JP2019197577A - Anaphora-omission analyzer - Google Patents

Abstract

To provide an anaphora-omission analyzer which can analyze anaphora and omission by using an item sharing relation knowledge.SOLUTION: An anaphora-omission analyzer 1150 includes: a predicate pair extraction part 1166 and an item sharing determination part 1168 for analyzing a text 250 input by an item sharing knowledge and generating information on the sharing of an item of a predicate pair in the input text 250; and an item sharing predicate network generation part 1174, a sharing item determination part 1176, and an antecedent update part 1184 for performing an analysis of anaphora and omission by using the analysis result obtained by the item sharing determination part 1168 on the input text by automatic processing, and outputting a text 254 annotated for the part of anaphora and omission and the indication destination.SELECTED DRAWING: Figure 30

Description

  The present invention relates to a natural language processing technique, and more particularly to a learning device for a term sharing discriminator used when analyzing an anaphoric relationship of an abbreviation that frequently appears in a natural language sentence, pronouns, and term sharing knowledge. The present invention relates to a linguistic knowledge collection device and an anaphora / omission analysis device using term sharing knowledge.

  Abbreviations and directives appear frequently in natural language text. For example, consider the example sentence 30 shown in FIG. The example sentence 30 includes a first sentence and a second sentence. The second sentence includes an instruction word (pronoun) 42 “it”. In this case, the instruction word “it” indicates the expression 40 “Month history New Year date” in the first sentence. The process of specifying the word indicated by the instruction word is called “anaphoric analysis”. On the other hand, consider the example sentence 60 of FIG. This example sentence 60 includes a first sentence and a second sentence. In the second sentence, the subject of the predicate “with self-diagnosis function” is omitted, but the word 72 “new switch” in the first sentence is omitted in the omitted part 76 of the subject. Similarly, the subject of the predicate “scheduled to install 200 systems” is omitted. In the omitted portion 74 of the subject, the word 70 “N company” in the first sentence is omitted. The process of detecting omission of a subject or the like and complementing it is called “omission analysis”. Hereinafter, the anaphoric analysis and the omitted analysis are collectively referred to as “anaphoric / omitted analysis”.

  In so-called artificial intelligence, natural language processing is an indispensable technology for communicating with humans. There are automatic translation and question answering as important problems in natural language processing. The anaphora and omission analysis technique is an essential element technique in such automatic translation and question answering.

  In the existing anaphoric / abbreviated analysis technology, in addition to information output by morphological analysis, syntax / dependency analysis, etc., for example, a dictionary that summarizes the characteristics that the expression of “food” is included in the object of “eating” An automatic analysis technology that uses GIS has been developed. Non-Patent Document 1 described later discloses an anaphora / omission analysis algorithm that automatically performs anaphora / omission analysis in a predetermined procedure. The technique disclosed in this document uses information output by morphological analysis, syntax / dependency analysis (hereinafter simply referred to as “dependency analysis”), and an external resource tool such as a dictionary. The dictionary used here is a collection of characteristics that the object of the verb “eat” includes an expression corresponding to “food”, for example. In the technique disclosed in Non-Patent Document 1, in addition to this, the indication of pronouns and supplementation of omissions are performed with reference to information obtained from text.

  However, according to Non-Patent Document 1, in the case of Japanese anaphora analysis, the number of recalls is 44%, the accuracy is 42%, and in the case of subject omission analysis, the numbers are both 35% and recall. Has been. Abbreviations and pronouns are frequently used in documents found in daily life. In order to perform accurate natural language processing, it is necessary to accurately detect the omitted part and the destination of the omitted part and the pronoun designation. For this purpose, it is necessary to obtain an accurate anaphoric / omitted analyzer. However, as described above, the accuracy of existing anaphoric / omitted analyzers has not reached the practical level.

Ryu Iida, Massimo Poesio.A Cross-Lingual ILP Solution to Zero Anaphora Resolution.The 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies (ACL-HLT2011), pp.804-813.2011.

  There are various reasons why the performance of anaphora and omission analysis is not improved. The main reason is that such analysis requires common sense. As general common sense, various knowledge such as a causal relationship, an implication relationship, an order relationship of occurrence of a situation, a relationship in which the occurrence of another situation is invalidated or prohibited is considered. With regard to anaphora / abbreviation analysis, it is expected that the accuracy of the analysis will be higher if there is knowledge about the relationship in which not only these but two predicates share the subject or the object.

For example, “The Socialist Party has adopted a framework document that defines philosophies and policies. A new party preparatory committee has been established as a party, and a plan to form a new party in May immediately after the unified district election (φ moth) has emerged.” Think of the text. Here, “φ Ga” represents omission of the subject of “aiming”. To specify that this “φ” is a “Social Party”, knowledge about whether two cases of two predicates share a term, such as “← adopted by X ← → X aims” (This is called term sharing knowledge.) ("← →" indicates that the left and right expressions share terms such as the main case and objective case). The same applies to anaphora analysis. For example, consider the text “A Club has three types of bedrock baths and beauty services. It can be used by a wide range of ages from young to senior.” In this text, “it” refers to “three kinds of bedrock baths and services such as esthetics”. Thus, in order to specify the phrase indicated by a certain word, the term sharing knowledge that “with X → ← X can be used” is required.

  However, such knowledge is not used in the existing automatic analysis technology for anaphora and omission analysis. This may be one reason why the accuracy of the anaphoric / omitted analyzer is not improved.

  Therefore, an object of the present invention is to learn a term sharing classifier that can determine whether two predicates share a term in order to efficiently collect term sharing relationship knowledge that can be used in anaphora and omission analysis. It is to provide a term sharing classifier learning device that can perform.

  Another object of the present invention is to provide a linguistic knowledge collection device that can efficiently collect linguistic knowledge including term sharing relation knowledge that can be used in anaphora and omission analysis.

  Still another object of the present invention is to provide an anaphora / abbreviation analysis apparatus that can perform an anaphora / abbreviation analysis using term sharing relation knowledge.

  The term sharing discriminator learning device according to the first aspect of the present invention performs learning of a term sharing discriminator that discriminates a term sharing relationship between predicate pairs by using annotated natural language text data related to an anaphoric / omitted relationship. Do. The term shared classifier learning device is configured to select a plurality of predicate pairs from annotated natural language text data, and morpheme analysis and syntax analysis for each predicate pair selected by the predicate pair selection unit. And at least the character strings of the predicates that constitute the predicate pair, the part of speech, the relation source, the relation destination, the number of words that appear between the predicate pair, and the predicate pair appear from the analysis result. A positive example and a negative example according to whether or not a feature quantity extraction unit that extracts a feature quantity including a function word and whether the predicate pair from which the feature quantity is extracted by the feature quantity extraction unit is annotated as sharing a term When a predicate pair is given by learning data generation means for generating learning data using feature amounts and machine learning using learning data generated by the learning data generation means, Pair and a learning means for performing learning of the term shared classifiers consisting statistical models to indicate whether they share the term.

  A term sharing classifier learning device according to a second aspect of the present invention uses an annotated natural language text data related to anaphora and omission relationships and a text archive including a plurality of natural language texts. To learn. The term shared classifier learning device uses an existing term shared classifier for each of the predicate pair extracting means for extracting a predicate pair from each sentence included in the text archive and the predicate pair extracting means. For each type of predicate pair for which the presence or absence of term sharing is determined by the term sharing determination unit, a term sharing determination unit for determining whether or not the predicate pair shares a term. A term sharing ratio calculation means for calculating a term sharing ratio, which is a ratio determined to be shared, a predicate pair selection means for selecting a plurality of predicate pairs from annotated natural language text data, and a predicate pair selection. For each predicate pair selected by the means, morphological analysis and syntactic analysis are performed, and at least the character string, part of speech, source of relation, relation destination, predicate pair of the predicate constituting the predicate pair from the result of the analysis A feature quantity extracting means for extracting a feature quantity including a term share ratio calculated by the term share ratio calculation means for the number of words appearing between them, a function word appearing between the predicate pairs, and the predicate pair; and feature quantity extraction Learning data generation means for generating learning data of positive examples and negative examples using feature quantities according to whether or not the predicate pair from which the feature quantity has been extracted by means is annotated as sharing a term; When a predicate pair is given by machine learning using learning data generated by the data generation means, learning of a term sharing classifier consisting of a statistical model so as to indicate whether or not the predicate pair shares a term Learning means.

  Preferably, the term sharing rate calculating means determines whether or not the appearance frequency of the predicate pair in the learning data is greater than a predetermined threshold for each type of predicate pair for which the presence or absence of term sharing is determined by the term sharing determining means. Accordingly, a process of calculating a term sharing ratio, which is a rate at which the type of predicate pair is determined to share a term, and a process of assigning a predetermined constant to the term sharing rate of the type of predicate pair. Selective calculation means for selectively executing is included.

  The language knowledge collection apparatus according to the third aspect of the present invention collects language knowledge about predicate pairs sharing a term from a text archive including a plurality of natural language texts. This linguistic knowledge collection device calculates a predetermined feature amount from each of the predicate pair selection means for selecting a predicate pair from each text of the text archive, and the predicate pair selected by the predicate pair selection means. A term sharing discriminator for determining whether or not the predicate pair shares a term by inputting to the term discriminator learned by the term sharing discriminator learning device and obtaining the discrimination result; And a term sharing knowledge storage unit for storing term sharing knowledge indicating that the predicate pair shares a term.

  Preferably, the term sharing discriminator outputs a score representing the reliability of the discrimination result together with the discrimination result as to whether or not the predicate pair corresponding to the input feature quantity shares the term, and the term sharing The knowledge accumulating means selects and accumulates the predicate pairs determined to share the term by the term sharing discriminating means and the score output by the term sharing discriminating means is equal to or greater than the threshold as the term sharing knowledge. Means for doing so.

  The anaphoric / omitted analysis apparatus according to the fourth aspect of the present invention performs an anaphoric / omitted analysis of input text using term sharing knowledge regarding a predicate pair sharing a term. This anaphoric / abbreviated analysis device relates to term sharing knowledge storage means for storing term sharing knowledge, and sharing of terms of predicate pairs in the input text by analyzing the text input using the term sharing knowledge. The term sharing analysis means for generating information, and the input text is subjected to anaphora and omission analysis using the analysis result by the term sharing analysis means by automatic processing. And an anaphoric / abbreviated analyzing means for outputting the annotated text.

  The anaphora / omission analysis device performs an anaphora / omission analysis of the input text using the term sharing relation knowledge regarding the predicate pairs sharing the terms. This anaphoric / abbreviated analyzer performs an anaphoric / abbreviated analysis by automatic processing, and outputs an annotated / abbreviated analysis means and an anaphoric / abbreviated analyzing means that outputs the annotated text regarding the anaphoric / abbreviated part and the pointing destination. A predicate pair selection means for selecting a predicate pair in the annotated text to be output, wherein the predicate pairs are annotated so that the specified term points of both predicates are different from each other. In response to the search means for searching for the term shared knowledge regarding the term of the predetermined case of the predicate pair selected by the pair selection means, and the fact that the term shared knowledge regarding the term of the predetermined case of the predicate pair is found by the search means, Among the predicates that make up the predicate pair, the point of one specified case selected by a predetermined method is rewritten with the point of the other specified case and a new annotation is added. And outputs the Deployment candidate and a rewriting means.

  Preferably, the anaphoric / abbreviated analysis means includes a network generation means for generating a term shared predicate network composed of a plurality of predicates sharing the same term for the predicate pair determined to be shared by the term sharing analysis means, and network generation For each of the term shared predicate networks generated by the means, a term determining means for determining a value of a term shared by the predicates belonging to the term shared predicate network, and a value determined by the term determining means for the term shared predicate network Value propagation means for annotating by propagating to a term shared by the predicates within.

  More preferably, the term determining means, for each predicate belonging to each of the term shared predicate networks generated by the network generating means, candidate determining means for determining a candidate value of a term that the predicate shares with other predicates; When there is one value candidate determined by the candidate determining means, means for determining the value of a term shared by predicates belonging to the term shared predicate network as the candidate, and the value determined by the candidate determining means When there are a plurality of candidates, the candidate having the highest reliability when each candidate is determined by the candidate determination unit is selected, and the value of the term shared by the predicates belonging to the term shared predicate network is determined as the candidate. Means.

  More preferably, the anaphoric / abbreviation analysis apparatus further includes a predicate belonging to the term shared predicate network when each of the predicates belonging to a certain term shared predicate network has no value candidate determined by the candidate determining means. Means for determining the value of the term to be shared by a technique different from the anaphora and omission analysis means.

  The anaphoric / abbreviation analysis device further points to a predicate that is determined not to share a term with another predicate by the term sharing analysis unit in the input text by a method different from that of the anaphora / omission analysis unit. Means for specifying the value of the term may be included.

  A computer program according to the fifth aspect of the present invention causes a computer to function as all means of any of the above-described devices.

It is a schematic diagram for demonstrating anaphora analysis. It is a mimetic diagram for explaining omission analysis. It is a schematic diagram for demonstrating the omission analysis for learning data generation, and the manual selection process with respect to the result. 1 is an overall block diagram of a system for anaphora and omission analysis including a learning data generation assisting device according to an embodiment of the present invention. It is a figure for demonstrating schematic structure of the rule which comprises linguistic knowledge. It is a block diagram which shows the structure of a term sharing discriminator learning device. It is a flowchart which shows the control structure of the program which produces the learning data for a term shared discriminator. It is a block diagram which shows schematic structure of the term sharing knowledge collection part using a term sharing discriminator. It is a schematic diagram explaining schematic structure of question type DB. It is a schematic diagram which shows the outline of the annotation method using a question answering system. It is a block diagram of a learning data generation auxiliary device. It is a block diagram of the 1st candidate production | generation part which produces | generates an annotation candidate. It is a block diagram of the 2nd candidate production | generation part which produces | generates an annotation candidate. It is a figure explaining the production | generation process of the annotation candidate using linguistic knowledge. It is a figure explaining the production | generation process of the annotation candidate using linguistic knowledge. It is a figure explaining the production | generation process of the annotation candidate using linguistic knowledge. It is a figure explaining the production | generation process of the annotation candidate using linguistic knowledge. It is a figure explaining the example of the production | generation process of the annotation candidate using term sharing knowledge. It is a block diagram of the 3rd candidate production | generation part which produces | generates an annotation candidate. It is a block diagram of the 4th candidate production | generation part which produces | generates an annotation candidate. FIG. 6 is a state transition diagram of a program for generating learning data by executing annotation on a text through interaction with a user. 3 is a detailed functional block diagram of a detector learning device 222. FIG. It is a conceptual diagram of the term sharing relationship verification process which performs annotation verification using term sharing relationship knowledge. It is a schematic flowchart of the control structure of the program which implement | achieves a term sharing relationship verification apparatus. It is a functional block diagram of the term sharing discriminator relearning device according to the second embodiment. It is a graph which shows the performance of the term sharing discriminator concerning a 2nd embodiment compared with other methods. In a 4th embodiment, it is a mimetic diagram for explaining three types of term sharing. It is a block diagram of the learning apparatus of the term sharing discriminator according to three types shown in FIG. It is a block diagram of the term shared knowledge collection device concerning a 4th embodiment. FIG. 10 is a block diagram of an anaphoric / omitted analyzer according to a fourth embodiment. It is a flowchart which shows the control structure of the computer program which implement | achieves the anaphora and omission analyzer shown in FIG. It is a figure which shows the external appearance of the computer which performs the program for producing | generating learning data. FIG. 33 is a hardware block diagram of a computer whose appearance is shown in FIG. 32.

  In the following description and drawings, the same parts are denoted by the same reference numerals. Therefore, detailed description thereof will not be repeated.

<First Embodiment>
[Outline]
With reference to FIG. 3, the outline of the operation | work which attaches an annotation to a text in the learning data generation assistance apparatus using the term shared knowledge collected by the term shared discriminator based on the 1st Embodiment of this invention is demonstrated. With reference to the upper part of FIG. 3, the annotation target text 90 is displayed on the screen of the display device. This text 90 consists of two sentences. The first sentence includes a predicate “introducing”, a word 100 “N company” that is the subject of the predicate, and a word 110 “new switch” that is the object of the predicate. The second sentence essentially consists of two sections. At the beginning of the first half of the sentence, there is a portion (omitted portion) 114 in which the subject “with self-diagnosis function” is omitted. There is an abbreviated portion 104 where the subject “200 system is scheduled to be installed” is omitted in the middle portion between the first half sentence and the latter half sentence.

  In the anaphora / omission analysis, as described in the rectangle 102 and the rectangle 112 in FIG. Thereafter, an expression to be supplemented for the omitted portion is searched and determined. In the case of the omitted portion 104, the word 100 is an expression to be complemented. In the case of the omitted portion 114, the word 110 is an expression to be complemented.

  Conventionally, humans have performed both the discovery of such omitted parts and the search and determination of candidates for complementary expressions. However, such work is inefficient and has been a factor in raising the cost of generating learning data. In the present embodiment, the above-mentioned omission spot discovery and complementary expression candidate search are automatically performed by a computer program. For such processing, various methods are used as described later. In particular, by utilizing linguistic knowledge including term-sharing knowledge, it is possible to efficiently find abbreviations and search for and determine candidates for complementary expressions, similar to work performed by humans.

  When the discovery of the omitted part and the search for the candidate for the complementary expression are completed, the processing as shown in the lower part of FIG. In this example, it is assumed that the word 110 “new switch” is found for the supplementary expression for the omitted portion 114 shown in the upper part of FIG. 3, and a plurality of words including the word 100 are found in the omitted portion 104. . In this case, the learning data generation assisting device inserts a character string 130 composed of the word “new type switch”, which is the same as the word 110, and the main case particle “c” in the omitted portion 114, and a so-called drop on the right side thereof. An icon “▼” indicating that a down list exists is displayed. When the user clicks on this icon, a drop-down list 132 is displayed. In the drop-down list 132, two options (YES, NO) for asking whether or not the displayed candidate can be added as an annotation to the omitted portion 114, and an option for manual input by a human ("manual input") Is displayed. When the user selects YES, the character string 130 is added as an annotation to the omitted portion 114. If NO is selected, no annotation is added to the omitted portion 114. When “Manual Input” is selected, an input dialog (not shown) is opened and annotations can be directly input.

  In the case of the omitted portion 104, a character string 134 including one of the candidates, for example, the same word as the word 100 and the case particle “c” is displayed, and an icon “▼” indicating a drop-down list is displayed on the right side. When the user clicks on this icon, a drop-down list including a plurality of candidate lists and “direct input” as options is displayed. When the user selects one of the candidates, the candidate word is added as an annotation for the omitted portion 104. The case where the direct input is selected is the same as the omitted portion 114.

[Overall configuration of anaphora / analysis system]
With reference to FIG. 4, first, the overall configuration of the anaphora / omission analysis system 150 that uses the learning data generation assisting device according to the present embodiment will be described.

  This anaphoric / abbreviated analysis system 150 detects abbreviations and anaphors for a large amount of text prepared in advance, and presents candidates for expressions (preceding antecedents) pointed to by a user to select them with a simple operation. Thus, the annotation data generation auxiliary system 160 that assists the user in generating the model learning data (annotation data) necessary for anaphora and omission analysis, and the learning generated by the user using the annotation data generation auxiliary system 160 A learning data DB 162 for storing data, and a learning system 164 for learning an abbreviation detector 166, an anaphor detector 168, and an antecedent identifier 170 described later using the learning data stored in the learning data DB 162. Including. The abbreviation detector 166 detects abbreviations for any predicate in the text. The omission detector 166 includes a statistical model for detecting omissions where the subject, object, etc. in the text are omitted. The anaphoric detector 168 includes a statistical model for detecting an anaphor for indicating another word or the like with an arbitrary noun phrase (anaphor candidate) in the text as a classification target. The antecedent specifying unit 170 classifies a predicate having an abbreviation and a noun phrase that is an antecedent candidate in a text as a classification target, a noun in the text, and a noun in the text. A phrase pair (anaphoric and antecedent candidate) is classified, and a detector for specifying the antecedent pointed to by the anaphor is included. Each of these detectors has a separate statistical model. In the present embodiment, each of the omission detector 166, the anaphoric detector 168, and the antecedent identifier 170 uses SVM (Support Vector Machine) as a statistical model.

  The anaphoric / abbreviated analysis system 150 further detects an abbreviation and an anaphor in the input text using the abbreviation detector 166, the anaphor detector 168, and the antecedent identifier 170, and identifies the antecedent pointed to by them. And an anaphoric / omitted analysis device 172 that outputs information attached thereto.

<Annotation Data Generation Auxiliary System 160>
The annotation data generation assisting system 160 is a text archive 192 using an input / output device 202 that performs interactive input / output with a user using a display device, a keyboard, a mouse, and the like, and various resources including language knowledge. A learning data generation assisting device 204 that generates learning data and outputs it to the learning data DB 162 by assisting the user with annotating work on a large amount of text stored in.

  The resources used by the learning data generation assisting device 204 include language knowledge in a format applicable to the existing small-scale learning data DB 190 storing learning data similar to the learning data DB 162 and text stored in the text archive 192. When the language knowledge DB 194 stored in the form of a rule and the learning data generation assisting device 204 apply the language knowledge stored in the language knowledge DB 194 to text, the expression in the rule is rephrased to another expression, A paraphrase / category dictionary 196 storing a paraphrase rule and category used to replace a word in the rule with another word, and an abbreviated portion accompanying the pattern when the text has a specific pattern A pattern dictionary 198 for storing patterns prepared in advance.

  The annotation data generation assisting system 160 further includes a question answering system 206 that inquires about a word candidate indicated by the omitted portion when the learned data generation assisting device 204 finds an omitted portion in the text, and the learning data generation assisting device 204. Is a question type DB 200 that is referred to when creating a question for the question answering system 206 based on the text, and an anaphoric / omitted analysis executed by the learning data generation auxiliary device 204 by performing an anaphoric / omitted analysis on the text. In addition, an existing anaphoric / abbreviated analyzer 208 for presenting candidates such as omitted parts, anaphors, antecedents, etc. to the learning data generation assisting device 204 is included.

  The configuration of the existing small-scale learning data DB 190 may be in any format as long as it can specify an abbreviation, an anaphor, and an antecedent in the text.

  The text stored in the text archive 192 is a simple text in the present embodiment, and it is assumed that morphological analysis and structure / dependency analysis have not been performed. However, the present invention is not limited to such an embodiment. A morphological analysis and / or a structure / dependency analysis performed by another system and the information to which such information is attached may be used.

<Language knowledge DB 194>
Referring to FIG. 5, the language knowledge rules stored in language knowledge DB 194 are roughly divided into five types. That is, causal relationship knowledge 270, implication relationship knowledge 272, event occurrence order relationship knowledge 274, event occurrence invalidation relationship knowledge 276, and term sharing knowledge 278. Each knowledge may include a variable (hereinafter, represented by “X”, “Y”, etc.) that can be replaced with an arbitrary noun phrase or the like that plays a role of a wild card.

  The causal relationship knowledge 270 is an ordered pair of an expression corresponding to the cause, for example, “X smokes tobacco” and an expression corresponding to the result, “X increases the possibility of lung cancer”. Here, it is assumed that the expression corresponding to the cause is located first and the expression corresponding to the result is located later.

  The implication relation knowledge 272 indicates that the expression “X wrote Y” is implied by the expression “X is the author of Y”, for example, by an ordered pair of these two expressions. Here, it is assumed that an implied expression comes first and an implied expression comes later.

  The event occurrence order relationship knowledge 274 expresses the normal order relationship of an event that an event “X announces Y” can occur when an event “X discovers Y” occurs. This is expressed as an ordered pair. Here, it is assumed that expressions are arranged according to the order relation of events.

  In the event occurrence invalidation relation knowledge 276, for example, when an event “Prohibit X” occurs, an occurrence of another event occurs due to the occurrence of an event such that an event “X cannot be disclosed” occurs. An ordered pair of expressions that are in a relationship that is invalidated. Here, it is assumed that an event that occurs is positioned first, and an event whose generation is invalidated by the event is positioned later.

  As will be described later, these knowledge can be applied to the text alone, but not only that. For example, in the two knowledges, the latter half of one knowledge matches the first half of the other knowledge. In some cases, knowledge can be chained and applied. Here, “match” includes, for example, synonyms, paraphrases, implications, and the like.

  On the other hand, the term sharing knowledge 278 has a slightly different personality from other knowledge. As shown in FIG. 5, in the term sharing knowledge 278, two predicates (for example, “deepen” and “discuss”, “aim” and “continue”, etc.) share the subject independently of their meanings. It is knowledge that there are many things. With such knowledge, when a certain term (for example, a prominent word) related to one (first predicate) of two predicates with high probability of sharing a term is known in a sentence. If the same term related to the other predicate (second predicate) is omitted, it is highly probable that the term of the second predicate is the same as the term of the first predicate. Therefore, the word can be a candidate for the word indicated by the omitted part of the second predicate.

<Term sharing classifier>
In order to collect such term sharing knowledge efficiently, it is determined whether or not two cases of two given predicates share a term by using a corpus that is annotated with anaphora and omission relations. The discriminator is effective. For the discriminator itself, machine learning is used, and for example, SVM is used as a statistical model. Of course, the statistical model is not limited to SVM, and a neural network or the like can also be used.

  Referring to FIG. 6, in the present embodiment, term sharing discriminator learning device 282 that performs learning of term sharing discriminator 283 using learning data obtained from annotated text storage unit 281 is used. The annotated text storage unit 281 stores, for example, the following sentence.

"Tomiichi Murayama Prime Minister met the Cabinet press conference and twenty-eight days in the Prime Minister's Office Upon the beginning of the year, even if defections's stomach for the defection issue of the New Democratic Coalition lawmakers of the Socialist Party, said that I think that remains in that range, a large amount defections It showed the prospect that does not lead to. "
Several predicates appear in this text. In the text stored in the annotated text storage unit 281, the expression of the term is annotated for each case of each predicate. In this example, it is annotated that the subject of “consult”, “think”, “state”, “show”, “show” is “Prime Minister Tomiichi Murayama”. For other predicates “is”, “stays” and “to”, other expressions are the subject. When such an annotation result is given, an arbitrary predicate pair is extracted, and the case of the pair, for example, the subject of the predicate is shared, a positive example, otherwise a negative example Learning data is created, and the term sharing discriminator 283 is trained so as to determine whether or not two predicates share a case based on the learning data.

  The term sharing discriminator learning device 282 generates, for each predicate pair in the text stored in the annotated text storage unit 281, a predicate character string, part of speech, information on a relation source, information on a relation destination, a distance between predicates, and an appearance between them. A learning data generation unit 284 that extracts learning function words and the like as feature quantities, generates learning data for each case, and a learning data storage unit 285 that stores learning data for each case output by the learning data generation unit 284 A learning processing unit 286 that learns the term sharing discriminator 283 for each case using the learning data for each case stored in the learning data storage unit 285. As predicate pairs, only predicate pairs in the same sentence may be considered, or even predicate pairs appearing beyond the sentence boundary in the text may be considered. In the present embodiment, the latter is adopted, and learning data is also generated from a predicate pair appearing beyond a sentence boundary.

  Referring to FIG. 7, the computer program for realizing learning data generation unit 284 includes step 291 for performing the following processing 292 for each of all texts in annotated text storage unit 281.

  The process 292 includes a step 293 for performing the process 294 for each of all predicate pairs included in the document to be processed.

  The process 294 includes a step 295 of extracting the above-described feature quantity from the predicate pair to be processed, and a process 296 of generating and outputting learning data using the extracted feature quantity. Specifically, in step 295, after performing morphological analysis and syntactic analysis for each of the statements including the predicate pair to be processed, feature quantities are extracted from the obtained morpheme strings and dependency relationships. In the present embodiment, feature amounts are only a morpheme sequence including part-of-speech information and a dependency structure of sentences, and thus feature amount extraction can be realized by a simple process. Process 296 is executed for each combination of predicate cases. In FIG. 7, only the processing for the subject is shown as processing 296 for easy understanding of the drawing. The processing for the object or the like is not shown, but can be performed in parallel with the processing of the subject after step 295.

  Process 296 includes step 297 for determining whether or not the predicate pair to be processed shares the subject, and the feature quantity extracted in step 295 as the learning data marked as a positive example when the determination in step 297 is affirmative Generated at step 298, and when the determination at step 297 is negative, the learning data marked as a negative example is generated from the feature amount extracted at step 295, and generated at step 298 and step 299. Outputting the learning data to a file corresponding to the case currently being processed.

  With reference to FIG. 8, the term shared knowledge collection device 301 that extracts the term shared knowledge from the unannotated text archive 302 using the term shared discriminator 283 that has been learned as described above will be described. Here, as the term sharing discriminator 283, three term sharing discriminators 306, 307, and 308 are used for each term to be discriminated. The term sharing knowledge collection apparatus 301 extracts the predicate pair extraction unit 303 that extracts all predicate pairs from the unannotated text archive 302 and the predicate pair extracted by the predicate pair extraction unit 303 in step 295 of FIG. The feature quantity extraction unit 304 that extracts the same feature quantity as that obtained, and the feature quantity extracted by the feature quantity extraction unit 304 are transferred to the three term sharing classifiers 306, 307, and 308 according to the classification of the term to be discriminated. The sorting unit 305 for sorting and the term sharing knowledge storage unit 310 that stores the discrimination results of the term sharing classifiers 306, 307, and 308, the predicate pairs to be processed, and the information of the case to be discriminated together as the term sharing knowledge. Including. The determination result includes information indicating whether or not the predicate pair to be processed shares a term, and its score. The term sharing discriminators 306, 307, and 308 are all made of SVM as described above. As the score here, a score corresponding to the distance from the separation plane serving as a criterion for discrimination to the data point to be discriminated, which is output from the SVM, is used.

  This embodiment further includes a verification processing unit 309 for verifying and selecting the term shared knowledge stored in the term shared knowledge storage unit 310 by manual interactive processing 312. For example, the verification processing unit 309 displays, on the display device, a discrimination result having a low score among the term sharing knowledge stored in the term sharing knowledge storage unit 310, and performs manual selection. However, the present invention is not limited to performing the discriminating determination of the term sharing knowledge using the interactive processing 312 by hand. For example, the term sharing discriminators 306, 307, and 308 may discard (not output) discrimination results with low reliability. Alternatively, when using the term shared knowledge stored in the term shared knowledge storage unit 310, only those having a score higher than a certain value may be used. This score may be changed for each application, or may be a different value for each case.

<Question type DB200>
Referring to FIG. 9, question type DB 200 is used when learning data generation assisting device 204 uses a question answering system 206 to obtain antecedent candidates. The question type DB 200 has a plurality of entries. For example, each entry asks what form a verb is asked to determine the grammatical role of the part omitted in the dependency relationship of the verb in an expression and the expression to be inserted into the omitted part. Information (question type) indicating whether a sentence should be generated. For example, for the verb “eating”, if the subject is omitted in an expression, if “who” is placed at the beginning of the sentence and the sentence is generated, the answer should be the subject of “eating” A representation is obtained from the question answering system.

<Question answering system 206>
With reference to FIG. 10, a process for obtaining a candidate for an expression pointed to by an omitted part using the question answering system 206 will be described. First, it is assumed that the input text 320 includes two sentences. The first sentence is "Let's drink red wine for dinner", and the second sentence is "Prevent heart disease". It is assumed that the subject of the verb 332 “Prevent” is omitted at the beginning of the second sentence, and a tag indicating the omission is attached to the omitted portion 334. Here, the word 330 “red wine” should be placed in this omitted part.

  Here, in order to obtain from the question answering system candidates for expressions to be included in the omitted part 334 from this text, it is necessary to generate an appropriate question sentence and give it as an input to the question answering system 206. Here, it is necessary to know the subject of “prevent” in the sentence of the omitted part of the input text 320. Although not shown, it is assumed that the paraphrase / category dictionary 196 stores information that “prevent” and “prevent” can be paraphrased. Then, an entry with the verb “prevent” and the grammatical role “subject” is found in the question type DB 200 shown in FIG. “What” is stored in the “question type” column. The learning data generation assisting device 204 generates a question sentence 336 “what prevents heart disease” from these pieces of information. In generating a question sentence, the end of the sentence is suitable for the question sentence, and is converted according to a rule prepared in advance. The conversion rule at the end of the sentence may be stored in advance in each entry of the question type DB 200.

  Referring to FIG. 10, upon receiving this question sentence 336, the question answering system 206 searches and extracts a word candidate group 342 suitable as an answer to the question sentence 336 from the database in the system. The question answering system 206 further transmits the word candidate group 342 obtained in this way to the learning data generation assisting device 204 so that each category 344, 346, and 348 includes one or more answer candidates for each category. . Here, it is assumed that the category 344 includes “red wine”, “sake”, and “beer”. The same applies to the other categories 346 and 348. The learning data generation assisting device 204, for example, of the word candidates obtained in this way, an expression that matches the word in the position before the omitted part 334 in the input text 320 (in this example, the word 330 “red wine”). Is selected, and the omitted portion 334 and the word 330 as a candidate for the destination are set as the target of the annotation 350. The above is the outline of the annotation addition work using the question answering system 206.

<Existing anaphoric / abbreviated analyzer 208>
Referring to FIG. 11, the existing anaphora / omission analyzer 208 performs existing anaphora / omission analysis. Here, the existing anaphoric / abbreviated analyzer 208 is used to present a candidate for an abbreviation, a candidate for an anaphor, and a candidate for an antecedent of the abbreviation and an anaphor. Eventually, a human selects a correct candidate from these candidates. Therefore, the accuracy of the existing anaphoric / omitted analyzer 208 used here is not necessarily high.

<Learning data generation auxiliary device 204>
FIG. 11 shows an annotation data generation auxiliary system 160. The learning data generation auxiliary device 204 performs morpheme analysis on each text included in the text archive 192, and outputs a morpheme string with various grammatical information. A morpheme analysis system 380 that performs a sentence structure and dependency analysis on a morpheme sequence output from the morpheme analysis system 380, and outputs a morpheme sequence to which structure / dependency information is attached; The post-analysis text DB 384 stores the morpheme string output by the dependency relationship analysis system 382 together with the grammatical information and the structure / dependency relationship. Here, the dependency relationship analysis system 382 attaches dependency information indicating at least which word is related to which word to each morpheme string. Although not shown in FIG. 11 for the sake of brevity, in the present embodiment, each element in the learning data generation auxiliary device 204 operates according to a user instruction via the input / output device 202. To do.

  The learning data generation assisting device 204 further detects an abbreviated portion in the analyzed text stored in the analyzed text DB 384 and uses the question type DB 200 and the question answering system 206 to indicate an annotated candidate group indicating the antecedent of each omitted portion. In the first candidate generation unit 388 that generates and the post-analysis text stored in the post-analysis text DB 384, the omission is detected in the same manner as the first candidate generation unit 388, and the language knowledge DB 194 is detected for each omission. The second candidate generation unit 390 that generates an annotation candidate group related to the omitted part using the paraphrase / category dictionary 196 and the pattern dictionary 198, and the learning data stored in the existing small-scale learning data DB 190, Create new learning data by changing some of the anaphoric relationships and omission relationships that exist inside Using the third candidate generation unit 392 that generates a new annotation candidate from the learning data and the existing anaphora / omission analyzer 208, the anaphora / omission analysis is performed on the text stored in the post-analysis text DB 384, and the result And a fourth candidate generation unit 394 that generates as an annotation candidate.

  The learning data generation assisting device 204 further selects the annotation candidate groups output from the first candidate generation unit 388, the second candidate generation unit 390, the third candidate generation unit 392, and the fourth candidate generation unit 394, respectively. Candidate DB 386 to be stored in association with an abbreviation, an anaphor, etc. to be annotated, and each abbreviation and anaphor of each text stored in post-analysis text DB 384, with reference to candidate DB 386, select a candidate And an interactive annotation device 396 that displays it on the display device of the input / output device 202 in a possible format, accepts a selection input from the user, and outputs it as learning data to the learning data DB 162.

  Referring to FIG. 4 again, the learning system 164 includes a question type DB 220 and a question answering system 226 similar to the question type DB 200 and the learning data generation auxiliary device 204 used in the annotation data generation auxiliary system 160, respectively, as will be described later. In addition, when learning by the antecedent identifier 170, the selection restriction DB 224 describing the restrictions imposed on the antecedent selected is similar to the language knowledge DB 194 used in the annotation data generation auxiliary system 160. Using the language knowledge DB 228 storing language knowledge, the learning data stored in the learning data DB 162, the question type DB 220, the question answering system 226, the language knowledge DB 228, and the selection restriction DB 224, an omission detector 166, an anaphoric detector 168 , And the learning process of the antecedent identifier 170 And a detector learning unit 222 of the row to.

<First candidate generation unit 388>
Referring to FIG. 12, the first candidate generation unit 388 includes a predicate search unit 420 that searches for a predicate in each text data stored in the post-analysis text DB 384, and each predicate that the predicate search unit 420 outputs. A predicate list storage unit 422 that stores the list together with the appearance position of each predicate, and a portion that is omitted from the dependency relationship involving the predicate is detected among the predicates stored in the predicate list storage unit 422. The question sentence for obtaining, as an answer, a word that complements the omitted part by referring to the question type DB 200 for each of the omitted candidates detecting section 424 that is output as an omitted candidate and each of the omitted candidates stored in the omitted candidate detecting section 424. Is generated and given to the question answering system 206 as a question sentence. The question sentence from the question sentence automatic generation section 426 is assigned to the question sentence. The answer receiving unit 428 receives the answer candidate group from the question answering system 206 and outputs the combination of the question sentence and the answer candidate group, and receives the pair of the question sentence and the answer candidate group in which the answer receiving unit 428 appears, Check candidates appearing in the post-analysis text sentence stored in the post-analysis text DB 384 among candidates included in the answer candidate group, and select all appearing before the omitted part that is the subject of the question Then, the appearance position confirming unit 430 that is output as a candidate for the omitted part and the omitted part that is output from the appearance position confirming unit 430 and a candidate group that complements the omitted part are added to the candidate DB 386 as an annotation candidate. And a candidate addition unit 432 for

<Second candidate generation unit 390>
Referring to FIG. 13, second candidate generation unit 390 includes predicate search unit 450 and predicate list storage unit similar to predicate search unit 420, predicate list storage unit 422, and abbreviation candidate detection unit 424 shown in FIG. 452, and abbreviation candidate detection unit 454, and among the abbreviation candidates detected by the abbreviation candidate detection unit 454, a search is performed for other predicates that appear before the predicate having the abbreviation candidate in the text. The target predicate search unit 456 that outputs each predicate and a predicate having a candidate for the omission as a pair, and for each predicate pair that the target predicate search unit 456 outputs, There is language knowledge in the language knowledge DB 194 such that predicates paraphrased with equivalent expressions obtained by referring to the substitution / category dictionary 196 appear in the first sentence and the second sentence, respectively, or It is determined whether or not the pattern of such expression is in the pattern dictionary 198, and if there is, the language knowledge search unit 458 that outputs the language knowledge or pattern, and the language knowledge or pattern output by the language knowledge search unit 458 are temporarily stored. Match language knowledge storage unit 460, and language knowledge or patterns stored in the match language knowledge storage unit 460, using expressions including each predicate of the predicate pair output by the target predicate search unit 456, and included in these expressions A candidate adding unit 462 that estimates a point of the omitted part to be stored and stores the pair of the omitted part and the pointing destination in the candidate DB 386 as an annotation candidate.

  The candidate addition unit 462 estimates the destination of the omitted portion as follows. Referring to FIG. 14, language knowledge search unit 458 searches language knowledge 480 for a predicate pair appearing in text 482 (“install” in character string 494, “install” in character string 500). Shall be. Text 482 is the same sentence as example sentence 60 shown in FIG. The language knowledge 480 includes a first sentence “X introduces Y” and a second sentence “X installs Y”. In the text 482, the expression corresponding to the variable X of the first sentence is a character string 490 “N company is”. The expression corresponding to “Y” in the first sentence is a character string 492 “new switch” in the text 482. The expression “introducing” in the first sentence of the language knowledge 480 matches the character string 494 “introducing” in the text 482. From this relationship, it can be seen that the candidate indicated by the variable X is the word “N company” in the character string 490, and the candidate indicated by the variable Y is the word “new switch” in the character string 492.

  On the other hand, what appears in the text 482 as an expression that matches the expression “install” in the second sentence of the language knowledge 480 is the character string 500 “install” in “install 200 systems”. A tag indicating an abbreviated portion 496 is attached to the head of this expression. By comparing this sentence ("(<subject>) installs 200 system") and the second sentence of the language knowledge 480, Y corresponds to "system" and X is omitted. I understand. Using the knowledge obtained from the first sentence of the language knowledge 480, the omitted part 496 of the text 482 indicates X of the language knowledge 480, and therefore the character string 490 “N company” is the subject of “installation”. I understand that there is. Further, the second sentence “system” of the text 482 corresponds to the variable Y of the language knowledge 480, and therefore the character string 492 “new switch is” in the text 482 corresponds to the character string 498 “system” in the text 482. It turns out that it is a thing. In this way, the candidate for the destination of the omitted portion 496 is obtained by comparing the language knowledge 480 with the text.

  Another example using language knowledge will be described with reference to FIG. In this example, two or more rules are combined sequentially to obtain a continuous causal relationship for three or more sentences, and the relationship is used to relate to three or more predicates appearing in the text. Represents a process for identifying two or more abbreviated point candidates that are detected in this manner.

  As shown in FIG. 15, it is assumed that the first language knowledge 520 has knowledge that “desertification proceeds → yellow dust increases”, and the second language knowledge 522 has “yellow dust increases → asthma worsens”. The second sentence of the first language knowledge 520 matches the first sentence of the second language knowledge 522. Therefore, by linking these two knowledge, the third knowledge 524 that “desertification advances → yellow sand increases → asthma worsens” is obtained. The third knowledge 524 is compared with, for example, the text 526 shown in FIG. Here, it is assumed that the omitted portion of the text 526 is estimated, and tags indicating the omitted portion 542, the omitted portion 548, and the like are added. In the text 526, “(φ1)” represents the first omitted portion, “(φ2)” represents the second omitted portion, and so on. Then, in the text 526, the expression 540 “desertification is progressing”, the expression 544 “(φ1 is increasing)”, and the expression 546 “(φ2 is getting worse)” are the third knowledge 524. Appear in the same order as the three predicate parts. The predicate part of these expressions is the same as the predicate of the third knowledge 524 or its utilization form. Therefore, by comparing the third knowledge 524 and the text 526 where the upper descriptive word appears, it can be seen that φ1 indicates yellow sand and φ2 indicates asthma.

  By making use of knowledge obtained by connecting two or more pieces of knowledge through such comparison, it is possible to efficiently search for a point to be omitted. The extent to which knowledge is linked is related to design matters. As the knowledge becomes longer, the maximum value of the number of points to be omitted that can be searched with one knowledge increases, but the number of texts to which the rules can be applied decreases. In this embodiment, it is assumed that up to two pieces of knowledge are connected. However, the present invention is not limited to such an embodiment. More knowledge may be concatenated and used to search for the abbreviation point. Alternatively, the maximum value of the number of connections may be changed depending on the type of sentence, or the user may specify the maximum value of the number of connections of knowledge.

  With reference to FIG. 16, an example in which more points are searched with one piece of knowledge by applying paraphrasing of knowledge predicates will be described. It is assumed that there is a situation occurrence order relation language knowledge 570 “X is born → X is sold” as knowledge. On the other hand, it is assumed that a paraphrase rule 574 that “sell” can be paraphrased as “sell” is stored in the paraphrase / category dictionary 196. As a text to be processed, a text 572 “Whisky“ W ”was born in Hokkaido and sold by Company A (φwo)” is considered. The text to be processed includes an abbreviated portion 582 represented by (φ). Comparing the first half of the text 572 with the first sentence of the occurrence order relational language knowledge 570, it can be seen that the variable X points to the word 580 of “whiskey“ W ”” in the text 572. On the other hand, the object of the predicate “sell” in the text 572 is X. As a result, and the predicate in the latter half of the text 572 is the expression 584 “sold”, the omitted part 582 indicated by “φ wo” in the text 572 corresponds to the word 580 “whiskey“ W ””. I understand that Therefore, such an annotated text 572 is first obtained. Further, the text 572 thus obtained is replaced with the expression 586 of “sold” by “replaced” 576 to which the rephrase rule 574 is applied, and the annotation is left as it is. Annotated text 578 is obtained.

  With reference to FIG. 17, the process of specifying the omitted part and / or the pointing destination of the anaphor using an empirically obtained sentence pattern will be described. For example, when a pattern such as “X is to be” appears, there is an example in which the subject of “<predicate>” and the subject of “<predicate>” are both X. Many have been empirically known. Such knowledge is stored in the pattern dictionary 198 as a grammatical pattern. Information specifying a portion corresponding to the omitted portion and information indicating the pointing destination may be inserted into the pattern. Then, when an expression that matches the pattern stored in the pattern dictionary 198 appears in the text to be processed, the expression is compared with the pattern, so that the omitted part (pointer) in the input part, Can identify candidates for the point of omission.

  For example, assume that a pattern 612 is stored in the pattern dictionary 198 as shown in FIG. In this pattern 612, when an expression 650 “X is”, an expression 652 “... do”, and an expression 654 “... do” appear in this order, the subject of the expression 654 is omitted at the head of the expression 654. It is assumed that a tag 656 indicating the location is attached and that the fingertip of the tag 656 is X.

  Assume that text 610 is given as input. In the text 610, an expression 630 “Typhoon No. 20”, an expression 632 “Move north ...”, and an expression 634 “approach” appear in this order. By collating 614 this text 610 with the pattern 612, it is found that X is equivalent to Typhoon No. 20, the subject of “approaching” is omitted, and the destination is Typhoon No. 20. Therefore, a tag 636 indicating an abbreviated portion immediately before “approaching” and “typhoon No. 20” as the pointing destination are paired as an annotation candidate 616. In this way, by applying the pattern derived from the human experience to the anaphora / omission analysis, the anaphora / omission analysis utilizing the knowledge of the person can be realized.

  Referring to FIG. 18, candidate search using term sharing knowledge is performed as follows, for example. It is assumed that the term sharing knowledge 660 includes a term sharing knowledge 660 in which an expression 662 “X is ... northward” and an expression “X is ... approaching” are paired. Consider a case where an annotation candidate 616 shown in FIG. 18 is given as input text. In this example, there may be a case where another word is mistakenly cited as the destination of omission indicated by the tag 636, or the destination candidate is not found in the first place. In that case, an annotation candidate 668 indicating the correct pointing destination is obtained using the term sharing knowledge 660 as follows.

  That is, for the expressions “Kitakami” and “approach” which are predicate pairs appearing in the annotation candidate 616, the term shared knowledge 660 can be found by searching for the term shared knowledge and collating 666. According to the term sharing knowledge 660, “go north” and “approach” often share the main character. Therefore, there is a high possibility that the tag 636 in which the main character is omitted in the annotation candidate 616 indicates the same subject (typhoon No. 20) as the predicate expression 632, and the tag 636 and the “typhoon 20” as the pointing destination. “No.” is taken as a pair of annotation candidates 668. In this way, using the term sharing knowledge derived from human experience, it is possible to generate an annotation candidate by pairing an abbreviated portion and its pointing destination. In addition to the clues obtained at the time of other analysis, the possibility of sharing the terms using the term sharing knowledge is also a clue, so that the possibility of obtaining a correct analysis result is increased. Instead of searching for a predicate pair that shares a term using term sharing knowledge, the term sharing classifier described above may be used.

<Third candidate generation unit 392>
Referring to FIG. 19, the third candidate generation unit 392 searches for an annotated anaphoric relationship and an abbreviated relationship from the existing small-scale learning data DB 190, and distinguishes and outputs the anaphoric / omitted relationship search unit 680. And for each of the noun phrase pairs constituting the anaphoric relation output by the anaphoric / abbreviated relation searching unit 680, the noun phrase (those located behind in the text) is automatically omitted. An abbreviated sentence generation unit 682 that generates a new text by generating a sentence, and among the sentences generated by the specified abbreviated sentence generation unit 682, an anaphoric relationship is annotated with the omitted part and the omitted part. An annotation adding unit 684 that adds a new omission relationship annotation candidate between the noun phrase and the text portion including the omission relationship output by the anaphora / omission relationship search unit 680 Among the omitted parts, the omitted element restoration unit 688 that generates a new sentence by restoring the pointed element as a pronoun in the omitted part based on the annotation, and the new part generated by the omitted element restoration unit 688 A new annotation generated by the annotation adding unit 684 and the annotation adding unit 690. The annotation adding unit 690 adds an annotation candidate including the pronoun pointing source and the pointing destination restored by the omitted element restoring unit 688. A candidate adding unit 686 for adding a sentence to the candidate DB 386 together with the annotation candidates.

  When the third candidate generation unit 392 omits the pointing source, the omission is automatically performed according to a certain standard by the machine. For this reason, there is a case where a person cannot determine what the abbreviated part is pointed to. In such a case, the annotation is added to the learning data as a negative example.

  Note that if the text output from the anaphora / omission relation retrieval unit 680 and the omission element restoration unit 688 does not exist in the analyzed text DB 384, the candidate addition unit 686 adds these texts to the analyzed text DB 384, and adds annotation candidates. Is added to the candidate DB 386 in association with the text added to the post-analysis text DB 384.

<Fourth candidate generation unit 394>
Referring to FIG. 20, the fourth candidate generation unit 394 executes the anaphora / omission analysis by applying the existing anaphora / omission analyzer 208 to the post-analysis text stored in the post-analysis text DB 384. An anaphoric / omission analysis execution unit 710 that obtains the result and adds it as an annotation to the analyzed text, and an analysis result storage that stores the annotated anaphoric / omission analysis result annotation output by the anaphoric / omission analysis execution unit 710 A candidate adding unit 714 that adds an annotation included in the analysis result stored in the analysis result storage unit 712 to the candidate DB 386 as an annotation candidate for the text to be analyzed in the post-analysis text DB 384.

<Interactive annotation device 396>
The interactive annotation apparatus 396 shown in FIG. 11 is realized by a program that performs state transition as shown in FIG. Referring to FIG. 21, this program starts from a state (initial state) 740 in which initial processing is executed at the start of program execution and when a document is closed and the initial screen is displayed. In this state, it is possible to select either a process for selecting a document to be processed from the text archive or a process for terminating the execution of the program. When document selection is selected, the program state transitions to the document selection state 742. In the document selection state 742, a document file selection dialog is displayed. Here, selection of a document file and cancellation of processing can be selected. If cancel processing is selected, the state returns to the initial state 740. When a document is selected in the document selection state 742, the program transitions to a state (document display state) 744 that displays the contents of the document. In the document display state 744, processing for canceling the display of the document, returning to the document selection state 742 again without reflecting the update, processing for closing the document reflecting the update, and learning data for anaphora and omission analysis are created. In order to do this, it is possible to select one of the processes for instructing selection of annotation candidates. When the selection of the annotation candidate is instructed, the annotation candidate is displayed after passing through the state (candidate search state) 746 for searching for an annotation candidate from the current processing position of the document toward the end of the document. Transition is made to a state of waiting for selection of an annotation candidate by the user (candidate selection standby state) 748.

  In the candidate search state 746, the next position to which the annotation candidate is attached is searched. If there is only one annotation candidate attached to the searched position, the annotation candidate as shown in the lower part of FIG. A drop-down list for designating whether or not to approve is generated, and an icon “▼” for displaying the drop-down list is displayed at the target location. When there are a plurality of annotation candidates, a drop-down list for displaying all of them is generated, and an icon for displaying the drop-down list is displayed at the target location. When the user moves the pointer to any of the lists when the drop-down list is displayed, the candidate and the expression on the document corresponding to the candidate are highlighted in the same color. By such processing, the user can easily understand the correspondence between the pointing source and the pointing destination. In this state, the annotation selection processing is interrupted and the state is changed to the document display state 744, the annotation selection performed so far is reflected in the document, the document is saved, and the document file is closed. Either of these can be selected.

  When any of the candidates is selected by operating the drop-down list, the program transits to the annotation addition state 750. In the annotate addition state 750, the selected annotation is added to the designated place in the document according to the selection, and the document on the memory is updated so as to add information indicating that the other candidate is not selected. The program again moves to candidate search state 746 and transitions to candidate selection wait state 748 when the next candidate is found.

  If it is selected in the candidate selection standby state 748 that the annotation process is interrupted while the document is displayed, the program transitions to the document display state 744. If the current document is selected to reflect the annotation processing performed on the document, the program transitions to a document close state 752. In the document close state 752, the program overwrites and saves the updated document data loaded in the memory as a document file, closes the open document file, and does not save the document, and enters the document selection state 742. One of the transition processes can be selected. If it is selected not to save the document, the state transitions directly to the document selection state 742. When saving the document is selected, the state transits to a state (overwrite saving state) 754 in which the file on the memory is overwritten and saved on the storage medium as the document file, and when saving is completed, the state transits to the document selection state 742.

  By executing this program, the omitted part and the anaphor are automatically searched on the apparatus side, and each time any one is searched, the annotation candidate attached to the part is displayed. If there is one annotation candidate, the user may input whether to approve the annotation candidate. In some cases, annotation candidates may be entered manually. In any case, it is possible to annotate much more easily than when the user visually finds the deletion position or the anaphor, and further finds the pointing destination visually.

<Detector learning device 222>
Referring to FIG. 4 again, as described above, the detector learning device 222 uses the selection restriction DB 224, the question type DB 220, and the question answering system 226 from the learning data stored in the learning data DB 162, and the omission detector 166, Learning of the anaphor detector 168 and the antecedent identifier 170 is performed separately. Further, the antecedent identifier 170 is further independently learned for the abbreviation antecedent identifier and the antecedent antecedent identifier.

  With reference to FIG. 22, the detector learning device 222 includes an abbreviation detector learning unit 770, an anaphor detector learning unit 772, and an antecedent identifier learning unit 774 for this purpose.

<Omitted Detector Learning Unit 770>
In the present embodiment, the omission detector 166 includes a subject omission detector 800, an object omission detector 802, and an indirect object omission detector 804. Since the abbreviated detector learning unit 770 performs learning of these three detectors individually, each of the omitted parts of the learning data stored in the learning data DB 162 includes the subject of the predicate (ga) and the direct object (W Class) or indirect object (second case), and classifying them to perform learning of the subject omission detector 800, the object omission detector 802, and the indirect object omission detector 804, respectively. An exceptional feature vector generation unit 780 for generating feature vector groups 782, 784 and 786, and a subject omission detector 800, an object omission detector 802, and an indirect purpose using these feature vector groups 782, 784 and 786, respectively. And an SVM learning processing unit 788 for learning SVMs constituting the word abbreviation detector 804.

  Here, the predicate is a verb, an adjective, and a character string in the form of “noun +“ da ””. As a learning label included in the feature vector, 1 is used when the predicate has omission, and 0 is used otherwise.

  The feature vector generated by the exceptional feature vector generation unit 780 includes the following elements.

-Regarding the predicate of the omission detection target, 1 if the grammatical role of the detection target (for example, subject) is in a dependency relationship, 0 otherwise.
-1 if the subject expression (noun phrase marked with "ha") appears in the same sentence as the predicate, 0 otherwise
-1 if the omission detection target predicate appears in the first sentence of the target text, 0 otherwise
-1 if the omission detection target predicate is after the beginning of the statement, 0 otherwise
-Headwords and parts of speech of words that have a dependency relationship with the predicate to be omitted

<Anaphoric detector learning unit 772>
The anaphoric detector learning unit 772 uses the learning data stored in the learning data DB 162 to perform the SVM learning of the anaphoric detector 168, and is annotated as an anaphoric relationship indicator (anaphoric). An anaphor candidate selection unit 830 for selecting, a feature vector generation unit 832 for generating a feature vector as described later, and a feature vector generation unit 832 for each of the anaphors selected by the anaphor candidate selection unit 830. And an SVM learning processing unit 834 for performing learning of SVMs constituting the anaphor detector 168 using the feature vector group generated by.

  In the feature vector generated by the feature vector generation unit 832, the learning label is 1 if the noun phrase to be classified has a pointing destination that has an anaphoric relationship in front of the text, and 0 otherwise. Elements of the feature vector for anaphoric detection learning include:

・ Part of speech, headword character string of anaphoric candidate, and case particle following the candidate ・ Part of speech, headword, and case particle following the subject of anaphoric candidate ・ Before anaphor candidate in text 1 if the noun phrase that appears at the position exactly matches the anaphor candidate as a character string, 0 otherwise
-1 if the anaphora candidate partially matches in the text, 0 otherwise

<Predecessor identifier learning unit 774>
The antecedent identifier 170 includes an anaphoric antecedent identifier 900 and an abbreviation antecedent identifier 902. Therefore, the antecedent specifying device learning unit 774 has a configuration in which learning of the two specifying devices 900 and 902 is performed separately.

  Specifically, the antecedent specifier learning unit 774 is an anaphoric antecedent specifier learning unit 840 that performs learning of the anaphoric antecedent specifier 900 and an abbreviation antecedent identifier 902 that is omitted. An antecedent classifier learning unit 842.

  The antecedent antecedent identifier specifier learning unit 840 includes an anaphoric selector 850 that selects an expression annotated as an anaphoric relation indicator (anaphor) from the learning data stored in the learning data DB 162, and an anaphor An antecedent candidate selecting section 852 that selects an antecedent candidate composed of an annotated phrase that is actually annotated as an antecedent and an expression that may be another antecedent for the anaphor selected by the selecting section 850; Feature vector for generating a feature vector for learning the anaphor antecedent identifier 900 for each of the combinations of the anaphor selected by the verse selection unit 850 and the anaphor candidate selected by the antecedent candidate selection unit 852 The learning unit 854 and the learning of the SVM that constitutes the antecedent antecedent identifier 900 using the feature vector generated by the feature vector generation unit 854 And a VM learning processing unit 856.

  On the other hand, the abbreviation antecedent identifier learning unit 842 selects an expression (omitted) annotated as an abbreviation relationship from the learning data in the learning data DB 162, and the omission selected by the abbreviated selection unit 870. Including an annotated expression as an antecedent for an antecedent, an antecedent candidate selecting unit 872 for selecting candidates that may become an antecedent in the learning data, an abbreviation selected by the abbreviation selecting unit 870, and an antecedent candidate selection For each combination with the antecedent candidate selected by the unit 872, a feature vector generation unit 874 for generating a feature vector and a feature vector generated by the feature vector generation unit 874 are used to omit an antecedent antecedent identifier And an SVM learning processing unit 876 that performs learning of the SVM that constitutes 902. As will be described later, the feature vector generation unit 874 is connected to the selection restriction DB 224, the question type DB 220, the question answering system 226, and the language knowledge DB 228, and uses information obtained from these when generating the feature vectors.

  In the feature vector generated by the feature vector generation unit 854 and the feature vector generation unit 874, the learning label is 1, if the abbreviation / antecedent candidate or the anaphoric / precedent candidate pair to be classified has an anaphoric / abbreviated relationship. Otherwise it is 0.

  The elements of the feature vector are as follows.

-Part of speech candidate, headword character string, case particle following the candidate-1 if the antecedent candidate appears in the first sentence of the text, 0 otherwise
-1 if the antecedent candidate is the first antecedent candidate mentioned in the sentence, 0 otherwise
-1 if the antecedent candidate and anaphor are exactly the same string, 0 otherwise
-1 if the antecedent candidate and the anaphor are partially the same string, 0 otherwise
Uses the selection restriction DB 224, 1 if the selection restriction is satisfied between the predicate having an abbreviation and the noun phrase as an antecedent, 0 otherwise
For a sentence including an abbreviation and the omitted part, the question type DB 220 is used to generate a question sentence in which the part is an answer and give to the question answering system 226 that one of the answers is an antecedent candidate 1 if they match, 0 otherwise
1 if the predicate having an abbreviation or the predicate's predecessor predicate matches the predicate pair included in the causal relationship knowledge registered in the detector learning device 222, and 0 otherwise.

<Anaphoric / Omitted Analysis Device 172>
Referring to FIG. 4, the anaphora / omission analyzer 172 is connected to an omission detector 166, an anaphoric detector 168, and an antecedent identifier 170, and uses these to perform an anaphora / omission analysis and automatically An anaphoric / omitted analyzer 252 for adding an annotation to the input 250 and outputting it, and a language knowledge DB 256 and a question type DB 258 used for feature vector generation when the anaphoric / omitted analyzer 252 performs an anaphoric / omitted analysis. , A question answering system 260, and a selection restriction DB 262. The language knowledge DB 256 is a database storing language knowledge similar to the language knowledge DB 194 used in the anaphora / abbreviation analysis system 150 and the language knowledge DB 228 used in the learning system 164. The question type DB 258 has the same configuration as the question type DB 200 used in the annotation data generation auxiliary system 160 and the question type DB 220 used in the learning system 164, and the storage contents thereof are also the same. The question answering system 260 is a system having functions similar to the question answering system 206 used in the annotation data generation assisting system 160 and the question answering system 226 used in the learning system 164. The selection restriction DB 262 is the same as the selection restriction DB 224 used when generating a feature vector in the learning system 164. The DBs described as having the same configuration here may be the same as each other or different from each other.

[Operation]
The anaphora / abbreviation analysis system 150 whose configuration has been described above operates as follows. The description will be divided into five phases: learning of the term sharing classifier 283, collection of term sharing knowledge, creation of learning data, learning of the detector, and automatic annotation for input.

<Learning of the term sharing classifier 283>
Referring to FIG. 6, annotated text is stored in advance in annotated text storage unit 281. This text is often prepared manually. The amount may not be so large. The learning data generation unit 284 generates a positive example composed of predicate pairs sharing a term and a negative example not sharing a term as learning data for each case for the predicate, and stores it in the learning data storage unit 285 for each case. The Referring to FIG. 7, specifically, the following processing 292 is executed for all the texts stored in the annotated text storage unit 281 (step 291). In the process 292, the following process 294 is executed for all predicate pairs included in the document being processed.

  In the process 294, the feature amount described in the configuration column is extracted for the predicate pair to be processed. The following processing 296 is performed for each case using the extracted feature amount. First, according to the annotation, it is determined in step 297 whether or not the predicate pair to be processed shares the target case term. If the term is shared, the learning data consisting of the feature quantity is marked as a positive example in step 298 and output to the target case file in step 300. If it is determined in step 297 that the predicate pair does not share a term, the learning data is marked as a negative example in step 299 and is output to the target case file in step 300.

  The processing described above is executed for all predicate pairs of all texts in the annotated text storage unit 281, and the learning data is stored in the learning data storage unit 285.

  The learning processing unit 286 learns the term sharing classifier 283 prepared for each case. As a result, when a given feature vector is given for the predicate pair, a term sharing discriminator 283 that determines whether or not the predicate pair shares a certain term is obtained.

  If the term sharing discriminator 283 is obtained by the configuration of FIG. 6, the term sharing knowledge collection device 301 collects the term sharing knowledge from the unannotated text archive 302 and stores it in the term sharing knowledge storage unit 310 as shown in FIG. it can. That is, the predicate pair extraction unit 303 of the term shared knowledge collection device 301 extracts predicate pairs from the unannotated text archive 302 and supplies the predicate pair to the feature amount extraction unit 304. The feature amount extraction unit 304 calculates the same feature amount as that extracted by the learning data generation unit 284 for each predicate pair, and provides the same to the distribution unit 305. For each case to be processed for each predicate pair, the sorting unit 305 gives a feature vector including feature amounts to the term sharing classifiers 306, 307, and 308 corresponding to the case. Each of the term sharing classifiers 306, 307, and 308, when given a feature vector, outputs a determination as to whether or not the predicate pair shares a term with the score for the case that each shared classifier is in charge of. Are stored in the term shared knowledge storage unit 310.

  The term shared knowledge stored in the term shared knowledge storage unit 310 in this way is verified by the operator using the interactive processing 312 and the verification processing unit 309, and the inappropriate or low score is deleted. The term sharing knowledge can be obtained.

<Creating learning data>
Creation of learning data is performed by the annotation data generation auxiliary system 160. Prior to this processing, it is necessary to prepare the existing small-scale learning data DB 190, the text archive 192, the language knowledge DB 194, the paraphrase / category dictionary 196, the pattern dictionary 198, the existing anaphora / abbreviation analyzer 208, and the question type DB 200. is there. The question answering system 206 is prepared in advance in the annotation data generation auxiliary system 160 and is made accessible from the learning data generation auxiliary device 204, or prepared so that a question sentence can be sent to an external question answering system. It is necessary to keep.

  Referring to FIG. 11, the user activates learning data generation auxiliary device 204, designates text archive 192, and instructs the start of learning data generation. The morpheme analysis system 380 reads each text stored in the text archive 192, performs morpheme analysis, and gives a morpheme sequence with various grammatical information to the dependency relationship analysis system 382. The dependency relationship analysis system 382 performs grammatical structure analysis and dependency analysis for each sentence including a given morpheme sequence, and outputs the morpheme sequence to which the structure information and dependency information are attached to the text DB 384 after analysis. . In this way, each text stored in the text archive 192 is analyzed, and the analyzed text is accumulated in the analyzed text DB 384.

  Referring to FIG. 12, predicate search unit 420 of first candidate generation unit 388 reads each post-analysis text stored in post-analysis text DB 384, searches for a predicate, and writes it to predicate list storage unit 422. The abbreviation candidate detection unit 424 has, for each predicate in the predicate list stored in the predicate list storage unit 422, any of the subject of the predicate, the object, the indirect object, etc. in the dependency relationship including the predicate. If it does not exist, that portion is detected as an omission candidate and given to the question sentence automatic generation unit 426. The question sentence automatic generation unit 426 generates a question sentence from a question sentence whose answer is a part of the omission candidate. At this time, the question sentence automatic generation unit 426 accesses the question type DB 200 using the predicate to be processed and the grammatical role (subject, object, etc.) of the omission candidate as keys, and reads out the question type. The question sentence automatic generation unit 426 further generates a question sentence such as the question sentence 336 shown in FIG. 10 by modifying the sentence including the omitted part using the read expression of the question type. The question sentence automatic generation unit 426 gives this question sentence to the question answering system 206 and the answer receiving unit 428. When receiving the question message from the question message automatic generating unit 426, the answer receiving unit 428 waits until an answer to the question message is given from the question answering system 206.

  In this embodiment, the question answering system 206 generates a plurality of answer candidates for each category for each given category and gives the answer to the answer receiving unit 428. When the answer receiving unit 428 receives an answer from the question answering system 206, the answer receiving unit 428 confirms to which question sentence the answer is sent from the question sentence automatic generating unit 426, and becomes an object to be processed by the appearance position confirmation unit 430. Give information and answer to identify the predicate.

  Based on the information from the answer reception unit 428, the appearance position confirmation unit 430 confirms the position of the abbreviated pointing destination candidate that appears before the predicate that is the processing target in the post-analysis text to be processed. Then, using these as candidates for omission of complementation, the predicates that are the processing targets and each of the candidates are combined and output to the candidate adding unit 432. The candidate addition unit 432 adds the candidate given from the appearance position confirmation unit 430 to the candidate DB 386.

  Referring to FIG. 13, second candidate generation unit 390 operates as follows. The predicate search unit 450 searches each predicate in the analyzed text stored in the analyzed text DB 384 and accumulates it in the predicate list storage unit 452. Similar to the abbreviation candidate detection unit 424 shown in FIG. 11, the abbreviation candidate detection unit 454 has the subject of the predicate in the dependency relationship including the predicate for each predicate in the predicate list storage unit 452. It is determined whether or not any of the object and the indirect object is present, and when it does not exist, that part is detected as an omission candidate. The omission candidate detection unit 454 gives information indicating the position of the detected omission candidate to the target predicate search unit 456. The target predicate search unit 456 searches for other predicates that appear before the predicate having the abbreviation candidate in the text among the abbreviation candidates detected by the abbreviation candidate detection unit 454. The predicate and the predicate having a candidate for the omission are paired and given to the language knowledge search unit 458.

  For each predicate pair output by the target predicate search unit 456, the linguistic knowledge search unit 458 includes the first sentence in any of the linguistic knowledge in the linguistic knowledge DB 194 in which the predicate constituting the pair or the predicate rephrasing it is It is determined whether there is linguistic knowledge that appears in the second sentence, and if there is, linguistic knowledge is output. The linguistic knowledge search unit 458 also outputs an expression including the predicate pair given from the target predicate search unit 456, and if there is a match with the pattern stored in the pattern dictionary 198. The linguistic knowledge search unit 458 uses the paraphrase rules and categories stored in the paraphrase / category dictionary 196 when applying these linguistic knowledge and patterns to text, and paraphrases expressions in the rules to other expressions, Expand the scope of linguistic knowledge by replacing words in the rules with other words. The linguistic knowledge retrieved by the linguistic knowledge retrieval unit 458 is accumulated in the matching language knowledge storage unit 460 together with the predicate pair used in the retrieval of the linguistic knowledge. Finally, as described with reference to FIGS. 14 to 18, the candidate adding unit 462 collates the predicate pair stored in the matching language knowledge storage unit 460 with the matching language knowledge or pattern, The pointing candidate of the omission candidate is specified, and a pair of the omission candidate and each pointing destination candidate is added to the candidate DB 386 as an annotation candidate. In this case, as will be described later, by using term sharing knowledge, if there is a predicate that shares a term with a predicate that has an abbreviation candidate in the predicate pair, a pair of the abbreviation candidate and the content of the term that the predicate has Are added to the candidate DB 386 as annotation candidates. In this way, the term sharing knowledge can be utilized for generating annotation candidates.

  Referring to FIG. 19, the anaphoric / omitted relationship searching unit 680 of the third candidate generating unit 392 searches the existing small learning data DB 190 for the annotated anaphoric relationship and the omitted relationship, and a sentence including the anaphoric relationship. Is output to the source abbreviated sentence generation unit 682, and the sentence including the omission relation is output to the omitted element restoration unit 688.

  The ellipsis abbreviation generator 682 searches the text including the anaphoric relationship for a noun phrase pair that constitutes the anaphoric relationship, and for each of them, automatically identifies the locator located behind in the text within the noun phrase. A new omitted text is generated and given to the annotation adding unit 684. The annotation adding unit 684 creates a new line between the omitted part and the noun phrase in which the anaphoric relationship is annotated together with the omitted part of the sentence generated by the pointing source omitted sentence generating unit 682 in this way. Annotated annotation candidates are attached and output to the candidate adding unit 686.

  The candidate adding unit 686 generates new text by restoring the pointed element as a pronoun based on the annotation for the omitted part of the text including the omitted relationship output from the anaphoric / omitted relationship searching unit 680. To the annotation adding unit 690. The annotation adding unit 690 includes a new text generated by the omitted element restoring unit 688 and a pointing source composed of pronouns restored by the omitted element restoring unit 688 and a portion that was pointed to by the original annotation. A new anaphoric annotation candidate is added and output to the candidate addition unit 686.

  The candidate adding unit 686 adds the text output from the annotation adding unit 684 and the annotation adding unit 690 to the candidate DB 386 together with the annotation candidates added to them.

  Referring to FIG. 20, the anaphora / omission analysis execution unit 710 of the fourth candidate generation unit 394 reads the text from the post-analysis text DB 384 and executes the anaphora / omission analysis by the existing anaphora / omission analyzer 208. Although the existing anaphora / omission analyzer 208 has a limit in its performance, the anaphora / omission analysis is performed on the input text, and the annotated text is returned to the anaphora / omission analysis execution unit 710. The anaphora / omission analysis execution unit 710 accumulates the annotated analysis result in the analysis result storage unit 712. The candidate addition unit 714 adds the annotation included in the analysis result stored in the analysis result storage unit 712 to the candidate DB 386 as an annotation candidate.

  By executing the above processing for all the texts stored in the text archive 192 (see FIGS. 4 and 11), the text with the annotation candidates is finally stored in the candidate DB 386. The text accumulated in the candidate DB 386 includes abbreviation candidates and anaphoric candidates, and each abbreviation candidate and anaphoric candidate is added with one or more annotation candidates indicating their pointing destinations.

  The interactive annotation device 396 operates as follows for each of the omission candidates and the anaphoric candidates for each of the candidates accumulated in the candidate DB 386 by interactive processing with the user using the input / output device 202.

  With reference to the lower part of FIG. 3, the interactive annotation apparatus 396 displays text 92 including annotation candidates on the screen. At that time, the interactive annotation apparatus 396 creates a drop-down menu for selecting an annotation candidate as follows for each of the omission candidate and the anaphoric candidate. That is, when there is only one point where the omission candidate is pointed, the same word as the word 110 that is the pointed candidate is displayed at the position of the omission candidate indicated by the character string 130 in FIG. An icon “▼” indicating that there is a drop-down menu is displayed. In the drop-down menu, an option (YES / NO) indicating whether or not to accept the displayed candidate and an option indicating direct input are displayed. When the user selects YES, the selected annotation is confirmed as an annotation. If NO is selected, the annotation is left indeterminate. When direct input is selected, a dialog for directly specifying the destination of the omission candidate indicated by the character string 130 is displayed. When the user inputs the pointing destination, a new annotation according to the user's specification is generated and becomes a final annotation. In this case, the annotation that is initially displayed but not selected by the user is used to generate a negative example when generating a feature vector during learning. Therefore, a flag indicating a negative example is attached to this annotation candidate.

  When there are a plurality of annotation candidates, a drop-down list including the plurality of annotation candidates as options is generated. In the corresponding part, the top one of the annotation candidates or the annotation candidate with the highest score when scoring the annotation candidate in some form is displayed. On the right side, an icon “▼” indicating that there is a drop-down list is displayed. In this case as well, an option item for selecting to input directly is displayed in the drop-down list. When the user selects any option, the selected annotation candidate is confirmed as an annotation. The candidates that are not selected are used to generate a negative example when generating a feature vector during learning.

  In any case, the post-analysis text in which the annotation is confirmed is accumulated in the learning data DB 162. When the above-described annotation selection is completed for all the analyzed texts stored in the candidate DB 386, the learning data DB 162 is completed.

<Learning the detector>
Referring to FIG. 22, detector learning device 222 operates as follows. The exceptional feature vector generation unit 780 of the omission detector learning unit 770 reads out learning data including omissions from the learning data that has been confirmed annotation stored in the learning data DB 162. The case-specific feature vector generation unit 780 classifies each omitted part of the learning data when the predicate subject (ga case), direct object (wo case), or indirect object (second case). Thus, feature vector groups 782, 784 and 786 are generated and stored in a storage device (not shown). At this time, the exceptional feature vector generation unit 780 sets the learning label to 1 for the confirmed annotation and sets the learning label to 0 for the other annotations. The SVM learning processing unit 788 uses these feature vector groups 782, 784, and 786 to learn SVMs included in the subject omission detector 800, the object omission detector 802, and the indirect object omission detector 804.

  The anaphoric candidate selection unit 830 of the anaphoric detector learning unit 772 reads the text including the anaphoric candidate from the learning data accumulated in the learning data DB 162, and annotates it as the reference of the anaphoric relationship (anaphoric). Selected words. The feature vector generation unit 832 generates the above-described feature vector for each of the anaphor candidates selected by the anaphor candidate selection unit 830. At this time, the feature vector generation unit 832 sets the learning label to 1 for the confirmed annotation and sets the learning label to 0 for the other annotations. The feature vectors generated by the feature vector generator 832 are stored in a storage device (not shown). The SVM learning processing unit 834 performs learning of the SVM constituting the anaphoric detector 168 using the feature vector group accumulated in this way.

  In the antecedent classifier learning unit 774, the antecedent class antecedent classifier learning unit 840 operates as follows. The anaphoric selector 850 of the antecedent antecedent specifier learning unit 840 selects, from the learning data stored in the learning data DB 162, an expression that is annotated as an anaphoric relationship indicator (anaphor), and precedes it. This is given to the word candidate selection unit 852. The antecedent candidate selection unit 852 displays both an annotated actual annotator and an expression (noun phrase) that may be another antecedent for the anaphor selected by the anaphor selector 850. Select an antecedent candidate to contain. The feature vector generation unit 854 performs learning of the antecedent antecedent specifier 900 for the combination of the anaphor selected by the anaphor selection unit 850 and each of the anaphor candidates selected by the antecedent candidate selection unit 852. Generate feature vectors of. At this time, the feature vector generation unit 854 sets the learning label as 1 and the other learning labels as 0 for the annotated expression as the pointing destination of the anaphor in the anaphoric relation. The SVM learning processing unit 856 learns the SVM that constitutes the antecedent antecedent specifier 900 using the feature vector generated by the feature vector generation unit 854.

  On the other hand, the abbreviation antecedent identifier learning unit 842 operates as follows. The abbreviation selection unit 870 of the abbreviation antecedent identifier learning unit 842 selects an expression (omitted) that is annotated as an abbreviation relationship from the learning data in the learning data DB 162, and provides it to the antecedent candidate selection unit 872. The antecedent candidate selection unit 872 selects a candidate that may be an abbreviated antecedent in the learning data, including an annotated expression as an antecedent for the abbreviation selected by the abbreviation selection unit 870. The feature vector generation unit 874 generates a feature vector for each combination of the abbreviation selected by the abbreviation selection unit 870 and each of the antecedent candidates selected by the antecedent candidate selection unit 872 and stores it in a storage device (not shown). At this time, in the feature vector generated by the feature vector generation unit 874, the learning label is 1 if the omission / preceding candidate pair to be classified has an omission relationship, and 0 otherwise. Also, the feature vector generation unit 874 uses the selection restriction DB 224, the question type DB 220, the question answering system 226, and the language knowledge DB 228 as described above for generating the feature vector. The SVM learning processing unit 876 uses the feature vector generated by the feature vector generation unit 874 to learn the SVM that constitutes the abbreviation antecedent identifier 902.

  Through the above processing, the subject omission detector 800, the object omission detector 802, the indirect object omission detector 804, the anaphor detector 168, and the antecedent identifier 170 included in the omission detector 166 are included. Learning of the antecedent antecedent identifier 900 and the abbreviation antecedent identifier 902 is completed.

<Automatic annotation>
Returning to FIG. 4, the anaphora / omission analysis for the input 250 by the anaphora / omission analyzer 252 and the automatic annotation of the result are executed as follows. This analysis is divided into an abbreviation analysis, an anaphoric analysis, an abbreviation, and an antecedent analysis for the anaphor. In the omission analysis, an omission detector 166 is used. In the anaphora analysis, an anaphor detector 168 is used. In the antecedent analysis, an antecedent identifier 170 is used.

  First, morphological analysis and structure / dependency analysis are performed on the input 250, and a morpheme string to which structure / dependency information is attached is processed. For this morpheme sequence, in the omission analysis, feature vectors having the same structure as those obtained when the subject omission detector 800, the object omission omission detector 802, and the indirect object omission omission detector 804 shown in FIG. Generated and fed to these detectors. By the outputs of the subject omission detector 800, the object omission detector 802, and the indirect object omission detector 804, an annotation as an abbreviation is attached to each portion of the input 250 indicating the highest score.

  After that, to the input with such an annotation, a predicate that shares the term with the predicate is applied by applying the term sharing knowledge including the predicate having an abbreviation or an anaphor (hereinafter simply referred to as “abbreviation etc.”). It is determined whether or not is being input. If such a predicate is found, annotate the annotation so that the content of the term of the found predicate is complemented in the omitted part. In this case, instead of using the term sharing knowledge directly, the predicate sharing the term with the predicate having the omission etc. may be found in the input using the term sharing discriminator 283 shown in FIG. . If a word to be supplemented with a term omitted in such processing is found, the subsequent analysis is terminated for the omitted portion, and the processing for the next omitted portion may be performed. If a word to be complemented is not found, the subsequent anaphora and omission analysis is executed. Of course, it is also possible to annotate the word found here as one of the completion candidates and proceed to the anaphora / omission analysis process.

  In the anaphoric analysis, a phrase that can be an anaphor candidate is selected from the noun phrase pair included in the input 250, and each of the features has the same configuration as the feature vector generated by the feature vector generation unit 832 shown in FIG. A vector is generated from input 250. By providing this composition vector to the anaphor detector 168, a score is obtained from the anaphor detector 168. Based on this score, it is determined whether or not it is an anaphor, and an annotation to that effect is attached.

  The antecedent is estimated by the antecedent antecedent identifier 900 as follows. For each expression determined to be an anaphor by the anaphor detector 168, an expression that can be an antecedent is selected in the input 250, and a feature vector is generated in the same manner as the feature vector generation unit 854 in FIG. . Using this feature vector as an input to the antecedent antecedent identifier 900, the score of the antecedent antecedent identifier 900 is obtained, and the antecedent candidate that has the highest score with a score higher than the threshold Add an anaphoric annotation with the anaphor to be processed.

  The presumption of the antecedent by the abbreviation antecedent identifier 902 is the same. However, in this case, a feature vector having the same configuration as that created by the feature vector generation unit 874 is used instead of the feature vector generated by the feature vector generation unit 854. Therefore, at this time, the selection restriction DB 224, the question type DB 220, the question answering system 226, and the language knowledge DB 228 are required.

  As described above, according to the present embodiment, it is possible to easily generate learning data for learning a detector for anaphora and omission analysis from a text archive. For example, candidates for an anaphoric / abbreviated relationship are automatically searched, and the antecedent candidates are displayed as a list. The user may approve if there is one candidate displayed in the list and is correct, and select one of them if there are a plurality of candidates displayed in the list. As in the prior art, there is no need to visually search for an anaphoric / omission-related position, and then visually search for the indicated position and attach an anaphoric / omission-related annotation to both. If the correct pointer is not shown in the displayed list, the correct pointer can be designated manually.

  Therefore, it is possible to prepare learning data for learning the detector for anaphora and omission analysis much more easily than the conventional technique. As a result, both the learning data creation cost and creation time can be reduced. Since a large amount of learning data can be prepared at a low cost, text anaphora / abbreviation analysis is performed using the abbreviation detector 166, the anaphoric detector 168, and the antecedent identifier 170 that have been trained using the learning data. It becomes possible to carry out efficiently. In addition, since the amount of learning data is large, it can be expected that the accuracy of these detectors will be improved, and text adaptation and omission analysis can be performed almost automatically with high accuracy.

  In the above embodiment, the question answering system 206 has been described as being singular. However, the present invention is not limited to such an embodiment. If a plurality of question answering systems can be used, a plurality of question answering systems may be used. In this case, if each question answering system receives natural language texts as input, answer candidates can be obtained simply by sending the same question text to them.

  Further, in the above embodiment, in the anaphora / abbreviation analysis, if a word to be supplemented is found in a place such as an abbreviation in the process using the term sharing knowledge first, then the anaphora / abbreviation analysis for the abbreviated place is not performed. The anaphora and omission analysis is performed only when no word to be complemented is found. However, the present invention is not limited to such an embodiment. Information regarding the candidate for completion obtained by applying the term sharing knowledge may be used as a part of the feature for anaphora and omission analysis.

<Second Embodiment>
In the first embodiment, as shown in FIG. 6, the term sharing discriminator 283 learns from the text stored in the annotated text storage unit 281. However, when the amount of text in the annotated text storage unit 281 is small, the accuracy of the term sharing discriminator 283 is not so high. As a result, if a sufficient amount of text cannot be prepared in the annotated text storage unit 281, there is a possibility that the estimation accuracy of the pointing destination candidate based on the term sharing relationship may not be improved. Therefore, in this second embodiment, a method for learning a term sharing discriminator is shown in which the accuracy of term sharing discrimination is increased even if the amount of text in the annotated text storage unit 281 is small.

  Referring to FIG. 25, the term sharing discriminator learning device 972 adopting this method uses the text and text archive 192 stored in the annotated text storage unit 281 used in the first embodiment. The term sharing classifier 974 is learned. The term sharing discriminator learning device 972 performs a morphological analysis and a syntax analysis on each sentence of the text archive 192, and extracts a predicate pair extraction unit 980 that extracts a predicate pair in which a certain predicate is syntactically dependent on another predicate; The predicate pair extraction unit 980 determines whether or not there is a term sharing relationship for each predicate pair, and the term sharing discriminator 283 obtained in the first embodiment and the result obtained by the term sharing discriminator 283 Statistical processing unit 984 that calculates a rate (term sharing rate) determined to have a shared relationship for the same pair of predicates including a variation by statistical processing, and annotated text similar to that shown in FIG. Using the text stored in the storage unit 281 and the processing result of the statistical processing unit 984 as an input, the statistical processing unit 984 further calculates a feature amount similar to that of the learning data generation unit 284 shown in FIG. And a learning data generating unit 986 for generating learning data added as a feature section sharing ratio.

  The term sharing discriminator learning device 972 further includes a learning data storage unit 988 that stores the learning data output from the learning data generation unit 986, and a term sharing determination that includes SVM using the learning data stored in the learning data storage unit 988. And a learning processing unit 990 that performs learning of the device 974.

  In addition, in the processing by the statistical processing unit 984, it is considered that the reliability of the determination result regarding the presence / absence of term sharing is low for the predicate pair with low appearance frequency. Therefore, in the present embodiment, the term sharing rate is set to the default value of 0 for predicate pairs whose appearance frequency is less than a threshold value (for example, 5). The threshold value was changed to another value and the same processing was performed, but the final result was not greatly affected.

[Evaluation experiment]
An evaluation experiment was performed to confirm the performance of the term sharing discriminator 974. In the evaluation experiment, the subject of the predicate pair was shared, and 80% of the NAIST text corpus 1.4 was used for learning. The learning data included 42,806 predicate pairs. Among them, 8,496 predicate pairs shared the subject.

  In the evaluation experiment, 2,451,254 examples of causal relationships were sampled from 600 million web pages by the method described in Reference Document 1 described later. There were cases where the subject of the predicate was included in the original sentence, so they were excluded from the subject. More specifically, examples where the noun is followed by the particle “ha” or “ga” were excluded. As a result, the number of samples was changed from 2,451,254 to 1,321,172. 1000 of these samples were randomly sampled. For each of these, three determiners determined whether the predicate pair shared the subject. The average kappa coefficient (Cohen's Kappa) for these determinations was 0.683. The final annotation label was determined by the majority of the judgment results of the three judges. As a result, it was determined that 644 out of 1000 samples shared the subject. This percentage is higher than expected, but this is likely due to the fact that the underlying data represents a causal relationship.

  In the evaluation experiment, the performances of the following three methods and the term sharing classifier (related to the subject) according to the second embodiment were compared.

(1) BaselineKNP
This is to determine whether to share the subject based on the output of the Japanese parsing system, and is an analysis result obtained by incorporating the anaphora / abbreviation analysis system according to Reference 2 described later.

(2) BaselineAllShare
It is determined that all predicate pairs in the input share the subject. As mentioned earlier, this assumption is reasonable and seems to be a baseline because the original data represents a causal relationship.

(3) FirstStepClassifier
It is a term sharing discriminator according to the first embodiment.

  The results are shown in Table 1 below.

平均精度は、各手法の分類器により計算されたスコアの値によってサンプルをランキングした結果による。FirstStepClassifier及び本実施の形態による手法では、主語を共有するか否かはＳＶＭで判定している。BaselineAllShareでは、参考文献１によるＳＶＭによるスコアを用いている。

The average accuracy depends on the result of ranking the samples according to the score value calculated by the classifier of each method. In FirstStepClassifier and the method according to the present embodiment, whether or not to share a subject is determined by SVM. In BaselineAllShare, the score by SVM according to Reference 1 is used.

  From Table 1, the effect of the present embodiment can be seen as follows. First, in the F value, it can be seen that the value obtained by the method of the present embodiment shows a significant improvement over all other methods. Compared to FirstStepClassifier, the improvement in F value is large. This seems to be mainly due to the improvement of the reproducibility (0.155 → 0.826).

  FIG. 26 shows the recall-matching rate curve obtained by these methods. This graph is drawn by changing the threshold value of the score obtained by the SVM of each method. For BaselineKNP, one point of the recall and accuracy based on the output result is plotted on the graph.

  From FIG. 26, the term sharing discriminator according to the second embodiment has a relevance ratio of about 5 to 10% higher than the reproducibility values in a wide range as compared with the discrimination result by FirstStepClassifier and BaselineKBN. I understand that It can also be seen that this term sharing discriminator shows a precision of about 20% higher than BaselineKNP at the same recall rate.

  As described above, according to this embodiment, a term sharing discriminator with higher accuracy than the term sharing discriminator 283 used in the first embodiment can be obtained by almost automatic processing. If such a term sharing discriminator is used, not only the annotation verification but also the term sharing knowledge described in the first embodiment can be efficiently collected with high accuracy.

  The second embodiment has shown an example in which the term sharing classifier learned by the apparatus according to the present invention is applied to the anaphora / omission analysis. As a result, it is conceivable that the accuracy of anaphora and omission analysis is improved to a practical level. The effect is particularly evident in automatic translation from Japanese to English. In Japanese text, the subject is often omitted. On the other hand, English requires the presence of a subject. Therefore, when translating from Japanese to English, it is necessary to identify the subject omitted in Japanese with high accuracy. Therefore, it can be expected that the accuracy of automatic translation from Japanese to English will be improved by improving the accuracy of anaphora and omission analysis.

  Another example is a question answering system. In a question answering system, a document to be searched for an answer to a question is often a text written in a natural language. If the text contains an anaphoric / omitted relationship, it may not be possible to find the answer correctly. Therefore, if you apply anaphora / abbreviation analysis to the text in advance and identify anaphoric relations appropriately and supplement the omission appropriately, you will find appropriate answers in the text that could not be answered before. It becomes possible to do.

  Furthermore, when general linguistic knowledge is collected from a large amount of text, the coverage of knowledge acquisition from a single sentence will be improved if the anaphoric relationship is specified appropriately and omission complementation is appropriately performed in advance. It is thought that it can contribute to the expansion of the scale of knowledge that can be acquired.

<Third Embodiment>
In the above embodiment, the term sharing knowledge is used before the anaphora / omission analysis and during the anaphora / analysis process. However, the present invention is not limited to such an embodiment. For example, the analysis result can be verified and corrected by applying the term sharing knowledge to the output of anaphora and omission analysis that does not use the term sharing knowledge. FIG. 23 shows the function of a program that realizes such annotation processing in the form of a schematic block diagram.

  Referring to FIG. 23, the program for verifying / correcting the anaphora / abbreviation analysis result using the term sharing knowledge uses the term sharing knowledge 910 to verify the pointing destination for the output 911 of the anaphora / omission analyzer. For generating an output 912 corrected if necessary. For example, in the output 911 of the anaphora / abbreviation analyzer, it is annotated 913 that the designation of “shown” principal is “Prime Minister Tomi Murayama”, and the designation of omission of “description” is incorrect. It is assumed that it had been a “retired party” (annotation 914). In this case, since the program that verifies and corrects the term sharing relationship does not share the terms, first the term sharing knowledge including the two expressions “show” and “describe” is searched. . If there is such term sharing knowledge 910, the program for verifying / verifying the term sharing relationship collates 915 with the term sharing knowledge 910 and the output 911. Specifically, the program extracts a feature amount for determining whether “shown” and “description” share the main character from the output 911 of the anaphoric / omitted analyzer, and shares the term including SVM. It discriminates with a discriminator. This term sharing discriminator is the same as that shown in FIGS. If it does not share, the program does nothing further. If the result indicates that the result is shared, the program indicates the lower value of the scores of the predicate pairs to be processed that are attached to the annotation as the reliability of the target in the output of the anaphoric / omitting analyzer. Rewrite the tip with the higher point. As a result, for example, if the score of the annotation 914 shown in FIG. 23 is low, the new annotation 916 is changed to the same “Prime Minister Murayama Tomiichi” as that of the annotation 913, and is correctly output 912. However, when the reliability (score) of the term sharing determination by this program is low, such rewriting may not be performed. If one of the predicate pairs is not omitted, the pointer that is annotated as omitted may be rewritten with the pointer that is not omitted.

  Referring to FIG. 24, the program that executes this annotation includes the following processing 922 for each predicate pair that includes anaphora and omission among the outputs of the anaphora and omission analyzer and does not share a term. Step 920 is performed.

  The process 922 determines whether or not the pointed to by the annotation of the predicate pair to be processed matches. If the result of the determination is affirmative, the process for the predicate pair is terminated in step 924 and the determination of step 924 Is negative, step 926 that searches for the term shared knowledge that matches the predicate pair being processed, and ends the processing for the current predicate pair when the determination is negative, and when the determination of step 926 is affirmative And the step 928 of rewriting the pointing point with the lower score by the anaphoric / omitting analyzer of the two annotations to the pointing point with the higher score, and ending the processing for this predicate pair. The score of the annotation may be a score (a distance from the separation plane to the data point in the case of SVM) based on a machine learning model that realizes anaphora / omission analysis.

  By performing such processing, it is possible to verify and correct the results of the anaphora / analysis processing afterwards using the term sharing knowledge.

<Fourth embodiment>
<Outline>
In the first to third embodiments, the term sharing discriminators 306, 307, and 308 are learned using the annotations in the unannotated text archive 302, and the term sharing knowledge is acquired using these. . Each of the term sharing discriminators 306, 307, and 308 discriminates whether or not each predicate pair individually shares a term. However, the present invention is not limited to such an embodiment. There may be cases where there are three or more predicates that share a term. In such a case, if the anaphora analysis is performed by applying the term sharing knowledge to each predicate pair individually, the results may be inconsistent or the judgment conditions may be complicated. Therefore, it is desirable that predicate groups that share terms with each other can be collectively determined. In the fourth embodiment, this is solved by introducing the concept of a term shared predicate network.

  The term shared predicate network refers to a network obtained by using a group of predicates sharing terms as vertices and connecting the vertices with edges. Since it represents a group of predicates that share a term, it may be called a term sharing group. When such a network is considered, if the value of the term (for example, the subject) is known for one predicate that forms a vertex, the value of the term is propagated to other vertices of the same network, so that each The value of a predicate group term is defined. When the value of the term can exist for a plurality of predicates in the same network, the most reliable value among them may be determined as the value of the term of each predicate group.

  In this embodiment, unlike the term sharing discriminators 306, 307 and 308 used in the first embodiment, three types of term sharing discriminators corresponding to patterns in which predicates having a term sharing relationship appear are used. To do. There are three types of patterns. DEP type, ADJ type, and PNP type. Hereinafter, these will be described in order.

―DEP type―
The DEP type is a typical term sharing pattern, and refers to a case where predicate pairs are dependent on each other in a dependency relationship. For example, there is a relationship between two predicates “Done” and “Drinked” in the sentence “Tom got flu, so _I took medicine (φi)”. This is because they have the same subject (Tom, φ _i ). This dependency relationship is shown in FIG. Such a term sharing relationship is called a DEP type.

─ADJ type─
Predicates located adjacent to each other in a sentence tend to have the same subject, but they are not always in a DEP-type term sharing relationship. This is because there is a long-range dependency relationship between two predicates. For example, in the sentence "The plane has landed successfully but the control sticks have stopped working after entering the taxiway." The two predicates "Landed" and "Entered" have the same subject, but are directly related. Does not have a relationship. This relationship is shown in FIG. In order to clarify such a term sharing relationship, cases where two predicates appear adjacent to each other in the surface layer structure are also classified into the ADJ type.

─PNP type─
In the case of Japanese predicate pairs, the predicate often depends on a noun (or noun phrase) that is syntactically related, and the noun (or noun phrase) depends on another predicate in the syntactic relation. is there. For example, it is the case of “withdrawn” and “clarified” in the sentence “The governor has revealed the policy (φ _i ) has withdrawn the clause”. This relationship is shown in FIG. This type of term sharing relationship is called a PNP type.

In the present embodiment, such a term sharing relationship (particularly subject) is detected as a binary identification problem, and a discriminator that determines whether two predicates share the same subject is configured. Here, the discriminator is learned using supervised learning. For this purpose, for example, each type of learning data is extracted separately from an annotated text archive as shown in FIG. As an annotation, at least, a (zero) pronoun relationship, a coreference relationship, and a subject need only be annotated. If the two extracted predicates share a coreference relationship and a (zero) pronoun relationship, the label is positive, and if not, the label is negative, and learning data is created with a predetermined feature. In this embodiment, SVM ^light is used for the discriminator. The features used for learning each relationship are shown in Table 2 below.

<Constitution>
FIG. 28A shows a block diagram of a DEP-type term sharing discriminator learning device 1012 that performs learning of the DEP-type term sharing discriminator 1014 in this embodiment, and FIG. 28B shows an ADJ-type term sharing discriminator 1054. FIG. 28C is a block diagram of an ADJ-type term sharing discriminator learning device 1092 that performs learning of PNP type term sharing discriminator 1094, and FIG. Show. The DEP type term sharing discriminator learning device 1012 is manually extracted from, for example, the learning data DB 162 and stored in the DEP type text storage unit 1010 using DEP type term sharing relation learning data. The device 1014 is learned. Similarly, the ADJ type term sharing discriminator learning device 1052 performs learning of the ADJ type term sharing discriminator 1054 using ADJ type term sharing relation learning data stored in the ADJ type text storage unit 1050. The PNP type term sharing discriminator learning device 1092 performs learning of the PNP type term sharing discriminator 1094 using the PNP type term sharing learning data stored in the PNP type text storage unit 1090. The DEP type text storage unit 1010, the ADJ type text storage unit 1050, and the PNP type text storage unit 1090 are all marked as positive examples prepared for learning the classifiers of each type of term sharing relationship. It stores sentence examples and sentence examples marked as negative examples.

  The DEP type shared terminator learning device 1012, the ADJ type shared terminator learning device 1052, and the PNP type shared terminator learning device 1092 are all term shared discriminator learning according to the first embodiment shown in FIG. It has the same configuration as the device 282. For example, the DEP-type term shared discriminator learning device 1012 reads out the positive example sentence and the negative example sentence stored in the DEP type text storage unit 1010, extracts the features as shown in the upper part of Table 2, and extracts the positive examples. Or a learning data generation unit 1030 that generates and outputs learning data labeled with a negative example, a learning data storage unit 1032 that stores the learning data, and learning data stored in the learning data storage unit 1032. A learning processing unit 1034 that performs training of the DEP-type term sharing classifier 1014. Similarly, the ADJ-type term sharing discriminator learning device 1052 includes a learning data generation unit 1070, a learning data storage unit 1072, and a learning processing unit 1074. The PNP-type term sharing discriminator learning device 1092 includes a learning data generation unit 1110, a learning data storage unit 1112, and a learning processing unit 1114.

  Here, the learning data generation unit 1030 and the learning data generation unit 1070 generate the learning data by extracting the features shown in the upper part of Table 2. On the other hand, the learning data generation unit 1110 generates learning data by extracting the features added to the lower stage in addition to the features shown in the upper stage of Table 2.

  Referring to FIG. 29, the term sharing knowledge collection device 1130 according to this embodiment includes a DEP type term sharing discriminator 1014, an ADJ type term sharing discriminator 1054 that have been learned as described above, and The term shared knowledge is extracted from the unannotated text archive 302 by using the PNP type term shared discriminator 1094 or the like, and is replaced with the term shared knowledge collection device 301 (see FIG. 8) according to the first embodiment. Can be used.

  The term sharing knowledge collection device 1130 includes a predicate pair extraction unit 303 that extracts all predicate pairs from the unannotated text archive 302, and each predicate pair extracted by the predicate pair extraction unit 303 in a sentence in which they appear. It is classified into one of DEP type, ADJ type, and PNP type according to the dependency relationship and dependency relationship, and is classified into the DEP type term sharing classifier 1014, the ADJ type term sharing classifier 1054, and the PNP type term sharing classifier 1094, respectively. A distribution unit 1140 for determining whether there is a sharing relationship of each type, a term sharing relationship by a DEP type term sharing classifier 1014, an ADJ type term sharing classifier 1054, and a PNP type term sharing classifier 1094 A term sharing knowledge storage unit 310 that stores information representing the discrimination result of the item as a term sharing knowledge together with the predicate pair that is the basis of the determination

  On the other hand, referring to FIG. 30, in this embodiment, an antecedent identifier 1150 shown in FIG. 30 is used in place of the antecedent identifier 170 in the anaphoresis / omission analyzer 252 shown in FIG. The antecedent specifying unit 1150 stores a predicate extraction unit 1162 that extracts all predicates from the input 250 to be subjected to anaphora and omission analysis, a predicate DB 1164 that stores predicates extracted by the predicate extraction unit 1162, and a predicate DB 1164. 28. A predicate pair extraction unit 1166 that extracts all predicate pairs from the set of predicates thus obtained, and for each predicate pair extracted by the predicate pair extraction unit 1166, a DEP type term sharing discriminator 1014, ADJ as shown in FIG. Type term sharing discriminator 1054, PNP type term sharing discriminator 1094 and other term sharing discriminators are respectively determined to determine whether or not the predicate pair has a term sharing relationship. A term sharing determination unit 1168 for updating the predicate DB 1164 so as to attach the same label to both of the predicate pairs.

  The term sharing determination unit 1168 updates the predicate DB 1164 as follows when there is a term sharing relationship between the predicate pairs to be processed. If both of the predicate pairs in which the term sharing relationship is found are not yet labeled, a new predicate group is created and a label indicating the group is attached to both of the predicate pairs. Therefore, the term sharing determination unit 1168 needs to have a label DB (not shown) for storing the labels of the groups created so far. In the case where only one of the predicate pairs is labeled and the other is not labeled, the other label is attached to the unlabeled side. When labels are attached to both of the predicate pairs, the labels are unified to one of them, and the labels of the other predicates having the other label are unified to the one label. The other label is deleted from the label DB. By doing so, predicate pairs having a term sharing relationship are grouped by label. Thus, when the labels are sequentially unified, there is an effect that the processing becomes easier as compared with the case where the terms shared for a plurality of predicate pairs are individually determined and then further unified.

  The antecedent identifier 1150 further includes a term shared predicate group search unit 1170 that searches for the term shared predicate group with the same label from the predicate DB 1164 for each group after the update of the predicate DB 1164 by the term sharing determination unit 1168 is completed. A term shared predicate network generation unit 1172 that generates a term shared predicate network based on the predicates belonging to each group searched by the term shared predicate group search unit 1170.

  Here, the term shared predicate network refers to a group of a plurality of predicates sharing the same term. The data format of the term shared predicate network may be such that all predicates belonging to the same group can be traced from predicates belonging to a certain group. For example, it may be a list format or an array composed of arrays having group predicates as elements.

  The antecedent identifier 1150 further has to share a term with other predicates by the term shared predicate network DB 1174 that stores the term shared predicate network generated by the term shared predicate network generating unit 1172 and the term shared predicate network generating unit 1172. For each of the non-term shared predicate storage unit 1180 that stores the determined predicate and the term shared predicate network stored in the term shared predicate network DB 1174, a shared term that identifies a term shared by predicates belonging to the term shared predicate network By referring to the determination unit 1176 and the term shared predicate network DB 1174, the values of the terms determined by the shared term determination unit 1176 are propagated to all predicates belonging to the same term shared predicate network, and terms corresponding to those predicates. The shared term propagation unit 11 that substitutes (embeds) the value into 8 and the predicates stored in the non-term shared predicate storage unit 1180 and the predicates belonging to the term shared predicate network determined by the shared term determining unit 1176 that the value of the shared term (subject) cannot be specified The supplementary determination unit 1182 that supplementarily performs the process of specifying the term (subject) of each predicate using the existing term determination method, the processing result by the shared term propagation unit 1178, and the processing result by the supplemental determination unit 1182, An antecedent update unit 1184 that updates the input 250 and outputs the output 254 by embedding the term as an antecedent in the omitted part of the input 250.

  An existing one can be used as the replenishment determination unit 1182. For example, the technique described in Non-Patent Document 1 or the technique described in Reference Document 2 can be used. Also, as described in the description of the prior art, in addition to the information output by morphological analysis, syntax / dependency analysis, etc., for example, the characteristic that “food” is included in the object of “eating” is summarized It is also possible to apply automatic analysis technology using a dictionary.

  FIG. 31 shows a flowchart of a computer program for realizing the antecedent identifier 1150 shown in FIG. 30 by a computer. Referring to FIG. 31, this program performs step 1190 for performing morphological analysis, syntax analysis, and dependency analysis for input 250 (see FIG. 30), and uses all the predicates in input 250 by using the analysis result. Step 1192 for extracting and registering each predicate in the predicate DB 1164 together with an empty label, and determining whether or not there is a term sharing relationship for all predicate pairs, and using the determination result, the term shared predicate belonging to each predicate Step 1194 for performing a process of labeling a group, Step 1196 for generating a term shared predicate network using the result of Step 1194, and Step 1198 for determining a value of a term shared for each term shared predicate network Step 120 for determining an abbreviated portion that could not be determined by the processing up to Step 1198 using a conventional method or the like. Including the door.

  The processing in step 1194 includes step 1210 in which the processing in steps 1212 and 1214 is repeatedly executed for all predicate pairs.

  In step 1212, it is determined whether the predicate pair to be processed shares a term. If the predicate pair shares a term, in step 1214, the predicate constituting the predicate pair is attached with a label indicating that the term is shared, and the predicate DB is updated. The label here represents a group of predicates that share a term. If both predicate pairs are not labeled, a new label is added to update both, and the label is stored in the label DB. When only one of the predicate pairs is labeled, the same label is attached to the other predicate pair and the predicate DB is updated. If labels are attached to both predicate pairs, one of the labels is used to unify the labels. Labels not adopted here are deleted from the label DB.

  Step 1196 includes a step 1230 for searching for predicates by sorting from the predicate DB using labels as keys, and a step 1232 for executing the processing of step 1234 for each label on the searched predicates. In step 1232, a term sharing predicate network is constructed using predicates with the same label.

  Step 1198 includes a step 1250 that executes the processing of steps 1252, 1254, and 1256 for each term-shared predicate network generated in step 1196.

  In step 1252, an attempt is made to determine the value of the shared term for each predicate in the processing-target term shared predicate network. Here, if the value of only one term can be determined, that value is adopted. When there are two or more term value candidates, any one of them is determined by some criterion. For example, you may employ | adopt the thing with the highest reliability of determination. If no term value candidate is found, the process proceeds to the next process. In the subsequent step 1254, it is determined in step 1252 whether or not the value of the term has been determined. If the value of the term has been determined, the value is propagated to all predicates in the same term sharing predicate network. If the value of the term cannot be determined, the process proceeds directly to the next process.

  In step 1200, the value of the location where the value of the shared term could not be determined by the processing up to step 1198 is determined by another method (for example, a conventional method).

<Operation>
The operation of this term shared predicate network is roughly divided into three phases. The first phase is learning of the DEP type term sharing discriminator 1014, the ADJ type term sharing discriminator 1054, and the PNP type term sharing discriminator 1094 shown in FIG. The second phase is the collection of term sharing knowledge by the term sharing knowledge collection device 1130 using the DEP type term sharing discriminator 1014, the ADJ type term sharing discriminator 1054, and the PNP type term sharing discriminator 1094 shown in FIG. It is. The third phase is processing for specifying an antecedent at a place omitted in the input 250 by using the term sharing knowledge or the term sharing discriminator by the antecedent specifying device 1150 shown in FIG.

  The learning of the DEP type term sharing discriminator 1014, the ADJ type term sharing discriminator 1054, and the PNP type term sharing discriminator 1094 shown in FIG. 28 is very similar to each other. For example, the learning of the DEP type term sharing discriminator 1014 is performed as follows. First, learning data including text including a predicate pair for a DEP-type term sharing classifier is manually collected and stored in the DEP-type text storage unit 1010. At this time, a label indicating whether or not the predicate pair has a DEP-type term sharing relationship is attached to the learning data. The learning data generation unit 1030 generates learning data including features and labels as shown in Table 2 from these texts, and stores the learning data in the learning data storage unit 1032. The learning processing unit 1034 uses the learning data stored in the learning data storage unit 1032 to perform learning of the DEP type term sharing classifier 1014 made of SVM. The DEP-type term sharing discriminator 1014 after learning can determine whether or not they have a term sharing relationship when given a text including two predicate pairs.

  The same applies to learning of the ADJ type term sharing classifier 1054 by the ADJ type term sharing classifier learning device 1052 and learning of the PNP type term sharing classifier 1094 by the PNP type term sharing classifier learning device 1092.

  Referring to FIG. 29, the term shared knowledge collection by the term shared knowledge collection device 1130 is performed as follows. The predicate pair extraction unit 303 extracts all predicate pairs from the unannotated text archive 302. The sorting unit 1140 performs syntax analysis and dependency analysis of the text including each extracted predicate pair, and determines the predicate pair as one of the DEP type, the ADJ type, and the PNP type based on the dependency between the words. And is given to any one of the DEP-type term sharing discriminator 1014, the ADJ-type term sharing discriminator 1054, and the PNP-type term sharing discriminator 1094 to determine the presence or absence of the term sharing relationship. The DEP type term sharing discriminator 1014, the ADJ type term sharing discriminator 1054, and the PNP type term sharing discriminator 1094 determine whether the given predicate pair is a DEP type, an ADJ type, and a PNP type, respectively. If there is a relationship, the term is specified, and information (dependency relationship, information on words, parts of speech, etc.) that enables the term sharing relationship to be determined later is output to the term shared knowledge storage unit 310. The term shared knowledge storage unit 310 accumulates this information as term shared knowledge.

  On the other hand, when performing anaphora / abbreviation analysis on the input sentence, the antecedent identifier 1150 operates as follows. The predicate extraction unit 1162 extracts all predicates from the input 250, and registers them in the predicate DB 1164 together with empty labels (steps 1190 and 1192 in FIG. 31).

  When the predicate extraction by the predicate extraction unit 1162 is completed, the predicate pair extraction unit 1166 extracts all the predicate pairs included in the predicate DB 1164 and gives them to the term sharing determination unit 1168 (step 1210 in FIG. 31). The term sharing determination unit 1168 determines whether or not each predicate pair has a term sharing relationship (FIG. 31, step 1212). If there is a term sharing relationship, a term is added to both of the predicates constituting the predicate pair. The predicate DB 1164 is updated with a label indicating sharing. The labeling method at this time is as described above, and the same label is attached to predicates that share the same term.

  When the determination of the term sharing relationship is completed for all predicate pairs in this way, the term sharing predicate group search unit 1170 searches the predicate DB 1164 for predicates by label (FIG. 31, step 1230). The term shared predicate network generation unit 1172 generates a term shared network with predicates having the same label for each label (FIG. 31, steps 1232 and 1234), and registers them in the term shared predicate network DB 1174. Predicates that do not belong to the term shared predicate network, predicate pairs that do not form a group with other predicate pairs, and the like are output to the non-term shared predicate storage unit 1180.

  The shared term determination unit 1176 attempts to determine the value of the shared term for each term sharing predicate network (FIG. 31, steps 1250 and 1252). If the value is determined (YES in step 1254), the shared term propagation unit 1178 propagates the value of the determined term to all the predicates belonging to the same term shared network (step 1256). On the other hand, the supplement determination unit 1182 uses the predicate stored in the non-term shared predicate storage unit 1180 and the value of the predicate term belonging to the term shared predicate group whose shared term value cannot be determined by the shared term determination unit 1176. It is determined by some method such as a conventional method.

  The antecedent update unit 1184 integrates the propagation result from the shared term propagation unit 1178 and the determination result from the supplement determination unit 1182 to generate and output an output 254.

  As described above, according to the present embodiment, an appropriate term sharing determiner is configured according to a sentence pattern, and among predicates included in an input, those sharing a term are specified. Furthermore, a term sharing predicate network is generated by predicates that share the same term, a word shared by the term sharing predicate network predicate is determined by some method, and the value is assigned to all the terms belonging to the same term sharing predicate network. Propagate. Therefore, when there are a plurality of omitted parts in the input 250 and the values indicated by them are the same, the omitted parts can be identified efficiently.

  For example, consider the text “The government is planning to dispatch 50 people to the disaster area and is preparing for it”. In this text, three predicates “dispatch”, “plan”, and “progress” appear. By applying the term sharing relation discriminator, for example, it is possible to specify that a term is shared for two predicate pairs of “dispatch”, “plan”, “plan”, and “progress”. In that case, the three predicates related to these ("dispatch", "plan", "progress") are considered to be predicates having the same term. Therefore, these three predicates can be grouped. When a term can be specified for any of the three predicates belonging to this group, an omission analysis can be realized by performing a process of propagating it to other predicates belonging to this group. For example, when the subject “government” can be identified with respect to “advance”, the subject omission is complemented by complementing the subject to the positions of the subjects “dispatch” and “plan”.

  Omission analysis can be performed robustly by applying existing abbreviation analysis techniques to problems that cannot be analyzed by such abbreviation analysis techniques that focus on term sharing.

[Reference 1]
Chikara Hashimoto, Kentaro Torisawa, Julien Kloetzer, Motoki Sano, Istvan Varga, Jong-Hoon Oh, and Yutaka Kidawara. 2014.Toward future scenario generation: Extracting event causality exploiting semantic relation, context, and association features.In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 987-997.
[Reference 2]
Ryohei Sasano and Sadao Kurohashi. 2011.A discriminative approach to Japanese zero anaphora resolution with large-scale lexicalized case frames.In Proceedings of 5th International Joint Conference on Natural Language Processing, pages 758-766.

  The embodiment disclosed herein is merely an example, and the present invention is not limited to the above-described embodiment. The scope of the present invention is indicated by each claim of the claims after taking into account the description of the detailed description of the invention, and all modifications within the meaning and scope equivalent to the wording described therein are included. Including.

[Realization by computer]
Learning data generation assisting device 204, detector learning device 222, anaphora / omission analyzer 252, DEP type term sharing discriminator learning device 1012, ADJ type term sharing discriminator learning device 1052, PNP type term sharing according to the above embodiment Each of the discriminator learning device 1092, the term sharing knowledge collection device 1130, and the antecedent identifier 1150 can be realized by computer hardware and a computer program executed on the computer hardware. FIG. 32 shows the external appearance of this computer system 1330, and FIG. 33 shows the internal configuration of the computer system 1330.

  Referring to FIG. 32, this computer system 1330 includes a computer 1340 having a memory port 1352 and a DVD (Digital Versatile Disc) drive 1350, and a keyboard 1346, a mouse 1348, and a monitor 1342 that are all connected to the computer 1340. Including.

  Referring to FIG. 33, in addition to the memory port 1352 and the DVD drive 1350, the computer 1340 includes a CPU (Central Processing Unit) 1356, a bus 1366 connected to the CPU 1356, the memory port 1352, and the DVD drive 1350, and a boot program. A read-only memory (ROM) 1358 for storing etc., a random access memory (RAM) 1360 connected to the bus 1366 for storing program instructions, system programs, work data and the like, and a hard disk 1354 are included. The computer system 1330 further includes a network interface (I / F) 1344 that provides a connection to a network 1368 that allows communication with other terminals.

  A computer program for causing the computer system 1330 to function as each function unit of the learning data generation assisting device 204 according to the above-described embodiment is stored in the DVD 1362 or the removable memory 1364 mounted in the DVD drive 1350 or the memory port 1352, Further, it is transferred to the hard disk 1354. Alternatively, the program may be transmitted to the computer 1340 through the network 1368 and stored in the hard disk 1354. The program is loaded into the RAM 1360 when executed. The program may be directly loaded into the RAM 1360 from the DVD 1362, from the removable memory 1364, or via the network 1368.

  This program includes an instruction sequence including a plurality of instructions for causing the computer 1340 to function as each functional unit of the learning data generation assisting device 204 according to the above embodiment. Some of the basic functions necessary to cause computer 1340 to perform this operation are an operating system or third party program that runs on computer 1340 or various dynamically linkable programming toolkits or programs installed on computer 1340 Provided by the library. Therefore, this program itself does not necessarily include all functions necessary for realizing the system and method of this embodiment. This program can be used as a system as described above by dynamically calling the appropriate program in the appropriate function or programming toolkit or program library at run time in a controlled manner to achieve the desired result. It is only necessary to include an instruction for realizing the function. Of course, all necessary functions may be provided only by the program.

150 Anaphoric / Omission Analysis System 160 Annotation Data Generation Auxiliary System 162 Learning Data DB
164 Learning system 166 Omission detector 168 Anaphoric detector 170 Antecedent identification device 172 Anaphoric / omission analysis device 190 Existing small learning data DB
192 Text Archive 194, 228, 256 Language Knowledge DB
196 Paraphrase / Category Dictionary 198 Pattern Dictionary 200, 220, 258 Question Type DB
202 I / O device 204 Learning data generation auxiliary device 206, 226, 260 Question answering system 222 Detector learning device 224, 262 Selection restriction DB
252 Anaphoric / omission analyzer 278 Term sharing knowledge 282, 972 Term sharing discriminator learning device 283, 974 Term sharing discriminator 1014 DEP type term sharing discriminator 1054 ADJ type term sharing discriminator 1094 PNP type term sharing discriminator 1130 Term sharing Knowledge Collection Device 1150 Antecedent Identifier 1162 Predicate Extraction Unit 1164 Predicate DB
1168 term sharing determination unit 1170 term shared predicate group search unit 1172 term shared predicate network generation unit 1176 shared term determination unit 1178 shared term propagation unit 1182 supplement determination unit

Claims (6)

An anaphoric / elliptical analysis device that performs anaphoric / elliptical analysis of input text using term sharing knowledge about predicate pairs sharing terms,
A term sharing knowledge storage means for storing knowledge term sharing knowledge regarding whether or not two cases of two predicates share a term;
A term sharing analysis unit that generates information related to sharing a term of a predicate pair in the input text by analyzing the input text using the term sharing knowledge;
The input text is subjected to anaphora / omission analysis using the analysis result of the term sharing analysis means by automatic processing, and the annotated text regarding the anaphora / omission part and its pointing destination is output. An anaphoric / abbreviated analysis device including an anaphoric / abbreviated analyzing means. The anaphoric / abbreviated analysis means is:
Network generation means for generating a term sharing predicate network consisting of a plurality of predicates sharing the same term for the predicate pair determined to share the term by the term sharing analysis means;
For each of the term shared predicate networks generated by the network generating means, term determining means for determining a value of a term shared by predicates belonging to the term shared predicate network;
The anaphoresis / omission analysis device according to claim 1, further comprising value propagation means for performing annotation by propagating a value determined by the term determination means to a term shared by predicates in the term shared predicate network. The term determining means includes
For each predicate belonging to each of the term shared predicate networks generated by the network generating means, candidate determining means for determining a candidate value of a term that the predicate shares with other predicates;
Means for determining a value of a term shared by predicates belonging to the term sharing predicate network as the candidate, when there is one value candidate determined by the candidate determining means;
When there are a plurality of candidate values determined by the candidate determination unit, the candidate having the highest reliability when each candidate is determined by the candidate determination unit is selected, and the predicate belonging to the term shared predicate network is selected. The anaphoresis / omission analysis apparatus according to claim 2, further comprising: means for determining a value of a shared term as the candidate.   The anaphoric / abbreviated analysis device further shares, for each predicate belonging to a certain term shared predicate network, a predicate belonging to the term shared predicate network when there is no value candidate determined by the candidate determining means. 4. The anaphoresis / omission analysis device according to claim 3, further comprising means for determining a value of a term by a method different from the anaphora / omission analysis means.   The anaphoric / abbreviated analysis device further uses a method different from the anaphoric / abbreviated analyzing unit for the predicate determined not to share a term with another predicate by the term sharing analyzing unit in the input text, The anaphoresis / omission analysis device according to claim 3 or 4, further comprising means for specifying a value of a term indicated by the predicate. A computer program that causes a computer to function as all means of the anaphoric / omitted analysis device according to any one of claims 1 to 5.

Images