JP5229782B2 - Question answering apparatus, question answering method, and program - Google Patents

Description

  The present invention relates to a question answering apparatus or the like that accepts a non-factoid type question and outputs answer information for the question.

Conventionally, a question answering system that accepts a question sentence and outputs an answer to the question sentence has been developed (see, for example, Patent Document 1).
Japanese Patent No. 3861105

  However, the conventional question answering apparatus is often a system that mainly answers a What type question, and a system that answers a How and What type question has not been developed so much. Further, the conventional How, Why type question answering system has a low performance, and it has been a problem to improve the performance.

  The present invention has been made to solve the above-described problems, and an object of the present invention is to provide a high-performance question answering apparatus and the like that can appropriately answer a How and Why type question. To do.

  In order to achieve the above object, a question answering apparatus according to the present invention includes a question information receiving unit that receives question information, which is information indicating a non-factoid type question, and question information received by the question information receiving unit. A classification unit that assigns one of a plurality of classification information that is information indicating a classification of the question information and that includes a reason question that is a question for asking a reason as at least one classification, and the question Information that includes a term extraction unit that extracts terms from the question information received by the information reception unit, classification information that is information indicating classification, and additional terms that are added to the terms extracted by the term extraction unit A correspondence information storage unit for storing certain correspondence information, an additional term associated with the classification information provided by the classification unit in the correspondence information, and a term extraction unit The answer is information indicating the answer corresponding to the question information by using the term, the corpus stored in the accessible corpus storage unit, and the formula according to the classification information given by the classification unit A response information acquisition unit that acquires information from the corpus; and a response information output unit that outputs the response information acquired by the response information acquisition unit.

  With such a configuration, it is possible to appropriately acquire and output answer information even for non-factoid type question information. Furthermore, higher performance can be obtained by adding additional terms.

In the question answering apparatus according to the present invention, the first expression used by the answer information acquisition unit is an expression that becomes higher as the two terms are closer to each other in the document included in the corpus. Yes, the answer information acquisition unit includes, for the document included in the corpus, an additional term associated with the classification information given by the classification unit in the correspondence information, and the term extracted by the term extraction unit It is also possible to calculate the value of the formula using the two selected, and obtain answer information, which is information with a larger value of the formula than others.
With such a configuration, by acquiring the response information using the first equation, appropriate response information is acquired.

  Further, in the question answering apparatus according to the present invention, the answer information acquiring unit uses a second expression that is an expression indicating the degree to which a certain term characterizes a document, and the term extracted by the term extracting unit. Document acquisition means for acquiring a plurality of documents characterized to be high from the corpus, additional term acquisition means for acquiring additional terms corresponding to the classification information given by the classification unit, from the correspondence information, Answer candidate information that is information that is a candidate for answer information included in each document acquired by the document acquisition unit using the term extracted by the term extraction unit and the additional term acquired by the additional term acquisition unit The calculation means for calculating the value of the first equation according to the classification information given by the classification section, and the value calculated by the calculation means from the plurality of answer candidate information is compared with others. And answer information selecting means for selecting the answer information is a great value, may be provided.

  With such a configuration, a document is acquired using the second equation, and then the response information is selected from the answer candidate information included in the document using the first equation, thereby increasing the processing load. The amount of calculation related to the first equation can be reduced, the processing load can be reduced, and the processing time can be shortened.

  In the question answering apparatus according to the present invention, the classification unit asks at least a definition question that is a question asking for a definition, a reason question that is a question asking a reason, and a method of the question information received by the question information receiving unit. The question may be classified as a method question.

  With such a configuration, it is possible to appropriately acquire answer information by adding additional terms according to this classification and using an expression corresponding to this classification.

In the question answering apparatus according to the present invention, the term extraction unit also extracts a focus expression that is an expression for which a definition is being asked from question information classified as a definition question by the classification unit. The first expression is an expression that has a larger value when the focus expression is included in the document acquired by the document acquisition unit than when the focus expression is not included. When the focus expression included in the document acquired by the means is modified by a linkage modification clause, the expression is a value that is larger than the case where the focus expression is not modified. When the focus expression included in the acquired document is modified with a combination modification clause, the combination modification clause may be selected as response information from the answer candidate information.
With such a configuration, it is possible to more appropriately extract answer information for a definition question.

Further, in the question answering device according to the present invention, the answer information selecting means selects answer candidate information whose value calculated by the calculating means is larger than others, prepared in advance, Machine learning is performed using answer information that is information indicating an answer to the question indicated by the question information and information indicating the suitability of the answer information as teacher data, and the selected result is obtained using the result of the machine learning. The answer information may be extracted from the answer candidate information.
With such a configuration, it is possible to appropriately extract answer information from answer candidate information using machine learning.

  In the question answering apparatus according to the present invention, the classification unit holds classification correspondence information, which is information including classification information, which is information indicating classification, and word information, which is information indicating phrases, in a recording medium. In addition, when the question information includes a phrase indicated by the phrase information, classification information corresponding to the phrase information may be given to the question information.

  In the question answering apparatus according to the present invention, the classifying unit performs machine learning using prepared question information and classification information that is information indicating the classification of the question information as teacher data, and performs the machine learning. The question information received by the question information receiving unit may be classified using the result.

  In the question answering apparatus according to the present invention, the term extraction unit performs morphological analysis on the question information, and from the question information, (1) independent words, (2) nouns, (3) nouns and verbs, (4) Part of speech terms included in any one of (1) to (5) selected from nouns and adjectives, (5) nouns, verbs, and adjectives may be extracted.

  According to the question answering device or the like according to the present invention, it is possible to appropriately answer a non-factoid type question.

  Hereinafter, a question answering apparatus according to the present invention will be described using embodiments. In the following embodiments, components and steps denoted by the same reference numerals are the same or equivalent, and repetitive description may be omitted.

(Embodiment 1)
A question answering apparatus according to Embodiment 1 of the present invention will be described with reference to the drawings. The question answering device according to the present embodiment classifies non-factoid questions, and obtains and outputs answers according to the classification results.

  FIG. 1 is a block diagram showing a configuration of a question answering apparatus 1 according to the present embodiment. The question answering apparatus 1 according to the present embodiment includes a question information receiving unit 11, a classification unit 12, a term extracting unit 13, a correspondence information storage unit 14, a corpus storage unit 15, an answer information acquisition unit 16, and an answer And an information output unit 17.

  The question information receiving unit 11 receives question information that is information indicating a non-factoid type question. This question information is, for example, text data indicating a question. The question information receiving unit 11 may receive a question other than a non-factoid type question. Here, the factoid type question is a question in which a noun is an answer. For example, “Where is the capital of Japan?” And “When did humans reach the moon?” Are factoid questions. The answers to these questions are nouns such as “Tokyo” or “July 19, 1969”. A non-factoid type question is a question in which a document is an answer, unlike a factoid type question. For example, “Why do people oppose the Personal Information Protection Law?” And “What is the World Heritage Convention?” Are non-factoid questions. is there.

  The question information receiving unit 11 may receive question information input from an input device (for example, a keyboard, a mouse, a touch panel, etc.), for example, and receive the question information transmitted via a wired or wireless communication line. Alternatively, query information read from a predetermined recording medium (for example, an optical disk, a magnetic disk, a semiconductor memory, etc.) may be received. For example, the question information reception unit 11 may receive question information that is a result of voice recognition of a user's voice. Note that the question information receiving unit 11 may or may not include a device (for example, a modem or a network card) for receiving. The question information receiving unit 11 may be realized by hardware, or may be realized by software such as a driver that drives a predetermined device.

  Moreover, the question information received by the question information receiving unit 11 may be temporarily stored in a recording medium (not shown). Question information used in processing performed by the classification unit 12 and the term extraction unit 13 to be described later may be read from the recording medium. The accumulation of the question information in the recording medium (not shown) may be performed by the question information reception unit 11 or may be performed by another storage unit (not shown).

  The classification unit 12 assigns classification information corresponding to the question information to the question information received by the question information reception unit 11. The classification information is information indicating the classification of the question information. The classification information includes a reason question that is a question asking a reason as at least one classification. The classification unit 12 specifies classification information indicating the classification of the question information received by the question information reception unit 11 from among a plurality of classification information, and assigns the specified classification information to the question information. The classification unit 12 may add the classification information to the question information, for example, may store the classification information to be added in a predetermined recording medium, and set a predetermined flag in association with the classification information to be added. It may be to do. Although the case where the classifying unit 12 classifies the question information has been described here, classifying the question information is equivalent to classifying answer information indicating an answer corresponding to the question information. For example, classifying certain question information as “reason question” is equivalent to classifying answer information corresponding to the question information as “reason answer”. Therefore, it can also be said that the classification unit 12 classifies the response information (the classification information is given to the response information).

  The classifications other than the reason questions include definition questions that ask questions, method questions that ask questions, method questions that ask questions, questions to some extent questions, change questions that ask questions, and questions that ask questions. There are questions. That is, the classification unit 12 includes at least the question information received by the question information receiving unit 11 as a definition question that asks a definition, a reason question that asks a reason, a method question that asks a method, and a degree. The question to be asked may be classified as a question to some extent, a change question that is a question to ask for change, and a history question that is a question to ask about the story. Alternatively, the classification unit 12 may perform other classifications. For example, the classification unit 12 may classify the question information received by the question information reception unit 11 into a definition question that is a question asking a definition, a reason question that is a question asking a reason, and a method question as a question asking a method. Good.

  The definition question is a question such as “What is K-1?” Or “What is K-1?”.

  The reason questions are, for example, questions such as “Why are people against the Personal Information Protection Law?” “Why are the people opposed to the Private Information Protection Law?”.

  The method question is, for example, a question such as “How is World Heritage determined?” “How is a World Heritage determined?”.

  The degree question is, for example, a question such as “How much damage was caused by the Chernobyl nuclear accident?” “How extensible was the damage caused by Chernobyl nuclear accident?”

  The change question is, for example, a question such as “How did the juvenile law change?” “How was the juvenile row changed?”.

  The progress question is, for example, a question such as “How did the Ryukyu Kingdom become part of Japan?” Or “How did Ryukyu come to belong to Japan?”.

  The classification unit 12 may, for example, (1) give classification information according to a predetermined rule, or (2) give classification information by performing machine learning. These methods for assigning classification will be briefly described below.

(1) Method for Assigning Classification Information According to Rules It is assumed that the classification unit 12 holds classification correspondence information on a recording medium (not shown). Here, the classification correspondence information is information having classification information, which is information indicating classification, and word information, which is information indicating phrases, in association with each other. In the classification correspondence information, for example, the classification information “definition question” is associated with the phrase information “what”, “what”, “what”, “what”, “what”, “what”, etc. Also good. For example, the phrase information “why”, “why”, “why”, “why”, “why”, and the like may be associated with the classification information “reason question”. In addition, for example, the word information “how”, “how”, “how”, “how”, “how”, “how”, “how”, etc. are associated with the classification information “method question”. Good. For example, the phrase information “how much”, “how much”, “how much”, and the like may be associated with the classification information “degree question”. For example, the phrase information “change question” may be associated with the phrase information “what is different”, “how is changed”, “where is different”, and the like. Further, for example, the phrase information “what kind of history”, “what kind of circumstances”, “what kind of going” may be associated with the classification information “background question”. Then, when the question information received by the question information receiving unit 11 includes a phrase indicated by the phrase information, the classifying unit 12 assigns classification information corresponding to the phrase information to the question information. For example, when the phrase information “what” is included in the question information, the classification unit 12 can add the classification information “definition question” corresponding to the phrase information “what” to the question information. In addition, when the phrase indicated by the phrase information is included in a specific part of speech of the question information (for example, question pronouns, adjectives, adverbs, etc.), classification information corresponding to the phrase information is given to the question information May be.

  The rule may be information other than the classification correspondence information. For example, when the beginning of the question information is “why” and the rear end is “no?” Or “no?”, The classification unit 12 determines that the question information is the “reason question”. May be. Needless to say, the classification unit 12 may assign the classification to the question information according to other rules.

(2) Method of Assigning Classification Information by Machine Learning The classification unit 12 performs machine learning using the prepared question information and classification information that is information indicating the classification of the question information as teacher data. Using the learning result, classification information can be assigned to the question information received by the question information receiving unit 11. In machine learning, morpheme analysis may be performed on the question information included in the teacher data, and the morpheme obtained by the morpheme analysis may be used as a feature. Note that only the surface layer (character string itself) of the morpheme may be used as the feature, the surface layer and the part of speech may be used, or the surface layer, the part of speech and the utilization form may be used. Further, instead of using morphemes as features, a predetermined number of characters (characters) of question information may be used as features. In this case, any character string may be used as a feature, or a predetermined number of character strings (for example, three consecutive character strings) shifted by one character may be used as a feature, starting from the beginning of a sentence. Only a character string may be used as a feature, or only a character string ending at the end of a sentence may be used as a feature. For example, two or three consecutive characters may be acquired from the question information and used as features.
Various algorithms can be used for machine learning. Details of this algorithm will be described later in the “Explanation on machine learning” column.

  As a morphological analysis system, in the case of Japanese, for example, “ChaSen” (http://chasen.naist.jp) developed at Nara Institute of Science and Technology is known. . In the case of English, examples of software that gives parts of speech to English words include “TnT” (http://www.coli.uni-saarland.de/˜thorsen/tnt/) and “Brill Tagger” ( http://www.cs.jhu.edu/˜brill/) and the like are known. See, for example, the following document for the Brill version.

  Literature: Eric Brill, “Transformation-Based Error-Driving Learning and Natural Language Processing: A Case Study in Part-of-Speech Tagging”, Computational L. 21, no. 4, p. 543-565, 1995

  In addition, here, two cases of assigning a classification using a rule and assigning a class by machine learning have been described. However, if question information can be appropriately classified as a result, otherwise Needless to say, the question information may be classified by the above method.

  The term extracting unit 13 extracts terms from the question information received by the question information receiving unit 11. The terms extracted by the term extraction unit 13 are used in the acquisition of response information by the response information acquisition unit 16 described later. Therefore, it is preferable that the term extraction unit 13 extracts terms that characterize the question information.

  The term extraction unit 13 may perform morphological analysis on the question information, and extract a term that is an independent word from the question information that has been subjected to the morphological analysis. Moreover, the term extraction part 13 may extract the term which is a noun from the question information which carried out the morphological analysis. The term extraction unit 13 may extract a term that is a noun and a term that is a verb from the question information subjected to morphological analysis. Further, the term extraction unit 13 may extract a term that is a noun and a term that is an adjective from the question information subjected to morphological analysis. Further, the term extraction unit 13 may extract a term that is a noun, a term that is a verb, and a term that is an adjective from the question information subjected to morphological analysis. That is, the term extraction unit 13 performs morphological analysis on the question information, and from the question information, (1) independent words, (2) nouns, (3) nouns and verbs, (4) nouns and adjectives, (5) nouns A part of speech term included in any one of (1) to (5) selected from a verb and an adjective may be extracted.

  Further, the term extraction unit 13 may refer to a technical term held in advance in a recording medium (not shown) and extract a term when the term matching the technical term is included in the question information. . The technical terms held in the recording medium (not shown) may be collected manually, for example, from technical dictionaries, economic terms dictionaries, dictionaries of other technical terms, etc. It may be, or may be obtained mechanically from a large corpus. When acquiring technical terms from a corpus mechanically, for example, by extracting morphological analysis of technical literature, single nouns etc. are extracted, scoring indicating the possibility of being technical terms for each single noun etc. is performed, A single noun or the like given a high score may be acquired as a technical term. Here, as a method for assigning a score, a plurality of types such as scoring based on word-making ability and scoring based on appearance frequency are known. Further, other methods may be used. About the method of acquiring a technical term, since it is disclosed by the following some literature etc. and is conventionally known, the detailed description is abbreviate | omitted. Also, TermExtract has been released as a tool for creating a technical term list (http://gensen.dl.itc.u-tokyo.ac.jp/termextract.html).

  Literature: Hiroshi Nakagawa, “Automatic Term Recognition based on Statistics of Compound Nouns”, Terminology, Vol. 6, no. 2, p. 195-210, 2000

  Literature: Hirokazu Ohata, Hiroshi Nakagawa, “Extraction of technical terms by the number of words connected differently”, Information Processing Society of Japan Research Report, 2000-NL-136, p. 119-126

  References: Hiroshi Nakagawa, Masanori Mori, Yasuaki Yumoto, “Extraction of technical terms based on appearance frequency and connection frequency”, natural language processing, Vol. 10 No. 1, p. 27-45, January 2003

  In addition, the term extraction unit 13 may extract only the shortest keyword included in the question information as a term. For example, the term extraction unit 13 may divide the question information into word strings and extract each word (for example, a noun or an unknown word) as a term. Specifically, when “commercial merger” is included in the question information, the terms “corporate” and “merger” are respectively extracted.

  Further, the term extraction unit 13 may extract keywords of all patterns as terms. For example, the term extraction unit 13 divides the question information into word strings, extracts the words (for example, nouns and unknown word words) themselves as terms, and also extracts continuous word strings as terms. Also good. Specifically, when “commercial merger” is included in the question information, the terms “corporate”, “merger”, and “corporate merger” are respectively extracted. In this case, normalization is used to compensate for the difference from the viewpoint that it is unfair that the difference between the term obtained from “Merger” and the term obtained from “Merger” is unfair. You may make it do.

  In addition, the term extraction unit 13 may extract terms using a lattice. For example, the term extraction unit 13 specifies every term as in the case of extracting keywords of all patterns as the term described above, calculates a score for each specified term, and calculates the maximum score. It may be extracted as a term. Specifically, when “commercial merger” is included in the question information, the terms “corporate”, “merger”, and “corporate merger” are respectively specified. Scores are calculated for both “Company”, “Merger”, and “Company Merger”. If the former score is higher, the terms “Company” and “Merger” are extracted, and the latter If the score is high, the term “business merger” is extracted. The formula for calculating the score may be, for example, the TF / IDF formula, the formula (2) in the next document similar to the formula, or any other formula.

  References: Maki Murata, Ma Aoi, Kiyochi Uchimoto, Hiromi Osaku, Masao Uchiyama, Hitoshi Isahara, “Information Retrieval Using Location Information and Field Information”, Natural Language Processing (Journal of the Language Processing Society), Vol. 7, No. 2, p. 141-160, April 2000

  In addition, the term extraction unit 13 may extract terms using down-weighting. For example, the term extracting unit 13 identifies all terms as in the case of extracting keywords of all patterns as terms, and extracts the shortest keyword as it is, and the longer keyword has a smaller weight. After being weighted as such, it is extracted as a term. Specifically, when “commercial merger” is included in the question information, the terms “corporate” and “merger” are extracted as they are, and the term “corporate merger” is extracted after weighting so that the weight is reduced. It will be.

  In addition, about the method of extracting only the shortest keyword as a term, the method of extracting a keyword of any pattern as a term, the method of extracting a term using a lattice, and the method of extracting a term using down-weighting, See the above document "Information Search Using Location Information and Field Information".

The term extraction unit 13 may also extract the focus expression from the question information classified as a definition question by the classification unit 12. The extraction of the focus expression may be performed only when the focus expression is used in the process of acquiring response information by the response information acquisition unit 16 described later. Here, the focus expression is an expression for which the definition is requested in the question information, and can also be called a focus term. For example, when the question information that is the definition question is “What is genetic manipulation?”, The expression that is the target of the definition is “gene manipulation”. Extract the expression “gene manipulation”.
As a method by which the term extraction unit 13 extracts a focus expression, for example, there are (1) a method using a clue phrase and (2) a method using machine learning.

(1) Method of using a clue phrase If the focus expression is “X”, in the question information that is a definition question, expressions such as “What is X”, “X is”, “What is X”, and “What is X” appear. Will come. Accordingly, the term extraction unit 13 holds clue phrases such as “to”, “ha”, “no” and “te” in a recording medium (not shown) in advance, and the question is defined using the clue phrase as a search key. When the question information is searched and hit, the word preceding the clue phrase can be extracted as the focus expression. Note that when performing a search using a clue phrase, the word may be extracted as a focus expression only when the word preceding the clue phrase is a noun or an unknown word.

(2) Method Using Machine Learning The term extraction unit 13 performs machine learning using a prepared focus expression and question information that is a definition question including the focus expression as teacher data, and the result of the machine learning Can be used to extract a focus expression from question information classified as a definition question. At the time of machine learning, morphological analysis may be performed on the question information included in the teacher data and the focus expression, and the morpheme obtained by the morphological analysis may be used as the feature. Note that only the surface layer (character string itself) of the morpheme may be used as the feature, the surface layer and the part of speech may be used, or the surface layer, the part of speech and the utilization form may be used. Further, instead of using morphemes as features, a predetermined number of characters (characters) of question information may be used as features. In this case, any character string may be used as a feature, or a predetermined number of character strings (for example, three consecutive character strings) shifted by one character may be used as a feature, starting from the beginning of a sentence. Only a character string may be used as a feature, or only a character string ending at the end of a sentence may be used as a feature.
Various algorithms can be used for machine learning. Details of this algorithm will be described later in the “Explanation on machine learning” column.

  Here, two cases of extracting a focus expression using a clue phrase and extracting a focus expression by machine learning have been described. However, if a focus expression can be appropriately extracted as a result, Needless to say, the focus expression may be extracted by other methods.

  The terms extracted by the term extraction unit 13 and the focus expression may be temporarily stored in a recording medium (not shown). Terms and focus expressions used in processing performed by an answer information acquisition unit 16 to be described later may be read from the recording medium. Accumulation of terms and focus expressions in the recording medium (not shown) may be performed by the term extraction unit 13 or may be performed by another storage unit (not shown).

  The correspondence information storage unit 14 stores correspondence information. The correspondence information is information having classification information that is information indicating classification and an additional term that is a term added to the term extracted by the term extraction unit 13 in association with each other. This additional term is used together with the term extracted by the term extracting unit 13 in the process of acquiring answer information by the answer information acquiring unit 16 described later. By using this additional term, the answer information can be acquired more appropriately. In the correspondence information, one or more additional terms are associated with one classification information. This correspondence information is appropriately set by the system designer so that higher-performance answer information can be acquired. A specific example of the correspondence information will be described later. Note that the additional terms may not be associated with all the classification information. That is, correspondence information may not be associated with some classification information.

  The process in which the correspondence information is stored in the correspondence information storage unit 14 does not matter. For example, correspondence information may be stored in the correspondence information storage unit 14 via a recording medium, and correspondence information transmitted via a communication line or the like is stored in the correspondence information storage unit 14. Alternatively, the correspondence information input via the input device may be stored in the correspondence information storage unit 14. The storage in the correspondence information storage unit 14 may be temporary storage in the RAM or the like of correspondence information read from an external storage device or the like, or may be long-term storage. The correspondence information storage unit 14 can be realized by a predetermined recording medium (for example, a semiconductor memory, a magnetic disk, an optical disk, or the like).

  The corpus storage unit 15 stores a corpus. This corpus is preferably large-scale. Answer information is acquired from this corpus. This corpus may be, for example, information on newspaper articles, information on encyclopedias, etc., information published on the web, information on academic papers, It may be patent information, and any information can be used as long as it can be the basis for obtaining response information.

  The process of storing the corpus in the corpus storage unit 15 does not matter. For example, a corpus may be stored in the corpus storage unit 15 via a recording medium, or a corpus transmitted via a communication line or the like may be stored in the corpus storage unit 15; Alternatively, the corpus input via the input device may be stored in the corpus storage unit 15. The storage in the corpus storage unit 15 may be temporary storage in a corpus RAM read from an external storage device or the like, or may be long-term storage. The corpus storage unit 15 can be realized by a predetermined recording medium (for example, a semiconductor memory, a magnetic disk, an optical disk, etc.).

  In the present embodiment, the case where the question answering apparatus 1 includes the corpus storage unit 15 will be described. However, the question answering apparatus 1 may not include the corpus storage unit 15. Even when the question answering device 1 does not include the corpus storage unit 15, the question answering device 1 is assumed to be accessible to an external corpus. The external corpus that can be accessed by the question answering apparatus 1 may exist in one place or may be distributed. For example, when the corpus is information published on the web, it can be the latter.

  The response information acquisition unit 16 accesses the additional information associated with the classification information provided by the classification unit 12 using the correspondence information stored in the correspondence information storage unit 14, the term extracted by the term extraction unit 13, and the access By using a corpus stored in the possible corpus storage unit 15 and an expression corresponding to the classification information given by the classification unit 12, response information that is information indicating an answer corresponding to the question information is acquired from the corpus. . The method of acquiring the response information by the response information acquisition unit 16 will be described later with reference to the detailed configuration of the response information acquisition unit 16 shown in FIG. Note that the format of the response information acquired by the response information acquisition unit 16 does not matter. The answer information may be, for example, text data, or data described in a markup language such as HTML or XML. Further, when acquiring a document, the entire corpus may be used, or a part of the corpus may be used. In the latter case, for example, if the corpus is patent information, only the summary may be used.

  The response information output unit 17 outputs the response information acquired by the response information acquisition unit 16. When the answer information output unit 17 outputs the answer information, the user of the question answering apparatus 1 can know the answer corresponding to the question information. Here, the output may be, for example, display on a display device (for example, a CRT or a liquid crystal display), transmission via a communication line to a predetermined device, printing by a printer, or audio output by a speaker. Alternatively, it may be stored in a recording medium or delivered to another component. Note that the answer information output unit 17 may or may not include an output device (for example, a display device or a printer). The answer information output unit 17 may be realized by hardware, or may be realized by software such as a driver that drives these devices.

  FIG. 2 is a block diagram illustrating a configuration of the answer information acquisition unit 16. In FIG. 2, the reply information acquisition unit 16 according to the present embodiment includes a document acquisition unit 21, an additional term acquisition unit 22, a calculation unit 23, and a response information selection unit 24.

  The document acquisition unit 21 acquires, from the corpus, a plurality of documents that are highly characterized by the terms extracted by the term extraction unit 13 using the second expression. Here, the second expression is an expression indicating the degree to which a certain term characterizes a certain document. For example, the second expression may be an expression having a large value if the frequency of the term appearing in one document is high, and a small value if the term appears in many documents. The document acquisition unit 21 may calculate a value such as the appearance frequency using the corpus and the extracted term, and may calculate the value of the second expression using the value. The document acquisition means 21 calculates the value of the second expression for each document by calculating the sum of the terms extracted by the term extraction unit 13 in the second expression, and acquires a document having a higher value. Also good. The second expression may be, for example, a TF · IDF expression or the following expression obtained by improving the expression. In the present embodiment, a case where the second equation is the following equation (1) will be described.

Here, d is a document. A document is a collection of sentences. The document may be, for example, one article when the corpus is newspaper article information, or may be a commentary about one term when the corpus is information such as an encyclopedia. When the corpus is information published on the web, it may be one page, and when the corpus is patent information, it may be one gazette, or It may be one or more paragraphs included. t is a term extracted by the term extraction unit 13 from the question information. tf (d, t) is the appearance frequency (number of appearances) of t in the document d. df (t) is the number of documents in which t appears. N is the total number of documents. length (d) is the length of d. The length of d is indicated by, for example, the number of bytes of the document d, the number of characters, the number of words, and the like. Δ is the average of all document lengths. k _t and k ₊ are constants and values determined by experimental results. For _example, k t, a _{k +,} respectively may be used 0.00001,20. The above equation (1) is based on Robertson's Okapi weighting equation.

Further, the expression of TF · IDF is as follows.
Score (d) = Σ (tf (d, t) * log (N / df (t)))
Where Σ is a sum related to t. Further, tf (d, t), df (t) and the like are the same as described above.

  Even when the document acquisition unit 21 uses the TF / IDF formula, the term t is summed as in the formula (1). Further, in the present embodiment, the case where the summing term t is a term extracted by the term extraction unit 13 will be described. However, the document acquisition unit 21 includes a term extracted from the term extraction unit 13 and a classification unit. You may use the additional term matched with the provided classification | category information by correspondence information as a term (t in said Formula (1) or TF * IDF formula) which takes the sum.

  The document acquisition unit 21 calculates the value of the above equation (1), which is the second equation, for each document d included in the corpus storage unit 15. Then, the document d having a large value of Score (d) is acquired. The document d having a large value of Score (d) may be, for example, the document d having a value of Score (d) that is larger than a threshold value, and from the larger value of Score (d). There may be a predetermined number of documents d selected or a predetermined ratio of documents d. The value larger than the threshold value may or may not include the threshold value. The threshold value may be a preset value, for example, or may be determined according to the calculated Score (d). In the latter case, for example, the threshold value may be a value obtained by multiplying the maximum value of Score (d) by 0.9.

  The document acquired by the document acquisition unit 21 may be temporarily stored in a recording medium (not shown). A document used in processing performed by the calculation unit 23 to be described later may be read from the recording medium. The storage of the document in the recording medium (not shown) may be performed by the document acquisition unit 21 or may be performed by another storage unit (not shown). The document acquisition by the document acquisition unit 21 may be, for example, acquiring document identification information that is information for identifying a document.

  The additional term acquisition unit 22 acquires additional terms corresponding to the classification information given by the classification unit 12 from the correspondence information. The additional term acquisition unit 22 acquires, for example, classification information indicating the classification result by the classification unit 12, searches the correspondence information stored in the correspondence information storage unit 14 using the classification information as a search key, and is searched. By acquiring the additional terms associated with the classification information from the correspondence information storage unit 14, the additional terms can be acquired.

  The acquired additional terms may be temporarily stored in a recording medium (not shown). Additional terms used in processing performed by the calculation unit 23 described later may be read from the recording medium. The accumulation of additional terms in the recording medium (not shown) may be performed by the additional term acquisition unit 22 or may be performed by another storage unit (not shown).

  The calculation unit 23 is information that is a candidate of answer information included in each document acquired by the document acquisition unit 21 using the term extracted by the term extraction unit 13 and the additional term acquired by the additional term acquisition unit 22. For certain answer candidate information, the value of the first equation corresponding to the classification information given by the classification unit 12 is calculated. Here, the first expression is an expression having a higher value as the two terms are closer to each other in the document included in the corpus. The calculation means 23 calculates values such as the closeness of terms and the number of documents in which the terms appear using the extracted terms, additional terms, answer candidate information, corpus, etc. The value of the formula 1 may be calculated. The calculating means 23 may calculate the value of the first expression for each answer candidate information by taking the sum of the terms extracted by the term extracting unit 13 and the additional terms in the first expression. Specifically, the first formula may be a formula using the following formula (2).

Here, d is answer candidate information. The answer candidate information is a predetermined amount of text information included in the document acquired by the document acquisition unit 21. The answer candidate information may be, for example, one paragraph included in the document acquired by the document acquisition unit 21, two consecutive paragraphs, three consecutive paragraphs, one paragraph, It may be a set of two consecutive paragraphs and three consecutive paragraphs, and any combination of consecutive sentences (eg, only the first sentence included in the document, from the first sentence to the second sentence, from the first sentence to the first sentence) 3 sentences, 1st sentence to 4th sentence, ... 2nd sentence only, 2nd sentence to 3rd sentence, 2nd sentence to 4th sentence, 2nd sentence to 5th sentence, .. each of which may be the answer candidate information, etc., and any combination of consecutive phrases (for example, from the first section to the second section, from the first section to the third section included in the document, From the first section to the fourth section, ..., from the second section to the third section Then, each of the second to fourth sections, the second to fifth sections,... Is used as answer candidate information, etc. The answer candidate information may straddle sentences. It may be text information of other quantity. T is a term set including terms extracted from the question information and additional terms. w _dr2 (t2) is, for example, “0.5” if t2 is a verb, and “1” if it is a term of any other part of speech. Note that w _dr2 (t2) may be other complicated settings. dist (t1, t2) is an interval between t1 and t2. The interval may be, for example, the number of characters between t1 and t2, may be the number of bytes of characters included between them, or may be the number of words included between them. For convenience, it is assumed that dist (t1, t2) = 0.5 when t1 = t2. length (d) is the length of d. The length of d is indicated by, for example, the number of bytes of the document d, the number of characters, the number of words, and the like. The second term in the above equation (2) is used to increase the score when the answer candidate information is long. Moreover, T3 used in said Formula (2) is a set of t which satisfy | fills Formula (3).

  When two or more additional terms are acquired, all the two or more additional terms may be added to T, or one of the two or more additional terms acquired is added to T. Then, the value of the above formula (2) is calculated, then the added additional term is deleted, and a new additional term out of the two or more obtained additional terms is added to T, The calculation of the value of equation (2) may be repeatedly executed. Further, the same may be done for all combinations of two or more acquired additional terms.

  Further, the calculation unit 23 calculates the value of the first equation corresponding to the question information classification result by the classification unit 12. For example, when the question information is classified as a degree question, and the number information is included in the answer candidate information, the calculation unit 23 uses a formula that has a larger value than the case where the question candidate information is classified as a degree question. The value of the expression may be calculated. More specifically, the calculation means 23 is 1.1 times the result of the above formula (2) when the question information is classified as a degree question and the answer candidate information includes a numerical expression. This may be calculated as the value of the first equation. Here, the calculation means 23 may determine whether or not a numerical expression exists in the answer candidate information, for example, based on whether or not a number exists in the answer candidate information. The number is generally an Arabic number, but the calculation means 23 also determines whether or not a Chinese number or a Roman number is included, and if the answer candidate information includes a Chinese number or a Roman number, It may be determined that an expression exists.

  Further, for example, when the document acquired by the document acquisition unit 21 includes a focus expression (that is, the focus expression is extracted from the question information by the term extraction unit 13 and the extracted focus is calculated by the calculation unit 23. When the expression is included in the document acquired by the document acquisition unit 21), the value of the first expression may be calculated using an expression that has a larger value than when the focus expression is not included. . More specifically, when the document acquired by the document acquisition unit 21 includes a focus expression, the calculation unit 23 obtains a result obtained by multiplying the result of the above equation (2) by 1.1 by the first equation. A value may be calculated. Further, the calculation unit 23 extracts the focus expression from the question information by the term extraction unit 13 when the focus expression included in the document acquired by the document acquisition unit 21 is modified by the linkage modification clause (ie, In addition, the extracted focus expression is included in the document acquired by the document acquisition means 21, and the focus expression included in the document acquired by the document acquisition means 21 is modified with the combined modifier. In the case of (if present), the value of the first equation may be calculated using an equation that has a larger value than in other cases. More specifically, the calculation unit 23 multiplies the result of the above formula (2) by 1.1 when the focus expression included in the document acquired by the document acquisition unit 21 is modified by the combination modification clause. This may be calculated as the value of the first equation. Here, when the focus expression is included in the document acquired by the document acquisition means 21, the calculation means 23 performs dependency analysis on the document and uses the analysis result to relate to the focus expression. A clause can be specified as a combination modifier clause. Therefore, if there is a linkage modification clause that is a clause related to the focus expression, the calculation means 23 calculates the result of the above formula (2) multiplied by 1.1, It may be calculated as a value. As a system for performing the dependency analysis, for example, “Cabocha” (http://chasen.org/˜take/software/cabocha/) developed at the Nara Institute of Science and Technology is known. By using a system that performs the dependency analysis, it is possible to recognize a phrase and acquire information indicating a dependency relationship between phrases. In addition, although the case where said formula (2) is multiplied by a predetermined value was demonstrated here as a formula according to the provided classification information, the formula according to the provided classification information represents the provided classification. Expressions of completely different formats may be used for each piece of information. In addition, the formula corresponding to the assigned classification information may be held in advance in a recording medium (not shown) for each classification information, or in a recording medium (not shown) The formula (2)) is held, and the calculation unit 23 modifies the basic formula according to the assigned classification information (for example, 1 based on the classification result). (Variation such as 1 times) may be used.

  Note that the value calculated by the calculating means 23 may be temporarily stored in a recording medium (not shown). Values used in processing performed by the response information selection unit 24 described later may be read from the recording medium. The accumulation of values in a recording medium (not shown) may be performed by the calculation means 23 or may be performed by another storage unit (not shown).

  The answer information selecting unit 24 selects answer candidate information having a value calculated by the calculating unit 23 that is larger than others from the plurality of answer candidate information as answer information. The answer information whose value calculated by the calculating means 23 is larger than others may be answer candidate information whose value calculated by the calculating means 23 is larger than a threshold value, for example. Alternatively, the answer candidate information may be a predetermined number or a predetermined ratio of answer candidates selected from the larger values calculated by the calculation unit 23. The value larger than the threshold value may or may not include the threshold value. Further, the threshold value may be a preset value, for example, or may be determined according to the value calculated by the calculation unit 23. In the latter case, for example, the threshold value may be a value obtained by multiplying the maximum value calculated by the calculation means 23 by 0.9.

  Note that when selecting answer information from a plurality of answer candidate information, machine learning may be performed, and the selection may be performed using the result of the machine learning. For example, the answer information selecting unit 24 selects answer candidate information whose value calculated by the calculating unit 23 is larger than others from a plurality of answer candidate information, and from the selected answer candidate information, Answer candidate information may be selected using the result, and the selected answer candidate information may be used as answer information. As described above, the answer candidate information may be one paragraph or a plurality of continuous paragraphs, may be any combination of continuous sentences, may be any combination of continuous phrases, and others. The amount of text information may be sufficient. In the machine learning, machine learning is performed using at least teacher information prepared in advance as question information, answer information indicating an answer to a question indicated by the question information, and information indicating suitability of the answer information. In addition, the machine learning may be performed using the classification information corresponding to the question information included in the teacher data as the teacher data. Further, the value of the first equation corresponding to the answer candidate information may be included in the teacher data. Moreover, you may include the sentence, word, character string, etc. which are contained in parts other than answer candidate information among the documents containing answer candidate information in teacher data. When selecting candidate answer information using the results of machine learning, select candidate answer information using information other than information indicating the suitability of the answer information among the information used in the teacher data. become. For example, when the teacher data includes question information, classification information indicating the classification of the question information, answer information, and information indicating whether the answer is appropriate, the question data is also selected when selecting the answer candidate information. Using the information, the classification information given to the question information, the answer candidate information (information corresponding to the answer information in the teacher data), and the result of machine learning, the answer candidate information is appropriate as the answer information It is judged whether it is or is not appropriate. Also, in selecting answer candidate information using the results of machine learning, answer candidate information whose probability (confidence) value is larger than others is selected as answer information. May be. The answer candidate information whose probability value is larger than others is the same as the case of answer candidate information whose value calculated by the calculation means 23 is larger than others. In machine learning, morphological analysis may be performed on the question information and answer candidate information included in the teacher data, and the morphemes obtained by the morphological analysis may be used as the features. Note that only the surface layer (character string itself) of the morpheme may be used as the feature, the surface layer and the part of speech may be used, or the surface layer, the part of speech and the utilization form may be used. Further, instead of using morphemes as features, a predetermined number of characters (characters) such as question information may be used as features. In this case, any character string may be used as a feature, or a predetermined number of character strings (for example, three consecutive character strings) shifted by one character may be used as a feature, starting from the beginning of a sentence. Only a character string may be used as a feature, or only a character string ending at the end of a sentence may be used as a feature. Various algorithms can be used for machine learning. Details of this algorithm will be described later in the “Explanation on machine learning” column.

  In addition, when the focus expression included in the document acquired by the document acquisition unit 21 is modified with the combination modification clause, the response information selection unit 24 selects the combination modification clause from the answer candidate information as the response information. Also good. For example, the response information selection unit 24 performs dependency analysis on the document acquired by the document acquisition unit 21 and uses the analysis result to determine the clause related to the focus expression. Can be specified as a linkage modification clause. As described above, as a system for performing the dependency analysis, for example, “Cabocha” (http://chasen.org/˜take/software/cabocha/) developed at Nara Institute of Science and Technology is known. ing.

  The answer information selecting means 24 selecting the answer information may be, for example, accumulating the selected answer information itself in a recording medium (not shown), and specifying the selected answer information (for example, storing the answer information). Or a pointer indicating the position of the corpus being recorded) may be stored in a recording medium (not shown), or a flag or the like may be set in association with the selected answer information.

  In addition, the answer information selecting unit 24 selects answer candidate information whose value calculated by the calculating unit 23 is larger than the other from among a plurality of answer candidate information, performs machine learning, and the result of the machine learning May be used to extract the answer information from the selected answer candidate information. In the machine learning, machine learning is performed using at least teacher information prepared in advance as question information, answer information that is information indicating an answer to the question indicated by the question information, and information indicating whether the answer information is appropriate or not. In addition, classification information corresponding to the question information included in the teacher data may be used as teacher data to perform machine learning. Further, the answer candidate information from which the answer information is extracted may be included in the teacher data, and further, the value of the first equation corresponding to the answer candidate information may be included in the teacher data. In machine learning, morpheme analysis may be performed on question information and answer information included in teacher data, and the morpheme obtained by the morpheme analysis may be used as a feature. Note that only the surface layer (character string itself) of the morpheme may be used as the feature, the surface layer and the part of speech may be used, or the surface layer, the part of speech and the utilization form may be used. Further, instead of using morphemes as features, a predetermined number of characters (characters) such as question information may be used as features. In this case, any character string may be used as a feature, or a predetermined number of character strings (for example, three consecutive character strings) shifted by one character may be used as a feature, starting from the beginning of a sentence. Only a character string may be used as a feature, or only a character string ending at the end of a sentence may be used as a feature. Various algorithms can be used for machine learning. Details of this algorithm will be described later in the “Explanation on machine learning” column.

  The answer information selecting means 24 selects answer candidate information whose value calculated by the calculating means 23 is larger than the others from a plurality of answer candidate information, and from the selected answer candidate information, a predetermined rule. Depending on the response information, the answer information may be extracted. The predetermined rule may be, for example, a rule that extracts sentences or paragraphs that contain the largest number of additional terms acquired by the term extraction unit 13 or additional terms acquired by the additional term acquisition unit 22 as answer information. The rule may be that sentences and paragraphs including expressions that match the classification information given to the question information are extracted as answer information. In the latter case, for example, in a recording medium (not shown), classification / response correspondence information, which is information for associating the classification information with the expression that should be included in the response information, is stored, and using that information, Answer information may be extracted. More specifically, in the classification / answer correspondence information, the classification information “reason question” is associated with the expressions “that's why” and “that's why” that should be included in the answer information. When the means 24 selects the answer information corresponding to the question information to which the classification information “reason question” is assigned, the means 24 refers to the classification / answer correspondence information and selects any one of the information corresponding to the classification information “reason question”. A sentence or paragraph including an expression may be selected as answer information.

  In this embodiment, the case where the response information acquisition unit 16 has the configuration shown in FIG. 2 will be described. However, the response information acquisition unit 16 classifies the documents included in the corpus by the classification unit 12. The value of the first equation is calculated using two selected from the additional term associated with the information and the term extracted by the term extracting unit 13, and the value of the first equation is The configuration is not limited as long as the response information that is larger than the others can be obtained. For example, the answer information acquisition unit 16 does not include the document acquisition unit 21, and the calculation unit 23 sets the value of the first expression for each answer candidate information included in the corpus stored in the corpus storage unit 15. Answer candidate information that is calculated and whose value is larger than others may be acquired as answer information.

  The response information acquisition unit 16 further includes a reranking unit (not shown) that reranks the document acquired by the document acquisition unit 21, and the calculation unit 23 replaces the document acquired by the document acquisition unit 21 with the reranking unit. The value of the first equation may be calculated for the answer candidate information included in the document that has been reranked by the above. The higher-ranked document may be, for example, a document whose value calculated at the time of re-ranking is larger than a threshold value, and a predetermined number from the larger calculated value, Alternatively, it may be a document with a predetermined ratio. The value larger than the threshold value may or may not include the threshold value. The threshold value may be a preset value, for example, or may be determined according to a value calculated by the reranking unit. In the latter case, for example, the threshold value may be a value obtained by multiplying the maximum value calculated by the reranking means by 0.9.

Here, the reranking means (not shown) will be described. The reranking unit reranks the document acquired by the document acquisition unit 21 using the following formula.

  In this expression, d is a document acquired by the document acquisition means 21. Further, other dist (t1, t2) and the like are the same as those in the expressions (2) and (3). T may be a set of terms extracted from the question information, or may be a term set including terms extracted from the question information and additional terms. In the latter case, when two or more additional terms are acquired, all of the two or more additional terms may be added to T, or one of the two or more additional terms acquired Is added to T, and the value of the above equation (4) is calculated. Then, the added additional term is deleted, and a new additional term of two or more obtained additional terms is added to T. In addition, the calculation of the value of the above formula (4) may be repeatedly executed. Further, the same may be done for all combinations of two or more acquired additional terms.

  Thus, by providing the re-ranking means (not shown), the documents acquired by the document acquisition means 21 can be further narrowed down. For example, the document acquisition unit 21 acquires 300 documents, and the reranking unit (score calculation using the above formula (4)) by the reranking unit acquires 20 documents from the top, and the calculation unit 23 You may make it use by calculation by. As is clear from the above description, the calculation by the calculation means 23 is a process with a very large load. Therefore, narrowing down the number of documents used in the process in advance reduces the processing load and the process. Useful for reducing time.

  Further, in the present embodiment, the case where the question information receiving unit 11 receives non-factoid type question information will be mainly described. However, the question information receiving unit 11 also receives question information other than non-factoid type, After determining whether it is factoid type question information or other type of question information, if it is non-factoid type question information, it is processed by the method according to the present embodiment. If it is question information, you may make it process by the method of a prior art example. It may be further determined whether the question information other than non-factoid type is factoid type question information or other question information. Such determination may be, for example, a determination based on a rule or a determination based on machine learning. This determination may be performed by a question information determination unit (not shown). The following is a brief description of (1) a method for determining question information based on rules, and (2) a method for determining question information by machine learning, which can be performed by the question information determination unit.

(1) Method for Performing Determination Regarding Question Information Based on Rules Assume that a question information determination unit (not shown) holds determination information, which is information used for determination, in a recording medium (not shown). Here, the determination information is information indicating a rule for determining whether or not the question information is a non-factoid type question. The determination information may be, for example, information having information indicating the type of question information and information indicating words included in the question in association with each other. In the judgment information, for example, information indicating the type “non-factoid type”, information indicating the phrase “what”, “what” “what” “what” “what” “why” “why” “why” "What is the reason""What is the reason""How""How""How""How""How""How""Howmuch""What is different""How it changes""Where is different" “What kind of history”, “what kind of circumstances”, “what kind of way” may be associated with each other. In addition, information “when”, “who”, “where”, “what to read”, etc., indicating the phrase may be associated with the information “factoid type” indicating the type. Then, when the question information received by the question information receiving unit 11 includes information indicating a phrase, the question information determining unit determines that the question information is a type of question corresponding to the information indicating the phrase. May be. If none of the words is included, it may be determined that the question is not a factoid type question or a non-factoid type question, or any other question.

(2) Method for Performing Judgment on Question Information by Machine Learning A question information judgment unit (not shown) performs machine learning using prepared question information and information indicating the type of the question information as teacher data. Using the learning result, it may be determined whether the question information received by the question information receiving unit 11 is a factoid type question, a non-factoid type question, or another question. . In machine learning, morpheme analysis may be performed on the question information included in the teacher data, and the morpheme obtained by the morpheme analysis may be used as a feature. Note that only the surface layer (character string itself) of the morpheme may be used as the feature, the surface layer and the part of speech may be used, or the surface layer, the part of speech and the utilization form may be used. Further, instead of using morphemes as features, a predetermined number of characters (characters) of question information may be used as features. In this case, any character string may be used as a feature, or a predetermined number of character strings (for example, three consecutive character strings) shifted by one character may be used as a feature, starting from the beginning of a sentence. Only a character string may be used as a feature, or only a character string ending at the end of a sentence may be used as a feature. For example, two or three consecutive characters may be acquired from the question information and used as features.
Various algorithms can be used for machine learning. Details of this algorithm will be described later in the “Explanation on machine learning” column.

  Here, as described above, the question information determined by the question information determination unit (not shown) as a non-factoid type question is processed by the question answering apparatus 1 according to the present embodiment, that is, by the classification unit 12. Classification and term extraction processing by the term extraction unit 13 are performed. On the other hand, for question information that is determined to be a factoid type question, answer information may be acquired using a conventional method. In addition, as a method for obtaining answer information corresponding to factoid type question information, for example, the following method may be used.

[Method of extracting answer information corresponding to factoid type question information]
First, it is estimated what kind of solution expression the answer information corresponding to the question information has. For example, whether the response information is “person name”, “location”, “time expression”, “numerical expression”, “country name”, “capital name”, Estimate whether it becomes “Hiragana expression”. This estimation may be a rule-based method or a method using machine learning. In the case of the former, for example, information indicating the phrase is included in the question information by using information having information indicating the type of solution expression of the answer information and information indicating the phrase included in the question information. When it is included, it may be determined that the answer information is an answer expression corresponding to the information indicating the phrase. Specifically, the solution expression “person name” is associated with the word “who”, the solution expression “time expression” is associated with the word “when”, the solution expression “capital name”, The phrase “where is the capital” may be associated. In the latter case, for example, machine learning is performed using, as teacher data, question information prepared in advance and information indicating the type of solution expression of answer information corresponding to the question information, and the result of the machine learning is used. Thus, the answer expression type of the answer information corresponding to the question information may be determined. In machine learning, morpheme analysis may be performed on the question information included in the teacher data, and the morpheme obtained by the morpheme analysis may be used as a feature. Note that only the surface layer (character string itself) of the morpheme may be used as the feature, the surface layer and the part of speech may be used, or the surface layer, the part of speech and the utilization form may be used. Further, instead of using morphemes as features, a predetermined number of characters (characters) of question information may be used as features. In this case, any character string may be used as a feature, or a predetermined number of character strings (for example, three consecutive character strings) shifted by one character may be used as a feature, starting from the beginning of a sentence. Only a character string may be used as a feature, or only a character string ending at the end of a sentence may be used as a feature. For example, two or three consecutive characters may be acquired from the question information and used as features. Various algorithms can be used for machine learning. Details of this algorithm will be described later in the “Explanation on machine learning” column.

  Next, terms are extracted from the question information. The term extraction can be performed in the same manner as the term extraction unit 13. Then, similarly to the document acquisition unit 21, a document is extracted from the corpus using Expression (1). By extracting this document, a group of documents that are likely to be written is collected. For example, if the question information is “Where is the capital of Japan?”, For example, “Japan” and “capital” are extracted as terms, and a document group including them is acquired. Thereafter, processing similar to the above-described reranking is performed. Note that the reranking process is a process for narrowing down documents, and thus may not be performed.

  Next, a process of extracting a solution from the acquired document or the document ranked higher in the reranking is performed. More specifically, nouns and unknown word sequences are extracted from these documents and set as candidate solutions. Unlike the non-factoid type question described above, in the case of a factoid type question, the answer is a noun or a series of nouns, and thus the noun or the like may be extracted as a solution candidate. Then, Score (c) is calculated for the solution candidate “c”, and the answer with the larger value is selected and output. Score (c) is as follows.

Score (c) = Score _near (c) + Score _sem (c)
Score _near (c) is a score based on the proximity of the solution candidate and the keyword, and Score _sem (c) is a score based on whether the semantic constraint of the solution expression is satisfied. Score _near (c) is given by the following equation.

w _dr2 (t2) is a function determined by experiment. For example, it may be “0.5” if t2 is a verb, and “1” if it is a term of other parts of speech. .

Score _sem (c) gives a positive score if the type of solution representation of the candidate solution matches the estimated type of solution representation, otherwise gives no score, or It is a function that gives a negative score. For example, a score (for example, 1000) may be given to a solution candidate that matches the estimated type of solution expression (for example, a person name, place name, etc.).

  Whether the type of the solution expression of the solution candidate matches the estimated type of the solution expression may be determined by a rule or may be determined by machine learning. In the former case, for example, a country name dictionary, a personal name dictionary, a capital name dictionary, etc. are held in advance, and a candidate solution is determined by determining which dictionary the candidate solution is described in. The type of solution expression may be determined. In the latter case, that is, in the case of machine learning, machine learning is performed using the prepared solution and information indicating the type of solution expression of the solution as teacher data, and the result of the machine learning is used. Thus, the type of solution expression corresponding to the solution candidate may be determined. This machine learning method is, for example, a method known as a specific expression extraction technique using machine learning.

  In this way, it is possible to output answer information for factoid type question information. Similar processing may be performed in acquisition / output of answer information corresponding to factoid type question information and acquisition / output of answer information corresponding to non-factoid type question information. For example, extraction of terms from question information, acquisition of documents from a corpus, and the like. Accordingly, in the acquisition / output of answer information corresponding to factoid type question information and the acquisition / output of answer information corresponding to non-factoid type question information, common components (for example, the term extraction unit 13 and document acquisition) The processing may be performed using the means 21).

For extraction of answer information corresponding to factoid type question information, refer to the following document.
References: Maki Murata, Masao Uchiyama, Tamotsu Shirato, Hitoshi Isahara, “A Diminishing Addition Question Answering System Using Simple Combination Method for Series-type Question Sentences”, Journal of System Control Information Society, Vol. 20, no. 8, p. 18-26, 2007

  Of the correspondence information storage unit 14, the corpus storage unit 15, and the recording medium storing other information, any two or more storage units or recording media may be realized by the same recording medium. Alternatively, it may be realized by a separate recording medium. In the former case, for example, the area storing the correspondence information is the correspondence information storage unit 14, and the area storing the corpus is the corpus storage unit 15.

  In addition, a corpus, a term, and an additional term should be strictly described as information indicating a corpus, information indicating a term, and information indicating an additional term. However, for convenience of explanation, the corpus, term, It will be called additional terms.

Next, operation | movement of the question answering apparatus 1 by this Embodiment is demonstrated using the flowchart of FIG.
(Step S101) The question information receiving unit 11 determines whether question information has been received. If accepted, the process proceeds to step S102. If not, the process of step S101 is repeated until accepted.

  (Step S102) The classification unit 12 gives classification information to the question information received by the question information reception unit 11.

  (Step S103) The term extraction unit 13 extracts terms from the question information received by the question information receiving unit 11. Details of this processing will be described later with reference to the flowchart of FIG.

  (Step S104) The reply information acquisition unit 16 associates the classification information given by the classification unit 12, the terms extracted by the term extraction unit 13, and the classification information given by the classification unit 12 with correspondence information. The answer information indicating the answer corresponding to the question information received by the question information receiving unit 11 is acquired using the additional term and the corpus stored in the corpus storage unit 15. Details of this processing will be described later with reference to the flowchart of FIG.

(Step S105) The response information output unit 17 outputs the response information acquired by the response information acquisition unit 16. Then, the process returns to step S101.
In the flowchart of FIG. 3, the process ends when the power is turned off or the process is terminated.

  Next, FIG. 4 is a flowchart showing details of the process of extracting terms in the flowchart of FIG. 3 (the process of step S103). In the flowchart of FIG. 4, a case will be described in which the term extraction unit 13 extracts a focus expression as well as a term. In the flowchart of FIG. 4, a case where the focus expression is extracted using a clue phrase will be described.

  (Step S201) The term extracting unit 13 performs morphological analysis on the question information received by the question information receiving unit 11.

  (Step S <b> 202) The term extracting unit 13 extracts a specific part-of-speech term using a result of morphological analysis.

  (Step S203) The term extraction unit 13 temporarily stores the extracted specific part-of-speech term in a recording medium (not shown).

  (Step S <b> 204) The term extraction unit 13 determines whether or not the classification information “definition question” has been given by the classification unit 12 to the question information received by the question information reception unit 11. If the classification information “definition question” is given, the process proceeds to step S205, and if not, the process returns to the flowchart of FIG.

  (Step S205) The term extraction unit 13 reads a clue phrase that is stored in advance on a recording medium (not shown) and is used to extract the focus expression, and adds the question information to which the classification information “definition question” is assigned. Determine if a clue phrase is included. If it is included, the process proceeds to step S206. If not, the process returns to the flowchart of FIG.

  (Step S206) The term extraction unit 13 extracts words having a predetermined relationship with the clue phrase included in the question information. This extracted word is a focus expression.

  (Step S207) The term extraction unit 13 temporarily stores the extracted focus expression in a recording medium (not shown). And it returns to the flowchart of FIG.

  In the flowchart of FIG. 4, it goes without saying that the focus expression may be extracted using machine learning or other methods as described above. Further, in the later process of obtaining the answer information, when the focus expression is not used, the focus expression extraction process may not be performed.

  FIG. 5 is a flowchart showing details of the process of obtaining response information (the process of step S104) in the flowchart of FIG. In the flowchart of FIG. 5, processing when the answer information acquisition unit 16 has the configuration of FIG. 2 will be described.

  (Step S <b> 301) The document acquisition unit 21 stores a plurality of documents characterized by the terms extracted by the term extraction unit 13 in the corpus storage unit 15 using the above-described second formula. Get from the corpus you have. Details of this process will be described later with reference to the flowchart of FIG.

  (Step S302) The additional term acquisition means 22 acquires one or more additional terms corresponding to the classification information given by the classification unit 12 from the correspondence information stored in the correspondence information storage unit 14.

  (Step S303) The calculation unit 23 uses the terms extracted by the term extraction unit 13 and the additional terms acquired by the additional term acquisition unit 22 to determine the answer candidate information included in each document acquired by the document acquisition unit 21. The value of the first equation corresponding to the classification information given by the classification unit 12 is calculated. Details of this processing will be described later with reference to the flowchart of FIG.

  (Step S304) The answer information selection means 24 selects answer information from a plurality of answer candidate information whose values calculated by the calculation means 23 are larger than others. And it returns to the flowchart of FIG. Details of this process will be described later with reference to the flowchart of FIG.

  FIG. 6 is a flowchart showing details of the document acquisition processing (step S301 processing) in the flowchart of FIG. In the flowchart of FIG. 6, a case will be described in which the document acquisition unit 21 acquires a document using the above equation (1) as the second equation.

(Step S401) The document acquisition unit 21 sets the counter i to 1.
(Step S402) The document acquisition unit 21 acquires the i-th document from the corpus stored in the corpus storage unit 15.

  (Step S403) The document acquisition means 21 uses the term extracted by the term extraction unit 13, the above equation (1), and the corpus stored in the corpus storage unit 15, and the value of the second equation, That is, the value of equation (1) is calculated.

(Step S404) The document acquisition means 21 temporarily stores the calculated value of the expression (1) in a recording medium (not shown). In this storage, it is preferable to store information for identifying the document for which the value is calculated in association with the value of the expression (1). The information for identifying the document may be, for example, a counter value, a pointer indicating the position of the document stored in the corpus storage unit 15, or other document identification information. May be.
(Step S405) The document acquisition means 21 increments the counter i by 1.

  (Step S406) The document acquisition unit 21 determines whether or not the i-th document exists in the corpus stored in the corpus storage unit 15. And when it exists, it returns to step S402, and when that is not right, it progresses to step S407.

  (Step S407) The document acquisition unit 21 sorts the values of the formula (1) temporarily stored in Step S404.

  (Step S408) The document acquisition unit 21 selects a plurality of documents in which the value of the expression (1) is larger than the others in the sorting result. And it returns to the flowchart of FIG. Note that, as described above, the document acquisition unit 21 may select a document whose value of the expression (1) is equal to or greater than the threshold value. A document may be selected. Further, as described above, the selection of the document may be acquisition of information for identifying the document, or acquisition of the document information itself from the corpus storage unit 15.

  FIG. 7 is a flowchart showing details of the calculation processing of the first equation (processing in step S303) in the flowchart of FIG. In the flowchart of FIG. 7, a case where the value of the first equation is calculated using the above-described equation (2) will be described. Further, when the classification information given to the question information is a “degree question” and the answer candidate information has a numerical expression, the value of the expression (2) is multiplied by 1.1. Further, when the classification information given to the question information is a “definition question” and the answer candidate information includes a focus expression, the value of Expression (2) is multiplied by 1.1. In addition, when the classification information given to the question information is a “definition question”, the answer candidate information includes a focus expression, and the focus candidate expression is modified with a combination modification clause in the answer candidate information Is assumed to be 1.1 times the value of equation (2).

(Step S501) The calculation means 23 sets the counter i to 1.
(Step S502) The calculation means 23 sets the counter j to 1.

  (Step S503) The calculation unit 23 specifies j-th answer candidate information in the i-th document acquired by the document acquisition unit 21. For example, the calculation unit 23 may specify the j-th answer candidate information by acquiring the pointer, and the pointer indicating the position where the j-th answer candidate information is stored. As long as the answer candidate information specified in the subsequent process can be used as a result, the specific method may be used.

  (Step S504) The calculating unit 23 calculates the term extracted by the term extracting unit 13, the additional term acquired by the additional term acquiring unit 22, the answer candidate information specified in Step S503, and the above formulas (2) and (3 ) And the corpus stored in the corpus storage unit 15, the value of equation (2) is calculated. When the classification information given to the question information is not “degree question” or “definition question”, the value of the equation (2) is the value of the first equation.

  (Step S505) The calculating means 23 temporarily stores the calculated value of the formula (2) in a recording medium (not shown). In this storage, it is preferable to store information for identifying the answer candidate information for which the value is calculated in association with the value of the expression (2). The information for identifying the answer candidate information may be, for example, the values of the counters i and j, and includes information for identifying a document including the answer candidate information and a pointer indicating the position of the answer candidate information in the document. May be the answer candidate information itself, or may be other identification information.

  (Step S506) The calculation unit 23 determines whether the classification information given to the question information received by the question information receiving unit 11 is a degree question. And when it is a degree question, it progresses to Step S507, and when that is not right, it progresses to Step S513.

  (Step S507) The calculation unit 23 determines whether or not the number expression is included in the jth answer candidate information in the i-th document. If a numerical expression is included, the process proceeds to step S508, and if not, the process proceeds to step S509.

(Step S508) The calculation means 23 multiplies the value of the expression (2) temporarily stored in Step S505 by 1.1 and accumulates it by overwriting. The value after overwriting becomes the value of the first expression.
(Step S509) The calculation means 23 increments the counter j by 1.

(Step S510) The calculation means 23 determines whether or not the jth answer candidate information exists in the ith document. And when it exists, it returns to step S503, and when it does not exist, it progresses to step S511.
(Step S511) The calculation means 23 increments the counter i by 1.

  (Step S512) The calculation unit 23 determines whether or not the i-th document acquired by the document acquisition unit 21 exists. And when it exists, it returns to step S502, and when it does not exist, it returns to the flowchart of FIG.

  (Step S513) The calculation unit 23 determines whether the classification information given to the question information received by the question information receiving unit 11 is a definition question. If it is a definition question, the process proceeds to step S514, and if not, the process proceeds to step S509.

  (Step S514) The calculation means 23 determines whether or not the focus expression extracted by the term extraction unit 13 exists in the jth answer candidate information in the ith document. If it exists, the process proceeds to step S515. If it does not exist (including the case where the focus expression is not extracted), the process proceeds to step S509.

  (Step S515) The calculation means 23 multiplies the value of the expression (2) temporarily stored in step S505 by 1.1 and accumulates it by overwriting. If the value is not overwritten after that, the value after the overwriting becomes the value of the first expression.

  (Step S516) The calculation unit 23 determines whether or not the focus expression included in the jth answer candidate information in the i-th document is modified by the combination modification clause. If it is modified by the linkage modification clause, the process proceeds to step S517, and if not, the process proceeds to step S509.

  (Step S517) The calculation unit 23 further multiplies the value accumulated by overwriting in step S516 by 1.1 and accumulates it by overwriting. The value after overwriting becomes the value of the first expression.

  (Step S518) The calculation unit 23 replaces the jth answer candidate information in the i-th document with a linkage modification clause that modifies the focus expression. Then, the process proceeds to step S509.

  Note that the processing in step S518 in the flowchart of FIG. 7 may be performed by the calculation unit 23 or may be performed by the answer information selection unit 24. Further, in the flowchart of FIG. 7, the case where the value of the first expression is changed from the expression (2) is described only when the question information is classified into “degree question” and “definition question”. In the case where is classified into other classifications, the value of the first expression may be changed from Expression (2). Moreover, although the case where the degree of the change is “1.1 times” has been described, this need not be the case.

  FIG. 8 is a flowchart showing details of the answer information selection process (the process of step S304) in the flowchart of FIG. In the flowchart of FIG. 8, it is assumed that the answer information selection unit 24 performs a process of specifying the answer information included in the selected answer candidate information along with the selection of the answer candidate information. The answer information is specified using machine learning. It is assumed that machine learning processing is performed in advance before the processing from step S601 is executed.

  (Step S601) The answer information selecting unit 24 sorts the answer candidate information using the value calculated by the calculating unit 23. For example, the answer information selecting unit 24 sorts the answer candidate information so that the values corresponding to the answer candidate information (values calculated by the calculating unit 23) are in descending order.

  (Step S602) The answer information selection means 24 selects one or more answer candidate information whose values calculated by the calculation means 23 are larger than others in the sorting result. Note that, as described above, the answer information selecting unit 24 may select answer candidate information whose value calculated by the calculating unit 23 is equal to or greater than a threshold value, and a predetermined number from the larger value calculated by the calculating unit 23. Alternatively, a predetermined percentage of answer candidate information may be selected.

(Step S603) The answer information selection unit 24 sets the counter i to 1.
(Step S604) The answer information selection unit 24 sets the counter j to 1.

  (Step S605) The answer information selection unit 24 specifies the j-th part in the i-th answer candidate information. The unit of the specified part may be determined in advance. For example, when the specified part is a sentence unit, the answer information selecting unit 24 specifies the jth sentence in the i-th answer candidate information. For example, when the specified part is a paragraph unit, the answer information selecting unit 24 specifies the jth paragraph in the i-th answer candidate information.

  (Step S606) The answer information selecting unit 24 determines whether or not the part specified in Step S605 is answer information using the result of machine learning. If it is determined that the information is answer information, the process proceeds to step S607. If not, the process proceeds to step S608.

(Step S607) The response information selection unit 24 temporarily stores the portion determined to be response information as response information in a recording medium (not shown).
(Step S608) The answer information selection unit 24 increments the counter j by 1.

(Step S609) The answer information selecting unit 24 determines whether or not the j-th part exists in the i-th answer candidate information. And when it exists, it returns to step S605, and when it does not exist, it progresses to step S610.
(Step S610) The answer information selection means 24 increments the counter i by 1.

  (Step S611) The answer information selecting unit 24 determines whether or not the i-th answer candidate information exists in the answer candidate information whose value is calculated by the calculating unit 23. If it exists, the process returns to step S604, and if not, the process returns to the flowchart of FIG.

  In the flowchart of FIG. 8, the case where the selected answer candidate information portion is used as the answer information has been described, but the process may not be performed. For example, the answer candidate information selected in step S602 may be used as the answer information. Also, answer candidate information is selected from the answer candidate information selected in step S602 or the answer candidate information that is not subjected to the processing in step S602 using the result of machine learning, and the selected answer candidate information is used as the answer information. It is good. In that case, without using the counter j (that is, without specifying the j-th part in the i-th answer candidate information), the i-th answer candidate information is the answer information using the machine learning result. It may be determined whether or not. Further, in the flowchart of FIG. 8, the processing of only selection of answer information using the result of machine learning (steps S603 to S611) may be performed without performing the processing of steps S601 and S602. Even if the process of step S602 is not performed in the flowchart of FIG. 8, the answer information acquisition unit 16 indirectly uses the value of the first equation as teacher data during machine learning. The answer information is acquired using the additional term, the term extracted by the term extracting unit 13, and the formula (first formula) corresponding to the classification information. In the flowchart of FIG. 7, when the answer candidate information is replaced with a combination modification clause that modifies the focus expression (when the process of step S518 is executed), the answer candidate information itself is specified in step S605. You may make it do.

  Moreover, although the case where one or more answer information is selected has been described in the flowchart of FIG. 8, one answer information may be selected. For example, in step S602, only one answer candidate information may be selected, or in the determination using the result of machine learning in step S606, the answer information having the highest certainty (confidence level) is selected. You may make it select.

Next, the operation of the question answering apparatus 1 according to the present embodiment will be described using a specific example.
In this specific example, it is assumed that the question answering apparatus 1 according to the present embodiment is a stand-alone PC (Personal Computer).

  In this specific example, the correspondence information storage unit 14 stores the correspondence information shown in FIG. In the correspondence information in FIG. 9, classification information and additional terms are associated with each other. In this specific example, the classification unit 12 classifies the question information into six types of definition questions, reason questions, method questions, degree questions, change questions, and history questions, as shown in FIG. There may be classification information that does not correspond to the additional terms. In this specific example, the question information is classified using machine learning.

  First, it is assumed that the user operates the mouse, keyboard, etc. of the question answering apparatus 1 to display the question input screen as shown in FIG. Then, it is assumed that the user inputs the question information “How is World Heritage determined?” By operating a keyboard or the like, and clicks the “OK” button. Then, the question information reception unit 11 receives the question information “How is World Heritage determined?” (Step S101), and passes the question information to the classification unit 12 and the term extraction unit 13.

  It is assumed that the classification unit 12 performs machine learning related to classification in advance. When the question information “How is World Heritage determined?” Is received from the question information receiving unit 11, the question information is classified using the learning result. In this case, it is assumed that the classification unit 12 assigns the classification information “method question” to the question information (step S102). The classification unit 12 passes the classification result (assigned classification information) to the term extraction unit 13 and the answer information acquisition unit 16.

  When the term extraction unit 13 receives the question information from the question information reception unit 11 and receives the classification information from the classification unit 12, the term extraction unit 13 executes a process of extracting a term (step S103). Specifically, the term extraction unit 13 performs morphological analysis on the question information “How is World Heritage determined?” (Step S201), and extracts parts of speech that are predetermined as extraction targets. (Step S202). Here, it is assumed that nouns and verbs are extracted. Then, the term extraction unit 13 extracts the noun “world heritage” and the verb “determined” from the result of the morphological analysis (step S202), and temporarily stores them in a recording medium (not shown) (step S203). In this specific example, the case of extracting the compound noun “world heritage” will be described, but the noun “world” and the noun “world heritage” may be extracted. The noun “world” and the noun “heritage” may be extracted, and there are various methods by which the term extracting unit 13 extracts terms. In addition, as a method of extracting compound nouns, for example, a method of extracting consecutive nouns as compound nouns may be used, and a method of extracting using human rules or machine learning, as in the above-described proper expression extraction. You may use, the method of extracting using TermExtract mentioned above may be used, and the method will not be limited if compound noun can be extracted.

  The term extraction unit 13 determines whether the classification information received from the classification unit 12 is a “definition question”. In this case, since the classification information is a “method question”, the term extraction unit 13 determines that the classification information is not a “definition question” (step S204), and does not perform a process of extracting a focus expression thereafter. .

  Thereafter, the response information acquisition unit 16 executes a process of acquiring response information (step S104). In the process of acquiring the answer information, first, the document acquisition unit 21 executes a process of acquiring a document from the corpus storage unit 15 (step S301). Specifically, the document acquisition unit 21 reads the first document stored in the corpus storage unit 15 (steps S401 and S402). Here, it is assumed that the document ID for identifying the first document is “D001”. Next, the document acquisition unit 21 reads the formula (1) from a recording medium (not shown) in which the above formula (1) is stored, and the terms “world heritage” and “temporarily” stored by the term extraction unit 13. “Determined” is read out and the value of equation (1) is calculated using them (step S403). Here, it is assumed that the value of Expression (1) is “2.3”. Then, the document acquisition unit 21 temporarily stores the value “2.3” of the expression (1) in association with the document ID “D001” on a recording medium (not shown) (step S404). The first record in FIG. 11 is stored in this way. The document acquisition unit 21 sequentially executes the same processing for the second document, the third document,... (Steps S405, S406, S402 to S404).

  After calculating the value of equation (1) for all documents, the document acquisition means 21 sorts the calculated values in descending order (step S407). Then, in the sorted result, 300 document IDs having the largest value are acquired and stored in a recording medium (not shown) (step S408). In this way, the document acquisition unit 21 acquires the document. Here, it is assumed that document IDs “D002”, “D003”... Are selected.

  Next, the additional term acquisition unit 22 searches the correspondence information shown in FIG. 9 using the classification information “method question” received from the classification unit 12 as a key, and associates it with the searched “method question”. The additional terms “method”, “procedure” and “by” are read and stored in a recording medium (not shown) (step S302).

  After the acquisition of the document by the document acquisition unit 21 and the acquisition of the additional term by the additional term acquisition unit 22, the calculation unit 23 executes a process of calculating the value of the first expression (step S303). Here, it is assumed that the calculation unit 23 calculates the value of the first expression using one paragraph as answer candidate information in the document acquired by the document acquisition unit 21. Further, it is assumed that the document identified by the document ID “D002” acquired by the document acquisition unit 21 is the one shown in FIG.

  Then, the calculation means 23 first sets the first document, that is, the first paragraph of the document shown in FIG. 12, that is, “World Heritage has ...” as the answer candidate information. Specify (steps S501 to S503). Then, the calculation unit 23 reads the terms “world heritage” “determined” accumulated by the term extraction unit 13 and the additional terms “method”, “procedure”, and “according” accumulated by the additional term acquisition unit 22, and is not illustrated. The equation (3) stored in the recording medium is also read, and a set of T3 is calculated using these and the corpus. Then, using the calculated T3 set and the terms “world heritage”, “determined”, additional terms “method”, “procedure”, “by”, corpus, equation (2) read from a recording medium (not shown), and the like, The value of (2) is calculated (step S504). Here, it is assumed that 2.6. Then, the calculation unit 23 stores the value in a recording medium (not shown) in association with information for identifying the answer candidate information (step S505). The first record in FIG. 13 is information having the score of the first expression accumulated in this way and the identification information of the answer candidate information in association with each other. As identification information of the answer candidate information, a document ID and a paragraph number in the document are used. The paragraph number is a value indicating the number of paragraphs counted from the top of the document.

  In this case, since the classification information is a “method question” and not a “degree question” or a “definition question”, the identification of the answer candidate information, the calculation of the value of equation (2), and the accumulation thereof are performed. The steps are sequentially performed (steps S509, S510, S503 to S505). When the first document is completed, the value of the expression (2) for each paragraph included in the document is similarly calculated and accumulated for the second, third,. (Steps S502 to S505, S509 to S512).

  When the series of processes for calculating the value of the first expression is completed, the answer information selecting unit 24 executes a process of selecting answer information (step S304). Specifically, the answer information selection unit 24 reads out the calculation result shown in FIG. 13 from a recording medium (not shown), and sorts the records so that the scores in the first formula are in descending order (step S601). Then, the answer candidate information having a score of the first expression equal to or larger than a value (= 8.73) obtained by multiplying the maximum value (here, 9.7) of the score of the first expression by 0.9. Is selected (step S602). Here, it is assumed that the selected answer candidate information is only the answer candidate information identified by the document ID “D002” and the paragraph number “3”. In addition, the answer information selection unit 24 determines, for each sentence included in the answer candidate information, whether or not the sentence is appropriate as the answer information by machine learning. It is assumed that machine learning for the determination is performed in advance.

  Then, the answer information selecting unit 24 specifies the first sentence from the paragraph “Registering procedure as a registration procedure ...” shown in FIG. 12, which is the selected first answer candidate information (step S603). S605). Then, the response information selection unit 24 determines whether the specified sentence “As for the registration procedure, ... an expert group evaluates” is the response information, using the result of machine learning. Here, it is assumed that the sentence is determined to be answer information (step S606). Then, the answer information selection means 24 stores the sentence as answer information in a recording medium (not shown) (step S607), and the next second sentence “and determines based on the evaluation result”. Is determined and it is determined whether or not the sentence is answer information (steps S608, S609, S605, and S606). In this case, it is assumed that it is not answer information. Then, a series of processes for selecting the answer information ends.

  Thereafter, the answer information output unit 17 reads the answer information accumulated by the answer information selection unit 24 from a recording medium (not shown) and displays the answer information on the display of the question answering apparatus 1 (step S105). FIG. 14 is a diagram showing the answer information displayed as described above. The user who has input the question can know the answer to the question by viewing this display.

When the answer information is displayed on the display of the question answering apparatus 1, the entire document including the answer information is displayed on the display, and the answer information can be distinguished from the document other than the answer information. You may display as follows. Displaying in a distinguishable manner may be, for example, changing the display color of the response information and the portion of the document other than the response information, and changing the display font size and font type. It may be possible to add an underline only to the part of the response information, or to frame or shade only the part of the response information. The location may be displayed thinner than the answer information.
The numerical values used in this specific example are values set for explanation, and are not calculated by analyzing actual data.

  Further, in this specific example, the case of performing classification by machine learning has been described, but it goes without saying that classification may be performed using classification correspondence information as described above. The classification information used in that case may be, for example, that shown in FIG.

[Experimental example]
Next, experimental examples will be described. In this experiment, 100 pieces of non-factoid type question information were used. The question information is generated by the QAC-4 organizer and is a natural question generated without using the target document. In addition, the organizer of QAC-4 evaluated 4 or less answer information for each question information. The results are as shown in FIG. Method 1 is a method using one paragraph as answer candidate information. Method 2 is a method using one paragraph, two consecutive paragraphs, and three consecutive paragraphs as answer candidate information. A, B, C, and D are evaluation criteria. A is a case where the answer information describes the same content as the answer prepared by the organizer for the question information. Even if additional contents were included, those whose contents did not change were evaluated as A. B is a case where the answer information includes contents similar to the answer prepared by the organizer for the question information, but also includes contents that are different as a whole. C is a case where the answer information includes a part of the same content as the answer prepared by the organizer for the question information. D is a case where the answer information does not include the same content as the answer prepared by the organizer for the question information. In the table of FIG. 16, the values of A, B, C, and D are the number of pieces of question information whose answer information belongs to A, B, C, and D. The “correct answer” is the number of question information when the answer information belongs to any one of A, B, and C. This evaluation standard for correct answers was officially used in NTCIR-6 QAC-4.

From the results shown in FIG. 16, the following can be understood.
Method 1 has a higher score in evaluation of A than Method 2. From this, it can be seen that Method 1 extracts the complete correct answer more accurately than Method 2.

  In the evaluation of “correct answer”, the correct answer rate of Method 2 is 0.77, and Method 2 has a higher score than Method 1. From this, it can be seen that Method 2 tends to extract more partial correct answers than Method 1. Therefore, when it is desired to extract a complete correct answer, the method 1 may be used, and when more correct answers (including partial correct answers) are desired, the method 2 may be used.

  Of all 8 teams participating in NTCIR-6 QAC-4, the correct answer rate “0.77” of Method 2 using the question answering apparatus 1 according to the present embodiment was the highest value. From this, it can be seen that the question answering apparatus 1 according to the present embodiment was able to acquire the answer information most effectively as compared with the question answering apparatuses of other participating teams.

  Here, an example of the answer information actually acquired by the question answering apparatus 1 according to the present embodiment will be briefly described. In this example, Method 1 is used to evaluate A.

Question information: When is a fertilized egg diagnosis performed?
Answer information: Preventing mainly children with genetic illness QAC-4 organizers prepared example 1: Mainly genetic illness QAC-4 organizers prepared answer 2: Mainly born with genetically ill children When the possibility is high

  In addition, in the question answering apparatus 1 used when obtaining the result of this experimental example, the answer information using the formulas (1) to (3) was acquired. Then, 300 documents are obtained using the formula (1), and the answer has a value equal to or greater than a value obtained by multiplying the maximum value of the formula (2) calculated for the answer candidate information included in the document by 0.9. Candidate information was acquired and output as answer information. Therefore, the process of acquiring the answer information from the answer candidate information using the machine learning described in steps S604 to S609 in the flowchart of FIG. 8 is not performed.

  As described above, according to the question answering apparatus 1 according to the present embodiment, it is possible to appropriately extract and output answer information corresponding to non-factoid type question information. Further, as shown in the results of the experimental example, the method of the question answering apparatus 1 according to the present embodiment is a technique that can obtain the highest accuracy rate, and the question answering apparatus 1 according to the present embodiment is another conventional method. It can be seen that the performance is higher than that of the question answering apparatus.

  In the present embodiment, the question information, the answer information, and the language of the corpus are not limited. For example, Japanese, English, French, Russian, Chinese, Spanish, etc. may be used. Moreover, the language of question information and corpus may differ. For example, the question information is in Japanese and the corpus is in English. In that case, for example, the processing may be performed after the question information is translated into the language of the corpus, or the processing may be performed after the corpus is translated into the language of the question information. In general, the former is preferred because it is considered that the amount of translation is reduced. For the translation, for example, an existing machine translation method may be used. As machine translation, for example, statistical machine translation can be used.

[Explanation about machine learning]
Here, machine learning will be described. The machine learning method prepares many sets of problem-solution pairs, learns what kind of solution the problem will be by learning by using the set, and uses the learning results. It is a method that allows you to guess the solution even when it is a new problem. For example, see the following document.

  Literature: Maki Murata, “Language Processing Based on Machine Learning”, Faculty of Science and Engineering, Ryukoku University. Invited lecture, 2004 (http://www2.nict.go.jp/jt/a132/members/murata/ps/rk1-siryou.pdf)

  References: Masaki Murata, Ma Aoi, Kiyotaka Uchimoto, Hitoshi Isahara, “Japanese-English Translation of Tens / Aspect Modality Using Support Vector Machine”, IEICE Language Understanding and Communication, NLC 2000-78, 2001

  References: Maki Murata, Masao Uchiyama, Kiyotaka Uchimoto, Mao Ai, Hitoshi Isahara, "CRL's Effort in the NSEVAL2J Dictionary Task", IEICE Language Understanding and Communication Study Group, NLC 2001-40, 2001

  In order to operate a machine learning algorithm, when a problem situation is transmitted to a machine, a feature (each element constituting the problem by information used for analysis) is required. The problem is expressed by the feature. For example, in the problem of estimating the tense of Japanese sentence ending expressions, the problem: “He speaks.” --- If the solution “present” is given, an example of a feature is “He speaks.” . "" Speaking. "" Su "". "

  In other words, the machine learning method prepares many sets of feature set-solution pairs, and learns what type of feature set the solution will be by learning using that set. This is a method of using a learning result to extract a set of features from a problem and guessing a solution corresponding to the feature even when the problem is a new problem. Here, the “solution” is, for example, whether or not the answer information is described above, or classification information.

  As a machine learning method, for example, a k neighborhood method, a simple Bayes method, a decision list method, a maximum entropy method, a support vector machine method, or the like can be used. In the following description, machine learning in the case of classifying documents (when the problem-solution set is sentence-classification) will be mainly described, but the same applies to machine learning other than that. Needless to say.

  The k-nearest neighbor method is a technique for obtaining a solution (classification) by majority decision of k cases using the most similar k cases instead of the most similar case. k is a predetermined integer number, and generally an odd number between 1 and 9 is used.

The simple Bayes method is a method in which the probability of each solution (classification) is estimated based on Bayes' theorem, and the solution having the largest probability value is determined as a solution to be obtained.
In the simple Bayes method, the probability of outputting the classification a in the context b is given by the following equation.

Here, the context b is a set of features f _j (εF, 1 ≦ j ≦ k) set in advance. p (b) is the appearance probability of the context b. Here, since it is independent of the classification a and is a constant, it is not calculated. P (a) (where P is a tilde at the top of p) and P (f _i | a) are probabilities estimated from the teacher data, respectively, and the appearance probability of class a, and the feature f for class a means the probability of having _i . If the value obtained by performing maximum likelihood estimation as P (f _i | a) is used, the value is often zero, and it is difficult to determine the classification destination because the value of the expression in the second row is zero. Arise. Therefore, smoothing is performed. Here, a smoothed image using the following equation is used.

However, freq _(f i, a) has a feature _{f i,} and the number of cases classified is a, freq (a), the classification means the number of cases is a.
Needless to say, the smoothing is not limited to the method using the above formula, and other methods may be used.

  The decision list method uses features and combinations of classification destinations as rules, stores them in the list in a predetermined priority order, and when input to be detected is given, from the highest priority in the list This is a method in which input data is compared with the feature of the rule, and the classification destination of the rule having the same feature is set as the classification destination of the input.

In the decision list method, the probability value of each classification is obtained using only one of the features f _j (εF, 1 ≦ j ≦ k) set in advance as a context. The probability of outputting classification a in a context b is given by:

また、Ｐ（ａ_ｉ｜ｆ_ｊ）（ここでＰはｐの上部にチルダ）は、素性ｆ_ｊを文脈に持つ場合の分類ａ_ｉの出現の割合である。 However, f _max is given by the following equation.

Further, P (a _i | f _j ) (where P is a tilde at the top of p) is a rate of appearance of the classification a _i when the feature f _j is included in the context.

In the maximum entropy method, when a set of preset features f _j (1 ≦ j ≦ k) is F, the maximum entropy expression is satisfied while satisfying the following predetermined conditional expression. In this method, a probability distribution p (a, b) is obtained, and the classification having the largest probability value is obtained among the probabilities of the respective classifications obtained according to the probability distribution.

The predetermined conditional expression is given by the following expression.

Also, an expression meaning entropy is given by the following expression.

However, A and B mean a set of classification and context. Further, g _j (a, b) means a function that is 1 when the context b has a feature f _j and the classification is a, and is 0 otherwise. Further, P (a _i | f _j ) (where P is a tilde at the top of p) means the rate of occurrence of (a, b) in the known data.

  In the above conditional expression, the expected value of the frequency of the output and feature pair is obtained by multiplying the probability p and the function g meaning the appearance of the pair of output and feature. And the expected value calculated based on the probability distribution obtained on the left side is the constraint, and entropy maximization (smoothing of the probability distribution) is performed to obtain the probability distribution of the output and the context. For details of the maximum entropy method, refer to the following documents.

  Literature: Eric Sven Ristad, “Maximum Entropy Modeling for Natural Language”, (ACL / EACL Tutoral Program, Madrid, 1997)

  Literature: Eric Sven Ristad, “Maximum Entropy Modeling Tool, Release 1.6 beta”, (http://www.mnemonic.com/software/memt), 1998.

  The support vector machine method is a method of classifying data composed of two classifications by dividing a space by a hyperplane.

  FIG. 17 shows the concept of margin maximization in the support vector machine method. In FIG. 17, a white circle means a positive example, a black circle means a negative example, a solid line means a hyperplane that divides the space, and a broken line means a surface that represents the boundary of the margin area. 17A is a conceptual diagram when the interval between the positive example and the negative example is narrow (small margin), and FIG. 17B is a conceptual diagram when the interval between the positive example and the negative example is wide (large margin). is there.

  At this time, if the two classifications consist of positive examples and negative examples, the larger the interval (margin) between the positive examples and the negative examples in the learning data, the lower the possibility of incorrect classification with open data. As shown in FIG. 17B, a hyperplane that maximizes the margin is obtained, and classification is performed using the hyperplane.

  Basically, it is as described above. Usually, an extension of the method that the training data may contain a small number of cases in the inner area of the margin, or an extension that makes the linear part of the hyperplane nonlinear ( A kernel function introduced) is used.

This extended method is equivalent to classification using the following discriminant function (f (x)), and the two classes can be discriminated depending on whether the output value of the discriminant function is positive or negative.

Where x is the context (set of features) to be identified, and x _i and y _j (i = 1,..., L, y _j ε {1, −1}) mean the context and classification destination of the learning data And the function sgn is
sgn (x) = 1 (x ≧ 0)
-1 (otherwise)
In addition, each α _i is for maximizing Expression (M2) under the constraints of Expression (M3) and Expression (M4).

The function K is called a kernel function, and various functions are used. In this embodiment, for example, the following polynomial is used.
K (x, y) = (x · y + 1) d

Here, C and d are constants set experimentally. For example, C was fixed at 1 throughout all treatments. Moreover, two types of 1 and 2 are tried for d. Here, x _i satisfying α _i > 0 is called a support vector, and the portion taking the sum of the expression (M1) is normally calculated using only this case. That is, only actual cases called support vectors are used for actual analysis.
For details of the extended support vector machine method, refer to the following document.

  Literature: Nero Christianiani, John Shawe-Taylor, “An Introduction to Support Vector Machines and other kernel-based learning methods”, Cambridge Year 2000.

  Literature: Taku Kudoh, “Tinysvm: Support Vector machines”, (http://cl.aistnara.ac.jp/taku-ku//software/Tiny SVM / index.html)

  The support vector machine method handles data with two classifications. Therefore, when handling cases with three or more classifications, a pair-wise method or a one-VS rest method is usually used in combination with this.

  In the case of data having n classifications, the pairwise method generates every pair (n (n-1) / 2) of two different classification destinations, and determines which is better for each pair. That is, it is a method of obtaining a classification destination by a majority decision of the classification destination by n (n-1) / 2 binary classifications, which is obtained by the support vector machine method processing module.

  For example, when there are three classification destinations a, b, and c, the one VS rest method generates three sets of classification destination a and other, classification destination b and other, classification destination c and other, The learning process is performed on the set of the support vector machine method. Then, in the estimation process based on the learning result, the learning results of the three sets of support vector machines are used. See how the problem to be estimated is estimated in the three support vector machines, and among the three support vector machines, which is the other classification destination, and the separation plane of the support vector machine This is a method for obtaining a classification destination of a thing farthest from the object. For example, when a problem to be solved is farthest from the separation plane in the support vector machine created by the learning process of “classification destination a and others”, the classification destination of the problem to be solved is estimated as a. .

  How to find out what kind of solution (classification destination) is likely to be solved for a problem to be solved, which is estimated by a solution estimation unit (not shown) is determined by various methods used by a machine learning unit (not shown) as a machine learning method. Different.

  For example, when the machine learning means uses the k-nearest neighbor method as the machine learning method, the machine learning means uses the ratio of overlapping features (same features) among a set of features extracted from the examples of the teacher data. The degree of similarity between cases is defined based on the ratio of the number of cases and the defined degree of similarity and the case are stored in the learning result storage means as learning result information.

  Then, the solution estimation means refers to the similarity and case defined in the learning result storage means when the feature of the problem to be solved (sentence group attribute information) is extracted, and the feature of the extracted problem to be solved The k cases are selected from the learning result storage means in descending order of the similarity of the features of the problem to be solved, and the classification destination determined by the majority vote in the selected k cases is classified as the problem to be solved. Estimate as the first (solution). That is, in the solution estimation means, the degree of what kind of solution (classification destination) is likely to be obtained for the problem to be solved is set as the number of majority votes in the selected k cases.

  When the simple Bayes method is used as the machine learning method, the machine learning means (not shown) stores a combination of the solution of the case and a set of features as learning result information in the learning result storage means for the case of the teacher data. Remember. Then, when the features of the problem to be solved are extracted, the solution estimation means should solve based on the Bayes' theorem based on the set of the learning result information and the feature set in the learning result storage means. The probability of becoming each classification in the case of a set of problem features is calculated, and the classification having the largest probability value is estimated as the classification (solution) of the problem feature to be solved. That is, in the solution estimation means, the probability of becoming a solution in the case of a set of features of the problem to be solved is set as the probability of being classified.

  Further, when the decision list method is used as the machine learning technique, the machine learning means (not shown) is a method in which a list in which rules of features and classification destinations are arranged in a predetermined priority order in advance for example of teacher data. Thus, it is stored in the learning result storage means. Then, when the features of the problem to be solved are extracted, the solution estimation means compares the extracted features of the problem to be solved with the features of the rule in descending order of priority in the list of learning result storage means, Is estimated as the classification destination (solution) of the problem to be solved.

  When the maximum entropy method is used as the machine learning method, the machine learning means (not shown) specifies a class that can be a solution from the example of the teacher data, satisfies a predetermined conditional expression, and represents an expression indicating entropy. A probability distribution composed of a set of features when maximizing and a binomial classification that can be a solution is obtained and stored in the learning result storage means. Then, when the features of the problem to be solved are extracted, the solution estimation means uses the probability distribution of the learning result storage means to determine the probability of classification that can be the solution for the set of extracted feature features of the problem to be solved. Then, a class that can be a solution having the largest probability value is specified, and the specified class is estimated as a solution of the problem to be solved. That is, in the solution estimation means, the probability of becoming a solution in the case of a set of features of the problem to be solved is set as the probability of being classified.

  When the support vector machine method is used as the machine learning method, the machine learning means (not shown) identifies a class that can be a solution from the example of the teacher data, divides the class into a positive example and a negative example, In a space whose dimension is a set of case features according to a predetermined execution function using a function, the interval between the positive example and the negative example of the case is maximized, and the positive example and the negative example are divided by a hyperplane. A plane is obtained and stored in the learning result storage means. Then, when the features of the problem to be solved are extracted, the solution estimation means uses the hyperplane of the learning result storage means, and in the space where the set of the features of the problem to be solved is divided by the hyperplane, The classification determined based on the identified result is estimated as the solution of the problem to be solved. That is, in the solution estimation means, the degree of ease of becoming a solution in the case of a set of features of the problem to be solved is set as the magnitude of the distance from the separation plane to the case of the problem to be solved.

  In the above embodiment, the case where the question answering apparatus is stand-alone has been described. However, the question answering apparatus may be a stand-alone apparatus or a server apparatus in a server / client system. In the latter case, the output unit or the reception unit receives input or outputs information via a communication line.

  In the above embodiment, information related to processing executed by each component, for example, information received, acquired, selected, generated, transmitted, and received by each component. In addition, information such as threshold values, mathematical formulas, addresses, etc. used by each component in processing is retained temporarily or over a long period of time on a recording medium (not shown) even when not explicitly stated in the above description. It may be. Further, the storage of information in the recording medium (not shown) may be performed by each component or a storage unit (not shown). Further, reading of information from the recording medium (not shown) may be performed by each component or a reading unit (not shown).

  In the above embodiment, each process or each function may be realized by centralized processing by a single device or a single system, or may be distributedly processed by a plurality of devices or a plurality of systems. It may be realized by doing.

  In the above embodiment, when two or more components included in the question answering apparatus have a communication device, an input device, or the like, the two or more components may have a physically single device. Alternatively, it may have a separate device.

  In the above embodiment, each component may be configured by dedicated hardware, or a component that can be realized by software may be realized by executing a program. For example, each component can be realized by a program execution unit such as a CPU reading and executing a software program recorded on a recording medium such as a hard disk or a semiconductor memory. The software that realizes the question answering apparatus in the above embodiment is the following program. In other words, this program causes the computer to respond to the question information received by the question information receiving unit that receives question information, which is information indicating a non-factoid type question, and the question information received by the question information receiving unit. Information indicating a classification of information, a reasoning question that is a question for asking a reason, a classification unit that assigns one of a plurality of classification information that is information included as at least one classification, and the question information receiving unit A term extraction unit that extracts terms from question information, classification information that is information indicating a classification that is stored in the correspondence information storage unit in the classification information provided by the classification unit, and a term that is extracted by the term extraction unit Additional terms associated with correspondence information, which is information associated with additional terms to be added to the term, terms extracted by the term extraction unit, and access By using the corpus stored in the active corpus storage unit and an expression corresponding to the classification information given by the classification unit, answer information that is information indicating an answer corresponding to the question information is obtained from the corpus It is for functioning as an answer information acquiring unit to be acquired and an answer information output unit for outputting the answer information acquired by the answer information acquiring unit.

  In the program, the functions realized by the program do not include functions that can be realized only by hardware. For example, functions that can be realized only by hardware such as a modem or an interface card in an acquisition unit that acquires information, an output unit that outputs information, and the like are not included in at least the functions realized by the program.

  Further, this program may be executed by being downloaded from a server or the like, and a program recorded on a predetermined recording medium (for example, an optical disk such as a CD-ROM, a magnetic disk, a semiconductor memory, or the like) is read out. May be executed by

  Further, the computer that executes this program may be singular or plural. That is, centralized processing may be performed, or distributed processing may be performed.

  FIG. 18 is a schematic diagram showing an example of the external appearance of a computer that executes the program and realizes the question answering apparatus according to the embodiment. The above-described embodiment is realized by computer hardware and a computer program executed on the computer hardware.

  In FIG. 18, a computer system 100 includes a computer 101 including a CD-ROM (Compact Disk Read Only Memory) drive 105 and an FD (Flexible Disk) drive 106, a keyboard 102, a mouse 103, and a monitor 104.

  FIG. 19 is a diagram illustrating a computer system. In FIG. 19, in addition to the CD-ROM drive 105 and the FD drive 106, a computer 101 includes a CPU (Central Processing Unit) 111, a ROM (Read Only Memory) 112 for storing a program such as a bootup program, A CPU (Random Access Memory) 113 that is connected to the CPU 111 and temporarily stores application program instructions and provides a temporary storage space, a hard disk 114 that stores application programs, system programs, and data, a CPU 111 and a ROM 112. Etc. to each other. The computer 101 may include a network card (not shown) that provides connection to the LAN.

  A program for causing the computer system 100 to execute the function of the question answering apparatus according to the above embodiment is stored in the CD-ROM 121 or the FD 122, inserted into the CD-ROM drive 105 or the FD drive 106, and transferred to the hard disk 114. May be. Instead, the program may be transmitted to the computer 101 via a network (not shown) and stored in the hard disk 114. The program is loaded into the RAM 113 at the time of execution. The program may be loaded directly from the CD-ROM 121, the FD 122, or the network.

  The program does not necessarily include an operating system (OS) or a third-party program that causes the computer 101 to execute the function of the question answering apparatus according to the above embodiment. The program may include only a part of an instruction that calls an appropriate function (module) in a controlled manner and obtains a desired result. How the computer system 100 operates is well known and will not be described in detail.

  Further, the present invention is not limited to the above-described embodiment, and various modifications are possible, and it goes without saying that these are also included in the scope of the present invention.

  As described above, according to the question answering device or the like according to the present invention, the effect of being able to appropriately answer non-factoid type question information is obtained, which is useful as a device that outputs answer information corresponding to the question information. It is.

The block diagram which shows the structure of the question answering apparatus by Embodiment 1 of this invention. The block diagram which shows the structure of the reply information acquisition part by the embodiment The flowchart which shows operation | movement of the question answering apparatus by the embodiment The flowchart which shows operation | movement of the question answering apparatus by the embodiment The flowchart which shows operation | movement of the question answering apparatus by the embodiment The flowchart which shows operation | movement of the question answering apparatus by the embodiment The flowchart which shows operation | movement of the question answering apparatus by the embodiment The flowchart which shows operation | movement of the question answering apparatus by the embodiment The figure which shows an example of the correspondence information in the embodiment The figure which shows an example of the display in the embodiment The figure which shows an example of the score of the 2nd type | formula in the embodiment A figure showing an example of an acquired document in the embodiment The figure which shows an example of the score of the 1st formula in the embodiment The figure which shows an example of the display of the reply information in the embodiment The figure which shows an example of the classification | category corresponding | compatible information in the same embodiment The figure which shows an example of the experimental result in the same embodiment The figure for demonstrating the machine learning in the embodiment Schematic diagram showing an example of the appearance of the computer system in the embodiment The figure which shows an example of a structure of the computer system in the embodiment

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Question response apparatus 11 Question information reception part 12 Classification | category part 13 Term extraction part 14 Correspondence information storage part 15 Corpus storage part 16 Answer information acquisition part 17 Answer information output part 21 Document acquisition means 22 Additional term acquisition means 23 Calculation means 24 Answer information Selection means

Claims (9)

A question information receiving unit that receives question information, which is information indicating a non-factoid type question,
Any of a plurality of pieces of classification information that is information indicating the classification of the question information with respect to the question information received by the question information reception unit, and that is information including a reason question that is a question asking a reason as at least one classification A classification unit for assigning or
A term extraction unit that extracts terms from the question information received by the question information reception unit;
A correspondence information storage unit that stores correspondence information that is information that includes classification information that is information indicating classification and an additional term to be added to the term extracted by the term extraction unit;
An additional term associated with the classification information given by the classification unit in the correspondence information, a term extracted by the term extraction unit, a corpus stored in an accessible corpus storage unit, and a classification unit An answer information acquisition unit that acquires, from the corpus, answer information that is information indicating an answer corresponding to the question information by using an expression according to the assigned classification information;
An answer information output unit that outputs the answer information acquired by the answer information acquisition unit ,
The first formula used by the answer information acquisition unit is a formula that has a higher value as the two terms are closer to each other in the document included in the corpus,
The answer information acquisition unit includes a term including an additional term associated with the classification information given by the classification unit in the correspondence information and the term extracted by the term extraction unit for the document included in the corpus Calculate the value of the formula using two selected from the set, and obtain response information that is larger than the value of the formula,
The answer information acquisition unit
A plurality of documents that are highly characterized by the terms extracted by the term extraction unit are acquired from the corpus using a second expression that is an expression indicating the degree to which a certain term characterizes a document. Document acquisition means;
Additional term acquisition means for acquiring additional terms corresponding to the classification information given by the classification unit from the correspondence information;
Answer candidate information that is information that is a candidate for answer information included in each document acquired by the document acquisition unit using the term extracted by the term extraction unit and the additional term acquired by the additional term acquisition unit Calculating means for calculating the value of the first equation according to the classification information given by the classification unit;
Answer information selecting means for selecting answer information whose value calculated by the calculating means is a larger value than others from the plurality of answer candidate information,
The calculating means calculates the following formula for the answer candidate information d:

Score (d) is calculated, and the value of the first equation, which is a value obtained by changing the calculated value according to the classification information, is calculated (where d is the answer candidate information, and T is the term set Yes, dist (t1, t2) is the interval between the terms t1, t2, N is the total number of documents, df (t) is the number of documents in which the term t appears, and w _dr2 (t2) is , A function of the term t2 determined by experiment) . The classification unit classifies the question information received by the question information reception unit into at least a definition question that asks a definition, a reason question that asks a reason, and a method question that asks a method. Item 2. The question answering apparatus according to Item 1 . The term extraction unit also extracts a focus expression that is a target expression for which a definition is being asked from question information classified as a definition question by the classification unit,
The first equation is
In the case where the focus expression is included in the document acquired by the document acquisition unit, the value is larger than the case where the focus expression is not included.
When the focus expression included in the document acquired by the document acquisition unit is modified with a linkage modification clause, it is an expression having a value larger than that when it is not.
The answer information selecting means, when the focus expression included in the document acquired by the document acquisition means is modified with a combination modification clause, selects the combination modification clause from the answer candidate information as response information; The question answering apparatus according to claim 2 . The answer information selecting means includes
Select answer candidate information whose value calculated by the calculating means is larger than others,
Machine learning is performed using at least teacher data, which is prepared in advance, is question information, answer information that is information indicating an answer to the question indicated by the question information, and information indicating whether the answer information is appropriate,
Using the results of the machine learning, it extracts the response information from the selected answer candidate information, question answering apparatus according to any one of claims 1 to 3. The classification unit includes:
Classification information that is information indicating classification and classification correspondence information that is information having the phrase information that is information indicating a phrase in association with each other are held in the recording medium,
The question answering device according to any one of claims 1 to 4 , wherein when the question information includes a phrase indicated by the phrase information, classification information corresponding to the phrase information is given to the question information. . The classification unit includes:
Machine learning is performed using teacher information prepared in advance as question information and classification information that indicates the classification of the question information,
The question answering device according to any one of claims 1 to 4 , wherein classification information is given to question information received by the question information receiving unit using a result of the machine learning. The term extraction unit performs morphological analysis on the question information, and from the question information, (1) independent words, (2) nouns, (3) nouns and verbs, (4) nouns and adjectives, (5) nouns and verbs The question answering device according to any one of claims 1 to 6 , wherein a part-of-speech term included in any one of (1) to (5) selected from: Correspondence information that is information including a question information reception unit, a classification unit, a term extraction unit, classification information that is information indicating classification, and an additional term that is added to the term extracted by the term extraction unit. It is a question answering method processed using a stored correspondence information storage unit, an answer information acquisition unit having a document acquisition unit, an additional term acquisition unit, a calculation unit, and an answer information selection unit, and an answer information output unit. And
A question information receiving step in which the question information receiving unit receives question information which is information indicating a non-factoid type question;
A plurality of pieces of information in which the classification part is information indicating the classification of the question information with respect to the question information received in the question information reception step, and a reason question that is a question asking a reason is included as at least one classification A classification step for assigning any of the classification information of
The term extraction unit extracts a term from the question information received in the question information receiving step,
The answer information acquisition unit is stored in an accessible corpus storage unit, an additional term associated with the classification information given in the classification step by the correspondence information, a term extracted in the term extraction step A response information acquisition step of acquiring response information from the corpus, which is information indicating an answer corresponding to the question information, by using a corpus and an expression corresponding to the classification information given in the classification step;
The response information output unit includes a response information output step for outputting the response information acquired in the response information acquisition step ,
The first formula that is used in the answer information acquisition step is a formula that has a higher value as the two terms are closer to each other in the document included in the corpus,
In the answer information acquisition step, for the document included in the corpus, a term including an additional term associated with the classification information assigned in the classification step by the correspondence information and the term extracted in the term extraction step Calculate the value of the formula using two selected from the set, and obtain response information that is larger than the value of the formula,
The answer information acquisition step includes:
A plurality of documents having a high degree of being characterized by the terms extracted in the term extracting step by using the second expression that is an expression indicating the degree to which a certain term characterizes a certain document. Obtaining a document from the corpus;
The additional term acquisition unit acquires an additional term corresponding to the classification information given in the classification step from the correspondence information;
Information that is a candidate for answer information included in each document acquired in the document acquisition step, using the term extracted in the term extraction step and the additional term acquired in the additional term acquisition step. For certain answer candidate information, a calculation step for calculating the value of the first equation according to the classification information given in the classification step;
The answer information selecting means comprises, from the plurality of answer candidate information, an answer information selecting step for selecting answer information whose value calculated in the calculating step is a larger value than others,
In the calculating step, for the answer candidate information d,

Score (d) is calculated, and the value of the first equation, which is a value obtained by changing the calculated value according to the classification information, is calculated (where d is the answer candidate information, and T is the term set Yes, dist (t1, t2) is the interval between the terms t1, t2, N is the total number of documents, df (t) is the number of documents in which the term t appears, and w _dr2 (t2) is , A function of the term t2 determined by experiments) . Computer
A question information receiving unit that receives question information, which is information indicating a non-factoid type question,
Any of a plurality of pieces of classification information that is information indicating the classification of the question information with respect to the question information received by the question information reception unit, and that is information including a reason question that is a question asking a reason as at least one classification A classification unit for assigning or
A term extraction unit that extracts terms from the question information received by the question information reception unit;
Information having classification information provided by the classification unit in association with classification information stored in the correspondence information storage unit, which is information indicating classification, and an additional term to be added to the term extracted by the term extraction unit An additional term associated with certain correspondence information, the term extracted by the term extraction unit, a corpus stored in an accessible corpus storage unit, and an expression corresponding to the classification information given by the classification unit An answer information acquisition unit that acquires answer information, which is information indicating an answer corresponding to the question information, from the corpus, and
Function as an answer information output unit for outputting the answer information acquired by the answer information acquisition unit,
The first formula used by the answer information acquisition unit is a formula that has a higher value as the two terms are closer to each other in the document included in the corpus,
The answer information acquisition unit includes a term including an additional term associated with the classification information given by the classification unit in the correspondence information and the term extracted by the term extraction unit for the document included in the corpus Calculate the value of the formula using two selected from the set, and obtain response information that is larger than the value of the formula,
The answer information acquisition unit
A plurality of documents that are highly characterized by the terms extracted by the term extraction unit are acquired from the corpus using a second expression that is an expression indicating the degree to which a certain term characterizes a document. Document acquisition means;
Additional term acquisition means for acquiring additional terms corresponding to the classification information given by the classification unit from the correspondence information;
Answer candidate information that is information that is a candidate for answer information included in each document acquired by the document acquisition unit using the term extracted by the term extraction unit and the additional term acquired by the additional term acquisition unit Calculating means for calculating the value of the first equation according to the classification information given by the classification unit;
Answer information selecting means for selecting answer information whose value calculated by the calculating means is a larger value than others from the plurality of answer candidate information,
The calculating means calculates the following formula for the answer candidate information d:

Score (d) is calculated, and the value of the first equation, which is a value obtained by changing the calculated value according to the classification information, is calculated (where d is the answer candidate information, and T is the term set Yes, dist (t1, t2) is the interval between the terms t1, t2, N is the total number of documents, df (t) is the number of documents in which the term t appears, and w _dr2 (t2) is , A function of the term t2 determined by experiment) .

Images