimonious handling of word inflection via categorical stem + suffix N-gram language models |
US9668121B2 (en) | 2014-09-30 | 2017-05-30 | Apple Inc. | Social reminders |
US9501525B2 (en) | 2014-11-05 | 2016-11-22 | International Business Machines Corporation | Answer sequence evaluation |
US9361075B2 (en) | 2014-11-12 | 2016-06-07 | International Business Machines Corporation | Contraction aware parsing system for domain-specific languages |
US10558926B2 (en) * | 2014-11-20 | 2020-02-11 | Academia Sinica | Statistical pattern generation for information extraction |
US10552013B2 (en) | 2014-12-02 | 2020-02-04 | Apple Inc. | Data detection |
US10061842B2 (en) | 2014-12-09 | 2018-08-28 | International Business Machines Corporation | Displaying answers in accordance with answer classifications |
US9711141B2 (en) | 2014-12-09 | 2017-07-18 | Apple Inc. | Disambiguating heteronyms in speech synthesis |
US10152299B2 (en) | 2015-03-06 | 2018-12-11 | Apple Inc. | Reducing response latency of intelligent automated assistants |
US9865280B2 (en) | 2015-03-06 | 2018-01-09 | Apple Inc. | Structured dictation using intelligent automated assistants |
US10567477B2 (en) | 2015-03-08 | 2020-02-18 | Apple Inc. | Virtual assistant continuity |
US9886953B2 (en) | 2015-03-08 | 2018-02-06 | Apple Inc. | Virtual assistant activation |
US9721566B2 (en) | 2015-03-08 | 2017-08-01 | Apple Inc. | Competing devices responding to voice triggers |
US9899019B2 (en) | 2015-03-18 | 2018-02-20 | Apple Inc. | Systems and methods for structured stem and suffix language models |
US9842105B2 (en) | 2015-04-16 | 2017-12-12 | Apple Inc. | Parsimonious continuous-space phrase representations for natural language processing |
WO2016167796A1 (en) * | 2015-04-17 | 2016-10-20 | Hewlett Packard Enterprise Development Lp | Hierarchical classifiers |
CN109308357B (en) * | 2015-05-04 | 2023-07-18 | 上海智臻智能网络科技股份有限公司 | Method, device and equipment for obtaining answer information |
US10460227B2 (en) | 2015-05-15 | 2019-10-29 | Apple Inc. | Virtual assistant in a communication session |
US10083688B2 (en) | 2015-05-27 | 2018-09-25 | Apple Inc. | Device voice control for selecting a displayed affordance |
US10127220B2 (en) | 2015-06-04 | 2018-11-13 | Apple Inc. | Language identification from short strings |
US10101822B2 (en) | 2015-06-05 | 2018-10-16 | Apple Inc. | Language input correction |
US9578173B2 (en) | 2015-06-05 | 2017-02-21 | Apple Inc. | Virtual assistant aided communication with 3rd party service in a communication session |
US10255907B2 (en) | 2015-06-07 | 2019-04-09 | Apple Inc. | Automatic accent detection using acoustic models |
US10186254B2 (en) | 2015-06-07 | 2019-01-22 | Apple Inc. | Context-based endpoint detection |
US11025565B2 (en) | 2015-06-07 | 2021-06-01 | Apple Inc. | Personalized prediction of responses for instant messaging |
US20160378747A1 (en) | 2015-06-29 | 2016-12-29 | Apple Inc. | Virtual assistant for media playback |
US10671428B2 (en) | 2015-09-08 | 2020-06-02 | Apple Inc. | Distributed personal assistant |
US10747498B2 (en) | 2015-09-08 | 2020-08-18 | Apple Inc. | Zero latency digital assistant |
US9697820B2 (en) | 2015-09-24 | 2017-07-04 | Apple Inc. | Unit-selection text-to-speech synthesis using concatenation-sensitive neural networks |
US10366158B2 (en) | 2015-09-29 | 2019-07-30 | Apple Inc. | Efficient word encoding for recurrent neural network language models |
US11010550B2 (en) | 2015-09-29 | 2021-05-18 | Apple Inc. | Unified language modeling framework for word prediction, auto-completion and auto-correction |
US11587559B2 (en) | 2015-09-30 | 2023-02-21 | Apple Inc. | Intelligent device identification |
US10691473B2 (en) | 2015-11-06 | 2020-06-23 | Apple Inc. | Intelligent automated assistant in a messaging environment |
US10049668B2 (en) | 2015-12-02 | 2018-08-14 | Apple Inc. | Applying neural network language models to weighted finite state transducers for automatic speech recognition |
US10223066B2 (en) | 2015-12-23 | 2019-03-05 | Apple Inc. | Proactive assistance based on dialog communication between devices |
US10417346B2 (en) | 2016-01-23 | 2019-09-17 | Microsoft Technology Licensing, Llc | Tool for facilitating the development of new language understanding scenarios |
US10446143B2 (en) | 2016-03-14 | 2019-10-15 | Apple Inc. | Identification of voice inputs providing credentials |
US9934775B2 (en) | 2016-05-26 | 2018-04-03 | Apple Inc. | Unit-selection text-to-speech synthesis based on predicted concatenation parameters |
US9972304B2 (en) | 2016-06-03 | 2018-05-15 | Apple Inc. | Privacy preserving distributed evaluation framework for embedded personalized systems |
US10249300B2 (en) | 2016-06-06 | 2019-04-02 | Apple Inc. | Intelligent list reading |
US11227589B2 (en) | 2016-06-06 | 2022-01-18 | Apple Inc. | Intelligent list reading |
US10049663B2 (en) | 2016-06-08 | 2018-08-14 | Apple, Inc. | Intelligent automated assistant for media exploration |
DK179588B1 (en) | 2016-06-09 | 2019-02-22 | Apple Inc. | Intelligent automated assistant in a home environment |
US10586535B2 (en) | 2016-06-10 | 2020-03-10 | Apple Inc. | Intelligent digital assistant in a multi-tasking environment |
US10509862B2 (en) | 2016-06-10 | 2019-12-17 | Apple Inc. | Dynamic phrase expansion of language input |
US10067938B2 (en) | 2016-06-10 | 2018-09-04 | Apple Inc. | Multilingual word prediction |
US10490187B2 (en) | 2016-06-10 | 2019-11-26 | Apple Inc. | Digital assistant providing automated status report |
US10192552B2 (en) | 2016-06-10 | 2019-01-29 | Apple Inc. | Digital assistant providing whispered speech |
DK179343B1 (en) | 2016-06-11 | 2018-05-14 | Apple Inc | Intelligent task discovery |
DK179049B1 (en) | 2016-06-11 | 2017-09-18 | Apple Inc | Data driven natural language event detection and classification |
DK201670540A1 (en) | 2016-06-11 | 2018-01-08 | Apple Inc | Application integration with a digital assistant |
DK179415B1 (en) | 2016-06-11 | 2018-06-14 | Apple Inc | Intelligent device arbitration and control |
US10606952B2 (en) | 2016-06-24 | 2020-03-31 | Elemental Cognition Llc | Architecture and processes for computer learning and understanding |
CN109416803A (en) * | 2016-07-06 | 2019-03-01 | 万事达卡国际公司 | It is presented sales message the method and system with opinion by dialog interface |
US10474753B2 (en) | 2016-09-07 | 2019-11-12 | Apple Inc. | Language identification using recurrent neural networks |
US10043516B2 (en) | 2016-09-23 | 2018-08-07 | Apple Inc. | Intelligent automated assistant |
US11281993B2 (en) | 2016-12-05 | 2022-03-22 | Apple Inc. | Model and ensemble compression for metric learning |
US10593346B2 (en) | 2016-12-22 | 2020-03-17 | Apple Inc. | Rank-reduced token representation for automatic speech recognition |
US11204787B2 (en) | 2017-01-09 | 2021-12-21 | Apple Inc. | Application integration with a digital assistant |
US10334103B2 (en) | 2017-01-25 | 2019-06-25 | International Business Machines Corporation | Message translation for cognitive assistance |
CN108287858B (en) * | 2017-03-02 | 2021-08-10 | 腾讯科技(深圳)有限公司 | Semantic extraction method and device for natural language |
US11729120B2 (en) * | 2017-03-16 | 2023-08-15 | Microsoft Technology Licensing, Llc | Generating responses in automated chatting |
US10417266B2 (en) | 2017-05-09 | 2019-09-17 | Apple Inc. | Context-aware ranking of intelligent response suggestions |
DK201770383A1 (en) | 2017-05-09 | 2018-12-14 | Apple Inc. | User interface for correcting recognition errors |
DK201770439A1 (en) | 2017-05-11 | 2018-12-13 | Apple Inc. | Offline personal assistant |
US10726832B2 (en) | 2017-05-11 | 2020-07-28 | Apple Inc. | Maintaining privacy of personal information |
US10395654B2 (en) | 2017-05-11 | 2019-08-27 | Apple Inc. | Text normalization based on a data-driven learning network |
DK201770428A1 (en) | 2017-05-12 | 2019-02-18 | Apple Inc. | Low-latency intelligent automated assistant |
DK179745B1 (en) | 2017-05-12 | 2019-05-01 | Apple Inc. | SYNCHRONIZATION AND TASK DELEGATION OF A DIGITAL ASSISTANT |
DK179496B1 (en) | 2017-05-12 | 2019-01-15 | Apple Inc. | USER-SPECIFIC Acoustic Models |
US11301477B2 (en) | 2017-05-12 | 2022-04-12 | Apple Inc. | Feedback analysis of a digital assistant |
DK201770431A1 (en) | 2017-05-15 | 2018-12-20 | Apple Inc. | Optimizing dialogue policy decisions for digital assistants using implicit feedback |
DK201770432A1 (en) | 2017-05-15 | 2018-12-21 | Apple Inc. | Hierarchical belief states for digital assistants |
US10403278B2 (en) | 2017-05-16 | 2019-09-03 | Apple Inc. | Methods and systems for phonetic matching in digital assistant services |
US20180336892A1 (en) | 2017-05-16 | 2018-11-22 | Apple Inc. | Detecting a trigger of a digital assistant |
DK179560B1 (en) | 2017-05-16 | 2019-02-18 | Apple Inc. | Far-field extension for digital assistant services |
US10311144B2 (en) | 2017-05-16 | 2019-06-04 | Apple Inc. | Emoji word sense disambiguation |
US20180336275A1 (en) | 2017-05-16 | 2018-11-22 | Apple Inc. | Intelligent automated assistant for media exploration |
US10657328B2 (en) | 2017-06-02 | 2020-05-19 | Apple Inc. | Multi-task recurrent neural network architecture for efficient morphology handling in neural language modeling |
US10719539B2 (en) | 2017-06-06 | 2020-07-21 | Mastercard International Incorporated | Method and system for automatic reporting of analytics and distribution of advice using a conversational interface |
US11537644B2 (en) | 2017-06-06 | 2022-12-27 | Mastercard International Incorporated | Method and system for conversational input device with intelligent crowd-sourced options |
US10572963B1 (en) * | 2017-07-14 | 2020-02-25 | Synapse Technology Corporation | Detection of items |
US10445429B2 (en) | 2017-09-21 | 2019-10-15 | Apple Inc. | Natural language understanding using vocabularies with compressed serialized tries |
US10755051B2 (en) | 2017-09-29 | 2020-08-25 | Apple Inc. | Rule-based natural language processing |
US11010656B2 (en) * | 2017-10-30 | 2021-05-18 | Clinc, Inc. | System and method for implementing an artificially intelligent virtual assistant using machine learning |
US10636424B2 (en) | 2017-11-30 | 2020-04-28 | Apple Inc. | Multi-turn canned dialog |
US10685183B1 (en) * | 2018-01-04 | 2020-06-16 | Facebook, Inc. | Consumer insights analysis using word embeddings |
US10733982B2 (en) | 2018-01-08 | 2020-08-04 | Apple Inc. | Multi-directional dialog |
US10733375B2 (en) | 2018-01-31 | 2020-08-04 | Apple Inc. | Knowledge-based framework for improving natural language understanding |
US10789959B2 (en) | 2018-03-02 | 2020-09-29 | Apple Inc. | Training speaker recognition models for digital assistants |
US10592604B2 (en) | 2018-03-12 | 2020-03-17 | Apple Inc. | Inverse text normalization for automatic speech recognition |
US10831797B2 (en) * | 2018-03-23 | 2020-11-10 | International Business Machines Corporation | Query recognition resiliency determination in virtual agent systems |
US10818288B2 (en) | 2018-03-26 | 2020-10-27 | Apple Inc. | Natural assistant interaction |
US10909331B2 (en) | 2018-03-30 | 2021-02-02 | Apple Inc. | Implicit identification of translation payload with neural machine translation |
US11145294B2 (en) | 2018-05-07 | 2021-10-12 | Apple Inc. | Intelligent automated assistant for delivering content from user experiences |
US10928918B2 (en) | 2018-05-07 | 2021-02-23 | Apple Inc. | Raise to speak |
US10984780B2 (en) | 2018-05-21 | 2021-04-20 | Apple Inc. | Global semantic word embeddings using bi-directional recurrent neural networks |
US11386266B2 (en) | 2018-06-01 | 2022-07-12 | Apple Inc. | Text correction |
DK180639B1 (en) | 2018-06-01 | 2021-11-04 | Apple Inc | DISABILITY OF ATTENTION-ATTENTIVE VIRTUAL ASSISTANT |
US10892996B2 (en) | 2018-06-01 | 2021-01-12 | Apple Inc. | Variable latency device coordination |
DK179822B1 (en) | 2018-06-01 | 2019-07-12 | Apple Inc. | Voice interaction at a primary device to access call functionality of a companion device |
DK201870355A1 (en) | 2018-06-01 | 2019-12-16 | Apple Inc. | Virtual assistant operation in multi-device environments |
US10496705B1 (en) | 2018-06-03 | 2019-12-03 | Apple Inc. | Accelerated task performance |
AU2019289150B2 (en) * | 2018-06-17 | 2024-07-25 | Genesys Cloud Services Holdings II, LLC | Systems and methods for automating customer interactions with enterprises |
US10452959B1 (en) | 2018-07-20 | 2019-10-22 | Synapse Tehnology Corporation | Multi-perspective detection of objects |
US11010561B2 (en) | 2018-09-27 | 2021-05-18 | Apple Inc. | Sentiment prediction from textual data |
US11170166B2 (en) | 2018-09-28 | 2021-11-09 | Apple Inc. | Neural typographical error modeling via generative adversarial networks |
US11462215B2 (en) | 2018-09-28 | 2022-10-04 | Apple Inc. | Multi-modal inputs for voice commands |
US10839159B2 (en) | 2018-09-28 | 2020-11-17 | Apple Inc. | Named entity normalization in a spoken dialog system |
US11475898B2 (en) | 2018-10-26 | 2022-10-18 | Apple Inc. | Low-latency multi-speaker speech recognition |
US11638059B2 (en) | 2019-01-04 | 2023-04-25 | Apple Inc. | Content playback on multiple devices |
US11348573B2 (en) | 2019-03-18 | 2022-05-31 | Apple Inc. | Multimodality in digital assistant systems |
US11423908B2 (en) | 2019-05-06 | 2022-08-23 | Apple Inc. | Interpreting spoken requests |
DK201970509A1 (en) | 2019-05-06 | 2021-01-15 | Apple Inc | Spoken notifications |
US11307752B2 (en) | 2019-05-06 | 2022-04-19 | Apple Inc. | User configurable task triggers |
US11475884B2 (en) | 2019-05-06 | 2022-10-18 | Apple Inc. | Reducing digital assistant latency when a language is incorrectly determined |
US11140099B2 (en) | 2019-05-21 | 2021-10-05 | Apple Inc. | Providing message response suggestions |
DK201970511A1 (en) | 2019-05-31 | 2021-02-15 | Apple Inc | Voice identification in digital assistant systems |
US11289073B2 (en) | 2019-05-31 | 2022-03-29 | Apple Inc. | Device text to speech |
DK180129B1 (en) | 2019-05-31 | 2020-06-02 | Apple Inc. | User activity shortcut suggestions |
US11496600B2 (en) | 2019-05-31 | 2022-11-08 | Apple Inc. | Remote execution of machine-learned models |
US11360641B2 (en) | 2019-06-01 | 2022-06-14 | Apple Inc. | Increasing the relevance of new available information |
US11010605B2 (en) | 2019-07-30 | 2021-05-18 | Rapiscan Laboratories, Inc. | Multi-model detection of objects |
US11488406B2 (en) | 2019-09-25 | 2022-11-01 | Apple Inc. | Text detection using global geometry estimators |
CN113886571A (en) * | 2020-07-01 | 2022-01-04 | 北京三星通信技术研究有限公司 | Entity identification method, entity identification device, electronic equipment and computer readable storage medium |
US12008991B2 (en) | 2021-05-27 | 2024-06-11 | Soundhound Ai Ip, Llc | Enabling natural language interactions with user interfaces for users of a software application |
CN113822034B (en) * | 2021-06-07 | 2024-04-19 | 腾讯科技(深圳)有限公司 | Method, device, computer equipment and storage medium for replying text |
US20230401385A1 (en) * | 2022-06-13 | 2023-12-14 | Oracle International Corporation | Hierarchical named entity recognition with multi-task setup |
Citations (33)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US4706212A (en) * | 1971-08-31 | 1987-11-10 | Toma Peter P | Method using a programmed digital computer system for translation between natural languages |
US5128865A (en) * | 1989-03-10 | 1992-07-07 | Bso/Buro Voor Systeemontwikkeling B.V. | Method for determining the semantic relatedness of lexical items in a text |
US5299125A (en) * | 1990-08-09 | 1994-03-29 | Semantic Compaction Systems | Natural language processing system and method for parsing a plurality of input symbol sequences into syntactically or pragmatically correct word messages |
US5321608A (en) * | 1990-11-30 | 1994-06-14 | Hitachi, Ltd. | Method and system for processing natural language |
US5528491A (en) * | 1992-08-31 | 1996-06-18 | Language Engineering Corporation | Apparatus and method for automated natural language translation |
US5541836A (en) * | 1991-12-30 | 1996-07-30 | At&T Corp. | Word disambiguation apparatus and methods |
US5576954A (en) * | 1993-11-05 | 1996-11-19 | University Of Central Florida | Process for determination of text relevancy |
US5634024A (en) * | 1994-05-18 | 1997-05-27 | Nec Corporation | Definition execution system for operations and functions in computer system |
US5642502A (en) * | 1994-12-06 | 1997-06-24 | University Of Central Florida | Method and system for searching for relevant documents from a text database collection, using statistical ranking, relevancy feedback and small pieces of text |
US5675710A (en) * | 1995-06-07 | 1997-10-07 | Lucent Technologies, Inc. | Method and apparatus for training a text classifier |
US5677993A (en) * | 1992-08-31 | 1997-10-14 | Hitachi, Ltd. | Information processing apparatus using pointing input and speech input |
US5680511A (en) * | 1995-06-07 | 1997-10-21 | Dragon Systems, Inc. | Systems and methods for word recognition |
US5748974A (en) * | 1994-12-13 | 1998-05-05 | International Business Machines Corporation | Multimodal natural language interface for cross-application tasks |
US5805775A (en) * | 1996-02-02 | 1998-09-08 | Digital Equipment Corporation | Application user interface |
US5844302A (en) * | 1993-12-22 | 1998-12-01 | Siemens Aktiengesellschaft | Integrated semiconductor circuit with capacitors of precisely defined capacitance and process for producing the circuit |
US5895466A (en) * | 1997-08-19 | 1999-04-20 | At&T Corp | Automated natural language understanding customer service system |
US5918222A (en) * | 1995-03-17 | 1999-06-29 | Kabushiki Kaisha Toshiba | Information disclosing apparatus and multi-modal information input/output system |
US5960394A (en) * | 1992-11-13 | 1999-09-28 | Dragon Systems, Inc. | Method of speech command recognition with dynamic assignment of probabilities according to the state of the controlled applications |
US5987404A (en) * | 1996-01-29 | 1999-11-16 | International Business Machines Corporation | Statistical natural language understanding using hidden clumpings |
US5991712A (en) * | 1996-12-05 | 1999-11-23 | Sun Microsystems, Inc. | Method, apparatus, and product for automatic generation of lexical features for speech recognition systems |
US6029123A (en) * | 1994-12-13 | 2000-02-22 | Canon Kabushiki Kaisha | Natural language processing system and method for expecting natural language information to be processed and for executing the processing based on the expected information |
US6035338A (en) * | 1996-09-19 | 2000-03-07 | Kabushiki Kaisha Toshiba | Document browse support system and document processing system |
US6038527A (en) * | 1995-07-19 | 2000-03-14 | Daimler Benz Ag | Method for generating descriptors for the classification of texts |
US6092034A (en) * | 1998-07-27 | 2000-07-18 | International Business Machines Corporation | Statistical translation system and method for fast sense disambiguation and translation of large corpora using fertility models and sense models |
US6138085A (en) * | 1997-07-31 | 2000-10-24 | Microsoft Corporation | Inferring semantic relations |
US6161083A (en) * | 1996-05-02 | 2000-12-12 | Sony Corporation | Example-based translation method and system which calculates word similarity degrees, a priori probability, and transformation probability to determine the best example for translation |
US6167369A (en) * | 1998-12-23 | 2000-12-26 | Xerox Company | Automatic language identification using both N-gram and word information |
US20010009009A1 (en) * | 1999-12-28 | 2001-07-19 | Matsushita Electric Industrial Co., Ltd. | Character string dividing or separating method and related system for segmenting agglutinative text or document into words |
US6311152B1 (en) * | 1999-04-08 | 2001-10-30 | Kent Ridge Digital Labs | System for chinese tokenization and named entity recognition |
US6675159B1 (en) * | 2000-07-27 | 2004-01-06 | Science Applic Int Corp | Concept-based search and retrieval system |
US6839671B2 (en) * | 1999-12-20 | 2005-01-04 | British Telecommunications Public Limited Company | Learning of dialogue states and language model of spoken information system |
US6928407B2 (en) * | 2002-03-29 | 2005-08-09 | International Business Machines Corporation | System and method for the automatic discovery of salient segments in speech transcripts |
US6938024B1 (en) * | 2000-05-04 | 2005-08-30 | Microsoft Corporation | Transmitting information given constrained resources |
Family Cites Families (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5228491A (en) * | 1992-04-03 | 1993-07-20 | General Clutch Corporation | Monocontrol venetian blind |
US5642519A (en) * | 1994-04-29 | 1997-06-24 | Sun Microsystems, Inc. | Speech interpreter with a unified grammer compiler |
US5887120A (en) * | 1995-05-31 | 1999-03-23 | Oracle Corporation | Method and apparatus for determining theme for discourse |
US6178150B1 (en) | 1996-07-30 | 2001-01-23 | Seagate Technology Inc. | Offset optics for use with optical heads |
US5895456A (en) * | 1996-09-16 | 1999-04-20 | Inflight Ati, Inc. | Onboard currency and value card exchanger |
US6094635A (en) | 1997-09-17 | 2000-07-25 | Unisys Corporation | System and method for speech enabled application |
EP0962873A1 (en) | 1998-06-02 | 1999-12-08 | International Business Machines Corporation | Processing of textual information and automated apprehension of information |
US6421655B1 (en) | 1999-06-04 | 2002-07-16 | Microsoft Corporation | Computer-based representations and reasoning methods for engaging users in goal-oriented conversations |
US7181438B1 (en) * | 1999-07-21 | 2007-02-20 | Alberti Anemometer, Llc | Database access system |
US7158935B1 (en) * | 2000-11-15 | 2007-01-02 | At&T Corp. | Method and system for predicting problematic situations in a automated dialog |
WO2002073451A2 (en) * | 2001-03-13 | 2002-09-19 | Intelligate Ltd. | Dynamic natural language understanding |
-
2002
- 2002-03-13 WO PCT/IL2002/000202 patent/WO2002073451A2/en not_active Application Discontinuation
- 2002-03-13 US US10/097,537 patent/US7216073B2/en not_active Expired - Fee Related
- 2002-03-13 AU AU2002237495A patent/AU2002237495A1/en not_active Abandoned
- 2002-03-13 EP EP02703828A patent/EP1490790A2/en not_active Withdrawn
-
2006
- 2006-11-21 US US11/562,132 patent/US20070112555A1/en not_active Abandoned
- 2006-11-21 US US11/562,142 patent/US7840400B2/en not_active Expired - Fee Related
-
2008
- 2008-03-12 US US12/047,003 patent/US20080154581A1/en not_active Abandoned
Patent Citations (36)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US4706212A (en) * | 1971-08-31 | 1987-11-10 | Toma Peter P | Method using a programmed digital computer system for translation between natural languages |
US5128865A (en) * | 1989-03-10 | 1992-07-07 | Bso/Buro Voor Systeemontwikkeling B.V. | Method for determining the semantic relatedness of lexical items in a text |
US5299125A (en) * | 1990-08-09 | 1994-03-29 | Semantic Compaction Systems | Natural language processing system and method for parsing a plurality of input symbol sequences into syntactically or pragmatically correct word messages |
US5321608A (en) * | 1990-11-30 | 1994-06-14 | Hitachi, Ltd. | Method and system for processing natural language |
US5541836A (en) * | 1991-12-30 | 1996-07-30 | At&T Corp. | Word disambiguation apparatus and methods |
US5528491A (en) * | 1992-08-31 | 1996-06-18 | Language Engineering Corporation | Apparatus and method for automated natural language translation |
US5677993A (en) * | 1992-08-31 | 1997-10-14 | Hitachi, Ltd. | Information processing apparatus using pointing input and speech input |
US5960394A (en) * | 1992-11-13 | 1999-09-28 | Dragon Systems, Inc. | Method of speech command recognition with dynamic assignment of probabilities according to the state of the controlled applications |
US5694592A (en) * | 1993-11-05 | 1997-12-02 | University Of Central Florida | Process for determination of text relevancy |
US5576954A (en) * | 1993-11-05 | 1996-11-19 | University Of Central Florida | Process for determination of text relevancy |
US5844302A (en) * | 1993-12-22 | 1998-12-01 | Siemens Aktiengesellschaft | Integrated semiconductor circuit with capacitors of precisely defined capacitance and process for producing the circuit |
US5634024A (en) * | 1994-05-18 | 1997-05-27 | Nec Corporation | Definition execution system for operations and functions in computer system |
US5642502A (en) * | 1994-12-06 | 1997-06-24 | University Of Central Florida | Method and system for searching for relevant documents from a text database collection, using statistical ranking, relevancy feedback and small pieces of text |
US6088692A (en) * | 1994-12-06 | 2000-07-11 | University Of Central Florida | Natural language method and system for searching for and ranking relevant documents from a computer database |
US5893092A (en) * | 1994-12-06 | 1999-04-06 | University Of Central Florida | Relevancy ranking using statistical ranking, semantics, relevancy feedback and small pieces of text |
US5748974A (en) * | 1994-12-13 | 1998-05-05 | International Business Machines Corporation | Multimodal natural language interface for cross-application tasks |
US6029123A (en) * | 1994-12-13 | 2000-02-22 | Canon Kabushiki Kaisha | Natural language processing system and method for expecting natural language information to be processed and for executing the processing based on the expected information |
US5918222A (en) * | 1995-03-17 | 1999-06-29 | Kabushiki Kaisha Toshiba | Information disclosing apparatus and multi-modal information input/output system |
US5675710A (en) * | 1995-06-07 | 1997-10-07 | Lucent Technologies, Inc. | Method and apparatus for training a text classifier |
US5680511A (en) * | 1995-06-07 | 1997-10-21 | Dragon Systems, Inc. | Systems and methods for word recognition |
US6038527A (en) * | 1995-07-19 | 2000-03-14 | Daimler Benz Ag | Method for generating descriptors for the classification of texts |
US5987404A (en) * | 1996-01-29 | 1999-11-16 | International Business Machines Corporation | Statistical natural language understanding using hidden clumpings |
US5805775A (en) * | 1996-02-02 | 1998-09-08 | Digital Equipment Corporation | Application user interface |
US6161083A (en) * | 1996-05-02 | 2000-12-12 | Sony Corporation | Example-based translation method and system which calculates word similarity degrees, a priori probability, and transformation probability to determine the best example for translation |
US6035338A (en) * | 1996-09-19 | 2000-03-07 | Kabushiki Kaisha Toshiba | Document browse support system and document processing system |
US5991712A (en) * | 1996-12-05 | 1999-11-23 | Sun Microsystems, Inc. | Method, apparatus, and product for automatic generation of lexical features for speech recognition systems |
US6138085A (en) * | 1997-07-31 | 2000-10-24 | Microsoft Corporation | Inferring semantic relations |
US5895466A (en) * | 1997-08-19 | 1999-04-20 | At&T Corp | Automated natural language understanding customer service system |
US6092034A (en) * | 1998-07-27 | 2000-07-18 | International Business Machines Corporation | Statistical translation system and method for fast sense disambiguation and translation of large corpora using fertility models and sense models |
US6167369A (en) * | 1998-12-23 | 2000-12-26 | Xerox Company | Automatic language identification using both N-gram and word information |
US6311152B1 (en) * | 1999-04-08 | 2001-10-30 | Kent Ridge Digital Labs | System for chinese tokenization and named entity recognition |
US6839671B2 (en) * | 1999-12-20 | 2005-01-04 | British Telecommunications Public Limited Company | Learning of dialogue states and language model of spoken information system |
US20010009009A1 (en) * | 1999-12-28 | 2001-07-19 | Matsushita Electric Industrial Co., Ltd. | Character string dividing or separating method and related system for segmenting agglutinative text or document into words |
US6938024B1 (en) * | 2000-05-04 | 2005-08-30 | Microsoft Corporation | Transmitting information given constrained resources |
US6675159B1 (en) * | 2000-07-27 | 2004-01-06 | Science Applic Int Corp | Concept-based search and retrieval system |
US6928407B2 (en) * | 2002-03-29 | 2005-08-09 | International Business Machines Corporation | System and method for the automatic discovery of salient segments in speech transcripts |
Cited By (68)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8214196B2 (en) | 2001-07-03 | 2012-07-03 | University Of Southern California | Syntax-based statistical translation model |
US20030023423A1 (en) * | 2001-07-03 | 2003-01-30 | Kenji Yamada | Syntax-based statistical translation model |
US8234106B2 (en) | 2002-03-26 | 2012-07-31 | University Of Southern California | Building a translation lexicon from comparable, non-parallel corpora |
US8548794B2 (en) | 2003-07-02 | 2013-10-01 | University Of Southern California | Statistical noun phrase translation |
US20050228643A1 (en) * | 2004-03-23 | 2005-10-13 | Munteanu Dragos S | Discovery of parallel text portions in comparable collections of corpora and training using comparable texts |
US8296127B2 (en) | 2004-03-23 | 2012-10-23 | University Of Southern California | Discovery of parallel text portions in comparable collections of corpora and training using comparable texts |
US8977536B2 (en) | 2004-04-16 | 2015-03-10 | University Of Southern California | Method and system for translating information with a higher probability of a correct translation |
US20060142995A1 (en) * | 2004-10-12 | 2006-06-29 | Kevin Knight | Training for a text-to-text application which uses string to tree conversion for training and decoding |
US8600728B2 (en) * | 2004-10-12 | 2013-12-03 | University Of Southern California | Training for a text-to-text application which uses string to tree conversion for training and decoding |
US8886517B2 (en) | 2005-06-17 | 2014-11-11 | Language Weaver, Inc. | Trust scoring for language translation systems |
US10319252B2 (en) | 2005-11-09 | 2019-06-11 | Sdl Inc. | Language capability assessment and training apparatus and techniques |
US8943080B2 (en) | 2006-04-07 | 2015-01-27 | University Of Southern California | Systems and methods for identifying parallel documents and sentence fragments in multilingual document collections |
US8886518B1 (en) | 2006-08-07 | 2014-11-11 | Language Weaver, Inc. | System and method for capitalizing machine translated text |
US10510341B1 (en) | 2006-10-16 | 2019-12-17 | Vb Assets, Llc | System and method for a cooperative conversational voice user interface |
US10515628B2 (en) | 2006-10-16 | 2019-12-24 | Vb Assets, Llc | System and method for a cooperative conversational voice user interface |
US10755699B2 (en) | 2006-10-16 | 2020-08-25 | Vb Assets, Llc | System and method for a cooperative conversational voice user interface |
US11222626B2 (en) | 2006-10-16 | 2022-01-11 | Vb Assets, Llc | System and method for a cooperative conversational voice user interface |
US10297249B2 (en) | 2006-10-16 | 2019-05-21 | Vb Assets, Llc | System and method for a cooperative conversational voice user interface |
US8433556B2 (en) | 2006-11-02 | 2013-04-30 | University Of Southern California | Semi-supervised training for statistical word alignment |
US20080109209A1 (en) * | 2006-11-02 | 2008-05-08 | University Of Southern California | Semi-supervised training for statistical word alignment |
US9122674B1 (en) | 2006-12-15 | 2015-09-01 | Language Weaver, Inc. | Use of annotations in statistical machine translation |
US8468149B1 (en) | 2007-01-26 | 2013-06-18 | Language Weaver, Inc. | Multi-lingual online community |
US11080758B2 (en) | 2007-02-06 | 2021-08-03 | Vb Assets, Llc | System and method for delivering targeted advertisements and/or providing natural language processing based on advertisements |
US10134060B2 (en) | 2007-02-06 | 2018-11-20 | Vb Assets, Llc | System and method for delivering targeted advertisements and/or providing natural language processing based on advertisements |
US9406078B2 (en) | 2007-02-06 | 2016-08-02 | Voicebox Technologies Corporation | System and method for delivering targeted advertisements and/or providing natural language processing based on advertisements |
US8615389B1 (en) | 2007-03-16 | 2013-12-24 | Language Weaver, Inc. | Generation and exploitation of an approximate language model |
US8831928B2 (en) | 2007-04-04 | 2014-09-09 | Language Weaver, Inc. | Customizable machine translation service |
US8825466B1 (en) | 2007-06-08 | 2014-09-02 | Language Weaver, Inc. | Modification of annotated bilingual segment pairs in syntax-based machine translation |
US10347248B2 (en) | 2007-12-11 | 2019-07-09 | Voicebox Technologies Corporation | System and method for providing in-vehicle services via a natural language voice user interface |
US9620113B2 (en) | 2007-12-11 | 2017-04-11 | Voicebox Technologies Corporation | System and method for providing a natural language voice user interface |
US9305548B2 (en) | 2008-05-27 | 2016-04-05 | Voicebox Technologies Corporation | System and method for an integrated, multi-modal, multi-device natural language voice services environment |
US10553216B2 (en) | 2008-05-27 | 2020-02-04 | Oracle International Corporation | System and method for an integrated, multi-modal, multi-device natural language voice services environment |
US9711143B2 (en) | 2008-05-27 | 2017-07-18 | Voicebox Technologies Corporation | System and method for an integrated, multi-modal, multi-device natural language voice services environment |
US10089984B2 (en) | 2008-05-27 | 2018-10-02 | Vb Assets, Llc | System and method for an integrated, multi-modal, multi-device natural language voice services environment |
US9411800B2 (en) * | 2008-06-27 | 2016-08-09 | Microsoft Technology Licensing, Llc | Adaptive generation of out-of-dictionary personalized long words |
US20090326927A1 (en) * | 2008-06-27 | 2009-12-31 | Microsoft Corporation | Adaptive generation of out-of-dictionary personalized long words |
US20100017293A1 (en) * | 2008-07-17 | 2010-01-21 | Language Weaver, Inc. | System, method, and computer program for providing multilingual text advertisments |
US10553213B2 (en) | 2009-02-20 | 2020-02-04 | Oracle International Corporation | System and method for processing multi-modal device interactions in a natural language voice services environment |
US9953649B2 (en) | 2009-02-20 | 2018-04-24 | Voicebox Technologies Corporation | System and method for processing multi-modal device interactions in a natural language voice services environment |
US9570070B2 (en) | 2009-02-20 | 2017-02-14 | Voicebox Technologies Corporation | System and method for processing multi-modal device interactions in a natural language voice services environment |
US8990064B2 (en) | 2009-07-28 | 2015-03-24 | Language Weaver, Inc. | Translating documents based on content |
US8676563B2 (en) | 2009-10-01 | 2014-03-18 | Language Weaver, Inc. | Providing human-generated and machine-generated trusted translations |
US8380486B2 (en) | 2009-10-01 | 2013-02-19 | Language Weaver, Inc. | Providing machine-generated translations and corresponding trust levels |
US10417646B2 (en) | 2010-03-09 | 2019-09-17 | Sdl Inc. | Predicting the cost associated with translating textual content |
US10984429B2 (en) | 2010-03-09 | 2021-04-20 | Sdl Inc. | Systems and methods for translating textual content |
US11003838B2 (en) | 2011-04-18 | 2021-05-11 | Sdl Inc. | Systems and methods for monitoring post translation editing |
US8694303B2 (en) | 2011-06-15 | 2014-04-08 | Language Weaver, Inc. | Systems and methods for tuning parameters in statistical machine translation |
US20130054226A1 (en) * | 2011-08-31 | 2013-02-28 | International Business Machines Corporation | Recognizing chemical names in a chinese document |
US9575957B2 (en) * | 2011-08-31 | 2017-02-21 | International Business Machines Corporation | Recognizing chemical names in a chinese document |
US8886515B2 (en) | 2011-10-19 | 2014-11-11 | Language Weaver, Inc. | Systems and methods for enhancing machine translation post edit review processes |
US8942973B2 (en) | 2012-03-09 | 2015-01-27 | Language Weaver, Inc. | Content page URL translation |
US10261994B2 (en) | 2012-05-25 | 2019-04-16 | Sdl Inc. | Method and system for automatic management of reputation of translators |
US10402498B2 (en) | 2012-05-25 | 2019-09-03 | Sdl Inc. | Method and system for automatic management of reputation of translators |
US9152622B2 (en) | 2012-11-26 | 2015-10-06 | Language Weaver, Inc. | Personalized machine translation via online adaptation |
US9213694B2 (en) | 2013-10-10 | 2015-12-15 | Language Weaver, Inc. | Efficient online domain adaptation |
US9898459B2 (en) | 2014-09-16 | 2018-02-20 | Voicebox Technologies Corporation | Integration of domain information into state transitions of a finite state transducer for natural language processing |
US10430863B2 (en) | 2014-09-16 | 2019-10-01 | Vb Assets, Llc | Voice commerce |
WO2016044321A1 (en) * | 2014-09-16 | 2016-03-24 | Min Tang | Integration of domain information into state transitions of a finite state transducer for natural language processing |
US10216725B2 (en) | 2014-09-16 | 2019-02-26 | Voicebox Technologies Corporation | Integration of domain information into state transitions of a finite state transducer for natural language processing |
US9626703B2 (en) | 2014-09-16 | 2017-04-18 | Voicebox Technologies Corporation | Voice commerce |
US11087385B2 (en) | 2014-09-16 | 2021-08-10 | Vb Assets, Llc | Voice commerce |
US10229673B2 (en) | 2014-10-15 | 2019-03-12 | Voicebox Technologies Corporation | System and method for providing follow-up responses to prior natural language inputs of a user |
US9747896B2 (en) | 2014-10-15 | 2017-08-29 | Voicebox Technologies Corporation | System and method for providing follow-up responses to prior natural language inputs of a user |
US10503828B2 (en) * | 2014-11-19 | 2019-12-10 | Electronics And Telecommunications Research Institute | System and method for answering natural language question |
US20160140187A1 (en) * | 2014-11-19 | 2016-05-19 | Electronics And Telecommunications Research Institute | System and method for answering natural language question |
US10431214B2 (en) | 2014-11-26 | 2019-10-01 | Voicebox Technologies Corporation | System and method of determining a domain and/or an action related to a natural language input |
US10614799B2 (en) | 2014-11-26 | 2020-04-07 | Voicebox Technologies Corporation | System and method of providing intent predictions for an utterance prior to a system detection of an end of the utterance |
US10331784B2 (en) | 2016-07-29 | 2019-06-25 | Voicebox Technologies Corporation | System and method of disambiguating natural language processing requests |
Also Published As
Publication number | Publication date |
---|---|
WO2002073451A3 (en) | 2004-03-11 |
EP1490790A2 (en) | 2004-12-29 |
US20080154581A1 (en) | 2008-06-26 |
US20070112556A1 (en) | 2007-05-17 |
AU2002237495A1 (en) | 2002-09-24 |
WO2002073451A2 (en) | 2002-09-19 |
US20020196679A1 (en) | 2002-12-26 |
US7840400B2 (en) | 2010-11-23 |
US7216073B2 (en) | 2007-05-08 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US7840400B2 (en) | Dynamic natural language understanding | |
US11086601B2 (en) | Methods, systems, and computer program product for automatic generation of software application code | |
US11250842B2 (en) | Multi-dimensional parsing method and system for natural language processing | |
US9824083B2 (en) | System for natural language understanding | |
US7174507B2 (en) | System method and computer program product for obtaining structured data from text | |
US9710458B2 (en) | System for natural language understanding | |
US7822699B2 (en) | Adaptive semantic reasoning engine | |
US10467122B1 (en) | Methods, systems, and computer program product for capturing and classification of real-time data and performing post-classification tasks | |
EP2317507B1 (en) | Corpus compilation for language model generation | |
US7925506B2 (en) | Speech recognition accuracy via concept to keyword mapping | |
CN1871597B (en) | System and method for associating documents with contextual advertisements | |
US10503769B2 (en) | System for natural language understanding | |
US11907665B2 (en) | Method and system for processing user inputs using natural language processing | |
López-Cózar et al. | Using knowledge of misunderstandings to increase the robustness of spoken dialogue systems | |
Stoness et al. | Incremental parsing with reference interaction | |
Rocha | Supporting anaphor resolution in dialogues with a corpus-based probabilistic model | |
Yoon et al. | Building Korean linguistic resource for NLU data generation of banking app CS dialog system | |
JP3506058B2 (en) | Semantic analysis method and apparatus, and recording medium storing semantic analysis program | |
Gruenstein | Response-based confidence annotation for spoken dialogue systems | |
Li et al. | PRIS at TAC 2009: Experiments in KBP Track. | |
Nagaraj et al. | Automatic Correction of Text Using Probabilistic Error Approach | |
Ahmed | Detection of foreign words and names in written text | |
CN117076618A (en) | Intelligent question-answering method and device | |
Khosravi-Bardsirpour | Extracting pragmatic content from Email. | |
Ocelikova et al. | Processing of Anaphoric and Elliptic Sentences in a Spoken Dialog System |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
STCB | Information on status: application discontinuation |
Free format text: ABANDONED -- FAILURE TO RESPOND TO AN OFFICE ACTION |