RU2480822C2 - Coreference resolution in ambiguity-sensitive natural language processing system - Google Patents

Abstract

FIELD: information technology.

SUBSTANCE: in the method of integrating coreference resolution mechanisms, a portion of text is retrieved using the natural language mechanism of a server computer. Coreference within said portion of text is identified using the natural language mechanism of the server computer. A fact is retrieved from said portion of text using the natural language mechanism of the server computer, wherein the fact has a value. Said fact is expanded using the natural language mechanism of the server computer so that it includes a coreference value different from said value and based on the identified coreference.

EFFECT: improved indexing of documents in natural language.

20 cl, 5 dwg

Description

FIELD OF THE INVENTION

The present invention relates to methods for resolving coreference in an ambiguity sensitive natural language processing system, in particular to integrating the functionality of resolving coreference into a document processing system.

BACKGROUND

In a natural language, it is not uncommon for an object to be referenced using different descriptions. For example, pronouns are usually used instead of nouns. In addition, various other descriptions or various forms of reference may be used to refer to an object. Consider the following parts of text as an example:

"Pablo Picasso was born in Malaga."

"The Spanish painter has become famous thanks to his various styles."

“Among his paintings is the large-scale“ Guernica. ”

"He painted this exciting masterpiece during the Spanish Civil War."

"Picasso died in 1973."

Here we come across a sequence of linguistic change. For example, two different names are used, Pablo Picasso and Picasso. The defining description, the “Spanish painter” and the two pronouns “him” and “he”, are all used to refer to Picasso. Two different expressions are used to refer to the picture: the title of the work, "Guernica", and the indicative description, "this exciting masterpiece."

Of the two linguistic expressions, we can say that they are coreferent if they have the same referent. In other words, if they refer to the same object. The second phrase may be anaphora, which is anaphoric to the first phrase. Also, the first phrase is the antecedent of the second phrase. Knowing the referent of the antecedent may be necessary to determine the referent of the anaphora. The general task of finding coreferential expressions, anaphores and their antecedents within a document may be referred to as resolution of coreference. Resolving coreference is the process of establishing that two expressions refer to the same referent without the need to establish what this referent is. Referencing is the process of establishing what the referent is.

For groups of expressions that are referential, regardless of their anaphoric relationship, expressions may be referred to as alternate names for each other. In accordance with the above example, the expressions "Pablo Picasso", "Spanish painter" "him", "he" and "Picasso" form a group of alternative names related to Picasso.

Natural language expressions often contain ambiguity (ambiguity). Ambiguity occurs when an expression can be interpreted with more than one value. For example, the sentence “Duck is ready to eat” can be interpreted as saying that the duck is either properly cooked or that the duck is hungry and needs food.

Resolving coreference and resolving ambiguity are two examples of natural language processing operations that can be used to mechanically support a language, as is commonly used by human users. Information processing systems, such as text indexing and requesting information search support, can benefit from the increased use of natural language processing systems.

The disclosure made here relates to these and other considerations.

SUMMARY OF THE INVENTION

Here, technologies for resolving coreference in an ambiguity-sensitive natural language processing system are described. In particular, technologies are described for integrating the functionality of resolving coreference into a document processing system that should be indexed in a search and retrieval system. This integration can improve indexing with information that supports corereference resolution and ambiguous value within natural language documents.

In accordance with one aspect presented here, information provided by the core resolution system can be integrated into a natural language processing system and increase its productivity. An example of such a system is a document indexing and retrieval system.

In accordance with another aspect presented herein, ambiguity awareness features, as well as ambiguity resolution functionality, may work in coordination with coreference resolution within a natural language processing system. Annotation of coreference objects, as well as ambiguous interpretations, can be supported by built-in markup inside text expressions or, alternatively, by external representations of objects.

In accordance with another aspect presented here, the facts can be extracted from the text to be indexed. Information expressed within the text can be formally organized in terms of facts. The fact used in this sense can be any information contained in the text, and does not have to be true. A fact can be represented as relations between objects. A fact can be stored in a semantic index as a relationship between objects stored inside an index. In a fact-based retrieval system, a document can be retrieved if it contains a fact that matches the fact determined through analysis of the request.

In accordance with yet another aspect presented here, the expansion process may support the application of numerous alternative names or ambiguities to the indexed object. Such an extension may support additional possible references or interpretations for a given object, fixed in a semantic index. Alternative stored descriptions may support the extraction of fact through either the original description or the core description.

It should be understood that the above discussion subject can also be implemented as a computer-controlled device, a computer process, a computing system, or as a manufacturing product, such as a computer-readable medium. These and various other features should become apparent from reading the following detailed description and consideration of the accompanying drawings.

This section of the "Summary of the invention" is intended to introduce selective concepts in a simplified form, which are further described below in the detailed description. The present "Summary of the invention" is not intended to identify key features or essential features of the claimed subject matter, nor to use this section of the "Summary of the invention" to limit the scope of the claimed subject matter. Additionally, the claimed subject matter is not limited to embodiments that eliminate any or all of the disadvantages noted in any part of this disclosure.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a diagram of a network architecture showing an information retrieval system in accordance with aspects of an embodiment presented herein;

FIG. 2 is a functional block diagram showing various components of a natural language text index and query system corresponding to aspects of the embodiment presented here;

FIG. 3 is a functional block diagram showing resolution of coreference and resolution of ambiguity within a natural language processing system in accordance with aspects of the embodiment presented herein;

FIG. 4 is a flowchart showing aspects of processes for ambiguity-sensitive indexing with corereference resolution in accordance with aspects of an embodiment presented herein; and

FIG. 5 is a computer architecture diagram showing an example of computer architecture of hardware and software for a computing system capable of implementing aspects of the embodiment presented here.

DETAILED DESCRIPTION

The following detailed detailed description relates to technologies for resolving coreference in an ambiguous-sensitive natural language processing system. Using the technologies and concepts presented here, the core resolution resolution functionality can be integrated into a natural language processing system that processes documents that must be indexed for use in the search and retrieval system. This integration can improve indexing with information that supports corereference for indexed natural language documents.

Although the subject matter described herein is presented in the general context of software modules that run in conjunction with running an operating system and application programs on a computer system, those skilled in the art will recognize that other embodiments may be implemented in combination with other types of software modules. In general, program modules contain routines, programs, components, data structures, and other types of structures that perform specific tasks or implement specific abstract data types. In addition, those skilled in the art should understand that the subject matter of the invention described herein can be implemented with other computer system configurations, including handheld devices, multiprocessor systems, programmable consumer electronics or based on microprocessors, minicomputers, general purpose computers, etc. .

In the following detailed description, reference is made to the accompanying drawings, which are part of the present description and show explanatory specific embodiments or examples. Now, with reference to the drawings, in which similar positions represent similar elements in different drawings, aspects of a computing system and a methodology for resolving coreference in an ambiguous-sensitive natural language processing system are described.

In FIG. 1, details will be provided regarding an example work environment for the presented embodiments. In particular, the network architecture circuit 100 shows an information retrieval system in accordance with aspects of an embodiment presented herein. Client computers 110A-110D may communicate through a network 140 with a server to obtain information related to natural language engine 130. Although four client computers 110A-110D are shown, it should be understood that any number of client computers 110A-110D can be used. Client computers 110A-110D may be geographically distributed over network 140, located in the neighborhood, or in any combination thereof. Although a single server 120 is shown in the drawing, it should be understood that the functionality of server 120 can be distributed to any number of multiple servers 120. Such multiple servers 120 can be located nearby, geographically distributed over network 140, or in any combination.

In accordance with one or more embodiments, the natural language engine 130 may support the functionality of the search engine. In a search engine scenario, a user request may be issued from the client computer 110A-110D via the network 140 and to the server 120. The user request may be in natural language format. On the server, the natural language engine 130 may process the natural language query to support a search based on the syntax and semantics extracted from the natural language query. The results of such a search can be provided from the server 120 through the network 140 back to the client computers 110A-110D.

One or more search indexes may be stored or associated with server 120. The information contained in the search index may be populated from a set of source information or from a collection. For example, when performing a web search, content may be collected and indexed from various websites on various web servers (not shown) over the network 140. Such collection and indexing may be performed by executing software on server 120 or on another computer (not shown). Collection can be done using applications, web crawlers, or spiders. The natural language engine 130 can be applied to the collected information in such a way that the natural language content collected from the totality can be indexed based on the syntax and semantics extracted by the natural language engine 130. Indexing and searching are discussed in more detail with reference to FIG. 2.

Client computers 110A-110D for server 120 may act as end clients, hypertext browser clients, graphic display clients, or other network clients. For example, a web browser application on client computers 110A-110D may support interaction with a web server application on server 120. To support interaction with server 120, such a browser may use controls, plugins, or applets. Client computers 110A-110D may also use other custom programs, applications, or modules to interact with server 120. Client computers 110A-110D may be desktop computers, laptops, PDAs, mobile terminals, mobile phones, television set-top boxes, kiosks, servers, terminals, thin clients (client terminals) or any other computerized devices.

Network 140 may be any communication network capable of communicating between client computers 110A-110D and server 120. Network 140 may be a wired, wireless, optical, radio network, packet switched, circuit switched, or any combination thereof. Network 140 can use any topology and communication lines 140 can support any network technology, protocol, or bandwidth such as Ethernet, DSL, cable modem, ATM, SONET, MPLS, PSTN, POTS, PONS, HFC, satellite, ISDN, WiFi WiMax, any combination of these, or any other data connection mechanism or network mechanism. Network 140 may be an intranet, Internet, Internet, web, LAN, WAN, MAN, or any other network for interconnecting computer systems.

It should be understood that in addition to the network environment shown, the natural language engine 130 can be controlled locally. For example, server 120 and client computer 110A-110D may be combined on a single computing device. Such a combined system may support search indexes stored locally or remotely.

In FIG. 2 is a functional block diagram of various components of a natural language mechanism 130 in accordance with one example embodiment. As discussed above, natural language engine 130 may support information searches. In order to support such searches, a content acquisition process 200 is performed. Operations associated with obtaining 200 content extract information from documents provided as text content 210. This information can be stored in semantic index 250, which can be used for search. Operations associated with a user search 205 may support the processing of a user-entered search request. A user request may take the form of a natural language question 260. The natural language engine 130 may analyze input from the user to translate the request into a presentation that must be compared with the information provided within the semantic index 250. The content and structuring of information in the semantic index 250 can support the quick matching and retrieval of documents or parts of documents that are relevant to the meaning of a query or question 260 in a natural language.

Text content 210 may contain documents in a very general sense. Examples of such documents can be web pages, text documents, scanned documents, databases, printouts of information, other Internet content or any other information source. This text content 210 may provide a collection of information that must be found. Processing of text content 210 can occur in two stages as parsing 215 and semantic display 225. Preliminary stages of language processing can be performed before or at the beginning of parsing 215. For example, text content 210 can be divided at the boundaries of sentences. Proper names can be identified as the names of specific people, places, objects or events. In addition, grammatical properties of meaningful verbal endings can be defined. For example, in English, a noun ending with the letter "s" is likely to be a plural noun, although a verb ending in "s" may be a third-person verb in the singular.

Parsing 215 can be performed using a parsing system such as Xerox Linguistic Environment (XLE), presented here only as a General example, but not limiting the possible options for implementing the present description. The parser 215 can transform sentences into representations that create clear syntactic relationships between words. The parser 215 may use the grammar 220 associated with the particular language used. For example, parser 215 may apply English grammar 220. Grammar 220 can be formalized, for example, as a lexical functional grammar (LFG) or other suitable parsing mechanism, such as those based on the grammar of direct components, controlled by headers (HPSG), combinatorial categorical grammar (CCG), probabilistic context-free grammar (PCFG) or any other formal grammar. Grammar 220 can identify possible ways of constructing meaningful sentences in a given language. The parser 215 may apply grammar rules 220 to lines of text content 210.

Grammar 220 may be provided for various languages. For example, LFG grammars were created for English, French, German, Chinese, and Japanese. Other grammars may also be provided. Grammar 220 can be developed by manual production, in which grammar rules are determined by a linguist or dictionary author. Alternatively, acquiring with a self-learning device may include automated observation and analysis of many examples of text from a large population to automatically determine grammar rules. Also, when obtaining grammar rules 220, a combination of manual determination and determination using a device with self-learning tools can be used.

The parser 215 may apply the grammar 220 to the text content 210 to determine the syntax structure. In the case of LFG-based parsing, the syntactic structures consist of constituent structures (c-structure) and functional structures (f-structure). A C-structure may represent a hierarchy of constituent phrases and words. An F structure can encode roles and relationships between various components of a c structure. An F-structure may also represent information derived from word forms. For example, in the f-structure, the plural of a noun or the verb tense can be defined.

During the semantic mapping process 225, which follows the parsing process 215, the information can be extracted from the syntax structures and combined with the word meaning information in the sentence. A semantic representation or semantic representation of a sentence can be provided as semantics 240 of the content. Semantic mapping 225 can improve the syntactic relationships provided by the parser 215 with the conceptual properties of individual words. The results can be converted into representations of the meaning of sentences from text content 210. Semantic display 225 may determine the roles played by words in a sentence. For example, a subject performing an action, something used to perform an action, or something affected by an action. For search indexing purposes, words can be stored in semantic index 250 along with their roles. Thus, extraction from the semantic index 250 may depend not only on a separately selected word, but also on the meaning of the word in sentences in which it appears within the textual content 210. Semantic display 225 can help to eliminate the ambiguity of terms, determine antecedent relations and expand terms through synonyms, hyperonyms or hyponyms.

Semantic mapping 225 may apply knowledge resources 230 as rules and techniques for extracting semantics from sentences. Knowledge resources can be obtained either through manual definition or through a device with self-learning tools, as discussed in relation to obtaining grammars 220. The process of semantic display 225 can provide semantics 240 of content in a semantic extensible markup language representation (semantic XML or semxml). Any suitable presentation language, such as expressions written in PROLOG, LISP, JSON, YAML or others, can also be used. Content semantics 240 may indicate the roles played by words in text content sentences 210. Content semantics 240 may be provided to indexing process 245.

An index can support the presentation of a large body of information so that the locations of words and phrases can be quickly identified within the index. The traditional search engine can use keywords as search terms, so that the index is displayed from keywords defined by the user in articles or documents where such keywords appear. The semantic index 250 may represent semantic meanings of words in addition to the words themselves. Semantic relationships can be assigned to words during acquisition of content 200, as well as during a search of 205 users. Requests for semantic index 250 can be based not only on words, but also on words in specific roles. Roles are what the word represents in a sentence or phrase, as it is stored in semantic index 250. Semantic index 250 can be considered as an inverted index, which is a quickly searchable database whose objects are semantic words (that is, words in a given roles) using pointers to documents, or web pages on which these words appear. Semantic Index 250 may support hybrid indexing. Such hybrid indexing can combine the features and functions of both keyword indexing and semantic indexing.

User input can be supported in the form of questions 260 in the natural language. The request can be analyzed through a source of information in a natural language, similar or identical to that used to obtain 200 content. That is, the natural language question 260 can be processed by the parser 265 to extract the syntax structure. After parsing 265, the natural language question 260 can be processed for semantic mapping 270. Semantic mapping 270 can provide question semantics 275, which should be used in the extraction process 280 with respect to semantic index 250, as discussed above. The extraction process 280 may support hybrid index queries, where both keyword index extraction and semantic index extraction can be provided separately or in combination.

In response to a user request, the results of extracting 280 from the semantic index 250 along with the semantics 275 of the question can inform the ranking process 285. Ranking can reinforce both keyword and semantic information. During ranking 285, the results obtained by extracting 280 can be arranged in order by various indicators in an attempt to bring the most desirable results closer to the top of the extracted information, which should be provided to the user as the resulting view 290.

Turning now to FIG. 3, a functional block diagram of resolving coreference and resolving ambiguity within a natural language processing system 300 is shown in accordance with aspects of an embodiment presented herein. As an example application, the natural language processing system 300 may support an information retrieval mechanism for indexing and retrieving a document. Such a natural language search engine can expand the information stored in its index based on linguistic analysis. The system may also support disclosure of intent within a user's request by analyzing the request linguistically. The indications of corereference permission and ambiguity resolution discussed here may work with respect to parsing 215, semantic mapping 225, and semantic indexing 245, as discussed with reference to FIG. 2. Resolution of the coreference can be performed directly on text content 210 or on usage information from analysis operations 215 or semantic display 225.

As shown in the drawing, the resolution 320, 370 of the reference can be performed directly on the segmented document and also as part of the semantic display 225. These two appearances of the permissions 320, 370 of the reference can be combined or their informational conclusions can be combined. It should be understood that corereference resolution may also occur between parsing 215 and semantic display 225. Corereference resolution may also occur at any other stage within the natural language processing pipeline. There may be one, two or more components of resolving coreference or steps at different positions within a natural language processing system. Text content 210 may be parsed for information intended to be stored in semantic index 250. The search may include a query on semantic index 250 for the desired information.

Content segmentation 310 can be performed on documents with text content 210. Documents can be segmented for more efficient and potentially more accurate resolution of 320 coreference. Reference permission 320 may consider potential reference relationships throughout the document. For long documents, a lot of time can be spent comparing expressions that are distant from each other. When processing speed is taken into account, document content segmentation 310 prior to coreference resolution 320 can significantly reduce the time used for processing. Content segmentation 310 can effectively reduce the amount of content text 210 that is examined when attempting to resolve corereference 320.

Content segmentation 310 may provide information for resolving 370 semantic corrections to indicate when a new document segment begins. Such information may be provided as a segmentation signal 312 or by introducing markup into a segment of the content document. An external file containing meta-information or other mechanisms may also be used.

The document structure can be used to identify segment boundaries that reference relationships are unlikely to cross. The structure of a document can be inferred from explicit markup, such as paragraph boundaries, chapters, or section headings. The structure of the document may also be disclosed through linguistic processing. Segments exceeding a given length can be further divided. The desired length with additional division can be expressed, for example, in terms of the number of sentences or the number of words.

When reliable document structuring is not available, heuristic or statistical criteria may be applied. Such criteria may be indicated in order to tend to maintain corereferences together, while limiting the size of the segment to a predetermined maximum. Various other approaches to segmentation of the text content of 210 documents may also be applied. Content segmentation 310 can also define an entire document as one segment.

The coreference permission 320, 370 can be used to identify coreferences and aliases within the text 210 of the content. For example, when indexing the sentence "He painted" Guernica "", it may be extremely important to decide that "he" refers to Picasso. This is especially true if fact-based extraction is used. Deciding to use a pronoun instead of Picasso may support the indexing of the fact that Picasso painted “Guernica”, instead of the less useful fact that some male man, “he,” painted “Guernica”. Without this ability to identify and index the pronoun referent, it can be difficult, using the fact-based extraction method, to retrieve a document in response to a query "Picasso drew." Recalling the system can improve when a document is returned relating to a request that could not otherwise be returned.

Annotation 330 can be applied to text content 210 to support tracking of objects and possible correlation relationships. Confidence values in permission decisions can also be annotated or labeled within the textual content 210. Permission definitions can be written by adding explicit annotation labels to the text. For example, the text “John visited Mary. He met her in 2003” is set (John visited Mary. He met her in 2003). Annotation 330 can be applied as "[E1: 0.9 John] visited [E2: 0.8 Mary]. [E1: 0.9 He] met [E2: 0.8 her] in 2003", where the words "John" and "He" can be related as one E1 object with a confidence value of 0.9. Similarly, the words "Mary" and "her" can be related as an object two E2 with a confidence value of 0.8. The confidence value may indicate a criterion of trust in the decision of the resolution 320 of the coreference. An annotation can encode coreference decisions directly or an annotation can function as identifiers connecting the corresponding terms in the annotated text with additional information in the offline annotation 325.

Solutions for resolution 320 of the reference can be used as part of the process of constructing the semantic display 225. Reference expressions used by the resolution system 320 of the reference can be integrated into the input representation for the semantic display of 225 in-line annotations within the text content. References can also be provided separately in the external, offline display of the object 325.

Within a set of large documents of textual content 210, such as a web network, the same sentence may appear multiple times in different contexts. These different contexts may provide different candidates for resolution 320 of the coreference. Since parsing 215 can be computationally expensive, it may be useful to store the parsing results for sentences in the cache. Such a caching mechanism 350 may support the quick extraction of parsing information when a sentence occurs in the future.

If corereference permission 320 applies to a single sentence appearing in different contexts, it can identify different corereference relations for the same reference expressions, since corereference may be contextual. Thus, various object identifiers can be inserted inside the text. For example, the text “He is smart” appearing in two different documents can be annotated with two different identifiers, “[E21 He] is smart” and “[E78 He] is smart” when the word “He "in the first document refers to a different person than the word" He "in the second document.

Various sources of information may exist for surface resolution 320 of coreference. For example, in addition to the expression detection performed during corereference permission 320, a system may exist for detecting proper names in text content 210. These various sources may identify conflicting resolution information. For example, conflicting resolution may occur at border crossings. For example, two systems could identify the following conflicting reference expressions:

"[John] told [George Washington] [Irving] was a great writer."

("[John] said that [George Washington] [Irving] was a great writer.")

"[John] told [George] [Washington Irving] was a great writer."

("[John] said that [George] [Washington, Irving] was a great writer.")

Consider the following border crossing conflicts: [George Washington] on the first line conflicts with [George] on the second line. Also [George Washington] in the first line conflicts with [Washington Irving] in the second line. Based on trusting information or contextual factors, various strategies can be applied iteratively to resolve this conflict or to preserve it. In a “reset” strategy, a decision for two or more conflicting boundaries can be made by resetting the one that has the least confidence. In a merge strategy, boundaries can be moved accordingly when two or more borders are equally plausible in compatible contexts. For example, “Mr. [John Smith]” (“[Mr. John] Smith”) and “Mr. [John Smith]” (“Mr. [John Smith]”) can be combined to obtain [Mr. John Smith] "(" [Mr. John Smith] ") In a" conservation "strategy, multiple boundaries can be saved, supporting them as an ambiguous conclusion when the boundary configuration and their confidence values support neither merging nor dumping. For example," [Alexander the Great] "(" [Alexander the Great] ") and" [Alexander] [the Great] "(" [Alexander] [Great] ") can be represented as alternative ambiguous resolutions.

Parsing component 215 215 may be an ambiguity-aware analyzer supporting direct ambiguous input parsing, in which parsing 355 can preserve ambiguity. Alternatively, ambiguous input permissions may need separate parsing and multiple output structures may be passed to semantic component 225 separately. Semantic processing 225, as discussed in more detail below, can be applied repeatedly to each output of the parser 215. This can lead to different semantic conclusions for different syntax inputs. Alternatively, semantic mapping 225 may combine the various inputs and process them together.

Semantic mapping 225 may have a semantic normalization of 360. Numerous findings of ambiguous parsing 355 sentences can share a value in various forms. For example, this can happen when the passive language is normalized. Consider "John gave Mary a present" and here the word "John" is subject, and "Mary" is an indirect addition. Consider "a present was given to Mary by John" ("the present was made by Mary John"), here the subject is "Mary", and "John" is the complement. Normalizing 360 may provide the conclusion that for these two examples it seems the same that “John” is semantically subject, and “Mary” is semantically an indirect complement. Alternatively, "John" can be identified as an agent, and "Mary" as a recipient. Similarly, identical representations can be provided for "Rome's destruction of Carthage" ("Rome's destruction of Carthage") and "Rome destroyed Carthage" ("Rome destroyed Carthage").

Semantic normalization can also add information about the different words of a parsed sentence. For example, words can be identified in a dictionary and associated with their synonyms, hyperonyms, possible alternative names and other lexical information.

The semantics-based resolution 370 of the coreference can decide on expressions based on syntactic and semantic information. For example, "John saw Bill. He greeted him" ("John saw Bill. He greeted him") can be defined as "he" ("he") in relation to "John" (John) and "him" ("him") in relation to "Bill" ("Bill"). Such permission may be accepted, since “he” and “John” are both subject, whereas “him” and “Bill” are both are add-ons.

The core resolution 320 may function by viewing a segment of a document in which terms appear. In contrast, semantic resolution 370 of coreference or deep resolution of coreference can process one sentence at a time. Possible offer antecedents can be stored in the repository of 375 antecedents, so that the semantic resolution 370 of the follow-up offer correlation can refer to previously entered elements. Antecedents can be stored with information about their grammatical function and roles in the sentence, their distance in the text, with information about their relationships with other antecedents and various other pieces of information.

A fusion of 380 expressions can combine expressions from the surface resolution 320 of the coreference, autonomously located annotations 325, and information from the semantic resolution 370 of the coreference. Information for terms to be combined can be identified using line alignment or annotations 330. Other mechanisms for combining two annotations on the same text can also be used.

Parsing 215 may be a natural integration point for arbitrary referenced expressions. The analyzer can support the allocation of structure in sentences, such as components or grammatical relationships, such as subject and complement. Parsing ambiguity 215 can identify numerous alternative structural representations of a sentence. In one example, information from corereference permission 320 can be used to filter the output of parser 215, storing only those views in which the left border of each reference expression matches the start of the compatible parsing part. For example, resolving coreference can set corereference as in the case of "[EO John] told [El George] [E2 Washington Irving] was a great writer"]. The parser 215 can separately provide four parsing capabilities:

1. [John] and [George] and [Washington Irving]

2. [John] and [George] and [Washington] and [Irving]

3. [John] and [George Washington] and [Irving]

4. [John] and [George Washington Irving]

The capabilities of the parser numbered three and numbered four can be filtered out due to incompatibility with the left border of the E2 Washington Irving object in accordance with a resolution of 320 references.

The extension process 385 may add additional information to the view. For example, for “John sold a car from Bill”, extension 385 might additionally display “Bill bought a car from John”. Likewise, for "John killed Bill", extension 385 may additionally display the representation "Bill died".

Traditional search engines can retrieve documents in response to user requests based on matching keywords or terms. In these traditional systems, documents can be ranked according to factors such as how many terms from the query are found within the documents, how often the terms appear, or how close the terms are to each other.

Consider the example query “Picasso painted” with the first example document containing “Picasso was born in Malaga. He painted Guernica” (Picasso was born in Malaga. He painted “Guernica”) and with the document of the second example, containing "Picasso's friend Matisse painted prolifically" ("Picasso's friend Matisse painted a lot"). If everything else is equal, the traditional system can assign the second document a higher rank than the first, because the words "Picasso" ("Picasso") and "painted" ("painted") are closer to each other in the second document. On the contrary, a system capable of deciding that the word “He” in the first document refers to Picasso (Picasso), can correctly determine the rank of the first document above, based on this knowledge. Assuming the query "Picasso painted" (Picasso painted) reflects the user's intention to find out what Picasso painted, the first document clearly gives a more consistent result.

The natural language processing system 300 may have a different architecture. In one embodiment, a conveyor may be provided in which information from one language processing step is passed as input to subsequent steps. It should be understood that these approaches can be implemented with any other architecture designed to extract facts that should be indexed from text content 210 in a natural language.

In FIG. 4 shows further details regarding the embodiments presented herein for resolving coreference in an ambiguous sensitive natural language processing system. In particular, in FIG. 4 is a flowchart showing aspects of processes 400 for ambiguity sensitive indexing with corereference in accordance with aspects of the embodiment presented herein.

It should be understood that the logical operations described here are carried out (1) as a sequence of actions performed by a computer or program modules operating in a computer system and / or (2) as interconnected machine logic circuits or circuit modules within a computer system. Implementation is a matter of choice, depending on the performance and other requirements of the computing system. Accordingly, the logical operations described herein are referred to differently as state operations, structural devices, actions, or modules. These operations, structural devices, actions and modules can be implemented in software, in embedded software, in specialized digital logic circuits, and any combination thereof. It should also be understood that more or fewer operations may be performed than shown in the drawings and described herein. These operations may also be performed sequentially, in parallel, or in an order different from that described herein.

Subroutine 400 begins at block 410, where a portion of the text content 210 can be extracted for analysis and indexing. At 420, the text content 210 may be segmented to relate areas of the text for which more search and analysis is performed during resolution processing. Segmentation can be based on a structure within the text, such as sentences, paragraphs, pages, chapters, or sections. Segmentation can also be based on the number of words, the number of sentences, or other indicators of volume or complexity.

At step 430, coreferencing decisions can be made within the textual content 210. By working with the boundaries established within step 430, corefeatures can be identified and agreed upon. Alternative title groups can be set. A surface structure may be used to provide “surface” resolution. The ambiguities that arise during the resolution of the reference can be annotated. Such annotation 340 may be provided in the form of markup within the textual content 210 or by using an external display of the object. A similar annotation can also be used to mark references and referents with object numbers. An annotation may also be provided to indicate confidence levels of established core permissions.

At step 440, parsing can transform sentences into representations that make syntactic relationships between words explicit. The parser 215 may apply a grammar 220 associated with a particular language to provide parsing information 355.

At 450, semantic representations can be extracted from text content 210. The information expressed in a document within text content 210 can be formally organized in terms of representations of the relationships between objects within the text. These relationships can generally be referred to as facts.

At step 455, parsing information 355 inferred from parsing 215 can be used to support the adoption of deep resolution 370 of coreference 370. The semantic representations created during step 450 can also be enhanced.

At step 460, the expressions obtained in step 430 of the surface resolution of the reference can be combined with the information obtained at step 455 of the deep resolution of the reference. The ambiguity syntax analyzer 215 can identify numerous alternative structural representations for the sentence. Information from resolving coreference can be used to filter the output of parser 215.

At step 470, the semantics of the text content 210 may be expanded to include selected implicit representations. At step 475, the facts can be extracted from semantic representations expressing the relationship between objects, events, and combinations of circumstances within the textual content. At step 480, facts and objects can be stored in semantic index 250.

Subroutine 400 may end after step 480. However, it should be understood that subroutine 400 can be applied repeatedly or continuously to extract pieces of text content 210 that should be applied to semantic index 250.

Shown in FIG. 5, an example of computer architecture 500 may execute the software components described herein for resolving coreference in an ambiguous sensitive natural language processing system. The computer architecture shown in FIG. 5 shows a conventional desktop computer, laptop, or server computer and can be used to execute any aspect of the software components presented here. It should be understood, however, that the described software components may also be executed on another example of computing environments such as mobile devices, television, set top boxes, kiosks, vehicle information systems, mobile phones, embedded systems, or any other. Any one or more client computers 110A-110D or server computers 120 may be implemented as a computer system 500 in accordance with embodiments.

The computer architecture shown in FIG. 5 may comprise a central processing unit 10 (CPU), a system memory 13 comprising a random access memory 14 (RAM) and a read-only memory 16 (ROM), and a system bus 11 that can connect the system memory 13 to the central processing unit 10. A basic input / output system containing basic routines that help transfer information between elements within the computer 500, such as those that act during startup, can be stored in ROM 16. Computer 500 may further contain compress a large storage device 15 for storing the operating system 18, software, data, and various software modules, such as those associated with the natural language engine 130. Natural language engine 130 may execute portions of the software components described herein. The semantic index 250 associated with the natural language engine 130 may be stored within a large volume storage device 15.

The mass storage device 15 can be connected to the CPU 10 through a mass storage device controller (not shown) connected to the bus 11. The mass storage device 15 and the associated computer-readable media can provide non-volatile storage for the computer 500. Although the description of the readable the computer storage medium contained herein relates to a mass storage device such as a hard disk or a CD-ROM drive, those skilled in the art Those skilled in the art should understand that computer-readable media can be any available computer storage media that can be accessed by computer 500.

By way of example, but not limitation, computer-readable media may include volatile and non-volatile, removable and non-removable media implemented by any method or technology for storing information, such as computer-readable instructions, data structures, program modules or other data. For example, computer-readable media contain RAM, ROM, EPROM, EEPROM, flash memory or other solid state technologies, CD-ROMs, digital versatile disks (DVDs), HD-DVDs, BLU-RAYs or other optical storage devices, magnetic tapes, magnetic tape, a magnetic disk storage device, or other magnetic storage devices or any other media that can be used to store the desired information and which can be accessed by the computer 500, but is not limited to this.

In accordance with various embodiments, computer 500 may operate in a network environment using logical connections to remote computers through a network such as network 140. Computer 500 may connect to network 140 through network interface unit 19 connected to bus 11. It should be understood that the network interface unit 19 can also be used to connect to other types of networks and remote computer systems. The computer 500 may also comprise an input / output controller 12 for receiving and processing input signals from many other devices, including a keyboard, mouse, or electronic pen (not shown). Similarly, the input / output controller 12 may provide an output signal to a video display, printer, or other type of output device (also not shown).

As briefly mentioned above, many program modules and data files can be stored in mass storage device 15 and in RAM 14 of computer 500, including an operating system 18 suitable for controlling the operation of a networked desktop computer, laptop, server computer, or other computing environment . Mass storage device 15, ROM 16 and RAM 14 may also store one or more program modules. In particular, the large storage device 15, ROM 16, and RAM 14 may store natural language engine 130 for executing CPU 10. Natural language engine 130 may include software components for implementing parts of the processes discussed in detail with reference to FIG. 2-4. Mass storage device 15, ROM 16 and RAM 14 may also store other types of software modules. The mass storage device 15, ROM 16 and RAM 14 may also store a semantic index 250 associated with the natural language engine 130.

Based on the foregoing, it should be understood that the technologies for resolving coreference in a sensitive to ambiguity natural language processing system are presented here. Although the subject matter presented here has been described in a language specific to computer structural features, methodological acts, and computer-readable media, it is intended that the invention as defined in the appended claims is not necessarily limited to the specific features, acts, or media described herein. Rather, specific features, acts, and carriers are disclosed as examples of a form of implementation of the claims.

The subject matter described above is presented by way of example only and should not be construed as limiting. In the subject matter described here, various modifications and changes can be made that are not following the shown and described examples of embodiments and applications and not deviating from the true nature and scope of the present invention, which are set forth in the following claims.

Claims (20)

1. A method for integrating mechanisms for resolving coreference, the method comprising the steps of:
retrieving, using the natural language mechanism of the server computer, part of the text;
identify, using the natural language mechanism of the server computer, the coreference within the mentioned part of the text;
extracting, using the natural language mechanism of the server computer, a fact from said part of the text, the fact being of significance; and
expanding, using the natural language mechanism of the server computer, the mentioned fact so that it includes a co-reference value different from the mentioned value and based on the identified core-reference. 2. The method according to claim 1, in which the identification of coreference within the said part of the text contains the identification of coreference in the said part of the text, using at least partially parsing. 3. The method according to claim 1, in which the identification of coreference within the said part of the text contains the identification of coreference in the said part of the text, using at least partially semantic display. 4. The method according to claim 1, in which the identification of coreference contains the identification of ambiguous coreference. 5. The method according to claim 1, further comprising identifying the ambiguity within the said part of the text. 6. The method according to claim 5, further comprising expanding the fact so that it includes an ambiguous value based on the identified ambiguity. 7. The method of claim 1, further comprising storing the expanded fact in an index suitable for supporting information retrieval. 8. The method according to claim 7, further comprising extracting the extended fact from the index in response to a search query. 9. The method according to claim 1, further comprising annotating the identified coreferences within the said part of the text. 10. The method according to claim 2, further comprising caching information from the parsing. 11. A computer storage medium having executable computer instructions stored on it which, when executed by a computer, cause the computer to:
Extract some text
identify coreference within the said part of the text;
extract a fact from part of the text, and the fact matters; and
expand the fact that it includes a co-reference value different from the mentioned value and based on the identified core-references. 12. The computer storage medium of claim 11, wherein the instructions for identifying coreferences comprise instructions for identifying coreferences in said part of the text, using at least partially parsing. 13. The computer storage medium of claim 11, wherein the instructions for identifying coreferences comprise instructions for identifying coreferences in said part of the text using at least partially a semantic mapping. 14. The computer storage medium of claim 11, wherein the instructions for identifying a coreference comprise instructions for identifying an ambiguous coreference. 15. The computer storage medium according to claim 11, further comprising instructions causing the computer to identify the ambiguity within the said part of the text. 16. The computer storage medium of claim 15, further comprising instructions forcing the computer to expand upon the fact that it includes an ambiguous value based on the identified ambiguity. 17. The computer storage medium according to claim 11, further comprising instructions causing the computer to store the expanded fact in an index suitable for supporting information retrieval. 18. The computer storage medium of claim 17, further comprising instructions for causing the computer to retrieve the expanded fact from the index in response to the search query. 19. The computer storage medium according to claim 11, further comprising instructions forcing the computer to annotate identified coreferences within a portion of the text. 20. A method for integrating mechanisms for resolving coreference, the method comprising the steps of:
retrieving, using the natural language mechanism of the server computer, part of the text;
identify, using the natural language mechanism of the server computer, the coreference within the mentioned part of the text;
identify, using the natural language mechanism of the server computer, the ambiguity within the said part of the text;
extracting, using the natural language mechanism of the server computer, a fact from said part of the text, the fact being of significance;
expand, using the natural language mechanism of the server computer, the fact that it includes a co-reference value different from the mentioned value and based on the identified core-reference;
store the expanded fact in an index suitable for supporting information extraction; and
retrieving the expanded fact from the index in response to a search query.

Images