JP2015028697A - Classification model learning device, classification determining device, method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To learn a classification model that accurately determines whether or not a predicate pair or an "argument-predicate" pair is synonymous, antonymous, or irrelevant.SOLUTION: A classification model learning device is configured to: extract, by a synonymous feature extraction unit 224, at least one of a dictionary definition sentence feature and a sematic attribute feature regarding each predicate pair to which any of synonym, antonym, or irrelevance stored in a learning corpus storage unit 222 is added in advance as classification; extract, by an antonymous feature extraction unit 226, a feature representing a substring and a feature representing co-occurrence likelihood within a sentence; construct, by a synonymous and antonymous-irrelevant classification feature construction unit 230, a feature for classifying synonym-and-antonym and irrelevance; construct, by a synonymous-antonymous classification feature construction unit 240, a feature for classifying synonym and antonym; learn, by a synonymous and antonymous-irrelevant classification model learning unit 234, a classification model for classifying synonym-and-antonym and irrelevance; and learn, by a synonymous-antonymous classification model learning unit 244, a classification model for classifying synonym and antonym.

Description

  The present invention relates to a classification model learning device, a classification determination device, a method, and a program, and more particularly, to a classification model learning device, a classification determination device, a method, and a program for classifying a relationship between predicate pairs.

  Currently, there is a demand for high-precision search technology for searching for desired information from a large amount of text information such as blogs and voice dialogue logs on the Web, and text mining technology for automatically extracting and counting only useful information. In order to realize these, it is necessary to understand the meaning of natural sentences by a computer.

  For example, if there are two sentences: (1) “I was satisfied with the lunch of XX” and (2) “I enjoyed the lunch of XX”, these were judged as “representing the same thing”. If it is not possible, the information required by the user cannot be searched correctly or the “same information gathering” necessary for text mining cannot be performed.

  Also, if there were two sentences (3) “XX lunch was delicious”, (4) “XX lunch was delicious”, these were “true opposites (conflicting opinions). Unless it can be discriminated as "representing", it is not possible to identify inconsistencies in information or conflicts of opinions with a computer.

  If the computer can recognize the synonymity between the sentence (1) and the sentence (2), or if the computer can identify the anomaly relationship between the sentence (3) and (4), Analysis from a new perspective such as analysis of opinions becomes possible.

  As a conventional method of acquiring words having a “synonymous” or “anti-sense” relationship, there is a method of acquiring a large number of word pairs having similar meanings based on similarity of context such as distribution similarity. In addition, there is a method in which a synonym and an antonym are sorted using an antonym list for a large number of word pairs acquired using the distribution similarity (Non-patent Document 1).

  Distribution similarity is based on the idea that words with similar meanings are similar in the context in which the word appears. Based on how often the element appears, it is calculated whether the word appears in a similar context.

Lin, D., Zhao, S., Qin, L., and Zhou, M. (2003). Identifying synonyms among distributionally similar words.Proceedings of the 18th International Joint conference on Artificial Intelligence (IJCAI-03), 1492-1493

  However, in the method of acquiring a large number of word pairs with similar meanings based on the similarity of contexts such as the above distribution similarity, the words of anomaly appear in the same context as the words of the same meaning Therefore, there is a problem that a word pair of “antegiance” representing the opposite meaning is also acquired as “similar in meaning”.

  Further, the method of Non-Patent Document 1 cannot cope with a word that is not in the antonym list, and as a result, there is a problem that an antonym representing the opposite meaning is erroneously treated as “synonymous”.

  In the present invention, the classification is made to solve the above-mentioned problem, and it is determined with high accuracy whether the predicate pair or the “term-predicate” pair is synonymous, anonymity, or irrelevant. It is an object of the present invention to provide a classification model learning apparatus, method, and program for learning a model.

  It is another object of the present invention to provide a classification determination apparatus, method, and program for highly accurately determining whether a predicate pair or a “term-predicate pair” is synonymous, affirmative, or irrelevant.

  In order to achieve the above object, the classification model learning device according to the first invention provides each predicate of a predicate pair obtained from a definition sentence set consisting of definition sentences for each of a plurality of predicates prepared in advance. The feature indicating the mutual complementarity of the definition statement indicating whether or not there is a predicate as a pair in the definition statement of each predicate of the predescription portion pair extracted based on the definition statement is the first feature. , A semantic attribute common to the previous description part pair extracted based on the semantic attribute of each predicate of the previous description part pair, obtained from a semantic attribute set consisting of semantic attributes for each of a plurality of predicates prepared in advance Is the second feature, and the combination of the partial character string in the character string representing one predicate of the preceding description pair and the partial character string in the character string representing the other predicate is the third feature. , A parallel predicate syntax generated containing each predicate in the predescription pair The likelihood is the fourth feature, and for each of the predicate pairs in which any one of synonym, anomaly, and irrelevance is previously given as a classification, at least of the first feature and the second feature On the other hand, the feature extraction unit extracts a feature set including the third feature and the fourth feature, and the feature extraction unit extracts each of the predicate pairs to which synonyms or anomalies are assigned as classifications. The feature set extracted and the feature set extracted by the feature extraction unit for each of the predicate pairs assigned as categorized as irrelevant are irrelevant to the anonymity that is the same or anonymity of the preceding description pair. And the feature set extracted by the feature extraction unit for each of the predicate pairs to which synonyms are given as classifications. , Anti For each predicate pair assigned as a classification, the feature set extracted by the feature extraction unit is constructed as a synonym-anonymity classification feature for classifying the previous description unit pair into synonyms and anomalies. Classification for classifying predicate pairs as irrelevant or irrelevant based on the synonym-irrelevant classification feature building unit and the same-irrelevant classification feature building unit Classify predicate pairs into synonyms and anomalies based on the synonym-anonymity classification model learning unit that learns the model and the synonym-anonymity classification feature construction unit constructed by the synonym-anonymity classification feature construction unit And a synonym-anonymity classification model learning unit that learns a model for this purpose.

  The classification model learning method according to the second aspect of the present invention includes a feature extraction unit, an anti-independent classification feature construction unit, a synonym-anti-discipline classification feature construction unit, an anti-invalid classification model learning unit, and a synonym-non-sense. A classification model learning method in a classification model learning device including a classification model learning unit, wherein the feature extraction unit is obtained from a definition sentence set including definition sentences for each of a plurality of predicates prepared in advance. Indicates mutual complementarity of definition statements that indicate whether or not there is a pair of predicates in the definition statement of each predicate of the preceding description pair extracted based on the definition statement of each predicate of the pair of pairs Pre-description extracted based on the semantic attribute of each predicate of the pre-description part pair obtained from the semantic attribute set consisting of semantic attributes for each of a plurality of predicates prepared in advance. Semantic attributes that are common to group pairs The combination of the partial character string in the character string representing one predicate of the preceding description part pair and the partial character string in the character string representing the other predicate is set as a third feature, The likelihood of the parallel predicate syntax generated including each of the predicates is set as the fourth feature, and each of the predicate pairs in which any one of synonym, anomaly, and irrelevance is previously given as a classification, The feature set including at least one of the first feature and the second feature, the third feature, and the fourth feature is extracted, and the anti-independent classification feature construction unit is synonymous. Or, the feature set extracted by the feature extraction unit for each predicate pair assigned an anomaly as a classification and the feature extraction unit extracted for each predicate pair assigned an irrelevance as a classification A feature set is synonymous with a previous description part pair. Constructed as an anti-unrelated classification feature to classify as unrelated and unrelated, and the synonym-non-defective classification feature construction unit described above for each predicate pair to which synonym is given as a classification The feature set extracted by the feature extraction unit, and the feature set extracted by the feature extraction unit for each predicate pair to which an anomaly is assigned as a classification, Is constructed as a synonym-anonymous classification feature, and the same-unrelated classification feature learning unit is based on the anti-irrelevant classification feature constructed by the anti-irrelevant classification feature construction unit, A model for classifying predicate pairs as irrelevant and unrelated is learned, and the synonym-anonymity classification model learning unit is constructed by the synonym-anonymity classification feature construction unit. Based on the features, the predicate page Learn a model for classifying a into synonyms and anomalies.

  Thus, according to the first and second inventions, the feature set extracted for each predicate pair to which synonym or anomaly is given as a classification, and each predicate pair to which unrelated is given as a classification The feature set extracted for is constructed as an anti-irrelevant classification feature, and a model for classifying the feature set as unrelated and unrelated is learned. Learn a model to classify feature sets and feature sets extracted for each predicate pair with anomaly as a classification into synonyms and anomalies, and classify them into synonyms and anomalies By doing so, it is possible to learn a classification model that can determine with high accuracy whether or not a predicate pair is synonymous, anomaly, or irrelevant.

  In the first invention, the feature extraction unit may extract an n-gram score of the parallel predicate syntax as a likelihood of the parallel predicate syntax.

  In the first invention, the feature extraction unit calculates the co-occurrence of words appearing around the previous description unit in the text corpus for each predicate of the previous description unit pair, The distribution similarity obtained by calculating the word similarity is set as the fifth feature, and the predicate of the predescription part pair obtained from the semantic label set made up of the semantic labels of the respective functional expressions of the plurality of predicates prepared in advance. The semantic label common to the previous description part pairs extracted based on the semantic labels of the respective functional expressions is the sixth feature, the overlapping degree of the common semantic labels of the previous description part pair is the seventh feature, The likelihood of the compound word generated by combining each predicate of the description part pair is the eighth feature, and the two predicates of the previous description part pair represent the situation in the sentence representing the relationship between the situations. The degree of co-occurrence as two predicates For each of the predicate pairs in which any one of the synonyms, anomalies, and irrelevance is given as a classification, at least one of the first feature and the second feature, and the first feature 3 features, the 4th feature, and the 5th feature, the 6th feature, the 7th feature, the 8th feature, and the 9th feature. A set may be extracted.

  In the first invention, the feature extraction unit includes, as the likelihood of the compound word, an n-gram score of the compound word generated for a previous description unit pair, and an appearance frequency of the compound word in a corpus At least one of them may be extracted.

  The classification model learning device of the third invention is based on a definition sentence of each predicate of a “term-predicate” pair obtained from a definition sentence set including a definition sentence for each of a plurality of predicates prepared in advance. Extracted from the above-mentioned "term-predicate" pair of predicates, the feature indicating mutual complementarity of the definition sentence indicating whether or not there is a pair of predicates, and the "term-predicate" There is at least a predescription part among the features indicating the mutual complementarity of the definition sentence that indicates whether there is a pair of "term-predicate" in each definition sentence of the predicate of the predicate pair. The above-mentioned “term-predicate” obtained from a semantic attribute set consisting of semantic attributes for each of a plurality of predicates prepared in advance is a feature indicating mutual complementarity of definition statements indicating whether or not to perform Common to the "term-predicate" pair extracted based on the semantic attributes of each predicate in the pair A combination of a partial character string in a character string representing one predicate of the "term-predicate" pair and a partial character string in a character string representing the other predicate, with a taste attribute as a second feature Is the third feature, and the likelihood of the parallel predicate syntax generated including each of the predicates of the “term-predicate” pair is the fourth feature, and any of synonym, anomaly, and irrelevant For each of the “term-predicate” pairs given in advance as a classification, at least one of the first feature and the second feature, the third feature, and the fourth feature, A feature extraction unit that extracts feature sets including, and the feature set extracted by the feature extraction unit for each of the “term-predicate” pairs to which synonyms or anomalies are assigned as classifications, and unrelated are assigned as classifications For each of the "term-predicate" pairs, In addition, the feature set is constructed as an anti-irrelevant classification feature for classifying the “term-predicate” pair into an unrelated or unrelated non-conformity. For each of the “term-predicate pair” to which the synonym is assigned as a classification, for each of the feature set extracted by the feature extraction unit and the “term-predicate” pair to which an anonymity is assigned as a classification A synonym-anonymity classification feature construction unit that constructs the feature set extracted by the feature extraction unit as a synonym-anonymity classification feature for classifying the "term-predicate" pair into synonyms and anomalies And learning the model for classifying the “term-predicate” pair as irrelevant and irrelevant based on the anti-irrelevant classification feature constructed by the same-irrelevant classification feature construction unit. Anti-irrelevant classification model learning unit and the same synonym-anti-sense A synonym-anonymity classification model learning unit that learns a model for classifying a “term-predicate” pair into a synonym and anomaly based on the synonym-anonymity classification feature constructed by the classification feature construction unit; , Including.

  A classification model learning method according to a fourth aspect of the present invention includes a feature extraction unit, a contradiction-independent classification feature construction unit, a synonym-anonymity classification feature construction unit, an anti-inverse classification model learning unit, and a synonym-anonymity. A classification model learning method in a classification model learning device including a classification model learning unit, wherein the feature extraction unit is obtained from a definition sentence set including a definition sentence for each of a plurality of predicates prepared in advance. Whether or not there is a paired predicate in the definition statement of each predicate of the “term-predicate” pair extracted based on the definition statement of each predicate of the “term-predicate” pair A definition indicating whether or not there is a term of “term-predicate” as a pair in the definition statement of each predicate of the “term-predicate” pair Definition statement indicating whether or not at least the preceding description part exists among the features indicating sentence mutual complementarity Meaning of each predicate of the “term-predicate” pair obtained from a semantic attribute set consisting of semantic attributes for each of a plurality of predicates prepared in advance as a feature indicating mutual complementarity A partial character string in a character string representing one predicate of the “term-predicate” pair, with the second feature being a semantic attribute common to the “term-predicate” pair extracted based on the attribute The likelihood of the parallel predicate syntax generated by including each of the predicates of the "term-predicate" pair, with the combination of the substring and the partial character string of the character string representing the other predicate as the third feature For each of the “term-predicate” pairs in which any one of synonym, contradiction, and irrelevance is previously assigned as a classification, and the first feature and the second feature A feature set including at least one of the third feature and the fourth feature is extracted. The same-unrelated classification feature construction unit is provided with the feature set extracted by the feature extraction unit for each “term-predicate” pair to which synonym or anomaly is given as a classification, and irrelevant as a classification. In order to classify the feature set extracted by the feature extraction unit for each of the “term-predicate” pairs as being independent of the same or opposite sense of the “term-predicate” pair. The synonym-anonymous classification feature construction unit constructs the feature sets extracted by the feature extraction unit for each of the "term-predicate pair" to which the synonym is assigned as a classification. And the feature set extracted by the feature extraction unit for each of the “term-predicate” pairs to which an anomaly is assigned as a classification, and the “term-predicate” pair is classified into a synonym and an anomaly Synonymous to anonymity-anonymity classification feature The reciprocal-irrelevant classification model learning unit reconstructs the term-predicate pair based on the reciprocal-irrelevant classification feature constructed by the reciprocal-irrelevant classification feature construction unit. The synonym-anonymous classification model learning unit learns a model for irrelevant and unrelated classification, and the synonym-anonymity classification feature construction unit constructs the `` term '' -Learn a model for classifying predicate pairs into synonyms and anomalies.

  Thus, according to the third and fourth inventions, feature sets extracted for each of the “term-predicate” pairs to which synonyms or anomalies are assigned as classifications, and unrelated are assigned as classifications. The feature set extracted for each of the “term-predicate” pairs is constructed as a contra-independent class feature, and a model for classifying the class as irrelevant and irrelevant is learned. A feature set extracted for each of the “term-predicate” pairs and a feature set extracted for each of the “term-predicate” pairs to which the anomaly is assigned as a classification are constructed as a synonym-anonymity classification feature Then, by learning a model for classifying synonyms and anomalies, it is possible to determine with high accuracy whether the "term-predicate" pair is synonymous, anomaly, or irrelevant You can learn the model.

  Further, in the third invention, the feature extraction unit is configured to determine, for each of the “term-predicate” of the “term-predicate” pair, a word appearing around the “term-predicate” in the text corpus. For each of the distribution similarity calculated by calculating the co-occurrence and the similarity of neighboring words of the “term-predicate” pair and the predicate of the “term-predicate” pair in the text corpus, the “term- The co-occurrence of words appearing around the predicate of the predicate is calculated, and at least the “term-predicate” of the distribution similarities of the prevalence of the predicate of the “term-predicate” pair is calculated. The “term” obtained from a semantic label set consisting of semantic labels of functional expressions of each of a plurality of predicates prepared in advance is set as a fifth feature, which is a distribution similarity obtained by calculating the similarity of neighboring words of the “part” pair. -Based on the semantic label of each functional expression of the predicate of the "predicate" pair The meaning label common to the “term-predicate” pair issued is the sixth feature, the overlapping degree of the common semantic labels of the “term-predicate” pair is the seventh feature, and the “term- The likelihood of a compound word generated by compounding each predicate in the “predicate” pair is the eighth feature, and the two predicates in the “term-predicate” pair indicate a relationship between situations. The degree of co-occurrence as two predicates representing the situation in the above is a ninth feature, and any of the above-mentioned synonyms, anomalies, and irrelevances is assigned as a classification to each of the “term-predicate” pairs. For each, at least one of the first feature and the second feature, the third feature, the fourth feature, the fifth feature, the sixth feature, the seventh feature The feature collection including at least one of a feature, the eighth feature, and the ninth feature It may be extracted.

  The classification determination apparatus of the fifth invention is extracted based on a definition sentence of each predicate of a predicate pair obtained from a definition sentence set including a definition sentence for each of a plurality of predicates prepared in advance. A plurality of predicates prepared in advance with the first feature as a feature indicating mutual definition of definition statements indicating whether or not there is a pair of predicates in the definition statement of each predicate of the predescription portion pair A semantic attribute common to the previous description part pair extracted based on the semantic attributes of each predicate of the previous description part pair obtained from the semantic attribute set consisting of the semantic attributes for each The combination of the partial character string in the character string representing one predicate of the part pair and the partial character string in the character string representing the other predicate is a third feature, and the predicate of the predescription part pair The likelihood of the parallel predicate syntax generated including each is the fourth feature, and is input A feature extraction unit that extracts a feature set including at least one of the first feature and the second feature, the third feature, and the fourth feature, and the feature extraction Based on the feature set extracted by the part and the pre-learned model for classifying the predicate pair as irrelevant or irrelevant, the classification of the predescription part pair is the same. If the classification of the previous description part pair is determined to be affirmative by the reciprocity-irrelevance determination part for determining whether it is affirmative or irrelevant, and the reciprocity-irrelevance determination part, it is extracted by the feature extraction part Based on the feature set and a pre-trained model for classifying predicate pairs into synonyms and anomalies Part.

  A classification determination method according to a sixth aspect of the present invention is a classification determination method in a classification determination apparatus including a feature extraction unit, a contradiction-irrelevance determination unit, and a synonym-anonymity determination unit, wherein the feature extraction unit The definition of each predicate of the predescription part pair extracted from the definition sentence of each predicate of the predicate pair obtained from the definition sentence set including the definition statements for each of the plurality of prepared predicates. A semantic attribute set consisting of semantic attributes for each of a plurality of predicates prepared in advance, with a feature indicating mutual definition of definition statements indicating whether or not there is a pair of predicates in the statement as a first feature A character string representing one predicate of the previous description part pair with the second feature as the semantic attribute common to the previous description part pair extracted based on the semantic attributes of each predicate of the previous description part pair obtained from Of the character string and the character string representing the other predicate Is a third feature, the likelihood of the parallel predicate syntax generated including each of the predicates of the previous description unit pair is a fourth feature, and for the input predicate pair, the first feature A feature set including at least one of a feature and the second feature, the third feature, and the fourth feature is extracted, and the same-irrelevance determining unit is extracted by the feature extracting unit. Based on the feature set and a pre-learned model for classifying predicate pairs as synonymous or anonymity and irrelevant, the classification of the predescription part pair is synonymous or irrelevant. And the synonym-antisense determination unit determines that the feature extracted by the feature extraction unit is determined by the anti-unity determination unit to determine that the classification of the previous description part pair is synonymous. To classify sets and pre-trained predicate pairs into synonyms and anomalies Based of on the model, determine the classification synonymous or anti justified before describing section pairs.

  As described above, according to the fifth and sixth inventions, based on the feature set extracted for the input predicate pair and the model classified as irrelevant to the pre-learned reciprocity, the predicate If the classification of the pair is determined to be synonymous or irrelevant, and determined to be synonymous, based on the extracted feature set and a model for classifying into pre-learned synonyms and anomalies, By determining whether the predicate pair classification is synonymous or anomaly, it is possible to determine with high accuracy whether the predicate pair is synonymous, anomaly, or irrelevant.

  According to the fifth invention, the feature extraction unit calculates the co-occurrence of words appearing around the previous description unit in the text corpus for each predicate of the previous description unit pair, The distribution similarity obtained by calculating the word similarity is set as the fifth feature, and the predicate of the predescription part pair obtained from the semantic label set made up of the semantic labels of the respective functional expressions of the plurality of predicates prepared in advance. The semantic label common to the previous description part pairs extracted based on the semantic labels of the respective functional expressions is the sixth feature, the overlapping degree of the common semantic labels of the previous description part pair is the seventh feature, The likelihood of the compound word generated by combining each predicate of the description part pair is the eighth feature, and the two predicates of the previous description part pair represent the situation in the sentence representing the relationship between the situations. The ninth feature determines the degree of co-occurrence as two predicates And about the inputted predicate pair, at least one of the first feature and the second feature, the third feature, the fourth feature, the fifth feature, The feature set including at least one of six features, the seventh feature, the eighth feature, and the ninth feature may be extracted.

  The classification judging device of the seventh invention is based on the definition sentences of the predicates of the “term-predicate” pair obtained from the definition sentence set including the definition sentences for each of the plurality of predicates prepared in advance. The extracted feature that indicates whether or not there is a pair of predicates in the definition statement of each predicate of the “term-predicate” pair, and the “term-predicate” There is at least a pre-description part among the features indicating the mutual complementarity of the definition sentence that indicates whether or not there is a pair of "term-predicate" terms in each of the definition sentences of the "part" pair predicates The above-mentioned “term-predicate” obtained from a semantic attribute set consisting of semantic attributes for each of a plurality of predicates prepared in advance is a feature indicating mutual definition of definition sentences indicating whether or not Semantic attributes common to the "term-predicate" pair extracted based on the semantic attributes of each predicate of the pair Is the second feature, and the combination of the partial character string in the character string representing one predicate of the “term-predicate” pair and the partial character string in the character string representing the other predicate is the first feature. And the likelihood of the parallel predicate syntax generated including each of the predicates of the “term-predicate” pair is a fourth feature, and the input “term-predicate” pair is A feature extraction unit that extracts a feature set including at least one of the first feature and the second feature, the third feature, and the fourth feature; and the feature extracted by the feature extraction unit Classification of the “term-predicate” pair based on a feature set and a pre-learned model for classifying “term-predicate” pairs as synonymous or anonymity. Are determined by the same-unrelated determination unit and the same-unrelated determination unit. When it is determined that the classification of the “predicate” pair is synonymous, the feature set extracted by the feature extracting unit and the previously learned “term-predicate” pair are defined as synonymous and anomalous. And a synonym-anonymity determination unit that determines whether the classification of the “term-predicate” pair is synonymous or anomaly based on a model for classification.

  A classification determination method according to an eighth aspect of the present invention is a classification determination method in a classification determination apparatus including a feature extraction unit, a contradiction-irrelevance determination unit, and a synonym-anonymity determination unit, wherein the feature extraction unit The “term-predicate”, which is extracted based on the definition statement of each predicate of the “term-predicate” pair obtained from the definition statement set including the definition statements for each of the plurality of prepared predicates. A feature indicating mutual complementarity of a definition statement indicating whether or not a predicate as a pair exists in the definition statement of each of the predicates of the pair, and the definition of each of the predicates of the “term-predicate” pair Definition sentence mutual complementarity that indicates whether or not at least the previous description part exists among the features that indicate the mutual complementarity of the definition sentence that indicates whether or not there is a pair of "term-predicate" terms in the sentence. The feature to be shown is the first feature, and the meaning attribute consists of a semantic attribute for each of a plurality of predicates prepared in advance. A semantic attribute common to the “term-predicate” pair extracted from the attribute set and extracted based on the semantic attribute of each predicate of the “term-predicate” pair is set as a second feature, and the “term” The combination of the partial character string in the character string representing one predicate in the “-predicate” pair and the partial character string in the character string representing the other predicate is a third feature, The likelihood of the parallel predicate syntax generated including each of the predicates of the “part” pair is set as a fourth feature, and the first feature and the second feature of the inputted “term-predicate” pair are used. The feature set including at least one of the third feature and the fourth feature is extracted, and the reciprocity-irrelevance determining unit includes the feature set extracted by the feature extracting unit, To classify learned “term-predicate” pairs as irrelevant or unrelated Based on Dell, it is determined whether the classification of the “term-predicate” pair is synonymous or irrelevant. When the pair classification is determined to be synonymous, the feature set extracted by the feature extraction unit and the previously learned “term-predicate” pair are classified into synonym and anomaly. Whether the classification of the “term-predicate” pair is synonymous or affirmative.

  As described above, according to the seventh and eighth inventions, based on the feature set extracted for the inputted “term-predicate” pair and the model classified as irrelevant and learned in advance. Then, it is determined whether the classification of the “term-predicate” pair is synonymous or irrelevant. If it is determined to be synonymous, it is classified into the extracted feature set and the previously learned synonyms and anomalies. To determine whether the term-predicate pair is synonymous, affirmative, or irrelevant by determining whether the classification of the term-predicate pair is synonymous or affirmative It can be determined with high accuracy.

  In the classification determination device according to a seventh aspect of the invention, the feature extraction unit appears in the vicinity of the “term-predicate” in the text corpus for each “term-predicate” of the “term-predicate” pair. For each of the distribution similarity calculated for the word co-occurrence and the similarity of neighboring words of the “term-predicate” pair and the predicate of the “term-predicate” pair in the text corpus, the “ The co-occurrence of words appearing around the predicate of “term-predicate” is calculated, and at least the “term” of the distribution similarities calculated of the similarities of the peripheral words of the predicate of the “term-predicate” pair The distribution similarity calculated for the similarity of the words in the predicate pair is a fifth feature, and is obtained from a semantic label set made up of semantic labels of each functional expression of a plurality of predicates prepared in advance. Based on the semantic label of each functional expression of the predicate of the "term-predicate" pair A semantic label common to the extracted “term-predicate” pair is defined as a sixth feature, an overlapping degree of the common semantic labels of the “term-predicate” pair is defined as a seventh feature, and the “term” The likelihood of the compound word generated by combining each of the predicates in the “-predicate” pair is the eighth feature, and the two predicates in the “term-predicate” pair represent the relationship between the situations. The degree of co-occurrence as two predicates representing the situation in the sentence is a ninth feature, and at least one of the first feature and the second feature for the input “term-predicate” pair And at least one of the third feature, the fourth feature, the fifth feature, the sixth feature, the seventh feature, the eighth feature, and the ninth feature It is also possible to extract the feature set including one.

  Note that the first feature, the second feature, the fifth feature, the sixth feature, and the seventh feature are synonymous features. Further, the third feature, the fourth feature, the eighth feature, and the ninth feature are anonymity features.

  Moreover, the program of this invention is a program for functioning a computer as each part which comprises said classification | category model learning apparatus or classification | category determination apparatus.

  As described above, according to the classification model learning device, method, and program of the present invention, it is highly accurate whether or not a predicate pair or a “term-predicate” pair is synonymous, affirmative, or irrelevant. A classification model that can be determined can be learned.

  Furthermore, according to the classification determination apparatus, method, and program of the present invention, it is possible to determine with high accuracy whether a predicate pair or a “term-predicate” pair is synonymous, anonymity, or irrelevant. .

It is a block diagram which shows the functional structure of the feature vector construction apparatus which concerns on embodiment of this invention. It is a figure which shows the example of the analysis result in a basic analysis part. It is a figure which shows the example of the feature extracted in a feature extraction part. It is a figure which shows the example of the feature vector constructed | assembled in the feature vector construction part. It is a block diagram which shows the functional structure of the classification model learning apparatus which concerns on embodiment of this invention. It is a figure which shows the example of a learning corpus. It is a block diagram which shows the functional structure of the synonym feature extraction part which concerns on embodiment of this invention. It is a figure which shows the example of the distribution similarity calculated in a distribution similarity calculation part. It is a figure which shows the example of the definition sentence extracted in a definition sentence extraction part (definition sentence mutual complementarity). It is a figure which shows the example of the definition sentence extracted in a definition sentence extraction part (overlap of vocabulary). It is a block diagram which shows the functional structure of the dictionary definition sentence feature extraction part which concerns on embodiment of this invention. It is a figure which shows the list of the features extracted in a dictionary definition sentence feature extraction part. It is a figure which shows the example of the feature extracted in a dictionary definition sentence feature extraction part. It is a figure which shows the example of a semantic attribute. It is a figure which shows the example of the hierarchy of a semantic attribute. It is a block diagram which shows the functional structure of the semantic attribute feature extraction part which concerns on embodiment of this invention. It is a figure which shows the example of the semantic attribute provided in a semantic attribute feature extraction part. It is a block diagram which shows the functional structure of the function representation feature extraction part which concerns on embodiment of this invention. It is a figure which shows the example of a semantic label. It is a figure which shows the example of the semantic label provided in a semantic label provision part. It is a block diagram which shows the functional structure of the anonymity extraction part which concerns on embodiment of this invention. It is a figure which shows the example of the n-gram score of a compound word. It is a figure which shows the example of the partial character string extracted in the partial character string feature extraction part. It is a figure which shows the example of the score between situations extracted in the score calculation part between situations. It is a figure which shows the example of a database between situations. It is a figure which shows the example of the feature memorize | stored in the feature set memory | storage part. It is a figure which shows the example of the feature constructed | assembled in the anti-irrelevant classification | category feature construction part. It is a figure which shows the example of the feature constructed | assembled in the synonym-anonymity classification | category feature construction part. It is a block diagram which shows the functional structure of the classification determination apparatus which concerns on embodiment of this invention. It is a figure which shows the example of the result of the same anti-irrelevance determination part. It is a figure which shows the example of the result of a synonym-antisense determination part. It is a flowchart which shows the feature vector construction process routine in the feature vector construction apparatus which concerns on embodiment of this invention. It is a flowchart which shows the classification model learning process routine in the classification model learning apparatus which concerns on embodiment of this invention. It is a figure which shows the extraction process routine of a synonym feature in the classification model learning apparatus which concerns on embodiment of this invention. It is a figure which shows the distribution similarity calculation process routine in the classification model learning apparatus which concerns on embodiment of this invention. It is a flowchart which shows the dictionary definition sentence feature extraction process routine in the classification model learning apparatus which concerns on embodiment of this invention. It is a flowchart which shows the semantic attribute feature extraction process routine in the classification model learning apparatus which concerns on embodiment of this invention. It is a flowchart which shows the extraction process routine of the function expression feature in the classification model learning apparatus which concerns on embodiment of this invention. It is a flowchart which shows the anonymity extraction process routine in the classification model learning apparatus which concerns on embodiment of this invention. It is a flowchart which shows the production | generation process routine of the compound word in the classification model learning apparatus which concerns on embodiment of this invention. It is a flowchart which shows the classification determination processing routine in the classification determination apparatus which concerns on embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

<Principle of the invention>
Embodiments of the present invention provide a relationship between two different “term-predicates” or predicates that are “synonymous” such as “lunch-enjoy” and “lunch-enjoy”. , “Lunch is delicious” and “Lunch is bad”, or “Lunch is expensive” and “Lunch is late”. It is related with the method of classifying by computer.

  In order to automatically classify the semantic relationship of the predicate pair into three classes, “synonymous”, “anti-righteous”, and “irrelevant”, in this embodiment, (a) predicate pairs in the synonymous relationship (B) a description in an anonymity relationship, using the four features “dictionary definition sentence feature”, “semantic attribute feature”, “distribution similarity”, and “functional expression feature” as synonym features. In order to identify a pair of pairs, the four features “partial string feature”, “compound word score”, “inter-situation score”, and “intra-sentence co-occurrence likelihood” are used as anonymity features, and (c) , Antagonism and irrelevant classification in a two-step manner. As a first step, “synonymous or anomalous vs. irrelevant” binary classification of “synonymous—irrelevant” is performed, and then “synonymous” and “anonymous” are identified. Perform value classification. As described above, classification is performed in two stages, thereby improving the accuracy of identifying semantically similar “synonyms” and “abuses”.

<Configuration of feature vector construction device>
A feature vector construction device according to an embodiment of the present invention will be described. As shown in FIG. 1, a feature vector construction apparatus 100 according to an embodiment of the present invention includes a CPU, a RAM, and a ROM that stores programs and various data for executing a feature vector construction processing routine described later. Can be configured with a computer. Functionally, the feature vector construction device 100 includes an input unit 10, a calculation unit 20, and an output unit 30 as shown in FIG.

  The input unit 10 receives a plurality of sentences written in a natural language and digitized from an input device such as a keyboard. This set of sentences is called a text corpus. Note that the input unit 10 may accept input from the outside via a network or the like.

  The calculation unit 20 includes a basic analysis unit 24, a feature extraction unit 26, and a feature vector construction unit 28.

A text corpus received by the input unit 10 is input to the basic analysis unit 24. The basic analysis unit 24 performs morphological analysis and dependency analysis for each sentence of the input text corpus, and features an analysis result including at least morpheme notation, standard form and part of speech, and dependency information for each clause. The data is output to the extraction unit 26. FIG. 2 shows an example of the output of the basic analysis unit 24 with respect to the sentence “planting flowers and flower beds completed”. Note that existing morphological analysis and dependency analysis may be used.

  The feature extraction unit 26 uses the analysis result of each sentence input from the basic analysis unit 24, and with respect to the “term-predicate” included in each sentence, Information (context information) is extracted and output as a term-predicate feature. In addition, the feature extraction unit 26 uses the analysis result of each sentence input from the basic analysis unit 24, and for the predicate included in each sentence, information (word context information) of words appearing around the predicate. Extract and output as predicate features. In this embodiment, for example, Non-Patent Document 2 (Tomohide Shibata, Ikuo Kurohashi (2010). Acquiring synonym relations of predicates depending on context. IPSJ SIG Notes 2010-NL-199 (13), 1-6.) Extract term-predicate features and predicate features in the same way. Specifically, another “term-predicate” and “predicate” related to the target “term-predicate” are extracted as a term-predicate feature. Furthermore, a term (noun phrase having a case particle) related to a single “predicate” and another “predicate” are extracted as predicate features. An example of the output of the feature extraction unit 26 in the present embodiment is shown in FIG. As shown in FIG. 3, “predicate” “planting” and “term-predicate” “flower-wo-planting” are compared to “term-predicate” “flowerbed-ga-complete”. As a term-predicate feature. For the predicate “complete”, the term “flowerbed-ga” and the predicate “plant” are extracted as predicate features.

  The feature vector construction unit 28 uses the analysis result of each sentence input from the basic analysis unit 24 to extract “term-predicate” included in each sentence as “term-predicate” of the feature vector construction target. . Then, the feature vector construction unit 28 uses, as elements, values calculated using the term-predicate features input from the feature extraction unit 26 for each “term-predicate” of the extracted feature vector construction target. A feature vector is constructed and output to the output unit 30. Also, the feature vector construction unit 28 extracts predicates included in each sentence as feature vector construction target predicates using the analysis result of each sentence input from the basic analysis unit 24. Then, the feature vector construction unit 28 constructs a feature vector whose elements are values calculated using the predicate feature input from the feature extraction unit 26 for each extracted predicate of the feature vector construction target, Output to the output unit 30. In the present embodiment, for example, a feature vector is constructed by the same method as in Non-Patent Document 2 described above. Specifically, construct a feature vector whose element value is the weight value calculated based on the mutual information (MI) of the term-predicate feature and the term-predicate feature. To do. In addition, a feature vector having a weight value calculated based on a predicate to be feature vector construction target and a mutual information (MI) of each predicate feature as an element value is constructed. The weight is calculated using the following formula (1). The mutual information (MI) is calculated using the following equation (2).

  A feature vector when the feature vector construction target is “term-predicate” is called a term-predicate feature vector. When the feature vector construction target is “term-predicate”, u represents “term-predicate”, and f represents the term-predicate feature. P (u) is the probability that the “term-predicate” of the feature vector construction target appears in the text corpus, P (f) is the probability that the term-predicate feature appears in the text corpus, and P (u, f) is Represents the probability that the “term-predicate” and term-predicate features that are feature vector construction targets will appear simultaneously. When MI is greater than 0, the value of weight is 1. When MI is 0 or less, the value of weight is 0. An example of the term-predicate feature vector constructed in the table on FIG. 4 is shown. In the example, when u is “flowerbed-ga-complete” and f is “flower-wo-plant”, it indicates that MI is greater than 0 and the weight value is 1. Further, when u is “flowerbed-ga-finished” and f is “time-over”, it indicates that MI is 0 or less and the weight value is 0.

  A feature vector when a feature vector construction target is a predicate is called a predicate feature vector. When the feature vector construction target is a predicate, u represents a predicate, and f represents a predicate feature. P (u) is the probability that the predicate of the feature vector construction target appears in the text corpus, P (f) is the probability that the predicate feature appears in the text corpus, and P (u, f) is the description of the feature vector construction target. Represents the probability that a part and predicate feature will appear simultaneously. When MI is greater than 0, the value of weight is 1. When MI is 0 or less, the value of weight is 0. An example of the predicate feature vector constructed in the lower table of FIG. 5 is shown. In the example, when u is “complete” and f is “flowerbed-ga”, MI is greater than 0 and the weight value is 1. Further, when u is “completed” and f is “house-ga”, it indicates that MI is 0 or less and the weight value is 0.

  As described above, in the present embodiment, the predicate feature vector in which each predicate included in the input text corpus is a construction target of the feature vector, and each “term-predicate” included in the input text corpus. 2 types of term-predicate feature vectors (hereinafter referred to as “feature vector” together) are created. In this embodiment, both the predicate feature vector and the term-predicate feature vector are created. However, only the feature vector used in the classification model learning device and the classification determination device described later may be created.

<Configuration of classification model learning device>
Next, the configuration of the classification model learning device according to the embodiment of the present invention will be described. As shown in FIG. 5, the classification model learning device 200 according to the embodiment of the present invention includes a CPU, a RAM, and a ROM that stores a program for executing a classification model learning processing routine described later and various data. It can consist of computers. Functionally, the classification model learning device 200 includes an input unit 210, a calculation unit 220, and an output unit 250 as shown in FIG.

  The input unit 210 accepts a plurality of predicate pairs and a plurality of “term-predicate” pairs to which information of any of the same meaning, anomaly, and irrelevant classification is given from an input device such as a keyboard. This predicate pair classification, which is synonymous, affirmative, or unrelated, is referred to as a “predicate class”. Also, a plurality of predicate pairs and a plurality of “term-predicate” pairs to which predicate class information is assigned are referred to as a learning corpus. FIG. 6 shows an example of a learning corpus. Note that the input unit 210 may accept input from the outside via a network or the like.

  The calculation unit 220 includes a learning corpus storage unit 222, a synonym feature extraction unit 224, an antonym feature extraction unit 226, a feature set storage unit 228, an anti-independent classification feature construction unit 230, and an anti-independent classification. A feature storage unit 232, an anti-independent classification model learning unit 234, an anti-independent classification model storage unit 236, a synonym-anonymity classification feature construction unit 240, and a synonym-anonymity classification feature storage unit 242; A synonym-anonymity classification model learning unit 244 and a synonym-anonymity classification model storage unit 246 can be used.

  The learning corpus storage unit 222 stores the learning corpus received by the input unit 210.

  A detailed configuration of the synonym feature extraction unit 224 is shown in FIG. The synonym feature extraction unit 224 includes a feature vector storage unit 300, a distribution similarity calculation unit 302, a definition sentence dictionary storage unit 304, a dictionary definition sentence feature extraction unit 306, a semantic attribute dictionary storage unit 308, and a semantic attribute feature. It can be expressed by a configuration including an extraction unit 310, a function expression dictionary storage unit 312, and a function expression feature extraction unit 314.

  The synonym feature extraction unit 224 determines the distribution similarity, dictionary definition sentence feature, meaning for each of all the learning corpus pairs (predicate pair and “term-predicate” pair) stored in the learning corpus storage unit 222. Extract attribute features and functional representation features. Note that the distribution similarity, dictionary definition sentence feature, semantic attribute feature, and functional expression feature are collectively synonymous.

  The feature vector storage unit 300 stores the feature vector output by the feature vector construction device 100.

  The distribution similarity calculation unit 302 obtains from the feature vector storage unit 300 for all of the learning corpus pairs (predicate pairs and “term-predicate” pairs) stored in the learning corpus storage unit 222. The distribution similarity is calculated using the corresponding feature vector, and the calculation result is output to the feature set storage unit 228 together with the calculation target pair (predicate pair and “term-predicate” pair). A distribution similarity calculated using a predicate feature vector is called a predicate distribution similarity, and a distribution similarity calculated using a term-predicate feature vector is called a term-predicate distribution similarity. When calculating the distribution similarity for a predicate pair, only the information about the predicate is used, so only the predicate distribution similarity is calculated, but the distribution similarity for the “term-predicate” pair. When calculating, it is possible to use not only information related to “term-predicate” but also information related to predicate, so it is possible to calculate both term-predicate distribution similarity and predicate distribution similarity. it can. The calculation of the distribution similarity is a process of calculating the distribution similarity using at least one of the predicate feature vector and the term-predicate feature vector obtained from the text corpus by the feature vector construction apparatus 100. In the present embodiment, the distribution similarity is calculated by the same method as in Non-Patent Document 2 described above. Specifically, the distribution similarity measure is calculated using the following equations (3) to (5).

However, in the numerator of the JACCARD coefficient in the above equation (4), when the distribution similarity is calculated using the term-predicate feature vector, when the two term-predicate feature vectors are compared for each element, The number of elements in which the value of the element of one term-predicate feature vector is 1 and the value of the element of the other term-predicate feature vector is also 1. In addition, the denominator of the JACCARD coefficient is obtained by comparing one term-predicate feature vector element and the other term-description when comparing two pairs of term-predicate feature vectors used for calculating the distribution similarity. This is the number of elements in which the value of at least one of the elements of the partial feature vector is 1.

In addition, the numerator of the SIMPSON coefficient in the above equation (5) is the same as the numerator of the JACCARD coefficient, and the denominator of the SIMPSON coefficient is an element whose element value is 1 in the pair term-predicate feature vector used for calculation. And the smaller of the number of elements whose element value is 1 in the other term-predicate feature vector.

Also, when calculating the distribution similarity using the predicate feature vector, the numerator of the JACCARD coefficient in the above equation (4) is one predicate when the two predicate feature vectors are compared for each element. This is the number of elements whose element value of the feature vector is 1 and whose element value of the other predicate feature vector is also 1. In addition, the denominator of the JACCARD coefficient is obtained by comparing the elements of one predicate feature vector and the elements of the other predicate feature vector when the two predicate feature vectors of the pair used for calculating the distribution similarity are compared for each element. This is the number of elements in which the value of at least one element is 1.

In addition, the numerator of the SIMPSON coefficient in the above equation (5) is the same as the numerator of the JACCARD coefficient, and the denominator of the SIMPSON coefficient is the number of elements whose element value is 1 in the pair predicate feature vector used for calculation. And the smaller of the number of elements whose element value is 1 in the other predicate feature vector.

  In this embodiment, both the predicate distribution similarity and the term-predicate distribution similarity are used. FIG. 8 shows an example of the distribution similarity of “flowerbed-ga-complete” and “flowerbed-ga-complete”. The upper part of the table of FIG. 8 is an example of the term-predicate distribution similarity, and the lower part is an example of the predicate distribution similarity. The predicate distribution similarity and the term-predicate distribution similarity are collectively referred to as distribution similarity. The distribution similarity is an example of a fifth feature.

  The definition sentence dictionary storage unit 304 stores a definition sentence dictionary that stores a definition sentence set corresponding to each of the plurality of predicates. As the definition sentence dictionary, an existing national language dictionary or a free dictionary on the Web that has been added and edited by a plurality of users may be used. The definition sentence dictionary is an example of a definition sentence set.

<Configuration of dictionary definition sentence feature extraction unit>
The dictionary definition sentence feature extraction unit 306 relates to the content words of all the predicate pairs of the learning corpus stored in the learning corpus storage unit 222 or the content words of the predicates of all “term-predicate” pairs. Based on the definition sentence dictionary stored in the definition sentence dictionary storage unit 304, a feature indicating “definition sentence mutual complementarity” and a feature indicating “overlapping vocabulary” are extracted, and a feature set storage unit 228 together with a pair to be extracted. Output to. The definition sentence dictionary includes one or more definition sentences corresponding to the content words of the plurality of predicates, and each set of one or more definition sentences extracted from the definition sentence dictionary is referred to as a definition sentence set. Note that the feature indicating the definition sentence mutual complementarity and the feature indicating the vocabulary overlap are collectively referred to as a dictionary definition sentence feature. The feature indicating the definition sentence mutual complementarity is an example of the first feature.

  The dictionary definition sentence feature extraction unit 306 indicates “definition sentence mutual complementarity” and “overlapping vocabulary” based on the feature that “synonymous predicates have similarities in mutual definition sentence sets explaining the meaning”. Extract features.

  Here, “definition statement mutual complementarity” means that one's predicate appears in the definition statement set of the other's predicate, and is synonymous with “complete” and “complete” shown in FIG. For example, the predicate “complete” appears in the synonymous “finished” dictionary definition sentence set, and the predicate “completed” is synonymous with “completed”. Appears in the dictionary definition sentence set.

  Further, “overlapping vocabulary” means that vocabulary overlaps between definition sentence sets, and is a statement having two synonyms of “price-ga-high” and “price-ga-high” shown in FIG. For example, the term “price” appears in both definition sentence sets.

  In this way, there is a feature that “synonymous predicates have similarities in mutual definition sentence sets that explain the meaning”, and this feature is expressed in the form of “definition mutual complementarity” or “vocabulary overlap”. Can be expressed. For the sake of explanation, the first “term-predicate” predicate (for the “term-predicate” pair such as “flowerbed-ga-finish” and “flowerbed-ga-complete”) ( That is, “complete”) is Pred1, “term” (ie, “flowerbed”) is Arg1, the second “term-predicate” predicate (ie, “completed”) is Pred2, and term (ie, “complete”). “Flowerbed”) is designated Arg2. Similarly, for predicate pairs such as “completed” and “completed”, the first predicate is Pred1, and the second predicate is Pred2.

  A detailed configuration of the dictionary definition sentence feature extraction unit 306 is shown in FIG. The dictionary definition sentence feature extraction unit 306 can be represented by a configuration including a definition sentence extraction unit 400, a definition sentence mutual complementarity extraction unit 402, and a vocabulary overlap extraction unit 404.

  The definition sentence extraction unit 400 defines a definition sentence for the content words of each predicate of all pairs (predicate pairs and “term-predicate” pairs) of the learning corpus stored in the learning corpus storage unit 222. Dictionary definition of the definition sentence dictionary stored in the dictionary storage unit 304 is performed, and a definition sentence set of each predicate is extracted for each predicate pair or “term-predicate” pair. Then, morphological analysis is performed on the extracted definition sentence set, and an analysis result including at least a notation, a standard form, a part of speech, and a reading for each morpheme is output to the definition sentence mutual complementarity extraction unit 402 and the vocabulary overlap extraction unit 404. .

  The definition sentence mutual complementarity extraction unit 402 defines, for each extraction target pair (predicate pair or “term-predicate” pair), from the result of the morphological analysis of the definition sentence set input from the definition sentence extraction unit 400. A feature indicating sentence mutual complementarity is extracted and output to the feature set storage unit 228 together with the pair to be extracted. Specifically, whether Pred2 appears in the definition sentence set of Pred1 or whether Pred1 appears in the definition sentence set of Pred2 is extracted by character string matching. Whether Pred1 appears in the definition sentence set of Pred2 is Pred1Match, and whether Pred2 appears in the definition sentence set of Pred1 is Pred2Match. In the present embodiment, the feature value is set to 1 when it appears. Similarly, whether Arg2 has appeared in the definition sentence set of Pred1 or whether Arg1 has appeared in the definition sentence set of Pred2 is extracted. Arg1Match indicates whether Arg1 appears in the definition sentence set of Pred2, and Arg2Match indicates whether Arg2 appears in the definition sentence set of Pred1. Note that these values may include real numbers such as the number of overlaps and the value obtained by normalizing the number of overlaps with the total number of words in the definition sentence set.

  In this embodiment, all of Pred1Match, Pred2Match, Arg1Match, and Arg2Match are used as features indicating mutual definition of definition sentences. However, Arg1Match and Arg2Match are not used when using features indicating mutual complementarity of definition sentences. May be. When predicate pairs are targeted, only Pred1Match and Pred2Match are used.

  The vocabulary overlap extraction unit 404 extracts, for each extraction target pair (predicate pair or “term-predicate” pair), the extraction target pair from the result of the morphological analysis of the definition sentence set input from the definition sentence extraction unit 400. A feature indicating whether there is a vocabulary overlap between the definition sentence sets of the predicates is output to the feature set storage unit 228 together with the pair to be extracted. In this embodiment, the number of vocabulary appearing in both definition sentence sets is assumed to be a feature, and the vocabulary part-of-speech (noun, non-independence excluding independence, adjective verb, and four types of part-of-speech) are aggregated. To do. Here, NounMatch is the number of vocabularies with part-of-speech nouns that appear in both definition sets, VerbMatch is the number of vocabularies with part-of-speech parts, AdjMatch is the number of vocabularies with part-of-speech parts, and adjective verbs The number of vocabularies with parts of speech is called AdjNMatch (FIG. 12). In the present embodiment, all of NounMatch, VerbMatch, AdjMatch, and AdjNMatch are used as features indicating vocabulary overlap, but at least one feature is sufficient. It should be noted that the presence or absence of vocabulary that appears in both definition sentence sets is not the number of vocabularies that appear in both definition sentence sets, and may be aggregated for each part of speech of the vocabulary. In this case, the value may be set to 1 and the value may be set to 0 if none. Furthermore, a value obtained by normalizing the number of vocabulary appearing in common in both definition sentence sets with the total number of words in the definition sentence set may be used. FIG. 13 shows an example of the feature list created.

  The semantic attribute dictionary storage unit 308 stores a semantic attribute dictionary that stores semantic attributes corresponding to the content words of the plurality of predicates. In the present embodiment, a prescriptive attribute dictionary is stored as a semantic attribute dictionary. The semantic attribute dictionary is an example of a semantic attribute set.

<Configuration of semantic attribute feature extraction unit>
The semantic attribute feature extraction unit 310 performs, for each of all pairs (predicate pair or “term-predicate” pair) of the learning corpus stored in the learning corpus storage unit 222, the content word of the predicate of the pair. A feature indicating the overlap of each abstract semantic attribute is extracted and output to the feature set storage unit 228 together with the pair to be extracted. In the present embodiment, two features, “overlapping semantic attribute” and “overlapping ratio with semantic attribute weight”, which will be described later, are extracted as features indicating the overlapping of abstract semantic attributes. The semantic attribute dictionary is composed of one or more semantic attributes corresponding to each of a plurality of predicates, and each set of one or more semantic attributes extracted from the semantic attribute dictionary is called a semantic attribute set. A semantic attribute that appears in common in the semantic attribute set of both predicates is extracted as a feature of the “overlapping semantic attribute”. In addition, a “semantic attribute weighted overlap rate” calculated by assigning a weight to a hierarchy to which a semantic attribute that appears in common in both can belong can be extracted as a feature. The semantic attribute feature extraction unit 310 outputs these two features together with the pair to be extracted to the feature set storage unit 228.

  The semantic attribute feature extraction unit 310 allows the abstract semantic attributes of the predicates themselves to overlap with the hierarchical overlap because of the feature that the synonymous predicates have similar abstract semantic attributes. By extracting “feature” in consideration of “depth”, it is possible to avoid the problem of “predicate indicating the passage of time is erroneously determined to be synonymous”.

  In this embodiment, as an abstract semantic attribute, “prescriptive attribute” (Non-patent Document 3: Satoru Ikehara, Masahiro Miyazaki, Satoshi Shirai, Akio Yokoo, Hiroaki Nakaiwa, Kentaro Ogura, Yoshifumi Oyama, Yoshihiko Hayashi (1999 ) Use Japanese Vocabulary System CD-ROM version. Iwanami Shoten.). An example of the predicate attribute set is shown in FIG. 14 with an example where the predicate is “completed” and “completed”. As FIG. 15 shows, these attributes often have a hierarchical structure, and the attributes are further refined as the hierarchy progresses downward. For example, a more detailed attribute such as “ownership movement” is specified for an upper attribute “action” via an intermediate attribute “physical behavior”.

  A detailed configuration of the semantic attribute feature extraction unit 310 is shown in FIG. The semantic attribute feature extraction unit 310 includes a semantic attribute overlap extraction unit 500 and a semantic attribute weighted overlap rate calculation unit 502.

  The semantic attribute overlap extraction unit 500, for each of all the pairs (predicate pair or “term-predicate” pair) stored in the learning corpus storage unit 222, the semantic attribute of the content word of each predicate of the pair Are extracted from the semantic attribute dictionary, semantic attributes appearing in both semantic attribute sets of the pairs are extracted as features, and output to the semantic attribute weighted overlap rate calculation unit 502 together with the extraction target pairs. In FIG. 17, there is an overlapping semantic attribute from the semantic attribute set (attribute change, generation) for “completed” of “flowerbed-ga-complete” and the semantic attribute set (generation) for “completed” of “flowerbed-ga-completed”. As an example, “generation” is extracted.

  The semantic attribute weighted overlap ratio calculation unit 502 extracts all semantic attributes input from the semantic attribute overlap extraction unit 500 and semantic attributes extracted from the semantic attribute dictionary for each pair (predicate pair or “term-predicate” pair). Based on the hierarchical information, the “meaning attribute weighted overlap rate” as a feature indicating the degree of overlapping of semantic attributes is calculated according to the following formulas (6) and (7), and input from the semantic attribute overlap extraction unit 500 It outputs to the feature set storage unit 228 together with the feature and the pair to be extracted. The two features of the overlapping semantic attribute and the semantic attribute weighted overlap rate are each characterized by "the more attributes are shared, the more predicates are similar" and "the more detailed attributes are shared, The predicates are similar to each other. The overlapping semantic attribute and the semantic attribute weighted overlapping rate are collectively referred to as a semantic attribute feature. The overlapping semantic attribute is an example of the second feature.

  For example, when calculating the weighted overlap ratio of “flowerbed-ga-complete” and “flowerbed-ga-finished”, it appears in both the “completed” semantic attribute set and the “finished” semantic attribute set. The semantic attribute to be generated is a semantic attribute of “generation”. Furthermore, since “generation” is the most detailed attribute of the hierarchy 4 as shown in FIG. 15, the weighted overlap ratio is calculated as in the following equation (8).

  The function expression dictionary storage unit 312 stores a function expression dictionary that stores semantic labels for function expressions of a plurality of predicates. In the present embodiment, the dictionary described in Non-Patent Document 3 is used. However, the present invention is not limited to this, and another sentence ending expression dictionary or modality expression dictionary may be used.

<Configuration of functional expression feature extraction unit>
The functional expression feature extraction unit 314 performs, for each of all pairs (predicate pairs or “term-predicate” pairs) of the learning corpus stored in the learning corpus storage unit 222, the functional expression of the predicate of the pair. A feature indicating an overlap of meaning is extracted and output to the feature set storage unit 228 together with the pair to be extracted. In the present embodiment, two features of “overlapping semantic label” and “semantic label overlapping rate” to be described later are extracted as features indicating the overlapping of meanings of function expressions.

  The feature expression feature extraction unit 314 extracts features indicating the overlap of the function expressions of the predicates, thereby increasing the synonym determination of the predicates even when the distribution similarity is calculated using only the predicate content words. Can be done with precision. In order to consider the meaning of the functional expression of the predicate, a huge amount of data is required when the distribution similarity calculation is performed by combining the predicate with each content word and the functional expression. By using the function expression feature, the meaning of the function expression can be incorporated into the synonym determination without requiring a huge amount of data.

  A detailed configuration of the function representation feature extraction unit 314 is shown in FIG. The functional representation feature extraction unit 314 includes a semantic label assignment unit 600, an overlapping meaning label extraction unit 602, and a semantic label overlap rate calculation unit 604.

  For each of all the learning corpus pairs (predicate pairs or “term-predicate” pairs) stored in the learning corpus storage unit 222, the semantic label assigning unit 600 represents the functional expression of each predicate of the pair. A semantic label is extracted from the functional expression dictionary, and is output to the overlapping semantic label extraction unit 602 together with the pair. In this embodiment, a statistical meaning labeling method is used (Non-Patent Document 4 (Kenji Imamura, Ikuko Izumi, Genichiro Kikui, Satoshi Sato (2011). Semantic Label Tagger of Predicate Functional Representation) See next competition. 518-521)). Specifically, the morpheme analysis of each pair of predicates stored in the learning corpus storage unit 222 is performed, and the most likely meaning is obtained by using the analysis result including at least the notation for each morpheme, the standard form, and the part of speech. Assign a label column. FIG. 19 shows an example of a function expression dictionary. FIG. 20 shows an example in which a meaning label string is given when the “term-predicate” pair “flowerbed-ga-completed” and “flowerbed-ga-completed” is input. In the example shown in FIG. 20, the predicates are “completed” and “completed”, respectively, and the content word part and the function expression part are identified, and “completed” is displayed in each function expression “ta”. The meaning label is given.

  In the present embodiment, the content label part and the function expression part of the predicate are identified by the semantic label assigning unit 600. However, a pre-processing that leaves only the function expression that affects the meaning of the event is performed. (Non-Patent Document 5 (Izumi T., Imamura K., Kikui G., & Sato S. (2010). Standardizing Complex Functional Expressions in Japansese Predicates: Applying Theoretically-Based Paraphrasing Rules. Proceedings of the Workshop on Multiword Exressions : From theory to applications (MWE 2010), 63-71)).

  The overlapping meaning label extraction unit 602 extracts, as overlapping meaning labels, meaning labels that appear in common in the meaning labels of both predicates from the pair to be extracted input from the meaning label assignment unit 600 and the respective meaning labels. , And output to the semantic label overlap calculator 604 together with the pair. In the example of the “term-predicate” pair “flowerbed-ga-completed” and “flowerbed-ga-completed”, the meaning label “completed” and predicate “completed” mean the predicate “completed”. From the label “completed”, “completed” is extracted as the overlapping meaning label.

  The semantic label overlap rate calculation unit 604 calculates the semantic label overlap rate according to the following equation (9) based on the extraction target pair and the overlap meaning label input from the overlap meaning label extraction unit 602, and the pair and overlap meaning The calculation result together with the label is output to the feature set storage unit 228. The overlapping meaning label and the meaning label overlapping rate are collectively referred to as a function representation feature. The overlapping meaning label is an example of a sixth feature, and the semantic label overlapping rate is an example of a seventh feature.

  FIG. 21 shows a detailed configuration of the anonymity extraction unit 226. The antonym feature extraction unit 226 includes a compound word score calculation unit 700, an n-gram model storage unit 706, a sentence co-occurrence likelihood calculation unit 708, a partial character string feature extraction unit 714, and an inter-situation score calculation unit. 716 and an inter-event database storage unit 718.

  The compound word score calculation unit 700 makes it easy to create a compound word of each pair of the learning corpora (predicate pair or “term-predicate” pair) stored in the learning corpus storage unit 222. Is calculated and extracted as a feature. The compound word score calculation unit 700 can be represented by a configuration including a compound word generation unit 702 and a compound word likelihood calculation unit 704. Note that the feature representing the compound word likelihood is an example of the eighth feature.

  The compound word generation unit 702 generates a compound word for each of all the learning corpus pairs (predicate pairs or “term-predicate” pairs) stored in the learning corpus storage unit 222. Specifically, for each of the predicates that are included in the predicate pair or the “term-predicate” pair, based on the morphological part-of-speech information obtained using an existing morpheme analyzer or morpheme analysis dictionary. Identify the part of speech for each predicate. When the part of speech of both predicates of the pair is a Japanese verb (a verb derived from Japanese) or an adjective, a compound word is generated by combining the consecutive forms of the pair of predicates. The compound word generates two types in which the order of one predicate and the other predicate is switched (for example, “sell” + “buy” → “sell / buy”, “buy / sell”). In the case of a Japanese verb, the part of speech is a “verb”. Verbs other than Japanese verbs become “motion nouns” that become verbs by adding the light verb “s”.

  Also, if the part-of-speech of both predicates in a pair is a behavioral noun and there is a “common character” for each of the paired predicates, the common character is retained. However, a compound word is generated by continuing different character strings. The compound word generates two types in which the order of one predicate and the other predicate is exchanged (for example, “hospitalization” + “discharge” → “entrance / exit”, “discharge”).

  In cases other than the above, compound words are generated by combining predicates that are paired as they are. Note that two types of compound words are generated by switching the order of one predicate and the other predicate (for example, “like” + “hate” → “like and dislike”, “dislike”).

  The compound word likelihood calculation unit 704 calculates the likelihood of each compound word of the two types of compound words for each of the pairs (predicate pair or “term-predicate” pair) generated by the compound word generation unit 702. It is calculated, extracted as a feature, and output to the feature set storage unit 228. Specifically, using the n-gram model stored in the n-gram model storage unit 706, an n-gram score representing the likelihood of the n-gram of the compound word as shown in FIG. 22 is calculated. In the example of FIG. 22, n-gram scores in the case of n = 2 are shown.

  The n-gram model storage unit 706 stores an n-gram model learned using a specific corpus.

  The intra-sentence co-occurrence likelihood calculation unit 708 generates each pair of learning corpora (predicate pair or “term-predicate” pair) stored in the learning corpus storage unit 222 including the pair. Intra-sentence co-occurrence likelihood that expresses the ease of creating a parallel predicate syntax is calculated and extracted as a feature. This is characterized by the property that “words in an antonym relation can appear in the same sentence”. The intra-sentence co-occurrence likelihood calculation unit 708 can be represented by a configuration including a parallel predicate syntax generation unit 710 and a parallel predicate syntax likelihood calculation unit 712. Note that the feature representing the intra-sentence co-occurrence likelihood is an example of the fourth feature.

  The parallel predicate syntax generation unit 710 generates a parallel predicate syntax including each pair of all the learning corpora (predicate pair or “term-predicate” pair) stored in the learning corpus storage unit 222. To do. Specifically, one predicate included in a predicate pair or “term-predicate” pair and the other predicate are connected by a connection particle “Tori” to generate a parallel predicate syntax. In this embodiment, two types of parallel predicate syntax in which the order of one predicate and the other predicate are exchanged are generated as the parallel predicate syntax. For example, when there are "item-ga-many" and "item-ga-many" as "term-predicate" pairs, two types of parallel predicates, "more or less" and "less or more" Generate syntax.

  For each pair (predicate pair or “term-predicate” pair), the parallel predicate syntax likelihood calculation unit 712 is the likelihood of the two types of parallel predicate syntax of the pair generated by the parallel predicate syntax generation unit 710. , And the intra-sentence co-occurrence likelihood is extracted based on the calculated likelihoods of the two types of parallel predicate syntax, and is output to the feature set storage unit 228. Specifically, an n-gram score representing the plausibility of the parallel predicate syntax is calculated using the n-gram model stored in the n-gram model storage unit 706. In this embodiment, n = 3 is used to obtain n-gram scores for two types of parallel predicate syntax, and the higher score is adopted as the intra-sentence co-occurrence likelihood. Note that the average value of n-gram scores for two types of parallel predicate syntaxes may be used as the intra-sentence co-occurrence likelihood.

  The partial character string feature extraction unit 714 performs, for each of all pairs (predicate pairs or “term-predicate” pairs) of the learning corpus stored in the learning corpus storage unit 222, the pair (predicate pair or “ The first word of each predicate included in the "term-predicate" pair), a character string combining the initials of one predicate and the initials of the other predicate, and the other predicate Is extracted as a partial character string feature and extracted as a feature. For example, when there is a predicate pair of “hospitalization” and “discharge”, two character strings “entrance / exit” and “exit” are extracted as features and output to the feature set storage unit 228 as shown in FIG. To do. The feature representing the partial character string feature is an example of the third feature.

  The inter-situation score calculation unit 716 stores each pair of learning corpora (predicate pair or “term-predicate” pair) stored in the learning corpus storage unit 222 in the inter-event database storage unit 718. Using the inter-event database, the inter-score score between the predicates from one predicate to the other predicate included in the pair as shown in FIG. 24, and the other predicate to one predicate The inter-situation score for the inter-event is extracted, extracted as a feature, and output to the feature set storage unit 228. Specifically, the inter-situation score is extracted from the inter-situation database in which the inter-situation score as shown in FIG. 25 is registered. In addition, when the inter-event score for the situation from one predicate to the other predicate included in the pair is not registered in the inter-event database, the inter-event score for the inter-event situation is “0”. . The feature representing the inter-situation score is an example of a ninth feature.

  The inter-event database storage unit 718 stores an inter-event database. Here, the inter-event score recorded in the inter-event database is to extract the predicate term structure pair from the large-scale text, and then calculate the co-occurrence degree of the predicate term structure pair in the sentence representing the relationship between the situations. Have won by. Since this co-occurrence degree is also calculated for a single predicate, for example, a score of 671.42 is given to the pair “hospitalization → discharge” (Non-Patent Document 6 (Tomohide Shibata, Ikuo Kurohashi ( 2011). Automatic acquisition of knowledge between situations using co-occurrence information of predicate term structure and case frames (see IPSJ SIG Technical Report, 1-8)).

  The feature set storage unit 228 extracts, for each of all the learning corpus pairs (predicate pairs or “term-predicate” pairs) stored in the learning corpus storage unit 222, Features representing pair distribution similarity, dictionary-defined sentence features, semantic attribute features, and functional expression features, features representing compound word likelihood of the pair extracted by the anonymity feature extraction unit 226, intra-sentence co-occurrence likelihood A feature set including a feature representing a degree, a feature representing a partial character string feature, and a feature representing an inter-situation score is stored. Note that a feature representing compound word likelihood, a feature representing intra-sentence co-occurrence likelihood, a feature representing a partial character string feature, and a feature representing an inter-situation score are combined as an anonymity feature. FIG. 26 shows an example of the feature set extracted.

  The anti-irrelevant classification feature construction unit 230 includes information on each predicate class of all pairs (predicate pairs or “term-predicate” pairs) stored in the learning corpus storage unit 222 and feature set storage. Classify predicate pairs or “term-predicate” pairs as “synonymous (synonymous or unsympathetic) vs. irrelevant” based on each feature set of all pairs stored in part 228 A feature for learning the model is constructed as an opposite-independent classification feature and stored in the opposite-independent classification feature storage unit 232. Hereinafter, the classification of “sameness (synonymous or affirmative)” and “unrelated” will be referred to as “same-unrelated classification”. Specifically, when the predicate class of the pair is “synonymous” or “disobedient”, the feature set of the pair is “positive example”, and the pair when the predicate class of the pair is “unrelated” A feature having a feature set of “a negative example” is constructed as an anti-independent classification feature. When there are “item-predicate” pair “item-ga-many” and “item-ga-minor”, the predicate class of the pair is “definience”, so the anti-unrelated classification feature building unit At 230, a “positive example” feature is constructed. An example is shown in FIG.

  The anti-irrelevant classification feature storage unit 232 stores the anti-irrelevant classification feature constructed for each pair (predicate pair or “term-predicate” pair) in the anti-irrelevant classification feature construction unit 230. ing.

  The anti-independent classification model learning unit 234 has a viewpoint that “because synonyms and anomalies are semantically similar, it is better to identify whether the semantic relationship of predicates is irrelevant first”. From the same-unrelated classification feature storage unit 232, based on each of the same-unrelated classification feature, “synonym” and “non-righteousness” are defined as one predicate class. Or the contradictory-irrelevant classification model, which is a model that performs binary classification into “or irrelevant” or “irrelevant”, and outputs it to the anti-irrelevant classification model storage unit 236 and output unit 250. Specifically, LIBLINERA (Non-Patent Document 7: R.-E. Fan, K.-W. Chang, C.-J. Hsieh, X.-R. Wang, and C.-J. Lin. (2008) LIBLINEAR: A library for large linear classification Journal of Machine Learning Research 9, 1871-1874.

  The anti-irrelevant classification model storage unit 236 stores the anti-irrelevant classification model learned by the anti-irrelevant classification model learning unit 234.

  The synonym-anonymity classification feature construction unit 240 includes information on each predicate class of all pairs (predicate pairs or “term-predicate” pairs) stored in the learning corpus storage unit 222 and a feature set storage. Based on the feature sets of each pair stored in the part 228, for each pair whose predicate class is “synonymous” or “dissent”, the pair stored in the feature set storage unit 228 (description Using the feature set of a part pair or “term-predicate” pair), a feature for learning a model that classifies a pair as “synonymous vs. anomaly” is constructed as a synonym—an anonymity classification feature, Store in the anonymity classification feature storage unit 242. Hereinafter, the classification of “synonymous” and “antinomy” will be referred to as “synonymous-antinomy classification”. Specifically, when the predicate class of the pair (predicate pair or “term-predicate” pair) is “synonymous”, the feature set of the pair is “positive example”, and the pair (predicate pair or predicate pair or When the predicate class of the “term-predicate” pair) is “antony”, a feature having the feature set of the pair as a “negative example” is constructed as a synonym-anonymity classification feature. In the case where there are “article-ga-many” and “article-ga-minor” as the “term-predicate” pair, the predicate class of the pair is “annoyance”, so the synonym-anonymity classification feature construction unit 240 The feature of “negative example” is constructed. An example is shown in FIG.

  The synonym-anonymity classification feature storage unit 242 includes a pair (predicate pair or “term-predicate” pair) whose predicate class is “synonymous” or “anonymity” in the synonym-anonymous classification feature construction unit 240. Remember the synonym-anonymity classification feature built for.

  The synonym-anonymity classification model learning unit 244 binarizes a pair as “synonymous” or “annoyance” based on each of the synonym-anonymity classification feature storage unit 242 stored in the synonym-anonymity classification feature storage unit 242. A synonym-anonymity classification model that is a model to be classified is learned and output to the synonym-anonymity classification model storage unit 246 and the output unit 250. Specifically, LIBLINERA (Non-Patent Document 7: R.-E. Fan, K.-W. Chang, C.-J. Hsieh, X.-R. Wang, and C.-J. Lin. (2008) LIBLINEAR: A library for large linear classification Journal of Machine Learning Research 9, 1871-1874.

  The synonym-anonymous classification model storage unit 246 stores the synonym-anonymous classification model learned by the synonym-anonymous classification model learning unit 244.

<Configuration of classification determination device>
Next, the configuration of the classification determination apparatus 800 according to the embodiment of the present invention will be described in detail. As shown in FIG. 29, a classification determination apparatus 800 according to an embodiment of the present invention is a computer including a CPU, a RAM, and a ROM that stores a program for executing a classification determination processing routine described later and various data. Can be configured. Functionally, the classification determination apparatus 800 includes an input unit 810, a calculation unit 820, and an output unit 850 as shown in FIG.

  The input unit 810 receives a classification determination target pair (predicate pair or “term-predicate” pair) from an input device such as a keyboard. Note that the input unit 810 may accept input from the outside via a network or the like.

  The calculation unit 820 includes a synonym feature extraction unit 822, an anonymity feature extraction unit 824, a feature construction unit 826, an anti-independence classification model storage unit 828, an anti-inverse relationship determination unit 830, and a synonym-inverse It can be expressed by a configuration including a determination unit 832 and a synonym-anonymity classification model storage unit 834.

  Similarly to the synonym feature extraction unit 224 of the classification model learning device 200, the synonym feature extraction unit 822 applies the classification determination target pair (predicate pair or “term-predicate” pair) received by the input unit 810 to the same. The “distribution similarity”, “dictionary definition sentence feature”, “semantic attribute feature”, and “functional expression feature” are extracted and output to the feature construction unit 826.

  The anonymity extraction unit 824 is similar to the anomaly feature extraction unit 226 of the classification model learning device 200 with respect to the classification determination target pair (predicate pair or “term-predicate” pair) received by the input unit 810. , “Feature representing compound word likelihood”, “feature representing intra-sentence co-occurrence likelihood”, “feature representing partial character string”, and “feature representing inter-situation score”, respectively, To 826.

  The feature construction unit 826 constructs a feature set including the features extracted by the synonym feature extraction unit 822 and the features extracted by the anonymity feature extraction unit 824 as features for predicate class determination. FIG. 30 shows an example of a feature set constructed by the feature construction unit 826 when “item-many” and “article-many” are input as “item-predicate” pairs.

  The anti-irrelevant classification model storage unit 828 stores the anti-irrelevant classification model previously learned by the classification model learning device 200.

  Based on the feature set input from the feature construction unit 826, the same-unrelated determination unit 830 is input using the same-unrelated classification model stored in the same-unrelated classification model storage unit 828. It is determined whether a predicate class that is most likely to be a pair of classification determination targets (predicate pair or “term-predicate” pair) is “synonymous (synonymous or synonymous)” or “unrelated”. When the most likely predicate class is determined to be “unrelated”, the predicate class of the input classification determination target pair (predicate pair or “term-predicate” pair) is “unrelated” Is output to the output unit 850. If the most likely predicate class is determined to be “synonymous”, the result is output to the synonym-antegiance determining unit 832. In FIG. 30, when “item-many” and “article-many” are input as “term-predicate” pairs, the predicate class is “conflict” in the reciprocity-irrelevant determination unit 830. An example in which it is determined to be present will be shown.

  Based on the feature set of pairs (predicate pairs or “term-predicate” pairs) in which the predicate class is determined to be “same” in the synonym-unrelated determination unit 830, Using a synonym-anonymity classification model stored in the synonym-anonymity classification model storage unit 834, a pair (predicate pair or It is determined whether the most likely predicate class of the “term-predicate” pair is “synonymous” or “dissent”, and the determination result is output to the output unit 850. In FIG. 31, when “article-ga-many” and “article-ga-minor” are input as the “term-predicate” pair, the predicate class is “annoyance” in the synonym-anonymity determination unit 832. An example in which it is determined to be present will be shown.

  The synonym-anonymity classification model storage unit 834 outputs the synonym-anonymity classification model obtained in advance by the classification model learning device 200 and stored in advance.

<Operation of feature vector construction device>
Next, the operation of the feature vector construction device 100 according to the embodiment of the present invention will be described. First, a text corpus is input by the input unit 10. Then, when the CPU executes the program stored in the ROM of the feature vector construction device 100, the feature vector construction processing routine shown in FIG. 32 is executed.

  First, in step S100, a text corpus that is a set of a plurality of sentences is read.

  First, in step S102, morphological analysis is performed on one sentence among the plurality of sentences received in step S100.

  Next, in step S104, dependency analysis is performed on the sentence for which morphological analysis was performed in step S100.

  Next, in step S106, the "term-predicate" included in the sentence from the analysis result including at least the notation for each morpheme, the standard form, the part of speech, and the dependency information for each clause obtained in steps S102 and S104. Information (context information) of words appearing around "is extracted as a term-predicate feature. Further, in step S106, the morpheme notation obtained in steps S102 and S104, the standard form and the part of speech, and the analysis result including at least dependency information for each clause appear around the predicate included in the sentence. Extract word information (context information) as predicate features. In the present embodiment, for example, the term-predicate feature and the predicate feature are extracted by the same method as in Non-Patent Document 2 described above. Specifically, another “term-predicate” and “predicate” related to the target “term-predicate” are extracted as a term-predicate feature. Furthermore, a term (noun phrase having a case particle) related to a single “predicate” and another “predicate” are extracted as predicate features.

  Next, in step S108, it is determined whether or not the processing in steps S102 to S106 has been executed for all of the plurality of sentences received in step S100. When the processes of steps S102 to S106 are executed for all sentences, the process proceeds to step S110, and when there is a sentence that does not execute the processes of steps S102 to S106, the process proceeds to step S102. Repeat the process.

  Next, in step S110, a feature vector is constructed based on the features extracted in step S106 for the “term-predicate” included in at least one sentence. In step S110, a feature vector is constructed based on the features extracted in step S106 for the predicates included in at least one sentence.

  Next, in step S112, it is determined whether or not a term-predicate feature vector has been constructed for each of all “term-predicate” included in at least one sentence read in step S100. In step S112, it is determined whether a predicate feature vector has been constructed for each of all predicates included in at least one sentence read in step S100. If a term-predicate feature vector is constructed for each of all “terms-predicates”, the process proceeds to step S114, and there is a “term-predicate” in which no term-predicate feature vector is constructed. In the case, the process proceeds to step S110 and each process is repeated. If a predicate feature vector is constructed for each of all predicates, the process proceeds to step S114. If there is a predicate that does not construct a predicate feature vector, the process proceeds to step S110. Repeat each process.

  Next, in step S114, all the feature vectors constructed in step S110 are output by the output unit 30, and the process is terminated.

In this embodiment, both the predicate feature vector and the term-predicate feature vector are created. However, only the feature vector used in the synonym learning device and the synonym judgment device may be created.

<Operation of classification model learning device>
Next, the operation of the classification model learning device 200 according to the embodiment of the present invention will be described. First, the feature vector output by the feature vector construction apparatus 100 in advance is stored in the feature vector storage unit 300. In addition, the n-gram model is received in advance and stored in the n-gram model storage unit 706. In addition, the inter-event database is received and stored in the inter-event database storage unit 718 in advance. A learning corpus is input by the input unit 210 and stored in the learning corpus storage unit 222. Then, when the CPU executes the program stored in the ROM of the classification model learning device 200, the classification model learning processing routine shown in FIG. 33 is executed.

  First, in step S200, a learning corpus is read.

In the subsequent steps, when learning a synonym determination model of predicate pairs, processing is performed using only the learning data of the predicate pair in the read learning corpus, and the classification of the “term-predicate” pair is performed. When learning the model, the processing is performed using only the learning data of the “term-predicate” pair in the read learning corpus.

  Next, in step S202, synonymity is extracted for each of the plurality of learning data obtained in step S200.

  The step S202 is realized by the synonymity extraction processing routine shown in FIG.

  In step S300 of FIG. 34, each of the learning data pairs (predicate pair or “term-predicate” pair) stored in the feature vector storage unit 300 for each of the plurality of learning data obtained in step S200. Based on the predicate feature vector, the predicate distribution similarity of the learning data pair is calculated. In step S300, for each of the plurality of learning data obtained in step S200, based on each term-predicate feature vector of the “term-predicate” pair of the learning data stored in feature vector storage unit 300. Thus, the term-predicate distribution similarity of the learning data pair is calculated. In the present embodiment, the predicate distribution similarity is calculated for the “term-predicate” pair, but only the term-predicate distribution similarity may be calculated.

  Next, in step S302, for each of the plurality of learning data obtained in step S200, the dictionary definition sentence feature of the learning data pair (predicate pair or “term-predicate” pair) is extracted.

  Next, in step S304, for each of the plurality of learning data obtained in step S200, semantic attribute features of the learning data pair (predicate pair or “term-predicate” pair) are extracted.

  Next, in step S306, for each of the plurality of learning data obtained in step S200, the function representation feature of the learning data pair (predicate pair or “term-predicate” pair) is extracted.

  Step S300 is realized by the distribution similarity calculation processing routine shown in FIG.

  First, in step S400, for each predicate pair of learning data to be calculated, each predicate feature vector of the pair stored in the feature vector storage unit 300 is read. In step S400, for the “term-predicate” pair of the learning data to be calculated, the predicate feature vector and each term-predicate feature vector of the pair stored in the feature vector storage unit 300 are read. In this embodiment, the predicate feature vector is read for the “term-predicate” pair of the correct answer data to be calculated.

  Next, in step S402, based on the feature vector read in step S400, the distribution similarity is calculated for the learning data pair (predicate pair or “term-predicate” pair).

  Next, in step S404, it is determined whether or not the distribution similarity of the learning data pair has been calculated for each of all the learning data read in step S200. If the distribution similarity is calculated for each pair of all learning data (predicate pair or “term-predicate” pair), the process is terminated, and the pair of learning data for which distribution similarity is not calculated If there is a (predicate pair or “term-predicate” pair), the process proceeds to step S400, and each process is repeated using the learning data as learning data to be calculated.

  Step S302 is realized by the dictionary definition sentence feature extraction processing routine shown in FIG.

  First, in step S500, a definition sentence set of each predicate of a pair of learning data to be extracted (predicate pair or “term-predicate” pair) is extracted from the definition sentence dictionary storage unit 304.

  Next, in step S502, morphological analysis is performed on each definition sentence of the definition sentence set extracted in step S500.

  Next, in step S504, based on the result of morpheme analysis including at least the notation, the standard form, the part of speech, and the reading for each morpheme obtained in step S502, a feature indicating definition sentence mutual complementarity is extracted. Pred1Match and Pred2Match are extracted as features indicating definition sentence mutual complementarity. If the pair of correct data to be extracted is a “term-predicate” pair, Arg1Match and Arg2Match can be further extracted.

  Next, in step S506, based on the result of morpheme analysis including at least the notation, the standard form, the part of speech, and the reading for each morpheme obtained in step S502, a feature indicating vocabulary overlap is extracted. At least one of NounMatch, VerbMatch, AdjMatch, and AdjNMatch is extracted as a feature indicating vocabulary overlap.

  Next, in step S508, a dictionary definition sentence feature is constructed based on the feature indicating the mutual definition of the definition sentences extracted in step S504 and the feature indicating the overlap of the vocabulary extracted in step S506.

  Next, in step S510, it is determined whether a dictionary definition sentence feature has been constructed for each pair (predicate pair or “term-predicate” pair) of all the learning data read in step S202. When the dictionary definition sentence feature is constructed for all the learning data pairs (predicate pair or “term-predicate” pair), the processing is terminated, and the learning data pair that does not construct the dictionary definition sentence feature ( If there is a predicate pair or “term-predicate” pair), the process proceeds to step S500, and each process is repeated using the learning data as extraction target learning data.

  Step S304 is realized by the semantic attribute feature extraction processing routine shown in FIG.

  First, in step S600, overlapping semantic attributes of each predicate of a pair of learning data to be extracted (predicate pair or “term-predicate” pair) are extracted.

  Next, in step S602, based on the overlapping semantic attribute obtained in step S600, a semantic attribute weighted overlap rate is calculated for a pair of learning data to be extracted (predicate pair or “term-predicate” pair). .

  Next, in step S604, based on the overlapping semantic attribute obtained in step S600 and the semantic attribute weighted overlapping rate obtained in step S602, the pair of learning data to be extracted (predicate pair or “term- Construct semantic attribute features for predicate pairs.

  Next, in step S606, it is determined whether or not a semantic attribute feature has been constructed for each pair (predicate pair or “term-predicate” pair) of all the learning data read in step S602. When a semantic attribute feature is constructed for each pair of all learning data (predicate pair or “term-predicate” pair), the processing is terminated, and a pair of learning data for which no semantic attribute feature is constructed If there is a (predicate pair or “term-predicate” pair), the process proceeds to step S600, and each process is repeated with the learning data as learning data to be extracted.

  The above step S306 is realized by the function expression feature extraction processing routine shown in FIG.

  First, in step S700, morpheme analysis of each predicate of a pair of learning data to be extracted (predicate pair or “term-predicate” pair) is performed, and an analysis including at least a notation, a standard form, and a part of speech for each morpheme. The most likely meaning label string is assigned using the result.

  Next, in step S702, overlapping semantic labels that appear in common in both semantic labels of the predicates of each pair obtained in step S700 are extracted.

  Next, in step S704, the semantic label overlap rate is calculated using the paired overlapping semantic labels obtained in step S702.

  Next, in step S706, based on the overlapping semantic label obtained in step S702 and the semantic label overlapping rate obtained in step S704, the pair of learning data to be extracted (predicate pair or “term-predicate”). “Pairs” are constructed with functional expression features.

  Next, in step S708, it is determined whether or not a function representation feature has been constructed for each pair (predicate pair or “term-predicate” pair) of all the learning data read in step S702. A pair of learning data that does not have a functional representation feature when processing of the functional representation feature is completed for each pair of learning data (predicate pair or "term-predicate" pair) If there is a (predicate pair or “term-predicate” pair), the process proceeds to step S700, and each process is repeated with the learning data as the correct answer data to be extracted.

  Next, in step S204 in FIG. 33, anonymity is extracted for each of the plurality of learning data obtained in step S200.

  Step S204 is realized by the anonymity extraction processing routine shown in FIG.

  Next, in step S800, the n-gram model stored in the n-gram model storage unit 706 is read.

  Next, in step S802, the event database stored in the event database storage unit 718 is read.

  Next, in step S804, morpheme analysis is performed on each predicate of the extraction target learning data pair (predicate pair or “term-predicate pair”) included in the learning corpus acquired in step S200.

  Next, in step S806, based on the morphological analysis results of each predicate acquired in step S804, two types of compound words are generated for the extraction target pair (predicate pair or “term-predicate” pair). To do.

  Next, in step S808, the compound word likelihood of each of the two types of compound words generated in step S806 is calculated using the n-gram model acquired in step S800, and extracted as a feature.

  Next, in step S810, two types of parallel predicate syntax are generated for the pair to be extracted (predicate pair or “term-predicate” pair).

  Next, in step S812, the intra-sentence co-occurrence likelihood of each of the two types of parallel predicate syntax generated in step S810 is calculated using the n-gram model acquired in step S800, and extracted as a feature.

  Next, in step S814, an inter-situation score is extracted from the pair to be extracted (predicate pair or “term-predicate” pair) using the inter-event database acquired in step S802, and extracted as a feature.

  Next, in step S816, one initial character of each predicate of the pair to be extracted (predicate pair or “term-predicate” pair) is extracted, and the initial character of one predicate and the other predicate is extracted. And the partial character string feature that combines the other predicate and the initial character of one predicate are extracted as features.

  Next, in step S818, it is determined whether or not the processing in steps S804 to S818 has been performed for all pairs (predicate pairs or “term-predicate” pairs) included in the learning corpus acquired in step S200. . When processing is performed for all pairs (predicate pair or “term-predicate” pair), processing is terminated and processing is performed for all pairs (predicate pair or “term-predicate” pair). If not, the pair to be extracted (predicate pair or “term-predicate” pair) is changed, and the processing from step S804 to step S818 is repeated.

  Step S806 is realized by the compound word generation processing routine shown in FIG.

  First, in step S900, based on the morphological analysis result of each predicate of the pair to be extracted (predicate pair or “term-predicate” pair) acquired in step S804, a pair (predicate pair or “term- It is determined whether the part of speech of both predicates of the predicate “pair” is a Japanese verb or an adjective. When the part of speech of both predicates of the pair (predicate pair or “term-predicate” pair) is a Japanese verb or adjective, the process proceeds to step S902, and the pair (predicate pair or “term-predicate”) If the part of speech of at least one predicate of the pair) is neither a Japanese verb nor an adjective, the process proceeds to step S904.

  Next, in step S902, two types of compound words are generated by combining the consecutive forms of the predicates of the extraction coping pair (predicate pair or “term-predicate” pair), and the process is terminated. The compound words are generated in two types: a compound word when one predicate appears first and a compound word when the other predicate appears first.

  Next, in step S904, based on the morphological analysis result of each predicate of the pair to be extracted (predicate pair or “term-predicate” pair) acquired in step S804, a pair (predicate pair or “term” -Part of speech of both predicates in the "-predicate" pair) is a behavioral noun, and there is a character common to each predicate of the pair to be extracted (predicate pair or "term-predicate" pair) Determine whether or not. If the part of speech of both predicates of the pair (predicate pair or “term-predicate” pair) is an action noun and there is a common character, the process proceeds to step S908, and the pair (predicate pair or When the part of speech of at least one predicate of the “term-predicate pair” is not a behavioral noun, or there is no character common to each predicate of the pair (predicate pair or “term-predicate” pair) Then, the process proceeds to step S910.

  Next, in step S908, a character common to each of the extraction target pairs (predicate pair or “term-predicate” pair) is held, and two types of compound words are generated by continuing different character strings. The process ends. The compound words are generated in two types: a compound word when one predicate appears first and a compound word when the other predicate appears first.

  Next, in step S910, the predicates of the pair to be extracted (predicate pair or “term-predicate” pair) are combined as they are to generate two types of compound words, and the process ends. The compound words are generated in two types: a compound word when one predicate appears first and a compound word when the other predicate appears first.

  Next, in step S206 of FIG. 33, for each of the plurality of learning data obtained in step S200, a feature set including each of the synonymous features acquired in step S202 and each of the anonymity features acquired in step S204, and Based on the predicate class assigned to the pair (predicate pair or “term-predicate” pair), an anti-irrelevant classification feature is constructed.

  Next, in step S208, among the plurality of learning data obtained in step S200, the predicate class given to the pair (predicate pair or “term-predicate” pair) is “synonymous” or “dissent”. ”For each of the synonym features acquired in step S202 and the feature set including each of the anonymity features acquired in step S204, and the pair (predicate pair or“ term-predicate ”pair) The synonym-anonymity classification feature is constructed based on the predicate class assigned to.

  Next, in step S210, learning of the anti-irrelevant classification model is performed based on the anti-irrelevant classification feature acquired in step S206. The learned anti-irrelevant classification model is stored in the anti-irrelevant classification model storage unit 236.

  Next, in step S212, the synonym-anonymity classification model is learned based on the synonym-anonymity classification feature acquired in step S208. The learned synonym-anonymity classification model is stored in the synonym-anomaly classification model storage unit 246.

  Next, in step S214, the same-unrelated classification model acquired in step S210 and the synonym-anonymous classification model acquired in step S212 are output to the output unit 250, and the process ends.

<Operation of classification judgment device>
Next, the operation of the classification determination apparatus 800 according to the embodiment of the present invention will be described. First, the same-unrelated classification model output by the classification model learning device 200 in advance is stored in the same-unrelated classification model storage unit 828. In addition, the synonym-anonymity classification model output by the classification model learning device 200 in advance is stored in the synonym-anonymity classification model storage unit 834. In addition, the same n-gram model as the n-gram model stored in the n-gram model storage unit 706 in advance is stored in the anonymity extraction unit 824. Further, the same event database stored in the event database storage unit 718 in advance is stored in the anonymity extracting unit 824. Further, the same feature vector as the feature vector stored in advance in the feature vector storage unit 300 is stored in the synonym feature extraction unit 822. When the determination target pair (predicate pair or “term-predicate” pair) is input by the input unit 810, the CPU executes the program stored in the ROM of the classification determination device 800. The classification determination processing routine shown in FIG. 41 is executed.

  First, in step S1000, a determination target pair (predicate pair or “term-predicate” pair) input in the input unit 810 is received.

  Next, in step S1002, the same-unrelated classification model stored in the same-unrelated classification model storage unit 828 is read.

  Next, in step S1004, the synonym-anonymity classification model storage unit 834 stores the synonym-anonymity classification model.

  Next, in step S1006, as in step S202, each of the synonymous features is extracted for the input judgment target pair (predicate pair or “term-predicate” pair).

  Next, in step S1008, as in step S204, each of the anonymity features is extracted for the input judgment target pair (predicate pair or “term-predicate” pair).

  Next, in step S1010, based on each of the synonyms acquired in step S1006 and each of the antonyms acquired in step S1008, a determination target pair (predicate pair or “term-predicate” pair) Create a feature set of.

  Next, in step S1012, based on the feature set created in step S1010 and the same-unrelated classification model acquired in step S1002, a pair of determination targets (predicate pair or “term-predicate” pair) is determined. It is determined whether the predicate class is “synonymous (synonymous or affirmative)” or “unrelated”.

  Next, in step S1014, it is determined whether the predicate class of the determination target pair determined in step S1012 is “irrelevant”. When the predicate class of the determination target pair is “unrelated”, the process proceeds to step S1018, and when the predicate class of the determination target pair is “synonymous (synonymous or affirmative)”, The process proceeds to step S1016.

  Next, in step S1016, based on the feature set created in step S1010 and the synonym-anonymity classification model acquired in step S1004, a pair of determination targets (predicate pair or “term-predicate” pair) is determined. It is determined whether the predicate class is “synonymous” or “antitrust”.

  Next, in step S1018, the determination result obtained in step S1012 or the determination result obtained in step S1016 is output as a predicate class of a determination target pair (predicate pair or “term-predicate” pair). The data is output to the unit 850 and the process is terminated.

  As described above, according to the classification model learning device according to the embodiment of the present invention, learning a classification model that determines with high accuracy whether a predicate pair is synonymous, anonymity, or irrelevant. Can do.

  In addition, according to the classification determination apparatus according to the embodiment of the present invention, it is possible to determine with high accuracy whether or not the predicate pair is synonymous, anomaly, or irrelevant.

  In addition, in the application such as text mining, which extracts useful information from a large amount of text by classifying the semantic relationship of predicates into three types: “synonymous”, “anti-righteous”, and “unrelated”. It is possible to summarize expressions and extract and analyze conflicting opinions that conflict.

  Also, when classifying the semantic relations of predicates, first synonymously synonymous “synonyms” and “antisenses” are combined into one predicate class, and the classification “sameness-unrelated classification” Next, since “synonymous” is a binary classification of “synonymous”, “synonymous-defensive classification”, it is possible to acquire a wide range of expressions with high accuracy and different from existing methods. , Improve the accuracy of applications that require natural language understanding, such as search and opinion analysis.

  Note that the present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  In the present embodiment, when generating a parallel predicate syntax, the case where the parallel predicate syntax is generated by connecting with a connection particle “Tari” has been described. For example, a parallel predicate syntax may be generated using other connection particles.

  In the present embodiment, the case where the intra-sentence co-occurrence likelihood is calculated using the n-gram score has been described, but the present invention is not limited to this. For example, the document frequency (df) on the Web, which is the number of hits generated by searching the parallel predicate syntax as the intra-sentence co-occurrence likelihood, may be used.

  Further, in the present embodiment, the case where learning is performed using LIBLINERA for learning the anti-independent classification model and the synonym-anti-sense classification model is not limited to this. For example, other discriminative learning such as Decision Tree may be used.

10, 210, 810 Input unit 20, 220, 820 Calculation unit 24 Basic analysis unit 26 Feature extraction unit 28 Feature vector construction unit 30, 250, 850 Output unit 100 Feature vector construction device 200 Classification model learning device 222 Learning corpus storage unit 224 , 822 Synonym feature extraction unit 226, 824 Anti-feature feature extraction unit 228 Feature set storage unit 230 Same-unrelated classification feature construction unit 232 Same-unrelated classification feature storage unit 234 Same-unrelated classification model learning unit 236, 828 Same Anti-unrelated classification model storage unit 240 Synonym-anonymity classification feature construction unit 242 Synonym-anonymity classification feature storage unit 244 Synonym-anonymity classification model learning unit 246, 834 Synonym-anonymity classification model storage unit 300 Feature vector storage unit 302 Distribution similarity calculation unit 304 Definition sentence dictionary storage unit 306 Dictionary definition sentence feature extraction unit 308 Semantic attribute dictionary storage unit 310 Semantic attribute feature extraction unit 312 Functional expression dictionary storage unit 314 Functional expression feature extraction unit 400 Definition sentence extraction unit 402 Definition sentence mutual complementation extraction unit 404 Lexical overlap extraction unit 500 Semantic attribute overlap extraction unit 502 Meaning Attribute weighted overlap rate calculation unit 600 Semantic label assignment unit 602 Semantic label extraction unit 604 Semantic label overlap rate calculation unit 700 Compound word score calculation unit 702 Compound word generation unit 704 Compound word likelihood calculation unit 706 Model storage unit 708 Occurrence likelihood calculation unit 710 Parallel predicate syntax generation unit 712 Parallel predicate syntax likelihood calculation unit 714 Partial character string feature extraction unit 716 Inter-situation score calculation unit 718 Inter-event database storage unit 800 Classification determination device 826 Feature construction unit 828 Unrelated classification model storage unit 830 Same-unrelated determination unit 832 Synonym-anonymity determination unit 34 synonymous - Anti defined classification model storage unit

Claims (16)

Each of the predicates of the predescription part pair extracted based on the definition sentence of each predicate of the predicate pair obtained from the definition sentence set including the definition sentences for each of the plurality of predicates prepared in advance. The feature indicating the mutual complementarity of the definition statement indicating whether or not there is a pair of predicates in the definition statement is the first feature,
Semantic attributes common to the predescription part pair extracted based on the semantic attributes of each predicate of the predescription part pair obtained from the semantic attribute set consisting of the semantic attributes for each of the plurality of predicates prepared in advance. The second feature,
The combination of the partial character string in the character string representing one predicate of the preceding description part pair and the partial character string in the character string representing the other predicate is a third feature,
The likelihood of the parallel predicate syntax generated including each predicate of the predescription part pair is a fourth feature,
For each of the predicate pairs in which any one of synonym, contradiction, and irrelevance is previously given as a classification, at least one of the first feature and the second feature, and the third feature, A feature extraction unit that extracts a feature set including the fourth feature;
The feature set extracted by the feature extraction unit for each predicate pair assigned synonym or anomaly as a classification and the feature extraction unit extracted for each predicate pair assigned an irrelevance as a classification An anti-irrelevant classification feature structuring unit that constructs the feature set as an anti-irrelevant classification feature for classifying a predescription part pair as synonymous or irrelevant.
The feature set extracted by the feature extraction unit for each predicate pair given synonym as a classification, and the feature extracted by the feature extraction unit for each predicate pair given anonymity as a classification A synonym-anonymity classification feature construction unit that constructs a set as a synonym-anonymity classification feature for classifying a predescription part pair into synonym and anomaly;
Reciprocal-irrelevant classification model learning for learning a model for classifying predicate pairs irrelevant and irrelevant based on the reciprocal-irrelevant classification feature constructed by the reciprocal-irrelevant classification feature construction unit And
A synonym-anonymity classification model learning unit that learns a model for classifying predicate pairs into synonyms and anomalies based on the synonym-anonymity classification feature constructed by the synonym-anonymity classification feature construction unit When,
Classification model learning apparatus including   The classification model learning device according to claim 1, wherein the feature extraction unit extracts an n-gram score of the parallel predicate syntax as a likelihood of the parallel predicate syntax. The feature extraction unit includes:
For each predicate part of the previous description part pair, the co-occurrence of words appearing in the vicinity of the previous description part in the text corpus is calculated, and the distribution similarity obtained by calculating the similarity of the surrounding words of the previous description part pair is the fifth distribution similarity. With features,
A predescription part pair extracted based on a semantic label of each predicate of the predescription part pair obtained from a semantic label set made up of semantic labels of the respective function expressions of a plurality of predicates prepared in advance. The semantic label that is common to
The overlapping degree of the common semantic labels of the previous description part pair is a seventh feature,
The likelihood of the compound word generated by compounding each predicate of the preceding description part pair is the eighth feature,
The degree to which the two predicates of the previous description part pair co-occur as two predicates representing the situation in the sentence representing the relationship between the situations is the ninth feature,
At least one of the first feature and the second feature, and the third feature, for each of the predicate pairs in which any one of the synonym, contradiction, and irrelevance is previously given as a classification And extracting the feature set including at least one of the fourth feature, the fifth feature, the sixth feature, the seventh feature, the eighth feature, and the ninth feature The classification model learning device according to claim 1 or 2.   The feature extraction unit extracts at least one of an n-gram score of the compound word generated for a previous description unit pair and an appearance frequency of the compound word in a corpus as the likelihood of the compound word. The described classification model learning device. The “term-predicate” is extracted based on the definition statement of each of the predicates of the “term-predicate” pair, which is obtained from a definition statement set including definition statements for each of a plurality of predicates prepared in advance. A feature indicating mutual complementarity indicating whether or not there is a pair of predicates in the definition statement of each of the predicates of the pair, and each of the predicates of the “term-predicate” pair Definition sentence mutual complementarity that indicates whether or not at least the previous description part exists among the features indicating the mutual complementarity of the definition sentence that indicates whether there is a pair of "term-predicate" terms in the definition sentence The feature indicating is the first feature,
The “term-predicate” extracted based on the semantic attribute of each predicate of the “term-predicate” pair obtained from a semantic attribute set consisting of semantic attributes for each of a plurality of predicates prepared in advance. The second attribute is the semantic attribute common to the pair,
A combination of a partial character string in the character string representing one predicate of the “term-predicate” pair and a character string representing the other predicate is a third feature,
The likelihood of the parallel predicate syntax generated including each of the predicates of the “term-predicate” pair is a fourth feature,
For each of the “term-predicate” pairs in which any one of synonym, contradiction, and irrelevance is previously given as a classification, at least one of the first feature and the second feature, and the first feature A feature extraction unit that extracts a feature set including three features and the fourth feature;
About each of the "term-predicate" pairs that are extracted by the feature extractor for each of the "term-predicate" pairs that are given synonyms or anomalies as a classification, and for each of the "term-predicate" pairs that are assigned an irrelevant as a classification The feature set extracted by the feature extraction unit is constructed as an anti-irrelevant classification feature for classifying the “term-predicate” pair as irrelevant or irrelevant. An anti-irrelevant classification feature construction unit;
The feature set extracted by the feature extraction unit for each “term-predicate pair” to which synonyms are assigned as classifications, and the feature for each “term-predicate” pair to which anonymity is assigned as classifications A synonym-anonymity classification feature construction unit that constructs the feature set extracted by the extraction unit as a synonym-anonymity classification feature for classifying the "term-predicate" pair into synonyms and anomalies;
Reciprocity learning a model for classifying a “term-predicate” pair as irrelevant and irrelevant based on the reciprocal-irrelevant classification feature constructed by the reciprocal-irrelevant classification feature construction unit An unrelated classification model learning unit;
A synonym-anonymity learning a model for classifying a “term-predicate” pair into a synonym and anomaly based on the synonym-anonymity classification feature constructed by the synonym-anonymity classification feature construction unit A classification model learning unit;
Classification model learning apparatus including The feature extraction unit calculates the co-occurrence of words appearing around the “term-predicate” in a text corpus for each of the “term-predicate” of the “term-predicate” pair. The distribution similarity that calculated the similarity of words around the term-predicate pair and the predicate of the term-predicate in the text corpus for each of the predicates of the term-predicate pair Calculate the co-occurrence of words appearing in the word, and calculate the similarity of the words surrounding the predicate of the “term-predicate” pair. The distribution similarity that calculated the degree is the fifth feature,
Extracted based on the semantic labels of each functional expression of the predicates of the “term-predicate” pair obtained from the semantic label set consisting of the semantic labels of the functional expressions of a plurality of predicates prepared in advance. A semantic label common to the “term-predicate” pair is a sixth feature,
The overlapping degree of the common semantic labels of the “term-predicate” pair is a seventh feature,
The likelihood of a compound word generated by compounding each of the predicates of the “term-predicate” pair is an eighth feature,
The ninth feature is the degree that the two predicates of the “term-predicate” pair co-occur as two predicates representing the situation in a statement representing the relationship between the situations.
For each of the “term-predicate” pairs, each of which is pre-assigned as a classification of any of the synonyms, anomalies, and irrelevance, and at least one of the first feature and the second feature, A third feature, the fourth feature, and at least one of the fifth feature, the sixth feature, the seventh feature, the eighth feature, and the ninth feature. The classification model learning device according to claim 5, wherein the feature set is extracted. Each of the predicates of the predescription part pair extracted based on the definition sentence of each predicate of the predicate pair obtained from the definition sentence set including the definition sentences for each of the plurality of predicates prepared in advance. The feature indicating the mutual complementarity of the definition statement indicating whether or not there is a pair of predicates in the definition statement is the first feature,
Semantic attributes common to the predescription part pair extracted based on the semantic attributes of each predicate of the predescription part pair obtained from the semantic attribute set consisting of the semantic attributes for each of the plurality of predicates prepared in advance. The second feature,
The combination of the partial character string in the character string representing one predicate of the preceding description part pair and the partial character string in the character string representing the other predicate is a third feature,
The likelihood of the parallel predicate syntax generated including each predicate of the predescription part pair is a fourth feature,
A feature extraction unit that extracts a feature set including at least one of the first feature and the second feature, the third feature, and the fourth feature for the input predicate pair;
Based on the feature set extracted by the feature extraction unit and a pre-learned model for classifying the predicate pair as synonymous or anonymity and irrelevant, An anti-irrelevant determination unit for determining whether the classification is the same or irrelevant;
When the same-unrelated determination unit determines that the classification of the previous description unit pair is synonymous, the feature set extracted by the feature extraction unit and the predicate pair learned in advance are synonymous. A synonym-anonymity determination unit that determines whether the classification of the previous description part pair is synonymous or anomaly based on a model for classifying an anomaly;
A classification determination apparatus comprising: The feature extraction unit includes:
For each predicate part of the previous description part pair, the co-occurrence of words appearing in the vicinity of the previous description part in the text corpus is calculated, and the distribution similarity obtained by calculating the similarity of the surrounding words of the previous description part pair is the fifth distribution similarity. With features,
A predescription part pair extracted based on a semantic label of each predicate of the predescription part pair obtained from a semantic label set made up of semantic labels of the respective function expressions of a plurality of predicates prepared in advance. The semantic label that is common to
The overlapping degree of the common semantic labels of the previous description part pair is a seventh feature,
The likelihood of the compound word generated by compounding each predicate of the preceding description part pair is the eighth feature,
The degree to which the two predicates of the previous description part pair co-occur as two predicates representing the situation in the sentence representing the relationship between the situations is the ninth feature,
For the input predicate pair, at least one of the first feature and the second feature, the third feature, the fourth feature, the fifth feature, the sixth feature, The classification determination apparatus according to claim 7, wherein the feature set including at least one of a feature, the seventh feature, an eighth feature, and a ninth feature is extracted. The “term-predicate” is extracted based on the definition statement of each of the predicates of the “term-predicate” pair, which is obtained from a definition statement set including definition statements for each of a plurality of predicates prepared in advance. A feature indicating mutual complementarity indicating whether or not there is a pair of predicates in the definition statement of each of the predicates of the pair, and each of the predicates of the “term-predicate” pair Definition sentence mutual complementarity that indicates whether or not at least the previous description part exists among the features indicating the mutual complementarity of the definition sentence that indicates whether there is a pair of "term-predicate" terms in the definition sentence The feature indicating is the first feature,
The “term-predicate” extracted based on the semantic attribute of each predicate of the “term-predicate” pair obtained from a semantic attribute set consisting of semantic attributes for each of a plurality of predicates prepared in advance. The second attribute is the semantic attribute common to the pair,
A combination of a partial character string in the character string representing one predicate of the “term-predicate” pair and a character string representing the other predicate is a third feature,
The likelihood of the parallel predicate syntax generated including each of the predicates of the “term-predicate” pair is a fourth feature,
A feature that extracts a feature set including at least one of the first feature and the second feature, the third feature, and the fourth feature for the input “term-predicate” pair An extractor;
Based on the feature set extracted by the feature extraction unit and a pre-learned model for classifying `` term-predicate '' pairs as synonymous or anonymity and irrelevant An anti-irrelevant determining unit that determines whether the classification of the “term-predicate” pair is the same or unrelated;
When the classification of the “term-predicate” pair is determined to be reciprocal by the anti-irrelevant determination unit, the feature set extracted by the feature extraction unit and a pre-learned “term” -A synonym-anonymity determination unit that determines whether the classification of the "term-predicate" pair is synonymous or anomaly based on a model for classifying predicate pairs into synonyms and anomalies;
A classification determination apparatus comprising: The feature extraction unit calculates the co-occurrence of words appearing around the “term-predicate” in a text corpus for each of the “term-predicate” of the “term-predicate” pair. The distribution similarity that calculated the similarity of words around the term-predicate pair and the predicate of the term-predicate in the text corpus for each of the predicates of the term-predicate pair Calculate the co-occurrence of words appearing in and calculate the similarity of neighboring words of the predicate of the “term-predicate” pair, and at least the similarity of the neighboring words of the “term-predicate” pair among the distribution similarities The distribution similarity that calculated the degree is the fifth feature,
Extracted based on the semantic labels of each functional expression of the predicates of the “term-predicate” pair obtained from the semantic label set consisting of the semantic labels of the functional expressions of a plurality of predicates prepared in advance. A semantic label common to the “term-predicate” pair is a sixth feature,
The overlapping degree of the common semantic labels of the “term-predicate” pair is a seventh feature,
The likelihood of a compound word generated by compounding each of the predicates of the “term-predicate” pair is an eighth feature,
The ninth feature is the degree that the two predicates of the “term-predicate” pair co-occur as two predicates representing the situation in a statement representing the relationship between the situations.
For the input “term-predicate” pair, at least one of the first feature and the second feature, the third feature, the fourth feature, and the fifth feature, The classification determination apparatus according to claim 9, wherein the feature set including at least one of the sixth feature, the seventh feature, the eighth feature, and the ninth feature is extracted. A classification model learning device including a feature extraction unit, a synonym-irrelevant classification feature construction unit, a synonym-anonymity classification feature construction unit, a contra-independence classification model learning unit, and a synonym-anonymity classification model learning unit A classification model learning method in
The feature extraction unit is a predescription unit pair that is extracted based on a definition statement of each predicate of a predicate pair obtained from a definition statement set including a definition statement for each of a plurality of predicates prepared in advance. The feature indicating the mutual complementarity of the definition statement indicating whether or not there is a pair of predicates in the definition statement of each of the predicates is a first feature,
Semantic attributes common to the predescription part pair extracted based on the semantic attributes of each predicate of the predescription part pair obtained from the semantic attribute set consisting of the semantic attributes for each of the plurality of predicates prepared in advance. The second feature,
The combination of the partial character string in the character string representing one predicate of the preceding description part pair and the partial character string in the character string representing the other predicate is a third feature,
The likelihood of the parallel predicate syntax generated including each predicate of the predescription part pair is a fourth feature,
For each of the predicate pairs in which any one of synonym, contradiction, and irrelevance is previously given as a classification, at least one of the first feature and the second feature, and the third feature, , Extract a feature set including the fourth feature,
The same-unrelated classification feature construction unit includes the feature set extracted by the feature extraction unit for each predicate pair to which synonym or anomaly is given as a classification, and a predicate pair to which irrelevant is given as a classification. Constructing the feature set extracted by the feature extraction unit for each of the above as an anti-independent classification feature for classifying the previous description unit pair as irrelevant or irrelevant as synonym or anomaly,
The synonym-anonymity classification feature construction unit includes the feature set extracted by the feature extraction unit for each predicate pair to which synonym is given as a class, and each predicate pair to which an anomaly is given as a class. Constructing the feature set extracted by the feature extraction unit for as a synonym-anonymity classification feature for classifying the previous description unit pair into a synonym and an anomaly,
The same anti-irrelevant classification model learning unit is configured to classify predicate pairs as irrelevant and irrelevant based on the same anti-irrelevant classification feature constructed by the same anti-irrelevant classification feature construction unit. Learn the model,
The synonym-anonymity classification model learning unit is a model for classifying predicate pairs into synonyms and anomalies based on the synonym-anonymity classification feature constructed by the synonym-anonymity classification feature construction unit. Classification model learning method. A classification model learning device including a feature extraction unit, a synonym-irrelevant classification feature construction unit, a synonym-anonymity classification feature construction unit, a contra-independence classification model learning unit, and a synonym-anonymity classification model learning unit A classification model learning method in
The feature extraction unit is extracted based on a definition statement of each predicate of the “term-predicate” pair, which is obtained from a definition statement set including a definition statement for each of a plurality of predicates prepared in advance. A feature indicating mutual complementarity indicating whether or not a predicate as a pair exists in each definition statement of each predicate of the “term-predicate” pair, and the “term-predicate” pair Whether or not at least the previous description part exists among the features indicating the mutual complementarity of the definition sentence indicating whether or not there is a pair of “term-predicate” terms in each definition sentence of the predicate. The feature indicating the mutual definition of the definition sentence to be expressed is the first feature,
The “term-predicate” extracted based on the semantic attribute of each predicate of the “term-predicate” pair obtained from a semantic attribute set consisting of semantic attributes for each of a plurality of predicates prepared in advance. The second attribute is the semantic attribute common to the pair,
A combination of a partial character string in the character string representing one predicate of the “term-predicate” pair and a character string representing the other predicate is a third feature,
The likelihood of the parallel predicate syntax generated including each of the predicates of the “term-predicate” pair is a fourth feature,
For each of the “term-predicate” pairs in which any one of synonym, contradiction, and irrelevance is previously given as a classification, at least one of the first feature and the second feature, and the first feature A feature set including 3 features and the fourth feature,
The same-unrelated classification feature construction unit is provided with the feature set extracted by the feature extraction unit for each “term-predicate” pair to which synonym or anomaly is given as a classification, and irrelevant as a classification. In order to classify the feature set extracted by the feature extraction unit for each of the “term-predicate” pairs as being independent of the same or opposite sense of the “term-predicate” pair. Is constructed as an anti-irrelevant classification feature of
The synonym-anonymity classification feature construction unit includes the feature set extracted by the feature extraction unit for each “term-predicate pair” to which synonym is provided as a classification, and an anomaly is provided as a classification. The feature set extracted by the feature extraction unit for each of the term-predicate pair is constructed as a synonym-anonymity classification feature to classify the term-predicate pair into synonyms and anomalies. And
The reciprocal-irrelevant classification model learning unit determines that the term-predicate pair is independent of reciprocity based on the reciprocal-irrelevant classification feature constructed by the reciprocal-irrelevant classification feature construction unit. Learn the model to classify,
The synonym-anonymity classification model learning unit classifies the “term-predicate” pair into synonyms and anomalies based on the synonym-anonymity classification feature constructed by the synonym-anonymity classification feature construction unit. Classification model learning method to learn a model to do. A classification determination method in a classification determination apparatus including a feature extraction unit, a contradiction-irrelevance determination unit, and a synonym-anonymity determination unit,
The feature extraction unit is a predescription unit pair that is extracted based on a definition statement of each predicate of a predicate pair obtained from a definition statement set including a definition statement for each of a plurality of predicates prepared in advance. The feature indicating the mutual complementarity of the definition statement indicating whether or not there is a pair of predicates in the definition statement of each of the predicates is a first feature,
Semantic attributes common to the predescription part pair extracted based on the semantic attributes of each predicate of the predescription part pair obtained from the semantic attribute set consisting of the semantic attributes for each of the plurality of predicates prepared in advance. The second feature,
The combination of the partial character string in the character string representing one predicate of the preceding description part pair and the partial character string in the character string representing the other predicate is a third feature,
The likelihood of the parallel predicate syntax generated including each predicate of the predescription part pair is a fourth feature,
For the input predicate pair, extract a feature set including at least one of the first feature and the second feature, the third feature, and the fourth feature;
The reciprocity-irrelevant determination unit includes a model for classifying the feature set extracted by the feature extraction unit and a predicate pair that is pre-learned as irrelevant to reciprocity that is synonymous or affirmative. Based on, determine whether the classification of the previous description pair is the same or irrelevant,
The synonym-antisense determination unit learns in advance the feature set extracted by the feature extraction unit when the classification of the previous description unit pair is determined to be synonymous by the anti-unity determination unit. A classification determination method for determining whether a classification of a predescription part pair is synonymous or anomaly based on a model for classifying predicate pairs into synonyms and anomalies. A classification determination method in a classification determination apparatus including a feature extraction unit, a contradiction-irrelevance determination unit, and a synonym-anonymity determination unit,
The feature extraction unit is extracted based on a definition statement of each predicate of the “term-predicate” pair, which is obtained from a definition statement set including a definition statement for each of a plurality of predicates prepared in advance. A feature indicating mutual complementarity indicating whether or not a predicate as a pair exists in each definition statement of each predicate of the “term-predicate” pair, and the “term-predicate” pair Whether or not at least the previous description part exists among the features indicating the mutual complementarity of the definition sentence indicating whether or not there is a pair of “term-predicate” terms in each definition sentence of the predicate. The feature indicating the mutual definition of the definition sentence to be expressed is the first feature,
The “term-predicate” extracted based on the semantic attribute of each predicate of the “term-predicate” pair obtained from a semantic attribute set consisting of semantic attributes for each of a plurality of predicates prepared in advance. The second attribute is the semantic attribute common to the pair,
A combination of a partial character string in the character string representing one predicate of the “term-predicate” pair and a character string representing the other predicate is a third feature,
The likelihood of the parallel predicate syntax generated including each of the predicates of the “term-predicate” pair is a fourth feature,
For the input “term-predicate” pair, extract a feature set including at least one of the first feature and the second feature, the third feature, and the fourth feature;
The reciprocity-irrelevance determination unit classifies the feature set extracted by the feature extraction unit and a pre-learned “term-predicate” pair as synonymous or irrelevant as irrelevant or irrelevant. And whether the classification of the “term-predicate” pair is a contradiction or irrelevant,
The synonym-antisense determination unit is configured to extract the feature set extracted by the feature extraction unit when the classification of the “term-predicate” pair is determined to be synonymous by the anti-nomination-unrelated determination unit. And whether the classification of the “term-predicate” pair is synonymous or anomaly based on a previously learned model for classifying “term-predicate” pairs into synonyms and anomalies Classification judgment method.   The program for functioning a computer as each part which comprises the classification model learning apparatus of any one of Claims 1-6.   The program for functioning a computer as each part which comprises the classification determination apparatus of any one of Claims 7-10.