JP2006244262A - Retrieval system, method and program for answer to question - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To highly accurately retrieve an example answer sentence to a question sentence. <P>SOLUTION: A talk unit candidate generating means 201 performs a morphological analysis and a syntax analysis of a question sentence to generate a candidate in a talk unit. A talk unit extracting means 202 applies a talk unit extraction rule stored in a storing means 301 to a candidate in a talk unit and extracts a talk unit from the question sentence. An example data registering means 203 registers a particular talk unit extracted from a question sentence example together with the question sentence and an answer sentence into a storing means 302. An example question answer sentence data retrieval means 204 retrieves, for each particular talk unit for a new question sentence, an example question answer sentence having a particular talk unit of the same talk type as the talk type of the particular talk unit from the storing means 302. An inter-question sentence collating means 205 determines whether a question sentence of the retrieved example question answer sentence and the new question sentence are question sentences about similar contents on the basis of the similarity between the sentences with each other. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a question answer search system, method, and program, and more particularly, to a system, method, and program for searching a case answer sentence for reference when an answerer answers a query from a questioner.

  With the spread of the Internet, inquiries by e-mail from customers regarding products and services are increasing. When an operator answers such an inquiry, it is desirable in terms of operational efficiency to appropriately search for a reference case answer sentence.

  Some examples of search systems for answer examples for such questions have been known. For example, Patent Document 1 discloses an automatic response device that automatically responds by using an answer to a past question when the same question as a past question is asked. This apparatus extracts a search term from a sentence subject of a question item, searches for a similar question using the search term, and calculates a similarity. Then, based on the similarity of the cases obtained by the similar question search, an automatic answer and a corrected answer are discriminated, and a query including a plurality of query items is divided by content dividing means and processed individually. Then, the inquiry source is made to determine whether the answer is correct or incorrect.

  Further, Patent Document 2 discloses a question answering device that automatically responds using an answer to a past question when the same question as a past question is asked. This apparatus determines a question element type in advance, extracts a question element by using a surface layer pattern for extracting the question element, and searches a question case using an independent word included in the question element as a search word. By retrieving a search word from only the extracted question element portion and performing a search, the search accuracy of related question cases is improved as compared with the apparatus of Patent Document 1. In addition, it is superior to the apparatus of Patent Document 1 in that an extraction pattern of question elements is created in advance and only correct answers are presented to the questioner.

  Furthermore, as a related technique, Patent Document 3 discloses an expression that automatically extracts a sensitivity expression corresponding to an opinion, an impression, a background, a question, a request, a situation explanation, a complaint, and the like in a free sentence described in a questionnaire from a text. An extraction device is disclosed. This device extracts a prescriptive phrase and its continuous phrase from the syntactic structure of the input sentence, and extracts those that are judged as emotional expressions. Then, judgment knowledge of sensitivity expression such as predicate classification, thesaurus, sensitivity expression classification, likelihood of prescriptive phrase + continuous clause sensitivity expression is acquired by user input. It is possible to automatically extract emotional expressions from a large number of documents, which is suitable for questionnaire analysis.

  The conventional device, for example, the question answering device of Patent Document 2, operates as follows. That is, the question answer case storage means reads the case in advance, divides it into a case question sentence and a case answer sentence, analyzes the format of the question case, extracts the description corresponding to the operation / state / question, and extracts the extracted description. Are converted into standard expressions, the question cases are classified into discourse types related to the questions, and stored in the question answer case storage means together with the answer cases. The extraction rule storage means is expressed by a specific discourse unit slot for extracting from a sentence a specific discourse unit for each predetermined discourse type such as “operation”, “state”, “question” and the surface layer of the sentence end expression. Store the extraction rules.

  The question extraction means collates the question sentence sent from the questioner by e-mail etc. with the surface layer of the sentence end expression of the extraction rule stored in the extraction rule storage means, and when the rule and the surface layer match, the specific discourse The question sentence portion corresponding to the unit slot is extracted as a specific discourse unit. The case search means searches the case question answer text storage means using the specific discourse unit related to the question extracted by the question extraction means as a search key, and outputs the case question text having the same discourse type to the answer presentation means. The answer sentence presenting means reads the answer case corresponding to the searched question case from the case question answer sentence storage means and presents it to the questioner.

Japanese Patent Laid-Open No. 2002-132661 JP 2002-278777 A JP 2004-62448 A

  The first problem of the prior art is that the accuracy of extracting a specific discourse unit from a question sentence is low. The reason is due to the fundamental limitation of the method using sentence end expressions as clues for extracting a specific discourse unit for each discourse type from a question sentence. Specifically, for example, an extraction rule “<Question>, please tell me” (see FIG. 6 of Patent Document 2) is “No, what should I do?” “What should I do (Solution)” is correctly extracted as the discourse type “Question”, but it is not the discourse type “Question”. This is because “I ’m sorry to trouble you,” but misunderstood it as a discourse type “question”.

  The second problem is that enormous costs are required to create an extraction rule. The reason is that the sentence end expression is used as a clue to extract a specific discourse unit for each discourse type from a question sentence. Therefore, in order to extract a specific discourse unit completely from inquiries by an unspecified number of questioners, a combination of numbers This is because it is necessary to cover the sentence end expressions that cause variations, and a large number of extraction rules must be described.

  The third problem is that when a plurality of question items are included in one question, a case search may not be performed properly. The reason for this is that when multiple questions are included in a question, it is a method of extracting without distinguishing which particular question unit of a particular question. This is because there is a high possibility that a case that is unrelated to any of the questions will be erroneously searched.

  The fourth problem is that there may be a case where a question case whose content matches the question cannot be obtained by the search. The reason is that, when the character string matching of the specific discourse unit between the question and the question example is used as the search condition (the method of Patent Document 2), the specific discourse unit becomes inconsistent due to a slight difference in expression. For example, when the specific discourse unit of the discourse type “operation” extracted from the question is “change address” and the specific discourse unit of the same discourse type extracted from the question example is “change address”, both are the same Despite showing the contents, they are inconsistent. In addition, when using keyword matching in a specific discourse unit between a question and a question case as a search condition, the tendency of the appearance keyword is similar when the topic is the same, so the case question sentence is similar even when the content of the question is different. This is because there is a possibility that it is mistakenly determined.

  The fifth problem is that when a question sentence includes omissions and anaphoric relations (estimation is what pronouns and omissions indicate), the original discourse unit is extracted to extract the original discourse unit. It means that the meaning content of may be impaired. The reason is that when extracting a specific discourse unit (a discourse is a plurality of consecutive sentences), it is not considered whether omission or anaphoric relationship is lost by the extraction. This is because the sentence case search may cause a problem that even if an appropriate related case exists in the database, it cannot be appropriately searched.

  An object of the present invention is to provide a question answer search system, method, and program capable of extracting a specific discourse unit for each discourse type from a question sentence with high accuracy.

  A method for searching for answers to questions according to one aspect of the present invention that achieves the above object is to perform discourse units based on the structure of a sentence or a section of a syntax structure after performing morphological analysis and syntax analysis on a case question sentence in advance. A step of generating candidates, a step of extracting and assigning syntactic and semantic attributes to the discourse unit candidates and the surface representation of the discourse unit candidates from the result of the syntax analysis, and a syntactic Applying the discourse unit extraction rule that describes the matching conditions for semantic attributes for each discourse type in a specific discourse unit, extracting a specific discourse unit from the case question sentence for each discourse type, and the extracted specific discourse Storing the unit as a question answer example in the database together with the case question sentence itself. In addition, when a new question sentence is entered, the step of extracting a specific discourse unit for the new question sentence in the same way as the case question sentence, and the identification of the same discourse type as the specific discourse unit for the new question sentence A case question based on the step of searching the database for question answer cases with discourse units and the similarity between specific discourse units where the discourse type matches between the case question sentence and the new question sentence in the searched question answer case Determining that the sentence is similar to the new question sentence.

  The question answer search system according to one aspect of the present invention includes a discourse unit that generates a discourse unit candidate based on a sentence or clause structure of a syntax structure after performing morphological analysis and syntax analysis on a question sentence. Candidate generation means, syntactic and semantic attributes for the discourse unit candidate and the surface representation of the discourse unit candidate are extracted and assigned, and the discourse unit candidate is used for each discourse type according to the matching conditions for syntactic and semantic attributes. Discourse unit extraction rule storage means for storing discourse unit extraction rules that describe rules for judging the extraction of discourses, and applying discourse unit extraction rules to discourse unit candidates. Discourse unit extraction means for performing extraction, and the specific discourse unit extracted from the case question sentence in the discourse unit extraction means as the answer sentence to the question sentence and the question sentence Case data registration means for registering in the case question answer sentence data storage means as case question answer sentence data, and case question answer sentence data storage means for storing case question answer sentence data registered by the case question answer sentence data registration means . Also, for each specific discourse unit extracted by the specific discourse unit extraction means for a new question sentence input from the input device, a case question answer having a specific discourse unit of the same discourse type as the discourse type of the new question sentence Case question answer sentence data retrieval means for retrieving sentences from case question answer sentence data stored in the case question answer sentence data storage means, and identification of the question sentence and new question sentence of the retrieved case question answer sentence Inter-question sentence collating means for judging whether or not the question sentences are question sentences related to similar contents based on the similarity in the discourse unit.

  A program according to an aspect of the present invention is based on the structure of a sentence or a section of a syntax structure after a morphological analysis and a syntax analysis are performed on a case question sentence in advance on a computer constituting a question answer search system. Processing to generate discourse unit candidates, processing to extract and add syntactic and semantic attributes to the discourse unit candidates and the surface representation of the discourse unit candidates from the result of syntax analysis, and syntax to the discourse unit candidates Extraction of specific discourse units from case question sentences for each discourse type by applying a discourse unit extraction rule that describes matching conditions for specific and semantic attributes for each discourse type. Processing to store specific discourse units as a question answer case in the database together with the case question sentence itself, and when a new question sentence is entered, it is new in the same way as the case question sentence Processing to extract a specific discourse unit for a question sentence, processing to search a question answer example having a specific discourse unit of the same discourse type as the discourse type of a specific discourse unit for a new question sentence, and a searched question answer example And a process of determining that the case question sentence is similar to the new question sentence based on the similarity between the specific discourse units in which the discourse types match between the case question sentence and the new question sentence.

  According to the present invention, a discourse unit candidate having a tree structure is generated from a syntax analysis result, and a discourse related to an intention or a request of a question sentence using an extraction rule based on a syntactic and semantic attribute. Search for cases by extracting all units. Therefore, the specific discourse unit for each discourse type can be extracted from the question sentence with high accuracy, and an appropriate answer sentence for the new question sentence can be searched.

  First, the main terms are explained.

  “Discourse unit”: A discourse is a plurality of sentences (utterances) that are configured to motivate a writer (speaker) to transmit information to a reader (listener) or to achieve a purpose through information transmission. The “discourse unit” is a sentence or a part of a sentence corresponding to the minimum unit of the content of communication that constitutes such a discourse and includes one or more predicates and modifiers. .

  “Sentence-style predicate sentence style”: The sentence style here refers to the attribute of a sentence that the sentence is a plain text, a command sentence, or a question sentence. Sentence style is given by the predicates of the sentence. For example, if there is a final particle "ka" at the end of the predicate, the sentence style becomes "question" by this predicate. Similarly, the discourse unit also has a predicate, and there is a sentence style determined by this predicate, as in the case of a sentence. This sentence style is described as “statement style of discourse predicate”.

  “Head predicate”: “Head” is a central part of the subject, predicate, object, and noun phrase that determines the overall meaning of a plurality of morphemes (words) constituting each phrase. It refers to a morpheme that plays a role. Among the predicates that make up the predicate, it refers to the predicates that play a role in determining the meaning of the predicate. Refers to the semantic classification of this idiom. The “semantic classification” refers to a classification name on a classification system assigned from a semantic classification system (thesaurus, semantic classification dictionary, etc.) based on the meaning of the morpheme.

  “Head morpheme for each case element”: “Head predicate” refers to a predicate that plays a central role in determining the meaning of a predicate, as described above. The “case element of the idiom” means a subject, object, noun that is a case element (a phrase that is related to the idiom with a case such as a case, a case, a case, etc.) Refers to a phrase. The “morpheme for each case element of the head idiom” refers to a morpheme constituting each case element of the head idiom.

  “Surface of (morpheme)”: “Surface” refers to a character string in the form (surface) as it is in a morpheme sentence.

  “Prototype (of morpheme)”: “Prototype” refers to a character string in a form that excludes the use of “surface layer”.

  “(Morphological) semantic classification”: “Semantic classification” refers to a classification name on the classification system assigned from a semantic classification system (thesaurus, semantic classification dictionary, etc.) based on the meaning of the morpheme.

  “Head of case element”: A case element refers to a subject, an object, or a noun phrase that is a phrase related to a predicate with a case such as a case, a case, a case, or the like. As described above, the “head” plays a central role in determining the meaning of the whole of a plurality of morphemes (words) constituting each phrase in the subject, predicate, object, and noun phrase. Refers to the morpheme that fulfills

  “Attribute”: A plurality of features that are important viewpoints (determination materials) when determining whether a discourse unit candidate corresponds to a specific discourse unit are extracted from the discourse unit candidates and collectively referred to as attributes. Specifically, for each discourse unit candidate, by morphological analysis and syntactic analysis, the sentence style of the predicate, the semantic classification of the head predicate, and the surface layers and representatives of all the morphemes belonging to the head predicate case element The notation, the part of speech, the semantic classification, and the flag to the morpheme that is the head of the case element are output as the feature amount of each discourse unit candidate. This output feature amount is called an attribute. These are the results obtained from morphological analysis and syntactic analysis, especially predicates that mainly determine the discourse type of the discourse unit, and important elements that form the semantic content of the discourse unit related to the predicate. It consists of syntactic and semantic information about the subject, object, and noun phrase for this predicate. By extracting this information (features) and collating it with the extraction rules, the discourse unit candidates form the specific discourse unit from the viewpoint of discourse type and semantic content, that is, the semantic content that the writer wants to convey in the question sentence It is judged whether or not the discourse unit has characteristics as a discourse unit that plays the most important role in conveying meaning content such as questions and requests.

[First Embodiment]
The question answer search system according to the first embodiment of the present invention generates a discourse unit candidate based on the structure of a sentence or a section of a syntax structure after performing morphological analysis and syntax analysis on a question sentence. The discourse unit candidate generator and the extraction rules for determining the discourse unit extraction from the discourse unit candidates for each discourse type, the sentence style (plain, command, question), tense aspect (currently, Past, future, progress, completion), semantic classification of head utterances, and if the head utterance has a subject, object or noun phrase as a case element, the surface layer of all morphemes belonging to those case elements , Prototype, part-of-speech, semantic classification, and whether or not the head of a case element is described by a combination of matching conditions for syntactic and semantic attributes for the discourse unit, and discourse unit extraction to store this Chromatography provided Le a storage unit, to apply such a discourse unit extraction rules for discourse unit candidates, a discourse unit extraction unit that performs discourse unit extraction from the question sentence for each discourse type. Then, instead of uniquely determining the discourse type extraction unit based on the extraction rule, it is generated as a discourse unit candidate that forms a tree structure from the parsing result, and for each of these candidates, syntactic and semantic The extraction rule described based on the feature is applied, the discourse unit to be extracted is determined, and this is extracted as the discourse type.

  By operating as described above, more accurate extraction of discourse types and effective coverage by extraction rules with higher descriptive power are possible. Therefore, the specific discourse unit for each discourse type can be extracted from the question sentence with high accuracy. In addition, extraction rules can be easily created.

  Further, in the collation of specific discourse units, the syntactic and semantic information extracted from the specific discourse unit is provided with a specific discourse unit collation unit that determines whether or not the specific discourse unit is collated based on the match of each feature of the syntax semantic information. Based on the above, it operates to perform collation for a specific discourse unit.

  In addition, the question answer search system matches the specific discourse units with syntactic and semantic viewpoints so that the content of the requested answer in the question sentence matches the answer content of the case question answer sentence. Search for example questions.

  Next, the question answer search system according to the first embodiment of the present invention will be described in more detail with reference to the drawings. Referring to FIG. 1, the question answer search system according to the first embodiment of the present invention includes an input device 100, a data processing device 200a that operates by program control to function the question answer search processing, and a storage device 300a. , And the output device 400.

  The data processing device 200 a includes a discourse unit candidate generation unit 201, a specific discourse unit extraction unit 202, a case data registration unit 203, a case question answer sentence data search unit 204, and an inter-question sentence collation unit 205. The storage device 300 a includes a specific discourse unit extraction rule storage unit 301 and a case question answer sentence data storage unit 302.

  The discourse unit candidate generation unit 201 divides the input character string into the smallest unit (morpheme) that can be meaningful for a new question sentence or case question sentence input from the input device 100, and a word A parsing process is performed to identify grammatical relationships such as nouns, verbs, objects, etc. from the positional relationship between each other. Thereafter, a discourse unit candidate is generated based on the sentence or clause structure of the syntax structure and output to the specific discourse unit extraction unit 202. Further, syntactic and semantic information for each discourse unit candidate is extracted as an attribute from the result of the parsing process.

  The specific discourse unit extraction unit 202 applies the specific discourse unit extraction rule stored in the specific discourse unit extraction rule storage unit 301 to each discourse unit candidate input from the discourse unit candidate generation unit 201, and asks the question sentence. From this, a specific discourse unit for each discourse type is extracted and output to the case data registration unit 203.

  The case data registration unit 203 registers the specific discourse unit extracted from the case question sentence by the specific discourse unit extraction unit 202 in the case question answer sentence data storage unit 302 together with the question sentence and the answer sentence to the question sentence.

  The case question answer sentence data search unit 204 inputs the specific discourse unit extracted by the specific discourse unit extraction unit 202 with respect to the new question sentence input from the input device 100. For each input specific discourse unit, the case question answer sentence having the specific discourse unit of the same discourse type as the discourse type of the specific discourse unit is stored in the case question answer sentence data storage unit 302. A search is performed from the data, and the search result is output to the inter-question sentence matching unit 205.

  The inter-question sentence matching unit 205 determines the ratio of the specific discourse units of the same discourse type in the specific discourse units of the question sentence and the new question sentence of the searched case question answer sentence, and between the specific discourse units of the same discourse type. Based on the similarity, it is determined whether or not the question sentences are related to similar contents. The case question answer sentence determined to be similar is output to the output device 400.

  The specific discourse unit extraction rule storage unit 301 includes a predicate sentence style, a tense aspect, a head categorization semantic classification, a morpheme surface layer, a prototype, and a part of speech for each head discourse for each specific discourse unit. , Semantic classification, and case element head as a feature quantity, and specifying a rule describing whether or not a discourse unit candidate is a specific discourse unit of each discourse type according to the condition related to the feature value It is stored as a discourse unit extraction rule.

  The case question answer sentence data storage unit 302 stores a specific discourse unit registered by the case question answer sentence data registration unit 203 and the case question answer sentence.

  Next, the overall operation of the system will be described in detail with reference to the flowcharts of FIGS. First, an operation from input of case question answer text data (step A1 in FIG. 2) to registration (step A4 in FIG. 2) will be described.

  Case question answer sentence data is input from the input device 100. In the case question answer sentence data to be input, the question sentence and the answer sentence corresponding thereto are paired (step A1 in FIG. 2).

  The discourse unit candidate generation unit 201 performs morphological analysis and syntax analysis on the question sentence of the input case question answer sentence data, and then generates discourse unit candidates based on the sentence or clause structure of the syntax structure. Is done. From the results of morphological analysis and syntactic analysis, sentence styles of predicates for each discourse unit, tense and aspect, semantic classification of head terms, morphological surface layer, prototype, part of speech, semantic classification for each case element of head terms , And the head of the case element are output as feature quantities (step A2 in FIG. 2).

  The specific discourse unit extraction unit 202 applies the specific discourse unit extraction rule stored in the specific discourse unit extraction rule storage unit 301 to each discourse unit candidate generated by the discourse unit candidate generation unit 201, and sets each discourse type. Extract a specific discourse unit. (Step A3 in FIG. 2).

  The case data registration unit 203 registers the specific discourse unit extracted by the specific discourse unit extraction unit 202 in the case question answer sentence data storage unit 302 together with the question sentence of the case question answer sentence data and the answer sentence (see FIG. 2). Step A4).

  Next, the operation from the input of a new question sentence (step A5 in FIG. 2) to the output of a similar case question answer sentence (step A10 in FIG. 2) will be described.

  A new question sentence is input from the input device 100 (step A5 in FIG. 2).

  The discourse unit candidate generation unit 201 generates a discourse unit candidate for the new question sentence by the same process as in step A2, and obtains a feature amount for each discourse unit candidate (step A6 in FIG. 2).

  The specific discourse unit extraction unit 202 extracts, for each discourse type, a specific discourse unit for each discourse unit candidate generated from the new question sentence in the discourse unit candidate generation unit 201 by the same process as in step A3 (FIG. 2). Step A7).

  In the case question answer sentence data search unit 204, the case question answer sentence data including the specific discourse unit of the same discourse type as each specific discourse unit extracted from the new question sentence in the specific discourse unit extraction part 202 is converted into the case question answer sentence data. Search is performed from the case question answer text data stored in the storage unit 302 (step A8 in FIG. 2).

  In the inter-question sentence collating unit 205, for each case question answer sentence searched by the case question answer sentence data searching unit 204, the similarity between the question sentence and the new question sentence is matched with the discourse type in both specific discourse units. And a case question answer sentence with a certain value or more is determined as a similar case question answer sentence (step A9 in FIG. 2).

  The question sentence matching unit 205 outputs the case question answer sentence determined as the similar case question answer sentence to the output device 400 (step A10 in FIG. 2).

  According to the question answer search system as described above, instead of extracting a specific discourse unit uniquely, a candidate for a discourse unit that forms a tree structure is generated from a syntax analysis result, and a syntactic and semantic feature is generated. All the discourse units related to the intent and request of the question sentence are extracted using the extraction rules with high descriptive power based on them, and the cases are searched based on this. Therefore, a specific discourse unit for each discourse type can be extracted from the question sentence with high accuracy, and an extraction rule can be easily created.

[Second Embodiment]
The question answer search system according to the second embodiment of the present invention includes a sentence style (statement, command, question), tense aspect (current, past, future, progress, completion), head of each discourse unit candidate. Semantic classification of predicates, and when head predicates have subject, object or noun phrases as case elements, the surface layers, prototypes, parts of speech, semantic classifications, and cases of all morphemes belonging to those case elements Discrimination unit extraction unit that extracts by using a specific discourse unit extraction model that determines whether it is a specific discourse unit of each discourse type by using the feature amount of that discourse unit as the element head or not And a specific discourse unit extraction model acquisition unit that mechanically learns a specific discourse unit extraction model using a question sentence provided with a discourse type to be extracted from the question sentence as learning data. And even if it does not describe a specific discourse unit extraction rule manually, it operates so that a specific discourse unit extraction model may be acquired by mechanical learning.

  Next, a question answer search system according to a second embodiment of the present invention will be described in more detail with reference to the drawings. Referring to FIG. 3, the question answer search system according to the second embodiment of the present invention includes an input device 100, a data processing device 200b that operates by program control to function the question answer search processing, and a storage device 300b. , And the output device 400.

  The data processing device 200b includes a discourse unit candidate generation unit 201a, a specific discourse unit extraction unit 202a, a case data registration unit 203, a case question answer sentence data search unit 204, a question sentence matching unit 205, and a specific discourse unit. An extraction model acquisition unit 206. The storage device 300b includes a case question answer sentence data storage unit 302 and a specific discourse unit extraction model storage unit 303. In FIG. 3, the case data registration unit 203, the case question answer text data search unit 204, the inter-question text matching unit 205, and the case question answer text data storage unit 302 are the same as those having the same reference numerals described in FIG. 1. Therefore, the description thereof is omitted.

  The discourse unit candidate generation unit 201a performs a morphological analysis process and a syntax analysis process on a new question sentence or a case question sentence input from the input device 100, and then a discourse based on the sentence or clause structure of the syntax structure. Unit candidates are generated, and syntactic and semantic information for each discourse unit candidate is extracted as a feature amount from the result of the parsing process.

  The specific discourse unit extraction unit 202a applies the specific discourse unit extraction model stored in the specific discourse unit extraction model storage unit 303 to the discourse unit candidate generated by the discourse unit candidate generation unit 201a, and uses the question sentence. A specific discourse unit for each discourse type is extracted and output to the case data registration unit 203.

  The specific discourse unit extraction model acquisition unit 206 receives, as learning data, specific discourse units of each discourse type when the case question answer sentence and the specific discourse unit to be extracted from the input are input from the input device 100. Train the model to be extracted. As a result of learning, the specific discourse unit extraction model acquired is stored in the specific discourse unit extraction model storage unit 303.

  The specific discourse unit extraction model storage unit 303 stores, for each specific discourse unit, the predicate sentence style, tense aspect, head semantics classification, morphological surface layer, prototype, part of speech for each head discourse case element. As a specific discourse unit extraction model, a model for determining whether or not a discourse unit candidate is a specific discourse unit of each discourse type by inputting this feature amount as a feature amount, and the head of the semantic classification and case element Remember.

  Next, the operation of the system will be described in detail with reference to the flowcharts of FIGS. In FIG. 4, steps A1, A2, A4 to A6, and A8 to A10 perform the same processing as the same reference numerals in FIG.

  When learning the specific discourse unit extraction model following step A2, the operation proceeds from the correct data input (step B1 in FIG. 4) to the storage of the specific discourse unit extraction model (step B4 in FIG. 4). When the specific discourse unit extraction model is stored in the specific discourse unit extraction model storage unit 303 and a new model is not learned, the process proceeds to extraction of a specific discourse unit (step B5 in FIG. 4).

  When learning the specific discourse unit extraction model, the specific discourse unit for each discourse type included in the question sentence of the example question answer sentence data (step A1 in FIG. 4) that has already been input is input from the input device 100 as correct data. (Step B1 in FIG. 4).

  The input correct answer of the specific discourse unit is collated with the discourse unit candidate generated by the discourse unit candidate generation unit 201a, and the discourse unit candidate with the closest extracted part is set as the correct answer of the specific discourse unit, and the discourse type is input. The correct answer type (step B2 in FIG. 4).

  In the specific discourse unit extraction model acquisition unit 206, each discourse unit candidate and its feature amount created from the question sentence of the case question answer sentence and the correct answer of the specific discourse unit, and the specific discourse unit to be extracted for each discourse type therefrom Is used as learning data to learn a specific discourse unit extraction model. The specific discourse unit extraction model inputs the features of each discourse unit candidate and outputs whether or not the candidate corresponds to the specific discourse unit of each discourse type. When a feature amount of a discourse unit candidate corresponding to a specific discourse unit is input, a specific discourse unit extraction model is determined that determines that the discourse type corresponds to a specific discourse unit (step B3 in FIG. 4).

  The specific discourse unit extraction model acquired by the specific discourse unit extraction model acquisition unit 206 is stored in the specific discourse unit extraction model storage unit 303 (step B4 in FIG. 4).

  The specific discourse unit extraction unit 202 applies the specific discourse unit extraction model stored in the specific discourse unit extraction model storage unit 303 to each discourse unit candidate generated by the discourse unit candidate generation unit 201, for each discourse type. Are extracted (step B5 in FIG. 4). Thereafter, the process proceeds to Step A4.

  Subsequent to step A6, the specific discourse unit extraction unit 202 applies a specific discourse unit for each discourse type by the same process as step B5 for each discourse unit candidate generated from the new question sentence in the discourse unit candidate generation unit 201. Extract (step B6 in FIG. 4).

  Subsequent to step B6, steps A8 to A10 are executed, and the case question answer sentence determined as the similar case question answer sentence is output from the output device 400.

  According to the question answer retrieval system as described above, a model for extracting discourse units related to the intention and request of a question sentence by mechanical learning from learning data is acquired. That is, the burden of describing the specific discourse unit rule manually is reduced by mechanical learning from learning data. Therefore, knowledge creation for extraction can be easily performed, the burden of manual description of specific discourse unit extraction rules can be reduced, and extraction rules can be easily created.

[Third Embodiment]
A question answer search system according to a third embodiment of the present invention includes a discourse block certifying unit that applies a discourse scheme to a specific discourse unit extracted from a question sentence and collects the specific discourse unit into a discourse block, and a discourse A discourse scheme storage unit for storing a scheme, and a question example collation unit that collates in units of discourse blocks, not in question sentence units, in collation of question sentences and case question sentence specific discourse units. When a plurality of question items are included in the question sentence, a discourse block is individually created for each question item, and the collation between the question sentence and the case is performed for each discourse block.

  Next, a question answer search system according to a third embodiment of the present invention will be described in more detail with reference to the drawings. Referring to FIG. 5, the question answer search system according to the third embodiment of the present invention includes an input device 100, a data processing device 200c that operates by program control to function the question answer search processing, and a storage device 300c. , And the output device 400.

  The data processing device 200c includes a discourse unit candidate generation unit 201, a specific discourse unit extraction unit 202, a case data registration unit 203a, a case question answer sentence data search unit 204, a question sentence matching unit 205a, and a discourse block generation. Part 207. The storage device 300c includes a specific discourse unit extraction rule storage unit 301, a case question answer sentence data storage unit 302a, and a discourse block generation schema storage unit 304. In FIG. 5, a discourse unit candidate generation unit 201, a specific discourse unit extraction unit 202, a case question answer sentence data search unit 204, and a specific discourse unit extraction rule storage unit 301 are the same as those in FIG. Description is omitted.

  The specific discourse unit extraction unit 202 applies the specific discourse unit extraction rule stored in the specific discourse unit extraction rule storage unit 301 to the discourse unit candidate generated by the discourse unit candidate generation unit 201, and uses the question sentence. A specific discourse unit for each discourse type is extracted and output to the discourse block generation unit 207.

  The discourse block generation unit 207 uses the discourse block generation schema stored in the discourse block generation schema storage unit 304 for a plurality of specific discourse units for each discourse type extracted from the question sentence by the specific discourse unit extraction unit 202. Apply. One or more specific discourse units are collected into a discourse block corresponding to a group of one question item, and the generated discourse block and information of a specific discourse unit belonging to the discourse block are stored as discourse block information in the case data registration unit 203a. Is output.

  The discourse block generation schema storage unit 304 stores, as a discourse block generation schema, a schema for collecting a plurality of specific discourse units for each discourse type extracted from the question sentence into discourse blocks corresponding to a set of one question item. . Here, the discourse block generation schema means a relationship in a discourse structure assumed between specific discourse units. By extracting each specific discourse unit belonging to a discourse type such as “situation”, “question”, “question”, “request”, etc. from the question sentence, the discourse structure of what each discourse unit constitutes I do n’t know. When different questions are included in the question text, there is a group of discourses about each question, and we want to search for question answer examples for each. Therefore, it is necessary to determine what sort of discourse is organized by the extracted specific discourse unit, that is, what constitutes a discourse block. For this reason, a relationship in the discourse structure assumed between specific discourse units, that is, a discourse block generation schema is defined, and what discourse block has a large question sentence from a discourse block (for example, FIG. 30 described later) generated thereby. It is determined from which discourse block each specific discourse unit is organized.

  Next, the operation of the system will be described in detail with reference to the flowcharts of FIGS. In FIG. 6, steps A1 to A3 and A5 to A7 perform the same processing as the same reference numerals in FIG.

  The discourse block generation unit 207 applies the discourse block generation schema stored in the discourse block generation schema storage unit 304 to a plurality of specific discourse units for each discourse type extracted by the specific discourse unit extraction unit 202. One or more specific discourse units are collected into a discourse block corresponding to a group of one question item, and the generated discourse block and the information of the specific discourse unit belonging thereto are output as discourse block information (FIG. 6). Step C1).

  In the case data registration unit 203a, the specific discourse unit extracted by the specific discourse unit extraction unit 202 is converted into the discourse block information generated by the discourse block generation unit 207, the question sentence of the inputted case question answer sentence data, and the answer sentence thereof. At the same time, it is registered in the case question answer text data storage unit 302a (step C2 in FIG. 6).

  The discourse block generation unit 207 generates a discourse block by the same process as step C1 for a plurality of specific discourse units for each discourse type extracted by the specific discourse unit extraction unit 202 for each new question sentence. Is output as discourse block information (step C3 in FIG. 6).

  In the case question answer sentence data search unit 204, the case question answer sentence data including the specific discourse unit of the same discourse type as each specific discourse unit extracted from the new question sentence in the specific discourse unit extraction part 202 is converted into the case question answer sentence data. Search is performed from the case question answer text data stored in the storage unit 302a (step A8 in FIG. 6).

  In the inter-question sentence matching unit 205a, for each case question answer sentence searched by the case question answer sentence data search unit 204, the same discourse block in each discourse block in the question sentence and each discourse block in the new question sentence are the same. Based on the ratio of specific discourse units of discourse type and the similarity between specific discourse units of the same discourse type, it is determined whether each discourse block of each question sentence is a discourse block related to similar contents ( Step C4 in FIG.

  The inter-question sentence matching unit 205a outputs the case question answer sentence determined as the similar case question answer sentence to the output device 400 together with the discourse block information used as the criterion for determination (step C5 in FIG. 6).

  According to the question answer search system as described above, when multiple question items are included in a question sentence, the extracted specific discourse units are collected into discourse blocks corresponding to each question item, and related case questions related to the discourse block Search for answers. In other words, when a plurality of question items are included in a question sentence, the problem of the conventional method that confuses and collates specific discourse units having the same discourse type that is extracted is solved. Therefore, even when a plurality of questions are included in one question, it is possible to correctly search for cases related to each question.

[Fourth Embodiment]
The question answer search system according to the fourth embodiment of the present invention does not perform collation between discourse units of the same discourse type when collating a specific discourse unit between a question sentence and a question example. A specific discourse unit collation unit that performs collation considering the correspondence of discourse types associated through specific discourse unit deformation rules for each type, and a specific discourse unit deformation rule storage unit that accumulates specific discourse unit deformation rules . And it operates so that it may be judged that the question sentence and question example which the corresponding specific discourse unit collates are related.

  Next, a question answer search system according to a fourth embodiment of the present invention will be described in detail with reference to the drawings. Referring to FIG. 7, a question answer search system according to a fourth embodiment of the present invention includes an input device 100, a data processing device 200d that operates by program control to function a question answer search process, and a storage device 300d. , And the output device 400.

  The data processing device 200d includes a discourse unit candidate generation unit 201, a specific discourse unit extraction unit 202, a case data registration unit 203, a case question answer sentence data search unit 204a, and an inter-question sentence matching unit 205b. The apparatus 300 includes a specific discourse unit extraction rule storage unit 301, a case question answer sentence data storage unit 302, and a discourse type conversion rule storage unit 305. In FIG. 7, the discourse unit candidate generation unit 201, the specific discourse unit extraction unit 202, the case data registration unit 203, the specific discourse unit extraction rule storage unit 301, and the case question answer sentence data storage unit 302 have been described with reference to FIG. Since it is the same as that of the same code | symbol, the description is abbreviate | omitted.

  The case question answer sentence data search unit 204a converts a discourse type or a discourse type in a specific discourse unit for each specific discourse unit extracted by the specific discourse unit extraction unit 202 with respect to a new question sentence input from the input device 100. A case question answer sentence having a specific discourse type of the same discourse type as the discourse type converted by the rule is searched from the case question answer sentence data stored in the case question answer sentence data storage unit 302.

  The discourse type conversion rule storage unit 305 stores discourse type conversion rules described for the purpose of discriminating the same type when the specific discourse units of different discourse types represent the intention and content of the same question sentence depending on the context. Remember.

  Next, the operation of the system will be described in detail with reference to the flowcharts of FIGS. In FIG. 8, steps A1 to A7 perform the same processes as those in FIG.

  In the case question answer sentence data search unit 204a, for each specific discourse unit extracted from the new question sentence by the specific discourse unit extraction unit 202, the discourse type or discourse type conversion rule stored in the discourse type conversion rule storage unit 305 is stored. A case question answer sentence having a specific discourse unit of the same discourse type as the discourse type after conversion by (2) is searched from the case question answer sentence data stored in the case question answer sentence data storage unit 302 (step of FIG. 8). D1).

  In the inter-question sentence matching unit 205b, for each case question answer sentence searched by the case question answer sentence data search unit 204a, the degree of similarity between the question sentence and the new question sentence is determined based on the discourse type discourse for each specific discourse unit. Case questions with a certain value or more, calculated based on the percentage of matches including matches after conversion by type conversion rules, and the similarity between specific discourse units with matching discourse types, including those after conversion by discourse type conversion rules The answer sentence is determined as a similar case question answer sentence (step D2 in FIG. 8).

  The inter-question sentence matching unit 205b outputs the case question answer sentence determined as the similar case question answer sentence to the output device 400 (step A10 in FIG. 8).

  According to the question answer search system as described above, it is possible to judge that the contents of the specific discourse unit differ between the question sentence and the question example, but appropriately match and match the same contents such as the request and intention of the question sentence. . Therefore, even if the expression style is different between the question text and the case question text, cases where the request contents and intentions by the question text are similar are appropriately searched, and the specific discourse unit for each discourse type from the question text Can be extracted with high accuracy.

[Fifth Embodiment]
A question / answer search system according to a fifth embodiment of the present invention includes a question / answer type estimation unit that determines a type related to a question content and a request content from a question text, and further estimates an expected answer type, and a question text And a question / answer type matching unit for matching the question type and answer type of the question case. Then, it operates so as to determine the match between the question sentence and the question example of the question case and the answer type.

  Next, a question answer search system according to a fifth embodiment of the present invention will be described in detail with reference to the drawings. Referring to FIG. 9, a question answer search system according to a fifth embodiment of the present invention includes an input device 100, a data processing device 200e that operates by program control to function a question answer search process, and a storage device 300e. , And the output device 400.

  The data processing device 200e includes a discourse unit candidate generation unit 201, a specific discourse unit extraction unit 202, a case data registration unit 203b, a case question answer sentence data search unit 204, an inter-question sentence matching unit 205c, and a request answer type. The storage device 300e includes a specific discourse unit extraction rule storage unit 301, a case question answer sentence data storage unit 302b, a semantic classification dictionary storage unit 306, and a request answer type extraction pattern storage unit 307. . In FIG. 9, a discourse unit candidate generation unit 201, a specific discourse unit extraction unit 202, a case question answer sentence data search unit 204, and a specific discourse unit extraction rule storage unit 301 are the same as those having the same reference numerals described in FIG. Since there is, explanation is omitted.

  The request answer type extraction unit 208 includes a request type extraction pattern and a semantic classification dictionary in an expression part related to the answer content that the question sentence expects as an answer included in the specific discourse unit extracted by the specific discourse unit extraction unit 202. Is applied in combination to obtain the request / response type of the question text.

  The semantic classification dictionary storage unit 306 stores a semantic classification dictionary in which a semantic classification is given to each word notation. Here, the semantic classification dictionary classifies vocabulary sets based on the same and similar meanings of each meaning, summarizes the lexical sets having similar meanings, defines the semantic classification (class) to which these belong, This is a vocabulary dictionary in which a semantic system is configured by a set of semantic classifications. Also called a thesaurus.

  The request answer type extraction pattern storage unit 307 collates with a description part related to request answer contents from a specific discourse unit, and sets a pattern for each request answer type described by the feature amount and semantic classification of the specific discourse unit as a request answer type extraction pattern. Remember as.

  Next, the operation of the system will be described in detail with reference to the flowcharts of FIGS. In FIG. 10, steps A1 to A3, A5 to A7, A8, and A10 perform the same processing as the same reference numerals in FIG.

  The request answer type extraction unit 208 stores the request answer type extraction pattern storage unit 307 in an expression part related to the answer content that the question text included in the specific discourse unit extracted by the specific discourse unit extraction unit 202 expects as an answer. The request type extraction pattern described for each stored request type and the semantic classification dictionary stored in the semantic classification dictionary storage unit 306 are applied in combination to extract the request answer type of the question sentence (step E1 in FIG. 10). .

  In the case data registration unit 203b, the specific discourse unit extracted by the specific discourse unit extraction unit 202 is added to the case question answer together with the request answer type extracted by the request answer type extraction unit 208 and the question sentence and the answer sentence to the question sentence. It is registered in the sentence data storage unit 302b (step E2 in FIG. 10).

  The request answer type extraction unit 208 applies a combination of the request answer type extraction pattern and the semantic classification dictionary to each specific discourse unit extracted from the new question sentence by the specific discourse unit extraction unit 202 by the same process as in step E1. The request answer type of the question sentence is extracted (step E3 in FIG. 10).

  In the inter-question sentence matching unit 205c, for each case question answer sentence searched by the case question answer sentence data search unit 204, the similarity between the question sentence and the new question sentence is set to match the discourse type in both specific discourse units. , The degree of similarity between specific discourse units with the same discourse type, and the matching of the requested answer type, and a case question answer sentence exceeding a certain value is determined as a similar case question answer sentence (FIG. 10). Step E4).

  The question answer search system as described above determines whether the question sentence matches the question example from the viewpoint of the question content and the request content. That is, it is considered that the contents of the requested question match in matching the question text of the new question text and the case question answer text. Therefore, it is possible to search for a question case in which the content of the request answer in the question sentence is consistent with the answer content of the case question answer sentence.

[Sixth Embodiment]
A question / answer search system according to a sixth embodiment of the present invention includes a question / answer type estimation unit that determines a type related to a question content and a request content from a question text, and further estimates an expected answer type, and a question text A question-answer correspondence unit that determines whether or not a part corresponding to the answer type estimated from the question exists in the answer sentence for the question, and a question-answer example storage unit that accumulates a pair of question examples and answer examples for the question With. Then, the question / answer type is estimated for the question sentence, the question case that matches the question / answer type is searched, and the part corresponding to the answer type estimated from the question sentence matches the question / answer type. It is determined whether or not it exists in the answer example for the question example to be answered, and if it exists, it is determined to be an answer example including the answer to the question.

  Next, a question answer search system according to a sixth embodiment of the present invention will be described in detail with reference to the drawings. Referring to FIG. 11, a question answer search system according to a sixth embodiment of the present invention includes an input device 100, a data processing device 200f that operates by program control to function a question answer search process, and a storage device 300f. , And the output device 400.

  The data processing device 200f includes a discourse unit candidate generation unit 201, a specific discourse unit extraction unit 202, a case data registration unit 203b, a case question answer sentence data search unit 204, a question sentence matching unit 205d, and a request answer type. And an extraction unit 208. The storage device 300f includes a specific discourse unit extraction rule storage unit 301, a case question answer sentence data storage unit 302b, a semantic classification dictionary storage unit 306, a request answer type extraction pattern storage unit 307, and an answer type collation pattern storage unit 308. including. In FIG. 11, a discourse unit candidate generation unit 201, a specific discourse unit extraction unit 202, a case data registration unit 203b, a case question answer sentence data search unit 204, a specific discourse unit extraction rule storage unit 301, and a case question answer sentence data storage unit 302b. The semantic classification dictionary storage unit 306 and the request answer type extraction pattern storage unit 307 are the same as those having the same reference numerals described in FIG.

  The question sentence matching unit 205d has similar question sentences based on the semantic sentence dictionary stored in the semantic classification dictionary storage unit 306 and the answer type verification pattern stored in the answer type verification pattern storage unit 308. It is a question sentence regarding the contents, and it is determined whether the request answer of the new question sentence and the answer contents of the case answer sentence are consistent. This determination is based on the ratio of specific discourse units of the same discourse type in the specific discourse unit of the question sentence of the searched case question answer sentence and the new question sentence, the similarity between specific discourse units of the same discourse type, This is performed based on the answer type matching pattern between the requested answer type of the new question sentence and the answer sentence of the case question answer sentence and the matching result by the semantic sentence dictionary. A case question answer sentence that has been determined that the question sentences are similar and the question sentence and the answer sentence match is output to the output device 400.

  The answer type collation pattern storage unit 308 stores a pattern for each answer type to be collated with a description part that can discriminate the answer type of the answer sentence as an answer type collation pattern.

  Next, the operation of the system will be described in detail with reference to the flowcharts of FIGS. In FIG. 12, steps A1 to A3, E1, E2, A5 to A7, E3, and A8 perform the same processing as the same reference numerals in FIG.

  In the inter-question sentence matching unit 205d, for each case question answer sentence searched by the case question answer sentence data search unit 204, the similarity between the question sentence and the new question sentence is set to match the discourse type in both specific discourse units. And a case question answer sentence with a certain value or more is determined as a similar case question answer sentence (step A9 in FIG. 12).

  Further, the inter-question sentence matching unit 205d stores the consistency between the request answer type of the new question sentence extracted by the request answer type extracting part 208 and the answer sentence of the case question answer sentence in the answer type matching pattern storage unit 308. The pattern corresponding to the request answer type of the new question sentence is determined based on whether or not to match the answer sentence of the case question answer sentence together with the semantic classification dictionary stored in the semantic classification dictionary storage unit 306 (see FIG. 12 steps F1). The question sentence matching unit 205d outputs the case question answer sentence determined as the similar case question answer sentence to the output device 400 (step A10 in FIG. 2).

  According to the question answer search system as described above, whether or not the correspondence between the question sentence and the answer sentence is semantically taken into consideration is taken into account in determining whether the case answer sentences match. That is, in matching the question text of the new question text and the case question answer text, it is considered that the content of the request question of the new question text matches the content of the answer text of the case question answer text. Therefore, it is possible to search for a question case in which the content of the request answer in the question sentence is consistent with the answer content of the case question answer sentence.

[Seventh Embodiment]
The question answer search system according to the seventh embodiment of the present invention generates a discourse unit candidate based on the sentence or clause structure of the syntax structure after performing morphological analysis and syntax analysis on the question sentence. The discourse unit candidate generator, the anaphoric relationship analysis unit that recognizes the anaphoric relationship included in the specific discourse unit, and the specified discourse unit contains a demonstrative pronoun or abbreviation and its antecedent exists outside the specific discourse unit In this case, an anaphoric element is replaced with an antecedent, and an anaphoric omission element complementing unit that supplements omission with the antecedent is provided. If the specified discourse unit extracted from the question sentence is supplemented by the anaphoric abbreviation element complementation part and the omitted pronoun and the abbreviated element, it is checked against the specific discourse unit case based on the completion result, If the element remains without being complemented, the discourse unit that includes the specific discourse unit is used as an alternative specific discourse unit so as to collate with the case.

  Next, a question / answer search system according to a seventh embodiment of the present invention will be described in more detail with reference to the drawings. Referring to FIG. 13, the question answer search system according to the seventh embodiment of the present invention includes an input device 100, a data processing device 200g that operates by program control to function the question answer search processing, and a storage device 300g. , And the output device 400.

  The data processing device 200g includes a discourse unit candidate generation unit 201, a specific discourse unit extraction unit 202, a case data registration unit 203, a case question answer sentence data search unit 204, an inter-question sentence matching unit 205, and an anaphoric analysis process. The storage device 300g includes a specific discourse unit extraction rule storage unit 301 and a case question answer sentence data storage unit 302. In FIG. 13, the discourse unit candidate generation unit 201, the specific discourse unit extraction unit 202, the case data registration unit 203, the case question answer text data search unit 204, the inter-question sentence collation unit 205, the specific discourse unit extraction rule storage unit 301, and the case Since the question answer sentence data storage unit 302 is the same as that of the same reference numerals described in FIG.

  When each discourse unit candidate generated by the discourse unit candidate generation unit 201 includes an demonstrative pronoun or an abbreviation, the anaphora analysis processing unit 209 identifies the demonstrative pronoun or the omitted antecedent in the question sentence. If the antecedent is successfully identified, each discourse unit candidate is replaced by an antecedent of the demonstrative pronoun or complemented by an antecedent of the omitted element. Furthermore, if there is a case element shared with other sections and it is not included in the discourse unit candidate, this complementation is also performed. If identification of the antecedent fails, the semantic classification that should be satisfied by the demonstrative pronoun or omitted element instead of the antecedent is estimated from the semantic restrictions on the type of the demonstrative pronoun and the case element of the omitted element, and this semantic classification information Substitution of omission pronouns and supplementation of omissions are performed with a dummy antecedent. Furthermore, the result of replacement or complementation in each of these discourse unit candidates is also reflected in the feature amount of each discourse unit candidate.

  Next, the operation of the system will be described in detail with reference to the flowcharts of FIGS. In FIG. 14, steps A1 to A10 perform the same processing as the same reference numerals in FIG.

  In the anaphora analysis processing unit 209, when each discourse unit candidate generated by the discourse unit candidate generation unit 201 includes a pronoun or abbreviation, the pronoun or abbreviation antecedent is identified in the question sentence. If the antecedent is successfully identified, each discourse unit candidate is replaced by an antecedent of the demonstrative pronoun or complemented by an antecedent of the omitted element. Furthermore, if there is a case element shared with other sections and it is not included in the discourse unit candidate, this complementation is also performed. If identification of the antecedent fails, the semantic classification that should be satisfied by the demonstrative pronoun or omitted element instead of the antecedent is estimated from the semantic restrictions on the type of the demonstrative pronoun and the case element of the omitted element, and this semantic classification information Substitution of omission pronouns and supplementation of omissions are performed with a dummy antecedent. Further, the result of replacement or complementation in each discourse unit candidate is reflected in the feature amount of each discourse unit candidate (step G1 in FIG. 14).

  Further, in the anaphora analysis processing unit 209, when each discourse unit candidate generated from the new question sentence has a pronoun, abbreviation, or a shared case element with another clause, the instruction is performed by the same process as in step G1. Identify pronouns, abbreviations, and shared elements. In addition, for each discourse unit candidate, replacement by an antecedent of the demonstrative pronoun, omission element, and complement of each shared element are performed. Furthermore, the results of these replacements and complements are reflected in the feature amounts of the respective discourse unit candidates (step G2 in FIG. 14).

  According to the question answer search system as described above, a question sentence and a question example are collated based on a specific discourse unit in which omitted elements and anaphoric relations are appropriately supplemented. Therefore, even when the question sentence includes omission or an anaphoric relationship, the specific discourse unit can be extracted without losing the original semantic content.

  Next, a first embodiment will be described with reference to FIGS. 1, 2, and 15 to 23. First, the preprocessing for the case question answer sentence data in which the question sentence and the answer sentence corresponding thereto are paired will be described. An example of case question answer sentence data is shown in FIG. Such case question answer sentence data is input from the input device 100 (step A1 in FIG. 2). As example question answer sentence data, for example, e-mail saved as a file is used.

  As shown in FIG. 16, after the morphological analysis and syntactic analysis are performed in the discourse unit candidate generation unit 201 on the question sentence of the inputted case question answer sentence data, based on the sentence or clause structure in the syntax structure A simple discourse unit candidate is generated. The candidate for the discourse unit has a tree structure with the candidate 7 as a root. The candidate 7 is connected to the candidates 5 and 6, the candidate 5 is connected to the candidates 1 and 2, and the candidate 6 is connected to the candidate 6. Candidate 3 and candidate 4 are connected. Further, in each candidate, a head by semantic processing of parsing is extracted and added.

  Furthermore, for each discourse unit candidate, by morphological analysis and syntactic analysis, the sentence style of the predicate, the semantic classification of the head predicate, and the surface layer of all the morphemes belonging to the case element of the head predicate, representative notation, part of speech , Semantic classification, and flag to the morpheme that is the head of the case element are output as attributes (features) of each discourse unit candidate. For example, for the first sentence discourse unit candidate 4 (FIG. 16), attributes as shown in FIG. 17 are extracted.

  In other words, the discourse unit candidate 4 “Are you sure to add a counter to the page?” Is “sentence style = question”, “semantic classification = state”, “representative notation =“ must ”” , “Part of speech = auxiliary verb”. Furthermore, the discourse unit candidate 4 is classified into surface layers such as “page”, “ni”, “counter”, “on”, “must have”, “no wonder” for each morpheme, and for each morpheme, “Representative notation”, “part of speech”, “meaning”, “case element”, and “flag indicating presence / absence of head” are attached. For example, the surface layer “page” is the representative notation “page”, the part of speech “noun”, the meaning “specific object, abstract object”, the case element “noun phrase”, and the flag is the head of the case element. “○” indicating that there is a mark is added.

  Such attribute extraction is similarly performed for each discourse unit candidate of each sentence of the question sentence (step A2 in FIG. 2).

  The specific discourse unit extraction unit 202 applies the specific discourse unit extraction rule stored in the specific discourse unit extraction rule storage unit 301 to each discourse unit candidate generated in this manner, and the specific discourse unit is classified by type. Extracted into The specific discourse unit extraction rule is described in advance for each type of specific discourse unit to be extracted as a conditional expression related to the attributes of the discourse unit candidate. An example of the specific discourse unit extraction rule is shown in FIG. In FIG. 18, two different rules are shown, an extraction rule and a suppression rule. The extraction rules are indicated by names “extraction 1” to “extraction 4” in the drawing, and are composed of discourse types and conditions. Of these, the condition means a condition for determining whether or not this extraction rule matches a certain discourse unit candidate, and the discourse type extracts a discourse unit candidate that satisfies the condition as a specific discourse unit of this discourse type. Means that. Application of these extraction rules will be described later using specific examples.

  On the other hand, the suppression rule is indicated by the names “suppression 1” and “suppression 2” in FIG. 18, and is configured only from conditions. This condition means a judgment condition as to whether or not this suppression rule is applicable to a certain discourse unit candidate. When a certain discourse unit candidate matches any of these suppression rules, it is not extracted as a specific discourse unit even if it matches the conditions of the extraction rule (extraction is suppressed). The suppression rule functions to suppress excessive extraction including an inappropriate specific discourse unit by the extraction rule, and control to extract only appropriate discourse unit candidates as specific discourse units.

  For example, as shown in FIG. 19, it is assumed that candidates 1 to 4 are obtained as discourse unit candidates when the question sentence is “I applied for an analysis pack but do not know where to put the analysis tag display”. Here, an attribute as shown in FIG. 20 is obtained for the discourse unit candidate 3. From the attribute “representative notation =“ know ”” shown in FIG. 20, it can be seen that the condition of the extraction 1 in FIG. 18 is met. However, in the discourse unit candidate 3, the verb “I understand” does not have any case element (as described in the above explanation of “morpheme for each case element of the head predicate”, the predicate itself is not a case element. Is the side that takes in). Therefore, since the capturing case element is 0, the condition of suppression 1 is also satisfied. In this case, the discourse unit candidate 3 is suppressed from being extracted as a specific discourse unit because it matches the extraction rule “extraction 1” and at the same time the suppression rule “suppression 1”.

  Without such a suppression rule, it is expected that a specific discourse unit like "I don't know" shown in this example will be extracted from a large number of question sentences. Because it does not contain information about the contents of things that we do not understand, searching for similar question cases based on such discourse units will search a large number of similar question sentences with completely different question contents Problems will arise. Accordingly, by appropriately setting the suppression rule as in the present embodiment, it is possible to control the extracted specific discourse unit to be an appropriate unit having information that can specify the content of the discourse.

  The attributes of the discourse unit candidate 4 shown in FIG. 17 are “sentence style = question” and “representative notation =“ must be ””. Therefore, the condition described in “extraction 3” in FIG. Meet. Satisfy the condition of the feature value for a specific discourse unit. Therefore, the discourse unit candidate 4 is recognized and extracted as a specific discourse unit of the discourse type “question”. The rules are similarly applied to each discourse unit candidate of each other sentence, and a specific discourse unit is extracted (step A3 in FIG. 2).

  Each specific discourse unit extracted in this way includes the discourse type, the question sentence of the inputted case question answer sentence data and the answer sentence (text information in FIG. 15), and the feature quantity of the discourse unit. The data registration unit 203 registers it in the case question answer sentence data storage unit 302 (step A4 in FIG. 2).

  Next, processing until a similar case is presented for a new question sentence will be described. A new question text is input from the input device 100. Here, it is assumed that the new question sentence is, for example, as shown in FIG. 21 (step A5 in FIG. 2).

  For the question sentence, the discourse unit candidate generation unit 201 generates a discourse unit candidate in the same manner as the process for the case question answer sentence (step A2), and outputs a feature amount for each discourse unit candidate. Then, two discourse unit candidates 1 and 2 as shown in FIG. 22 and their attributes are obtained (step A6 in FIG. 2).

  That is, the discourse unit candidate 1 “What does the green number that differs depending on the user?” Has, for example, “sentence style = question” and part-of-speech “sa change”. Further, the discourse unit candidate 2 “Do you have to provide a counter?” Has attributes including, for example, “sentence style = question” and a representative notation “must”.

  Further, for these discourse unit candidates, the specific discourse unit extraction unit 202 extracts specific discourse units for each type as in step A3. Since the discourse unit candidate 1 has the attributes “sentence style = question” and part of speech “indicative pronoun”, it is collated with the rule “extraction 2” in FIG. Further, the discourse unit candidate 2 includes “sentence style = question” and the representative notation “must be”, and is therefore collated with the rule “extraction 3” in FIG. It is extracted as a specific discourse unit regarding each discourse type “question” collated in this way (step A7 in FIG. 2).

  The case question answer sentence data search unit 204 stores, in the case question answer sentence data storage unit 302, case question answer sentences having specific discourse units of the same type as each specific discourse unit type extracted for the new question sentence. Search from the stored case question answer sentence data. Here, the two specific discourse units extracted from the new question sentence, "What does the different green number for each user mean?" And "Do you have to set up a counter?" Since the type is “question”, the specific discourse unit of the same discourse type is searched from the case question answer sentence data. For example, the specific discourse unit extracted from the first sentence of the example question answer sentence in FIG. 15 “Does a page have to be countered in order to analyze?” Therefore, it is included in this search result. (Step A8 in FIG. 2).

  The inter-question sentence matching unit 205 calculates the similarity between the searched case question sentence and the new question sentence based on the similarity between the specific discourse units. For example, the similarity between the new question sentence “Do I have to set up a counter?” And the searched case question sentence “Must I put a counter on the page?” Is shown in FIG. It is calculated by the similarity as shown. That is, in the new question sentence and the case question sentence, two morphemes of “counter (noun)” and “must be (auxiliary verb)” in the new question sentence match, so the number of matches / the number of morphemes in the new question sentence Is 2/5 and the degree of similarity is 0.4. Here, for example, the threshold is set to 0.3 as a condition for outputting as a similar case question sentence. If the case question sentence indicating the similarity exceeding the threshold is a similar case question sentence, the case question answer sentence of FIG. 15 is determined to be a related question answer example for the new question sentence, and the output device 400 A case question answer sentence is output (step A9 in FIG. 2).

  The calculation of the similarity is not limited to the above, and can be realized using various similarities. For example, not only the number of morphemes (words) but also the weights corresponding to the size of the morphemes that specify the content of the discourse can be given, and the degree of similarity considering these weights can be calculated. For example, there is a method as shown in FIG. 24 as a method of obtaining the similarity by evaluating the weight by a statistical method.

  Here, in the calculation of the weight, in order to measure the size of the morpheme that identifies the content of the discourse, a set of specific discourse units is created in advance to obtain the weight, and the contents of all discourse unit sets are It is determined whether or not they are similar, and this is used for calculating the weight. The weight calculation method itself can be various methods other than those shown here.

  There is also a method for obtaining the similarity by a machine learning method. For example, by using a support vector machine, it is possible to obtain a classifier that inputs a morpheme set common to two discourse units whose similarity is to be obtained and outputs the similarity between the discourse units. .

  Furthermore, not only common morphemes between two discourse units but also synonyms and synonyms can be combined in a method of considering similarity.

  According to the similarity calculation described above, “Do you have to provide an access meter?” And “Do you always need to add a counter to the page?” As shown in FIG. Become. “Access meter” and “counter” are synonyms, but in the case of similarity A that does not consider synonyms, the value of the similarity decreases from 0.4 in the above example to 0.2. Here, when the similarity B that similarly counts synonyms is used, the correspondence between the “access meter” and the “counter” is considered, and the value of the similarity is 0.4. Synonyms can also be taken into account for similarity values in a similar manner. For example, in the similarity C, the similarity between morphemes is evaluated with a value of 0 or more and less than 1 for the morphemes in the synonym relationship, and added to the value of the similarity B. As a result, if the verbs “provide” and “attach” have, for example, a similarity of 0.5, 0.1 is further added to the similarity value, and the similarity value becomes 0.5.

  Next, a second embodiment will be described with reference to FIGS. 3, 4, 26 and 27. FIG. The entire process flow is the same as that of the first embodiment except that an extraction model capable of machine learning is used for extracting a specific discourse unit. Therefore, description of common parts already described in the first embodiment is omitted.

  When the case question answer sentence data to be input is the same as that in the first embodiment and shown in FIG. 15, a discourse unit candidate as shown in FIG. 16 is generated, and further, the discourse unit candidate 4 of the first sentence is generated. For (FIG. 16), feature quantities as shown in FIG. 17 are extracted.

  Here, when learning the specific discourse unit extraction model, the operation proceeds from correct data input (step B1 in FIG. 4) to storage of the specific discourse unit extraction model (step B4 in FIG. 4).

  As the specific discourse unit for each discourse type included in the question sentence of the example question answer sentence data (step A1 in FIG. 4) already input in step B1 in FIG. 4, for example, a specific discourse unit question as shown in FIG. The extracted part in the sentence and its discourse type are input from the input device 100.

  In step B <b> 2 of FIG. 4, the input correct answer for the specific discourse unit is collated with the discourse unit candidate generated by the discourse unit candidate generation unit 201. For example, the discourse type “question” input to the first sentence in FIG. 26 is collated with the discourse unit candidate in FIG. The closest candidate 4 “Do I have to put a counter on the page?” Is changed to a correct answer for each specific discourse type of the discourse type “question”.

  In step B3 of FIG. 4, the specific discourse unit extraction model acquisition unit 206 creates each discourse unit candidate and its feature amount created from the question sentence of the case question answer sentence and the correct answer of the specific discourse unit, from each discourse type. The specific discourse unit extraction model is learned using the specific discourse unit to be extracted as the learning data. For example, a case where a classifier capable of machine learning (for example, a support vector machine) regarding a two-class classification problem is applied will be described.

  A separate classifier is learned for each discourse type to be determined. For example, a classifier that learns whether or not a feature amount of a discourse unit candidate is input and this corresponds to a specific discourse unit of the discourse type “question” is learned. As the feature vector expression for inputting the feature amount of the discourse unit candidate to the classifier, for example, an attribute expression as shown in FIG. 27 showing the same attributes as those in FIG. 17 is used. That is, it represents with the vector which makes the value of an element each element. In FIG. 27, all discourse types are individually learned to obtain classifiers.

  In step B5 of FIG. 4, the specific discourse unit extraction unit 202 selects the specific discourse unit extraction model stored in the specific discourse unit extraction model storage unit 301 for each discourse unit candidate generated by the discourse unit candidate generation unit 201. Is applied to extract a specific discourse unit for each discourse type.

  In step B6 of FIG. 4, for each discourse unit candidate generated from a new question sentence, an attribute expression using a feature vector is similarly obtained and input to all learned discourse type classifiers. As a result, if any of the discourse type classifiers is determined to be a specific discourse unit, the input discourse unit candidate is determined to be a specific discourse type of that discourse type.

  Next, a third embodiment will be described with reference to FIGS. 5, 6, and 28 to 30. FIG. After extracting a specific discourse unit, it is assembled into a discourse block using a discourse block generation schema, and the entire process flow is the same as in the first embodiment, except that this block is used as a unit to collate between question sentences. is there. Therefore, description of common parts already described in the first embodiment is omitted.

  In step C1 of FIG. 6, the discourse block generation unit 207 stores the discourse stored in the discourse block generation schema storage unit 304 for a plurality of specific discourse units for each discourse type extracted by the specific discourse unit extraction unit 202. Applying the block generation schema, one or more specific discourse units are grouped into discourse blocks corresponding to a set of question items. For example, assume that the extracted specific discourse unit is as shown in FIG. 28 and the discourse block generation schema is as shown in FIG. 29, for example. The discourse block generation schema of FIG. 29 is described by seven binary rules. For example, when the discourse block “situation” and the discourse block “situation” are adjacent to each other, the rule 1 “situation, situation → situation” collectively represents the discourse block “situation”. This rule is applied to a specific discourse unit preferentially from rule 1, and a discourse block is generated by generating a discourse block. For example, the rules 3, 7, and 5 are applicable to the array of five specific discourse units “problem”, “situation”, “question”, “problem”, and “request” in FIG. That is, “problem, situation → problem”, “problem, question → question”, and “problem, request → request” are matched, and these are sequentially adapted to obtain two discourse blocks as shown in FIG. The generated discourse block and information of a specific discourse unit belonging to the generated discourse block are output as discourse block information.

  Next, a fourth embodiment will be described with reference to FIGS. 7, 8, and 31 to 33. FIG. When searching for the case question answer sentence, not only the question sentence that includes the specific discourse unit that matches the discourse type extracted from the new question sentence but also the discourse type that matches the discourse type after applying the discourse type conversion rule The entire process flow is the same as that of the first embodiment except that a search is performed including a query and collation between question sentences is performed. Therefore, description of common parts already described in the first embodiment is omitted.

  In step D1 of FIG. 8, the case question answer sentence data search unit 204 stores the discourse type or the discourse type conversion rule storage unit 305 for each specific discourse unit extracted from the new question sentence by the specific discourse unit extraction unit 202. A case question answer sentence having a specific discourse unit of the same discourse type as the discourse type after conversion by the stored discourse type conversion rule is selected from the case question answer sentence data stored in the case question answer sentence data storage unit 302. Search for. Here, for example, it is assumed that the discourse type conversion rule is as shown in FIG. For example, rule 1 “situation, question → question, question” indicates that the arrangement of the discourse type “situation” and the discourse type “question” is replaced with the discourse type “question” and the discourse type “question”. . It should be noted that the original sequence before replacement is also left as it is. The specific discourse unit extracted from the new question sentence is as shown in FIG. 31, and a question answer example having the same content as this question sentence exists in the case question answer sentence data, and the specific answer extracted from the question sentence. Assume that the discourse unit is as shown in FIG. In this case, there is no specific discourse unit in which the discourse type matches between the question sentences, but the arrangement of the specific discourse units extracted from the new question sentence in FIG. 31 is “situation, question”. The first and second discourse type conversion rules can be applied. In addition to the original “situation / question”, “problem / question” and “problem / request” are obtained as the arrangement of specific discourse units after conversion. It is done. Among these converted discourse types, the “question, request” matches the specific discourse unit sequence for the question sentence in the case question answer sentence, so this case question answer sentence is searched.

  In step D2 of FIG. 8, the inter-question sentence matching unit 205 determines the similarity between the new question sentence and the question sentence of the case question answer sentence, the percentage of matching including the discourse type match after applying the discourse type conversion rule, Then, calculation is made based on the similarity between specific discourse units with the same discourse type, including after conversion by the discourse type conversion rule, and case question answer sentences with a certain value or more are determined as similar case question answer sentences.

  Next, a fifth embodiment will be described with reference to FIGS. 9, 10, 34 and 35. FIG. When matching between question sentences, in addition to the similarity based on the features of specific discourse units with matching discourse types, the question sentences are also matched in consideration of the matching of the requested question types of each question sentence The entire process flow is the same as in the first embodiment. Therefore, description of common parts already described in the first embodiment is omitted.

  When the specific discourse unit extracted from the new question sentence is as shown in FIG. 34 and the specific discourse unit extracted from the question sentence of the case question answer sentence is as shown in FIG. 35, both discourse types are “ The requested question type is extracted for each specific discourse unit of “Question”. Here, for example, the semantic classification assumed for a pronoun can be narrowed down by combining the semantic classification of a pronoun (question word) and the restriction on the case element of a predicate having the pronoun as a case element. Here, if the request answer type for a new question sentence is "concrete, abstract" and the request answer type for a question answer example question sentence is "location", these do not match, so both question sentences are required. It is judged that the content of the answer is different.

  A sixth embodiment will be described with reference to FIGS. 11, 12, and 34 to 36. FIG. When matching between question sentences, the request question type of the new question sentence and the contents of the answer sentence of the case question answer sentence should match in addition to the similarity based on the feature amount of the specific discourse unit with the same discourse type The entire processing flow is the same as that of the first embodiment except that the question text is checked in consideration of the above. Therefore, description of common parts already described in the first embodiment is omitted.

  It is assumed that the specific discourse unit extracted from the new question sentence is as shown in FIG. 34, and the answer sentence of the retrieved case question answer sentence is as shown in FIG. Here, for example, the semantic classification assumed for a pronoun can be narrowed down by combining the semantic classification of a pronoun (question word) and the restriction on the case element of a predicate having the pronoun as a case element. Here, the required question type extracted from the specific discourse unit whose discourse type of the new question sentence is “question” is presumed to be “concrete, abstract” from the pronoun “what” as the object of the verb “use” Suppose.

  This matches the semantic classification of the noun “FTP tool” as the object of the verb “use” in the first sentence of the answer sentence. In this case, it is determined that the answer contents are consistent with the request question. The

  A seventh embodiment will be described with reference to FIGS. 13, 14, and 37 to 42. FIG. Prior to the extraction of specific discourse units, the discourse unit candidates include anaphoric relationships, omissions, and sharing of case elements. If there are antecedents or shared case elements outside the discourse unit candidates, this is complemented. The overall processing flow is the same as in the first embodiment, except that a discourse unit candidate is created and the complement result is also reflected in the feature amount. Therefore, description of common parts already described in the first embodiment is omitted.

  In the anaphora analysis processing unit 209, when each discourse unit candidate generated by the discourse unit candidate generation unit 201 includes a pronoun or abbreviation, the antecedent is recognized in the question sentence.

  First, a case where omission completion and anaphoric relationship resolution (the antecedent pointed to by the demonstrative pronoun is recognized from the preceding and following contexts) will be described using an example. For example, assume that the question text input in step A1 of FIG. 14 is as shown in FIG. In step A2, a discourse unit candidate is generated and a feature amount is extracted from the input sentence. For example, it is assumed that the discourse unit candidate for the input sentence in FIG. 37 is as shown in FIG.

  In step G1 (FIG. 14), anaphoric analysis and omission supplement are performed for each of these discourse units. Here, anaphora analysis and abbreviation completion are the processes to determine the abbreviations, demonstrative pronouns, formal nouns included in each discourse unit candidate, and identify each antecedent from the entire input sentence including the discourse unit candidate. To do. A number of methods have already been disclosed for specific processing methods for such anaphora analysis and omission supplementation.

  First, in candidates 3, 4, and 5, it is determined that the object for the verb “attach” is omitted. Here, when “counter” is specified as an antecedent of the omitted object, the specified omitted element is complemented for each discourse unit candidate and feature amount. For example, when the attributes for the discourse unit candidate 3 are those shown in FIG. 39, the attributes after complementing the omitted elements are those shown in FIG. Shared relationship (Here, a plurality of predicates refers to a relationship that shares one case element. For example, in the case of subject sharing, a plurality of predicates share a single subject.) If a discourse unit that includes each word is extracted in a sentence with a unique sharing relationship, a discourse unit that apparently has no subject will be omitted.) In addition, the case elements that are shared can be identified and complemented.

  Next, a case where omission complementation and anaphoric relationship cannot be resolved will be described using an example. Further, it is assumed that the question sentence input in step A1 in FIG. 14 is as shown in FIG. 41, for example. In step A2, a discourse unit candidate is generated and an attribute is extracted from the input sentence. For example, it is assumed that the discourse unit candidate for the input sentence in FIG. 41 is as shown in FIG. In step G1 (FIG. 14), it is determined in candidates 4, 6, and 7 that the object for the verb “necessary” is omitted. Here, “Introduction procedure”, “Analysis target page”, “Counter”, and “Attach counter” are candidates as an antecedent of this omitted object, but it cannot be uniquely identified and is omitted. Suppose it failed. In this case, candidates 4 and 6 are excluded from the discourse unit candidates because there is no possibility of including an antecedent antecedent. As a result, candidates 1, 2, 5, and 7 are newly configured as discourse unit candidates.

As described above, new discourse unit candidates are constructed from discourse unit candidates that originally do not include omissions, anaphoric relationships, and shared relationships, and those that have omitted elements supplemented by anaphora analysis or omission complementation. Then, in the next step A3 (FIG. 14), a specific discourse unit is extracted.
Note that the extraction of the specific discourse unit from the discourse unit candidate is the same as that already described in the first embodiment, and therefore will be omitted. By such processing, if the omission completion is possible for the discourse unit candidate including the omission element, the specific discourse unit is extracted based on the complemented result, and if the omission cannot be completed, it is larger than the antecedent. A specific discourse unit is extracted from the discourse unit. This makes it possible to extract specific discourse units that do not cause information loss even if there are omissions, anaphoric relationships, and shared relationships in the sentence, and search for example question answer sentences based on more appropriate discourse units It becomes possible.

  According to the present invention, an operator support support device for an email in a contact center, an operator support support device for telephone voice, a similar case search device for question answer case data, and an analysis device for question answer case data. Applicable.

It is a block diagram which shows the structure of the question answer search system which concerns on the 1st Embodiment of this invention. It is a flowchart which shows operation | movement of the question answer search system which concerns on the 1st Embodiment of this invention. It is a block diagram which shows the structure of the question answer search system which concerns on the 2nd Embodiment of this invention. It is a flowchart which shows operation | movement of the question answer search system which concerns on the 2nd Embodiment of this invention. It is a block diagram which shows the structure of the question answer search system which concerns on the 3rd Embodiment of this invention. It is a flowchart which shows operation | movement of the question answer search system which concerns on the 3rd Embodiment of this invention. It is a block diagram which shows the structure of the question answer search system which concerns on the 4th Embodiment of this invention. It is a flowchart which shows operation | movement of the question answer search system which concerns on the 4th Embodiment of this invention. It is a block diagram which shows the structure of the question answer search system which concerns on the 5th Embodiment of this invention. It is a flowchart which shows operation | movement of the question answer search system which concerns on the 5th Embodiment of this invention. It is a block diagram which shows the structure of the question answer search system which concerns on the 6th Embodiment of this invention. It is a flowchart which shows operation | movement of the question answer search system which concerns on the 6th Embodiment of this invention. It is a block diagram which shows the structure of the question answer search system which concerns on the 7th Embodiment of this invention. It is a flowchart which shows operation | movement of the question answer search system which concerns on the 7th Embodiment of this invention. It is a figure which shows the example of the example question answer text data in 1st Example of this invention. It is a figure which shows the discourse unit candidate of the question sentence obtained with respect to the 1st sentence in the question sentence of the example question answer sentence data in 1st Example of this invention. It is a figure which shows the attribute of the discourse unit candidate extracted with respect to the 1st sentence and the discourse unit candidate 4 in the question sentence of the example question answer sentence data in 1st Example of this invention. It is a figure which shows the example of the specific discourse unit extraction rule in 1st Example of this invention. It is a figure which shows the discourse unit candidate for demonstrating a suppression rule. It is a figure which shows the attribute of the discourse unit candidate extracted with respect to discourse unit candidate 3 in description of the suppression rule. It is a figure which shows the example of the new question sentence in 1st Example of this invention. It is a figure which shows the discourse unit candidate with respect to the new question sentence in the 1st Example of this invention, and its attribute. It is a figure which shows the distance calculation method between the discourse units of the example question response sentence and new question sentence in 1st Example of this invention. It is a figure which shows the other distance calculation method between discourse units. It is a figure which shows the comparison of the various distance calculation methods between discourse units. It is a figure which shows an example of the correct answer data of the specific discourse unit for every discourse type with respect to the example question answer sentence in 2nd Example of this invention. It is a figure which shows an example of the feature vector expression for making the attribute of a specific discourse unit into the input of a classifier in 2nd Example of this invention. It is a figure which shows an example of the result of having extracted the specific discourse unit from the question sentence containing the several question matter in the 3rd Example of this invention. It is a figure which shows an example of the discourse block production | generation schema in the 3rd Example of this invention. It is a figure which shows an example of the discourse block obtained using the discourse block production | generation schema in the 3rd Example of this invention. It is a figure which shows an example of the specific discourse unit extracted from the new question sentence in the 4th Example of this invention. It is a figure which shows an example of the specific discourse unit extracted from the question sentence of the example question answer sentence in 4th Example of this invention. It is a figure which shows an example of the discourse type conversion rule in 4th Example of this invention. It is a figure which shows an example of the specific discourse unit extracted from the new question sentence in the 5th Example of this invention. It is a figure which shows an example of the specific discourse unit extracted from the question sentence of the case question answer sentence in 5th Example of this invention. It is a figure which shows an example of the reply sentence of the example question reply sentence in the 6th Example of this invention. It is a figure which shows the example of the new question sentence in the 7th Example of this invention. It is a figure which shows the discourse unit candidate with respect to the question sentence in the 7th Example of this invention. It is a figure which shows the attribute of the discourse unit candidate extracted with respect to the discourse unit candidate 3 in the 7th Example of this invention. It is a figure which shows the attribute after the complementation of the discourse unit candidate extracted with respect to the discourse unit candidate 3 in the 7th Example of this invention. It is a figure which shows the other example of the new question sentence in the 7th Example of this invention. It is a figure which shows the discourse unit candidate with respect to the other question sentence in the 7th Example of this invention.

Explanation of symbols

100 Input device 200a, 200b, 200c, 200d, 200e, 200f, 200g Data processing device 201, 201a Discourse unit candidate generation unit 202, 202a Specific discourse unit extraction unit 203, 203a, 203b Case data registration unit 204, 204a Case question answer Sentence data search unit 205, 205a, 205b, 205c, 205d Interrogation collation unit 206 Specific discourse unit extraction model acquisition unit 207 Discourse block generation unit 208 Request answer type extraction unit 209 Anaphoric analysis processing units 300a, 300b, 300c, 300d, 300e, 300f, 300g Storage device 301 Specific discourse unit extraction rule storage units 302, 302a, 302b Case question answer sentence data storage unit 303 Specific discourse unit extraction model storage unit 304 Discourse block generation schema storage unit 305 Discourse Type conversion rule storing unit 306 refers classification dictionary storage unit 307 requests the type of answer extraction pattern storage unit 308 answer type matching pattern storage unit 400 the output device

Claims (24)

Generating a discourse unit candidate based on the sentence or clause structure of the syntax structure after performing morphological analysis and syntax analysis on the case question sentence in advance;
From the result of parsing, extracting and assigning syntactic and semantic attributes to the discourse unit candidate and the surface layer representation of the discourse unit candidate;
Applying a discourse unit extraction rule in which matching conditions for the syntactic and semantic attributes are described for each discourse type for each discourse unit candidate for each discourse type, and a specific discourse unit from the case question sentence for each discourse type A step of extracting
Storing the extracted discourse unit as a question answer example in a database together with the example question sentence itself;
When a new question sentence is input, extracting a specific discourse unit for the new question sentence in the same manner as the case question sentence;
Searching the database for question answer examples having a specific discourse unit of the same discourse type as the discourse type of the specific discourse unit for the new question sentence;
It is determined that the case question sentence is similar to the new question sentence based on the similarity between specific discourse units whose discourse types match between the case question sentence in the searched question answer case and the new question sentence. And steps to
A method for searching for answers to questions characterized by including: In the question answer search method according to claim 1,
Enter the location information and discourse type of a specific discourse unit that should be extracted from the case question sentence in advance as learning data, perform morphological analysis and syntactic analysis, and then use the discourse unit based on the sentence or clause structure of the syntax structure. A candidate is generated, and a syntactic and semantic attribute for each discourse unit candidate is extracted and input from a result of the parsing, and a specific discourse unit extraction model for outputting a discourse type is learned.
When extracting a specific discourse unit from a question sentence for each discourse type, the specific answer unit of each discourse type is extracted from the discourse unit candidates by applying the specific discourse unit extraction model. retrieval method. In the question answer search method according to claim 1,
After extracting specific discourse units from case question sentences and new question sentences, a discourse scheme is a rule that summarizes multiple specific discourse units into discourse blocks that become the minimum unit for the same question content based on discourse type Including the step of applying to the discourse block
In the calculation of the similarity between the new question sentence and the case question sentence, the case question sentence and the new question are based on the similarity between the specific discourse units for which the discourse type matches in the new question sentence and the case question sentence for each discourse block. A question answer search method characterized by determining that the discourse blocks of sentences are similar. In the question answer search method according to claim 1,
When collating specific discourse units of case question sentences and new question sentences, the similarity between specific discourse units of the same discourse type and specific discourses between different discourse types mapped via specific discourse unit transformation rules A question answer search method characterized in that the similarity between a case question sentence and a new question sentence is determined based on the similarity between units. In the question answer search method according to claim 1,
Including a step of estimating an answer type, which is a classification related to an answer content expected as an answer to a question sentence,
When determining the similarity between a case question sentence and a new question sentence, the case type question sentence is matched with the similarity between the question type and the answer type for the question sentence based on the syntactic and semantic attributes between specific discourse units. And a question answer search method, characterized in that it is used to determine the similarity between a new question sentence and a new question sentence. In the question answer search method according to claim 5,
A question answer search method characterized in that the answer sentence case is stored in advance in a database together with the case question sentence, and the degree of coincidence between the answer type and the content of the case answer sentence with respect to the case question sentence is taken into consideration. In the question answer search method according to claim 1,
When generating a discourse unit candidate from a question sentence and extracting a specific discourse unit from it, the discourse unit candidate includes a demonstrative pronoun or abbreviation, and a demonstrative pronoun or abbreviation antecedent exists outside the specific discourse unit In this case, substitution or complementation with an anaphor or an antecedent antecedent is performed, or a larger discourse unit candidate including the target discourse unit candidate and antecedent is extracted as a specific discourse unit A question answer search method characterized by collating a question sentence with a question example based on a specific discourse unit that complements an anaphoric relationship.   In the question answer search method according to any one of claims 1 to 7, when a specific discourse unit of a case question sentence and a new question sentence is collated, a question case that is a syntactic and semantic attribute for each specific discourse unit Predicate sentence style, tense and aspect information, semantic classification of head utterances, and if head utterances have subject, object, or noun phrases as case elements, all belonging to those case elements A method for searching for answers to a question, wherein the similarity is obtained based on another match of a morpheme surface layer, prototype, part of speech, semantic classification, and case element head. A discourse unit candidate generating means for generating a discourse unit candidate based on the structure of a sentence or a section of the syntax structure after performing morphological analysis and syntax analysis on the question sentence;
Syntactic and semantic attributes are extracted and assigned to the discourse unit candidate and the surface representation of the discourse unit candidate, and the discourse unit candidate is changed from the discourse unit candidate to each discourse type according to the matching condition for the syntactic and semantic attribute. A discourse unit extraction rule storage means for storing a discourse unit extraction rule describing a rule for determining the extraction of
A discourse unit extraction unit that applies the discourse unit extraction rule to a discourse unit candidate and extracts a discourse unit from a question sentence for each discourse type;
Case data registration means for registering the specific discourse unit extracted from the question sentence case in the discourse unit extraction means in the question answer sentence case data storage means as question answer sentence case data together with the answer sentence to the question sentence and the question sentence;
Question answer sentence case data storage means for storing the question answer sentence case data registered by the question answer sentence case data registration means;
A question answer having a specific discourse unit of the same discourse type as the discourse type of the new question sentence, for each specific discourse unit extracted by the specific discourse unit extracting means for a new question sentence input from the input device Question answer sentence case data search means for searching sentence examples from the question answer sentence case data stored in the question answer sentence case data storage means;
Question sentence matching means for determining whether or not question sentences are question sentences related to similar contents based on the similarity in a specific discourse unit between the question sentence of the searched question answer sentence case and the new question sentence;
A question and answer search system comprising: A specific discourse unit extraction model acquisition means for acquiring a specific discourse unit extraction model by learning a specific discourse unit extraction model using a question sentence given a discourse type to be extracted from a question sentence as learning data,
Specific discourse unit extraction model storage means for storing the specific discourse unit extraction model;
In addition,
The specific discourse unit extraction means applies the specific discourse unit extraction model stored in the specific discourse unit extraction model storage means to each discourse unit candidate generated by the discourse unit candidate generation means, and for each discourse type 10. The question answer search system according to claim 9, wherein a specific discourse unit is extracted. A discourse block certification means that applies a discourse scheme to a specific discourse unit extracted from a question sentence and collects the specific discourse unit into a discourse block,
A discourse scheme storage means for storing a discourse scheme;
A question case matching means for matching in a discourse block unit in a specific discourse unit matching between a question sentence and a case question sentence,
When a question sentence includes a plurality of question items, a discourse block is individually created for each question item, and collation is performed for each discourse block in collation between a question sentence and an example. 9. The question answer search system according to 9. A specific discourse unit matching means for collating a discourse type that is matched using a specific discourse type modification rule for each discourse type when collating a specific discourse unit between a question sentence and a question example;
Specific discourse unit deformation rule storage means for storing specific discourse unit deformation rules for discriminating from the same discourse type when specific discourse units of different discourse types represent the intent and content of the same question sentence depending on the context,
The question answer search system according to claim 9, wherein the question sentence matched with the corresponding specific discourse unit and the question example are determined to be related. Question / answer type estimation means for judging the question type related to the question content and request content from the question sentence, and further estimating the expected answer type,
The system further comprises a question / answer type matching means for matching the question type and answer type between the question sentence and the question example,
The question answer search system according to claim 9, wherein a match between the question sentence and the question example of the question example and the answer type is determined based on the collation result. Question / answer type estimation means for determining the type of question content and request content from the question text and estimating the expected answer type;
A question-answer correspondence unit for determining whether or not a portion corresponding to the answer type estimated from the question sentence exists in the answer sentence for the question sentence;
A question answer case storage means for accumulating a pair of answer cases for the question case and the question case question,
The question / answer type is estimated for the question sentence, the question case with the matching question / answer type is searched, and the question / answer type matches the part corresponding to the answer type estimated from the question sentence. 10. The question answer search system according to claim 9, wherein it is determined whether or not it exists in an answer example for the case, and if it exists, it is determined as an answer example including an answer to the question. The anaphoric relationship included in the specific discourse unit generated by the discourse unit candidate generating means is recognized, and the specified discourse or abbreviation is included in the specific discourse unit and the specified pronoun or abbreviation is outside the specific discourse unit Further comprising an anaphoric analysis processing means for replacing an anaphor with an antecedent when it exists in the
When the specific discourse unit extracted from the question sentence is supplemented with the pronoun and abbreviation elements by the anaphora analysis processing means, the specific discourse unit and the abbreviation element are compared with the specific discourse unit case based on the completion result. 10. The question answer search system according to claim 9, wherein, when remaining, the collation unit including the specific discourse unit is collated with the case as an alternative specific discourse unit.   In the question answer search system according to any one of claims 10 to 15, when collating a specific discourse unit between a case question sentence and a new question sentence, a question example that is a syntactic and semantic attribute for each specific discourse unit Predicate sentence style, tense and aspect information, semantic classification of head utterances, and if head utterances have subject, object, or noun phrases as case elements, all belonging to those case elements A question answer search system, wherein the similarity is obtained based on another match of a morpheme surface layer, prototype, part of speech, semantic classification, and case element head. On the computer that composes the question answer search system,
After performing morphological analysis and syntactic analysis on the case question sentence in advance, a process of generating discourse unit candidates based on the sentence or clause structure of the syntax structure;
From the result of parsing, a process of extracting and assigning syntactic and semantic attributes to the discourse unit candidate and the surface layer representation of the discourse unit candidate;
Applying a discourse unit extraction rule in which matching conditions for the syntactic and semantic attributes are described for each discourse type for each discourse unit candidate for each discourse type, and a specific discourse unit from the case question sentence for each discourse type The process of extracting
A process of storing the extracted specific discourse unit as a question answer example in a database together with the example question sentence itself;
When a new question sentence is input, a process of extracting a specific discourse unit for the new question sentence in the same manner as the case question sentence;
Processing for searching a question answer example having a specific discourse unit of the same discourse type as the discourse type of the specific discourse unit for the new question sentence from the database;
It is determined that the case question sentence is similar to the new question sentence based on the similarity between specific discourse units whose discourse types match between the case question sentence in the searched question answer case and the new question sentence. Processing to
A program that executes The program according to claim 17, wherein
Enter the location information and discourse type of a specific discourse unit that should be extracted from the case question sentence in advance as learning data, perform morphological analysis and syntactic analysis, and then use the discourse unit based on the sentence or clause structure of the syntax structure. Generate a candidate and extract and input syntactic and semantic attributes for each discourse unit candidate from the result of parsing, and further execute a process of learning a specific discourse unit extraction model that outputs the discourse type,
A program characterized in that when a specific discourse unit is extracted from a question sentence for each discourse type, a specific discourse unit of each discourse type is extracted from discourse unit candidates by applying the specific discourse unit extraction model. The program according to claim 17, wherein
After extracting specific discourse units from case question sentences and new question sentences, a discourse scheme is a rule that summarizes multiple specific discourse units into discourse blocks that become the minimum unit for the same question content based on discourse type To further execute the process of grouping into discourse blocks by applying
In the calculation of the similarity between the new question sentence and the case question sentence, the case question sentence and the new question are based on the similarity between the specific discourse units for which the discourse type matches in the new question sentence and the case question sentence for each discourse block. A program characterized by determining that the discourse block of sentences is similar. The program according to claim 17, wherein
When collating specific discourse units of case question sentences and new question sentences, the similarity between specific discourse units of the same discourse type and specific discourses between different discourse types mapped via specific discourse unit transformation rules A program characterized in that the similarity between a case question sentence and a new question sentence is judged based on the similarity between units. The program according to claim 17, wherein
Execute the process of estimating the answer type, which is a classification related to the answer content expected as the answer to the question sentence,
When determining the similarity between a case question sentence and a new question sentence, the case type question sentence is matched with the similarity between the question type and the answer type for the question sentence based on the syntactic and semantic attributes between specific discourse units. A program characterized in that it is used to determine the similarity between a new question and a new question. The program according to claim 21, wherein
A program characterized in that an answer sentence case is stored in a database together with a case question sentence in advance, and the degree of coincidence between the answer type and the content of the case answer sentence with respect to the case question sentence is taken into consideration. The program according to claim 17, wherein
When generating a discourse unit candidate from a question sentence and extracting a specific discourse unit from it, the discourse unit candidate includes a demonstrative pronoun or abbreviation, and a demonstrative pronoun or abbreviation antecedent exists outside the specific discourse unit In this case, substitution or complementation with an anaphor or an antecedent antecedent is performed, or a larger discourse unit candidate including the target discourse unit candidate and antecedent is extracted as a specific discourse unit A program characterized by collating a question sentence with a question example based on a specific discourse unit that complements an anaphoric relationship. In the program according to any one of claims 17 to 23,
When collating the specific discourse unit of the case question sentence and the new question sentence, the syntactic and semantic attributes for each specific discourse unit are the sentence style of the predicate of the question case, tense / aspect information, and semantic classification of the head predicate In addition, if the head vocabulary has a subject, object, or noun phrase as a case element, it is the surface, prototype, part of speech, semantic classification, and case element head of all morphemes belonging to those case elements A program for obtaining the similarity based on another coincidence.

Images