JP4849596B2 - Question answering apparatus, question answering method, and question answering program - Google Patents

Description

  The present invention relates to a question answering technique for outputting an answer to a question expressed in a natural language as a natural language processing system by a computer, and more particularly, the input keyword is increased by a keyword extraction technique and is constituted by the increased keywords. The present invention relates to a question answering apparatus, a question answering method, and a question answering program for automatically obtaining and outputting solutions to a plurality of questions.

  A question answering device is a device that, when a question in natural language is input, outputs the answer itself. For example, if you enter the question “Which part of the brain is related to the death of Parkinson's disease?”, A large amount of electronic text containing data from the Web, newspaper articles, encyclopedias, etc. The disease is said to be caused by the degeneration of melanocytes in the substantia nigra and the absence of the neurotransmitter dopamine produced in the substantia nigra cells. " Output the answer to.

  Since the question answering apparatus can extract an answer from a normal sentence (text data) described in a natural language, not from a logical expression or a database, a large amount of existing document data can be used. Also, the question answering device outputs the answer itself, unlike an information search system in which the user himself / herself needs to find an answer from an article searched by a keyword. Therefore, the user can obtain the answer information earlier. Since the question answering apparatus is useful as described above, it is expected to realize a practical question answering apparatus that is easier to use.

  A general question answering apparatus (or question answering system) is roughly composed of three processing means: an answer expression estimation process, a document search process, and an answer extraction process.

  The answer expression estimation process is a process of estimating the answer expression based on the expression of the question pronoun in the input question. The answer expression is the type of the desired language expression of the answer, the type based on the meaning of the language expression as the answer (answer type), and the type based on the notation of the language expression as the answer (answer expression type) and so on. The question answering device estimates the answer type of the answer of the input question with reference to the correspondence relationship of what kind of question language expression requires what kind of answer expression. For example, when the input question is “How big is the area in Japan?”, The question answering device refers to a predetermined correspondence, and the answer type is “numeric expression” from the expression “how much” in the question. ". If the question is “Who is the prime minister of Japan?”, It is estimated from the expression “who” in the question that the answer type is “proprietary noun (person name)”.

  The document search process is a process of extracting a keyword from a question, searching a document data group for which an answer is to be searched using this keyword, and extracting document data that is considered that the answer is described. For example, when the input question is “Where is the capital of Japan?”, The question answering device extracts “Japan” and “capital” from the question as keywords, and extracts the keyword “ Search for document data including "Japan" and "Capital".

  The answer extraction process is a process of extracting a linguistic expression suitable for the estimated answer type from the document data including the keyword extracted in the document search process, and outputting it as an answer. The question answering device, for example, uses a linguistic expression “matching a proper noun (place name)” estimated in the answer expression estimation process from document data including the keywords “Japan” and “capital” searched in the document search process. Extract “Tokyo” as the answer.

  By performing the processing as described above, the question answering apparatus outputs the answer “Tokyo” to the question “Where is the capital of Japan?”.

In addition, as a specific conventional technique related to the question answering device (or question answering system), for example, in the following non-patent document 1, a question answering system that estimates answers using a plurality of articles is obtained from a plurality of articles. A technique is described in which candidates with the highest total score are output as an answer by adding while gradually reducing the scores of the answer candidates.
Masaki Murata, Hitoshi Isahara, Use of multiple article information based on the gradual addition point method in the question answering system, IPSJ Natural Language Processing Study Group 2004-NL-160, 2004. Kyushu University.

  In a conventional question answering apparatus, a linguistic expression that can be an answer is extracted from the retrieved document data as an answer candidate, and an answer type of each of the extracted answer candidates is determined. Then, increase the evaluation of the answer candidate determined to be the same or similar to the answer type estimated from the question, and in principle, the answer candidate is the same answer type and has a predetermined evaluation as the answer Output.

  However, the conventional question answering device is a system that outputs only the answer to the question inquired by inputting the question, and cannot output the answer to the question other than the inquired question.

  SUMMARY OF THE INVENTION It is an object of the present invention to provide a question answering apparatus, a question answering method, and a question answering program for solving the problems of the prior art and outputting an answer to a question asked and an answer to a question other than the question asked. .

  In order to solve the above problems, the present invention provides a question answering apparatus that outputs an answer to question data expressed in a natural language, a keyword input means for inputting a plurality of keywords as input keywords, Based on the keyword increase means for extracting more keywords than the number of input keywords and outputting them as output keywords, and answer candidates that are answer candidates for the questions configured by the output keywords are stored as answer candidates stored in advance. It is characterized by comprising answer candidate extracting means for extracting from a document data group to be searched, and answer table output means for outputting a table in which each extracted answer candidate is associated with a question as an answer table.

  In the question answering apparatus according to the present invention, the keyword increasing unit searches the input keyword in a keyword extraction database in which document data for keyword extraction is stored and appears around the input keyword. A pattern extraction means for extracting a pattern, and a keyword for full-text search of the pattern extracted by the pattern extraction means in the keyword extraction database, extracting an expression extracted by the pattern, and outputting the extracted expression as an output keyword And an extracting means.

  In the question answering apparatus according to the present invention, the keyword increasing means extracts a word in the same field as the input keyword from a database in which correspondence information between the word and the word field is stored, and outputs it as an output keyword. It is characterized by outputting.

  In the question answering apparatus according to the present invention, the keyword increasing means may be configured to input the input keyword and the input keyword based on thesaurus data that is preliminarily stored in a database and is word classification information based on semantic similarity. , A similarity calculation means for calculating a similarity to a word in the thesaurus data, and a keyword extraction means for extracting a keyword based on the calculated magnitude of the similarity and outputting it as an output keyword. Features.

  In the question answering apparatus according to the present invention, a first keyword and a second keyword are input as the input keyword to the keyword input means, and the keyword increasing means is the input first The third keyword is output as an output keyword based on the keyword, and the fourth keyword is output as an output keyword based on the input second keyword. The answer candidate extraction means is prepared in advance. Using a large number of sets of answer questions and answers to the problem, it is learned what kind of answer the problem is, and based on the learning result, the output third keyword and fourth answer An answer candidate that is a candidate for an answer to a question configured by keywords is extracted.

  In the question answering apparatus according to the present invention, a first keyword and a second keyword are input as the input keyword to the keyword input means, and the keyword increasing means is the input first The third keyword is output as an output keyword based on the keyword, and the fourth keyword is output as an output keyword based on the input second keyword. The document data including the output third keyword and the fourth keyword is extracted from a large amount of document data group stored in the document data, and from the linguistic expression of the extracted document data, Answers to questions composed of the output third keyword and fourth keyword using the frequency of appearance And extracting the complement.

  In the question answering apparatus according to the present invention, a first pronoun and a second keyword are input as the input keyword to the keyword input unit, and the question pronoun associated with the second keyword A type of linguistic expression of candidate answers to a question composed of an output keyword output by the keyword increase means based on the question pronoun input means input to the question pronoun input means Answer type estimating means for estimating a certain answer type, and the keyword increasing means outputs a third keyword as an output keyword based on the inputted first keyword, and the inputted second keyword. The keyword is output as an output keyword, and the answer candidate extraction means is a search target of the answer candidate. The document data group including the third keyword and the second keyword output by the keyword increasing means is searched from the document data group, and estimated by the answer type estimating means from the document data extracted by the search processing A linguistic expression suitable for the answer type is extracted as an answer candidate of a question composed of the third keyword and the second keyword.

  In the question answering apparatus according to the present invention, a first keyword and a second keyword are input as the input keyword to the keyword input unit, and the keyword is associated with the predetermined second keyword. Answer type estimation means for estimating an answer type that is a type of a linguistic expression of a candidate answer to a question configured by an output keyword output by the keyword increase means based on the given question pronoun, and the keyword increase The means outputs a third keyword as an output keyword based on the input first keyword, and outputs the input second keyword as an output keyword. The answer candidate extraction means includes the answer candidate extraction means. From the document data group that is the candidate search target, the keyword output unit outputs the first The document data including the keyword and the second keyword is searched, and from the document data extracted by the search process, a language expression that matches the answer type estimated by the answer type estimating means is obtained as the third keyword. And a second keyword as a candidate for answering a question.

  In the question answering apparatus according to the present invention, in the question answering apparatus, the keyword input means is configured to input a first keyword and a second keyword as the input keyword, and a question pronoun input means for inputting a question pronoun, Based on the question pronoun input to the question pronoun input means, an answer type estimation means for estimating an answer type that is a type of linguistic expression of answer candidates for a question configured by the output keyword output by the keyword increasing means And the keyword increasing means outputs a third keyword as an output keyword based on the input first keyword, and outputs a fourth keyword based on the input second keyword. Output as an output keyword, and the answer candidate extraction means is a document that is a search target of the answer candidate The document data including the third keyword and the second keyword output by the keyword increasing means is retrieved from the data group, and is estimated by the answer type estimating means from the document data extracted by this search processing. The linguistic expression suitable for the answer type is extracted as an answer candidate of a question composed of the third keyword and the fourth keyword.

  In the question answering apparatus according to the present invention, a first keyword and a second keyword are input to the keyword input unit as the input keyword, and the keyword is determined based on a predetermined question pronoun. Answer type estimation means for estimating an answer type that is a type of linguistic expression of an answer candidate for a question configured by the output keyword output by the increase means, and the keyword increase means includes the input first The third keyword is output as an output keyword based on the keyword, the fourth keyword is output as an output keyword based on the input second keyword, and the answer candidate extraction means The third key output by the keyword increasing means from the document data group to be searched Document data including a keyword and a second keyword is retrieved, and from the document data extracted by the retrieval process, a language expression suitable for the answer type estimated by the answer type estimating means is obtained as the third keyword. And a fourth keyword are extracted as answer candidates for the question.

  Further, the present invention provides an answer type input means for inputting an answer type, which is a type of linguistic expression of an answer candidate for a question constituted by an output keyword output by the keyword increasing means in the question answering apparatus. The keyword input means receives a first keyword and a second keyword as the input keyword, and the keyword increase means assigns a third keyword based on the input first keyword. Based on the input second keyword, a fourth keyword is output as an output keyword, and the answer candidate extraction unit is configured to output the answer keyword from the document data group that is the search target for the answer candidate. Document including the third keyword and the fourth keyword output by the keyword increasing means A linguistic expression suitable for the answer type input to the answer type input means from the document data extracted by the search process is obtained by the output third keyword and fourth keyword. It is characterized in that it is extracted as an answer candidate for a question that is composed.

  In the question answering apparatus according to the present invention, a first keyword and a second keyword are input as the input keyword to the keyword input means, and the keyword increasing means is the input first The third keyword is output as an output keyword based on the keyword, and the fourth keyword is output as an output keyword based on the input second keyword. The answer candidate extracting means is configured to output the answer candidate The document data including the third keyword and the fourth keyword output by the keyword increasing means is searched from the document data group that is the search target, and the document data extracted in the search process is determined in advance. The linguistic expression suitable for the answer type is composed of the output third keyword and fourth keyword. And extracting as the answer candidate in question is.

  Further, the present invention is the above question answering apparatus, wherein the keyword input means is configured by inputting the first keyword and the second keyword as the input keyword and outputting the keyword by the keyword increasing means. An answer type input means for inputting an answer type corresponding to the second keyword input to the keyword input means, which is an answer type that is a type of language expression of answer candidates for the question to be answered, The keyword increasing means outputs a third keyword as an output keyword based on the input first keyword, and outputs the input second keyword as an output keyword. The answer candidate extracting means From the document data group to be searched for the answer candidate, the keyword increasing means The document data including the output third keyword and the second keyword is searched, and from the document data extracted by the search process, a language expression suitable for the answer type input to the answer type input means is obtained. It is extracted as an answer candidate of a question constituted by the outputted third keyword and second keyword.

  In the question answering apparatus according to the present invention, a first keyword and a second keyword are input as the input keyword to the keyword input means, and the keyword increasing means is the input first The third keyword is output as an output keyword, the input second keyword is output as an output keyword, and the answer candidate extraction means is a document data group that is a search target for the answer candidate The document data including the third keyword and the second keyword output by the keyword increasing means is searched, and the second keyword is determined in advance from the document data extracted by the search process. The linguistic expression suitable for the associated answer type is converted into the output third keyword and second keyword. And extracts as the answer candidate questions made me.

  Further, the present invention is the above question answering apparatus, wherein the keyword input means is configured by inputting the first keyword and the second keyword as the input keyword and outputting the keyword by the keyword increasing means. An answer type input means for inputting an answer type corresponding to the second keyword input to the keyword input means, which is an answer type that is a type of language expression of answer candidates for the question to be answered, The keyword increasing means outputs a third keyword as an output keyword based on the input first keyword, and outputs a fourth keyword as an output keyword based on the input second keyword. And the second keyword similar to the output fourth keyword Each of the fourth keywords includes similar keyword determination means for determining as a similar keyword, and the answer candidate extraction means is output from the document data group that is the search target of the answer candidates by the keyword increase means. Document data including the third keyword and the fourth keyword is searched, and the answer type input is performed by associating the output fourth keyword with a similar keyword similar to the extracted document data. The linguistic expression suitable for the answer type input to the means is extracted as an answer candidate of a question constituted by the output third keyword and fourth keyword.

  In the question answering apparatus according to the present invention, a first keyword and a second keyword are input as the input keyword to the keyword input means, and the keyword increasing means is the input first The third keyword is output as an output keyword based on the keyword, the fourth keyword is output as the output keyword based on the input second keyword, and the output of the second keyword is output. Similar keyword determination means for determining a similar keyword to the fourth keyword for each of the fourth keywords is provided, and the answer candidate extraction means is a document data group that is a search target for the answer candidates. To include the third keyword and the fourth keyword output by the keyword increasing means. The document data is searched, and from the document data extracted by the search process, a language expression that matches an answer type that is previously associated with a similar keyword similar to the output fourth keyword is output as the output first data. It is characterized in that it is extracted as an answer candidate for a question composed of the third keyword and the fourth keyword.

  In the question answering apparatus according to the present invention, a first pronoun and a second keyword are input as the input keyword to the keyword input unit, and the question pronoun associated with the second keyword A type of linguistic expression of candidate answers to a question composed of an output keyword output by the keyword increase means based on the question pronoun input means input to the question pronoun input means Answer type estimating means for estimating a certain answer type, and the keyword increasing means outputs a third keyword as an output keyword based on the inputted first keyword, and the inputted second keyword. Based on the keyword, the fourth keyword is output as an output keyword, and the second keyword Of these, the fourth keyword includes similar keyword determination means for determining a similar keyword for each of the fourth keywords, and the answer candidate extraction means is a search target of the answer candidates. The document data including the third keyword and the fourth keyword output by the keyword increasing unit is searched from a certain document data group, and the output fourth keyword is extracted from the document data extracted by the search processing. A linguistic expression that matches the answer type estimated by the answer type estimating means based on the question pronouns input to the question pronoun input means associated with similar keywords that are similar to each other. It extracts as a candidate answer of the question comprised by 4 keywords.

  In the question answering apparatus according to the present invention, a first keyword and a second keyword are input as the input keyword to the keyword input unit, and the predetermined keyword corresponds to the second keyword. An answer type estimating means for estimating an answer type that is a type of a linguistic expression of an answer candidate for a question configured by an output keyword output by the keyword increasing means based on the attached question pronoun, the keyword The increasing means outputs a third keyword as an output keyword based on the input first keyword, and outputs a fourth keyword as an output keyword based on the input second keyword. Among the second keywords, those similar to the output fourth keyword are added to the fourth keyword. Similar keyword determination means for determining each keyword as a similar keyword is provided, wherein the answer candidate extraction means includes a third keyword output by the keyword increase means from the document data group that is the search target of the answer candidate, Document data including the fourth keyword is searched, and answer type estimation means based on the question pronoun associated with the similar keyword similar to the output fourth keyword from the document data extracted by the search processing The linguistic expression suitable for the answer type estimated by is extracted as an answer candidate of a question composed of the output third keyword and fourth keyword.

  Further, according to the present invention, in the question answering apparatus, the similar keyword determination unit is configured to share the fourth keyword output by the keyword extraction unit from a large amount of document data group stored in advance in the storage unit. Extracting a co-occurrence word that is a word that appears and appearing, and for each of the fourth keywords, the number of occurrences in the document data group co-occurring with each of the extracted co-occurrence words The degree of similarity between the co-occurrence vector for each fourth keyword and the co-occurrence vector for the same fourth keyword as the second keyword input to the keyword input means And a fourth keyword identical to the second keyword similar to each of the fourth keywords, determined based on the obtained degree of similarity, Characterized by a similar keyword.

  In the question answering apparatus according to the present invention, the similar keyword determination unit outputs the keyword increase unit based on thesaurus data that is preliminarily stored in the database and is the thesaurus data that is word classification information based on semantic similarity. Similarity calculation means for calculating the similarity between the same word as the fourth keyword and the same word as the second keyword input to the keyword input means for each of the fourth keywords, A second keyword similar to the fourth keyword determined based on the calculated degree of similarity is set as the similar keyword.

  The present invention is also a question answering method for outputting an answer to question data expressed in a natural language, the step of inputting a plurality of keywords as input keywords, and the number of input keywords based on the input keywords Extracting more keywords and outputting them as output keywords, and extracting answer candidates as answer candidates for a question configured by the output keywords from a document data group that is a search target for answer candidates stored in advance. And a step of outputting, as an answer table, a table in which each extracted answer candidate is associated with a question.

  Further, the present invention is a program for causing a computer provided in a question answering apparatus that outputs an answer to question data expressed in a natural language to execute a plurality of keywords as input keywords in the computer. Based on the input keywords, a process for extracting more keywords than the number of the input keywords and outputting them as output keywords, and answer candidates that are answer candidates for the questions constituted by the output keywords are stored in advance. A question answering program for executing a process of extracting from a document data group that is a search target of answer candidates and a process of outputting a table in which each of the extracted answer candidates is associated with a question as an answer table.

  According to the question answering device of the present invention, it is possible to output not only the answer to the inquired question but also the answer to the question other than the inquired question in a format associated with each question. That is, according to the question answering apparatus of the present invention, the user can automatically answer a large number of questions composed of keywords increased based on the input keywords by only inputting a small number of keywords of a genre that the user wants to know the answers. Can be obtained.

  For example, according to the question answering apparatus of the present invention, when the user inputs the first keyword and the second keyword, the third keyword more than the number of the first keywords is extracted based on the first keyword. At the same time, a fourth keyword larger than the number of the second keywords is extracted based on the second keyword, and an answer to the question configured based on the extracted third keyword and the fourth keyword is obtained. It is possible to output automatically using a machine learning technique.

  Further, for example, according to the question answering apparatus of the present invention, when the user inputs the first keyword, the second keyword, and the question pronoun associated with the second keyword, the number of the first keyword A large number of third keywords are extracted, and the answer type is estimated based on the input question pronoun, and the answer to the question configured based on the third keyword, the second keyword, and the question pronoun is It is possible to automatically output using the estimated answer type.

  For example, according to the question answering apparatus of the present invention, when the user inputs the first keyword, the second keyword, and the answer type, the number is larger than the number of the first keywords based on the first keyword. A third keyword is extracted, and a fourth keyword larger than the number of second keywords is extracted based on the second keyword, and is configured based on the extracted third keyword and fourth keyword. It is possible to automatically output the answer to the question to be asked using the inputted answer type.

  Further, for example, according to the question answering apparatus of the present invention, when the user inputs the first keyword, the second keyword, and the answer type associated with the second keyword, based on the first keyword A third keyword larger than the number of first keywords is extracted, and a fourth keyword larger than the number of second keywords is extracted based on the second keyword. A second keyword similar to the keyword (the same fourth keyword) is determined as a similar keyword. An answer to a question configured based on the extracted third keyword and fourth keyword can be automatically output using an answer type associated with the determined similar keyword. Become.

  First, before describing the embodiment of the present invention, the technique described in Non-Patent Document 1 will be described. Non-Patent Document 1 describes the use of multiple article information based on a gradual addition point method in a question answering system. The contents described in Non-Patent Document 1 will be described below.

  A question answering system is a system that outputs an answer to a given question. For example, when a question sentence is given, "Where is the capital of Japan?" "Tokyo is one of the 47 prefectures of Japan." Searches for electronic sentences such as the web and newspaper articles, and answers "Tokyo." The question answering system will be important as an alternative to information retrieval, and will be an important component of future artificial intelligence systems.

  Non-Patent Document 1 proposes a new method for adding points while reducing the scores of solution candidates obtained from a plurality of articles in order to improve the accuracy of the question answering system. This method is called a gradual addition point method.

  Question answers are often found in multiple articles, and in such cases, using multiple articles to estimate the answer may yield a better answer than using a single article Therefore, a method of using the information of a plurality of articles by adding the scores of solution candidates obtained from a plurality of articles can be considered. However, simply adding scores may reduce system performance.

Therefore, in Non-Patent Document 1, in order to deal with the problem that occurs when this addition is simply performed, a method of adding while reducing the score when adding the score is used. More specifically, according to the method of Non-Patent Document 1, the score of the i-th solution candidate is weighted with k ⁽ⁱ⁻¹⁾ and then the score is added. The final answer is determined by the total score. For example, it is assumed that “Tokyo” is extracted as a solution candidate from three articles, the scores are 26, 21, and 20, and k is 0.3. In this case, the total score of “Tokyo” is 34.1 (= 26 + 21 × 0.3 + 20 × 0.3 ² ). The score of each candidate is calculated by such a method, and the candidate having the highest total score is set as a solution.

  Next, the use of the progressive addition method in the use of multiple articles in Non-Patent Document 1 will be described in detail. Suppose you are given the question "Where is the capital of Japan?" At this time, the answer to be obtained is “Tokyo”. As shown in FIG. 21, the general question answering system can output the solution candidates and scores as a list, and can also output the article number indicating the article from which the solution candidates are extracted. The ranking shown in the figure indicates the ranking of the score size.

  In the example of the list shown in FIG. 21, the candidate with the highest score is “Kyoto”, and an incorrect solution is output. A method of simply adding the scores of solution candidates has already been proposed. When the list shown in FIG. 21 is used, the result shown in FIG. 22 is obtained according to the method of simply adding the scores of the solution candidates.

  In FIG. 22, the score of “Tokyo” is the highest, and the system can correctly output “Tokyo” as a solution. This method of simply adding the candidate scores of the solution could obtain the correct solution by using the information of multiple articles. However, this method has a problem that it is easy to pick out candidates for high-frequency solutions. This is a serious problem especially in high performance systems. In a system with originally high performance, the original score output by the system is often more reliable than a score obtained by simply adding, and the method of simply adding often degrades the performance of the system.

  In order to deal with this problem, the technique of Non-Patent Document 1 proposes a new method of adding while reducing the score. Instead of simply adding the candidate scores of the solution, the points are added with a weight that reduces the score. This method has the effect of reducing the bad effect of easily extracting high-frequency words and improving the performance of the system.

  The example which shows the effectiveness of the method proposed by this nonpatent literature 1 is given. It is assumed that the question “Where was the capital of Japan in the year 1000 AD” was given, and the system output the result shown in FIG. As shown in FIG. 23, “Kyoto” has the highest score. Here, the correct answer to the above question is “Kyoto”, and this system outputs the correct answer unless the score of the candidate solution is simply added. However, if a simple addition method is used, the result is as shown in the table of FIG. 24, and the wrong solution “Tokyo” is used as the system solution.

Here, a new method of Non-Patent Document 1 for adding while reducing the score will be used. Here, as a detailed system specification, the score of the i-th candidate is multiplied by 0.3 ^(i-1) . In that case, the score of “Tokyo” is 2.8 (= 2.1 + 1.8 × 0.3 + 1.5 × 0.3 ² + 1.4 × 0.3 ³ ), and the output result of the system is shown in FIG. Since the score of “Kyoto” is the highest, the correct answer “Kyoto” can be output correctly as a solution. That is, the method proposed in Non-Patent Document 1 can obtain a correct solution even in the first example (when a question sentence “where is the capital of Japan” is given). When applied to the first example, the score for “Tokyo” is 4.3 (= 3.2 + 2.8 × 0.3 + 2.5 × 0.3 ² + 2.4 × 0.3 ³ ), and the output result is shown in FIG. As shown in the table, “Tokyo” has the highest score and is output correctly as a solution.

  The method described in Non-Patent Document 1 for adding while reducing the score can improve the accuracy by using the information of a plurality of articles while reducing the disadvantage that it is easy to extract a candidate for a high-frequency solution.

The question answering system described in Non-Patent Document 1 is composed of the following three basic elements.
1. Estimation of solution expression The question answering system estimates the solution expression (what kind of language expression the solution is) based on the expression of interrogative pronouns. For example, if the input question sentence is “How big is the area in Japan?”, It is assumed from the expression “how much” the solution expression will be a numerical expression.
2. Document Search The question answering system retrieves keywords from a question sentence, and searches for documents using these keywords. This search collects documents that are likely to contain solutions. For example, if the input question sentence is “How big is Japan?”, “Japan” and “Area” are extracted as keywords, and documents including these are searched.
3. Solution Extraction The question answering system extracts a linguistic expression suitable for the estimated solution expression from a group of documents in which the solution is likely to be written, and outputs it as a solution. For example, if the input question is “How big is Japan?”, The language expression equivalent to the numerical expression estimated as the solution expression from the document group including “Japan” and “area” searched by the document search is used as the solution. Extract.

  Below, the technique proposed in Non-Patent Document 1 will be described in detail.

(Estimation of solution expression)
Estimate solution expressions using heuristic rules created manually. Create 16 rules. Some of them are shown below.
・ If there is an expression “who” in the question sentence, the answer is the name of the person.
・ If there is an expression “when” in the question sentence, the answer expression is a time expression.
・ If there is an expression “how much” in the question sentence, the solution expression is a numerical expression.

(Document search)
Keywords for document search are extracted by ChaSen, which is a known keyword extraction tool, and attached words and the like are excluded from the keywords. The document search is performed as follows.

First, a document search is performed using the following formula to extract the top k _dr1 articles.

Where d is an article, t is a keyword extracted from the question sentence, tf (d, t) is the frequency of the keyword t appearing in the article d, df (t) is the frequency of the keyword t appearing, and N Is the total number of articles, length (d) is the length of article d, and Δ is the average of article length. k _t and k ₊ are constants determined by experiments. This equation is based on the equation of Robertson's Okapi weighting (see, for example, the following document (1) and document (2)), and is often used in information retrieval (for example, the following document (3) Reference (4)). However, in the question and answer so important that the keywords are many types to match, using a large value as the value of k _t.

Reference (1): SE Robertson and S. Walker, Some simple effective approximations to the 2-poisson model for probabilistic weighted retrieval, Proceedings of the Seventeenth Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, (1994).
Reference (2): SE Robertson, S. Walker, S. Jones, MM Hancock Beaulieu, and M. Gatford, Okapi at trec-3, TREC-3, (1994).
Reference (3): Masaki Murata, Kiyotaka Uchimoto, Hiromi Osaku, Ma Aoi, Masao Uchiyama, Hitoshi Isahara, Information Retrieval using Location Information and Field Information, Journal of the Language Processing Society, Vol. 7, no. 2 (2000).
Reference (4): Masaki Murata, Qing Ma, and Hitoshi Isahara, High performance information retrieval using many characteristics and many techniques, Proceedings of the Third NTCIR Workshop (CLIR), (2002).
Next, the article is _reranked by the following formula, and the top k _dr2 articles are taken out.

However, T is a set of keywords, dist (t1, t2) is the distance between the keywords t1 and t2, and for convenience t1 = t2, dist (t1, t2) = 0.5. _wdr2 is a function of t2 and is determined by experiment.

  In general, in a question answering system, an article is divided into small units such as paragraphs in order to guarantee that a plurality of keywords extracted from a question sentence appear nearby. Since the formula for scoring when keywords are near by reranking is used, there is no need to divide the article, and the article can be used for document search as it is. In this document search, the top 20 articles are extracted and used for extracting the next solution.

(Solution extraction)
A series of nouns and unknown words are extracted from articles obtained by document search, and they are used as solution candidates. Each candidate is given two scores, Score _near (c) based on the proximity of the solution candidate and the keyword, and Score _sem (c) based on whether or not the semantic constraint of the solution expression is satisfied. The candidate with the largest sum is taken as the solution.

Score _near (c) is given by the following equation.

However, c is a solution candidate, and _wdr2 is a function determined by experiment.

The score Score _sem (c) based on the semantic constraint of the solution expression is given by a rule created manually. In Non-Patent Document 1, 45 rules were created. Some of them are shown below.
• 1000 is given to a candidate that matches the estimated solution expression (person name, place name, etc.). For example, a specific expression extraction technique based on SVM is used as a method for specifying whether a candidate for a solution is a person name or a place name. An example of the specific expression extraction technique will be described later.
When the solution expression is “country name”, 1000 is given when the solution candidate is the country name.
When the question sentence is “what + noun X”, 1000 is given to the candidate having the noun X at the end.

In the experiment in Non-Patent Document 1, the following score addition method is used.
(1) Original method A method that does not add points.
(2) Simple addition method The candidate scores of solutions extracted from a plurality of articles are added, and a solution is output based on the total score obtained by adding the scores as they are.
(3) Gradual addition point method Candidate scores taken from multiple articles are added. In this method, the score of the i th candidate is multiplied by the value of k ^{(i−1) and then} the score is added. That is, the addition result is expressed by the following equation.

However, Score _decreased is a number below 1000 of the final value after addition, and score _original (i) is a number below 1000 of the original value. n is the number of appearances of the same solution candidate obtained from a plurality of articles having the same number in the digits above 1000. k is a constant determined by experiments.
(4) Fusion method This method is a combination of the original method, simple addition method, and gradual addition point method. In this method, first, which of the methods gives the best accuracy is checked with the learning data, and the method with the highest accuracy is used to solve the problem.

  This method has the effect of improving accuracy through fusion and the effect of being able to make a fair evaluation.

  Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a diagram illustrating an example of a configuration of a question answering apparatus according to the first embodiment of the present invention. In the first embodiment, for example, when the first keyword “Japan” and the second keyword “area” are input, the first keyword “Japan” is selected based on the first keyword “Japan”. Increase to three third keywords: “Japan”, “USA” and “Germany”. Further, based on the second keyword “area”, the second keyword is increased to, for example, three fourth keywords “area”, “population”, and “latitude”. For each question such as "What is the area of Japan?", "What is the population of America?", "What is the latitude of Germany?" The answer candidate is obtained by using a machine learning technique and output as an answer.

  The question answering apparatus 1 is an apparatus that increases an input keyword and outputs an answer to a question configured by the increased keyword. The question answering apparatus 1 includes a keyword input unit 11, a keyword increase unit 12, a question creation unit 13, an answer candidate extraction unit 14, an answer table output unit 15, and a keyword extraction database (DB) 16. In the figure, reference numeral 17 denotes a learning database (DB) in which results of machine learning (learning results) by an answer candidate extraction unit 14 described later are accumulated.

  A keyword is input to the keyword input unit 11. For example, the first keyword “Japan” and the second keyword “area” are input. The keyword increasing unit 12 extracts keywords in the same field as the input keywords from the keyword extraction DB 16 by using a keyword extraction technique described later. As a result of keyword extraction, the total number of keywords increases. For example, the keyword increasing unit 12 outputs a third keyword that is larger than the number of first keywords based on the input first keyword. In addition, for example, the keyword increasing unit 12 outputs a fourth keyword that is larger than the number of second keywords based on the input second keyword.

  The question creating unit 13 creates a plurality of questions configured by the third keyword and the fourth keyword, which are increased keywords. For example, if one of the increased third keywords is “X” and one of the fourth keywords is “Y”, the question “What is Y of X?” create.

  The answer candidate extraction unit 14 extracts answer candidates that are answer candidates for the question created by the question creation unit 13 by a machine learning method described later. The answer table output unit 15 outputs a table in which each extracted answer candidate is associated with a question as an answer table. For example, an answer table as shown in FIG. 2 is output.

In the example of the answer table shown in FIG. 2, for example, as an answer corresponding to the question “What is the area of Japan?”, A row corresponding to the data item “Japan” in the answer table and a column corresponding to the data item “Area” “A1” (km ² ) is stored in the cell where “is crossed”, and as an answer corresponding to the question “What is the American population?”, The row corresponding to the data item “USA” and the data item “population” in the answer table “B2” (10,000 people) is stored in the cell where the corresponding column intersects.

In the embodiment of the present invention, an expression obtained by converting the extracted answer candidate into a predetermined unit (for example, km ² ) may be stored in the answer table, or the answer may be kept in the unit of the extracted answer candidate. It may be stored in a table.

Of course, the answer table output in the present invention is not limited to the one shown in FIG. 2. For example, “Area of Japan? → A1 (km ² )”, “American population? → B2 (10,000 It is also possible to adopt a configuration in which an answer table in a format in which data in which each answer candidate is associated with a question by an arrow, such as “person”, is stored as data in each row of the answer table.

  The keyword extraction DB 16 is a database that stores a certain amount of document data. The keyword extraction DB 16 stores, for example, data (a fixed amount of document data) extracted from newspapers, magazines, Web data (data on the network), and the like. The learning DB 17 stores learning results described later. For example, what kind of answer (“correct answer” or “incorrect answer”) is likely to occur when the set of features extracted from the question “Tokyo” is the question “What is the capital city of Japan?” Accumulated as a learning result.

  The keyword increasing unit 12 includes a pattern extracting unit 121 and a keyword extracting unit 122. The pattern extraction unit 121 performs a full-text search on the keyword input DB 16 for keywords input to the keyword input unit 11 and extracts patterns that appear around a plurality of input keywords. The keyword extraction unit 122 performs a full text search in the keyword extraction DB 16 for the pattern extracted by the pattern extraction unit 121, and outputs an expression extracted by the pattern as a keyword.

  In the embodiment of the present invention, the question creating unit 13 is omitted from the configuration shown in FIG. 1, and the answer candidate extracting unit 14 uses the machine learning technique to output the third keyword output by the keyword increasing unit 12. And a fourth keyword may be used to extract and output answer candidates for a question. In other words, the answer candidate extraction unit 14 learns what kind of problem and what kind of answer is obtained using a set of questions prepared in advance and answers to the problem, and based on the learning result. Alternatively, a configuration may be adopted in which answer candidates for the question configured by the third keyword and the fourth keyword output by the keyword increasing unit 12 are extracted.

Below, the keyword extraction process by the keyword increase part 12 is demonstrated. The pattern extraction unit 121 performs a full text search for a small number of input keywords in the keyword extraction DB 16 and extracts patterns c _i that appear around the small number of keywords. The keyword extraction unit 122 performs a full text search on the extracted pattern c _i in the keyword extraction DB 16 and extracts the expression exp extracted by the pattern c _i , and at the same time, extracts the expression exp based on the value of Score (score; evaluation value). Sort in descending order and output as keywords.

  In the embodiment of the present invention, the keyword extraction unit 122 may take a configuration in which a predetermined number of extracted expressions exp are extracted in order from the one with the highest Score value and output as a keyword. In addition, the keyword extraction unit 122 may take a configuration in which an extracted expression exp having a score value equal to or greater than a predetermined threshold is output as a keyword.

(Description of pattern example)
Hereinafter, the pattern extracted by the pattern extraction unit 121 will be described by taking as an example the case where the pattern is the country name A.

・ Input keywords:
Japan
China
Korea
Thailand
Korea ・ Example of extraction pattern (1): (Used at both ends, slow speed but good performance)
Sun, A army
A female
Japan is A
[A communication
Ministry. Ambassador to A ・ Example of extraction pattern (2): (Only one is used, one is Hiragana, and the speed is fast)
[..A country].

word. A
[..A country] side
[..A country]
A word input However, [..A ..] means that country name A itself matches. For example, [Country A] means that the end of the matched term is the country.

(Specific explanation of keyword extraction)
As a small number of keywords to be input, for example, it is assumed that five terms that are likely to be famous from those with a high frequency in daily newspapers are selected as representative forms of evaluation data. Further, for example, the CD Mainichi Newspaper (Mainichi Newspaper recorded on a compact disc), 1999-2000, is used as the keyword extraction DB 16. An example of the extraction procedure is as follows.

(1) a small number of the plurality of keywords and full-text search on the keyword extraction DB 16, when the appearance pattern around the plurality of keywords are extracted as c _i (pattern appearing around the keywords only that keyword (one) Is not extracted). (Definitions of patterns that appear in the vicinity are made as appropriate). For example, when using three character strings before and after (left and right) keywords as the patterns appearing in the vicinity, there are cases where there are one, two, and three characters respectively before and after, so there are nine patterns with one keyword. It will be possible. It can also be a pattern including a keyword (self).

(2) Next, the extracted pattern c _i is full-text searched in the keyword extraction DB 16 to extract the expression exp extracted from the pattern c _i .

  (3) The extracted expressions exp are sorted in descending order of Score values and output as keywords.

  There are the following as Score.

・ Method 1 (decision list method)
Method 1 is a method of using p _i of the pattern with the largest p _i in the pattern c _i as the score of the extracted expression exp. Here, p _i is the ratio of input keywords in the expression exp extracted with the pattern c _i (the probability, that is, the certainty level).

For example, as a result of the full text search for the pattern c _{1 in the} keyword extraction DB 16, five exps from exp1, exp2, exp3, exp4, and exp5 are extracted, and among these five exps, three from exp1 to exp3 Is an input keyword, p ₁ is 3/5.

・ Method 2 (Bayes method)
Method 2 uses a score obtained by multiplying all the patterns c _i by p _i as the score of the extracted expression exp.

In practice, the possibility of p _i = 0 is high, so in the embodiment of the present invention, instead of the above formula (8), the following formula (9)
Π ((1−Δ) / Δ * p _i +1) Equation (9)
It is also possible to take a configuration that uses. Here, Δ is a constant of a minute value, for example, 0.0001 is used.

For example, exp which calculates the Score is the case can not be acquired by the search process for the pattern c _i, as p _i = 0, is calculated using the above equation (9).

・ Method 3 (method based on similarity)
Method 3 uses the number (total number) of extracted patterns as the score of the extracted expression exp. That is, the score is increased as the number of patterns extracted is increased.

・ Method 4 (Refer to Research (1) below)
Method 4 as Score of the extracted expression exp, is to use a number of the extracted pattern plus the weight of p _i.

Here, f _i is the number of input keywords in which the pattern c _i appears.

  Study (1): Ellen Riloff and Rosie Jones "Learning dictionaries for information extraction by multi-level bootstrapping" Proceedings of AAAI-99, (1999).

・ Method 5 (Refer to the following document (5))
Method 5 uses at least one value that is likely to be the score of the extracted expression exp.

In the above equation (12), by multiplying (1− _pi ) which is not certain, no one is uncertain, and when subtracting this from 1, at least one becomes uncertain.

  Reference (5): Masaki Murata, Hitoshi Isahara "Automatic Acquisition of Knowledge about Paraphrases Based on Collation of Synonymous Texts" IPSJ SIG 2001-NL-142, (2001).

  In the above methods 1, 2, 4, and 5, when the score is the same, the score is sorted by the score of the method 3, and in the method 3, the score is sorted by the score of the method 5.

  FIG. 3 shows a predetermined number of types of correct data prepared in advance for keyword extraction results obtained by using a combination of 1 to 3 characters including either the left or the beginning of the keyword as a pattern and that on the right side. It is a figure which shows an example of the result of having used and calculated | required the precision and the recall. Here, as the correct answer data, for example, a data example as shown in FIG. 4 is prepared (FIG. 4 shows an example of country name data, the country name is divided into rows for each country, and the head of the line is represented. Other than that, it is stored in the same line as a variant of the representative form). A variety of correct data having a data format similar to the data format shown in FIG. 4 is prepared, for example, data related to satellites, holidays, solar system planets, world heritage, etc. in addition to country name data.

  In FIG. 3, AP is an average of average precision used in information retrieval (refer to the following document (6)), and is an average of relevance ratios obtained every time correct articles are taken from the top. In the case of the contents of the present application, it is an average of the relevance ratios obtained every time the correct keyword is taken from the top (however, the input keyword is excluded from the correct keyword).

  Reference (6): Masaki Murata, Ma Aoi, Kiyotaka Uchimoto, Hiromi Osaku, Masao Uchiyama, Hitoshi Isahara "Information Retrieval Using Location Information and Field Information" Journal of Language Processing Society, Vol.7, No.2, ( 2000).

  RP is an average of r-precision, and is a ratio of including correct articles when searching for the number of correct articles. In the case of the contents of the present application, it is a ratio in which correct keywords are included when only correct keywords are extracted. Note that the relevance rate is the same as the correct answer rate, and is the rate at which correct keywords are included. TP is the average accuracy of the top five.

(Explanation of extraction method based on constraints)
(A) Method of Using Character Type and KR In the example shown in FIG. 3, a method of further using character type and KR was used as the extraction method. Here, the character types are kanji, katakana, hiragana, symbols, numbers, and the like. For example, in English, alphabets, numbers, symbols, and whether the beginning of a word is capitalized or the like.

  The method using character types is a method that does not extract expressions including character types that were not found in a small number of input keywords (five in this example). For example, when there are no hiragana characters in five input keywords, an expression including hiragana characters is not extracted.

In the method using KR, p _i is replaced with p _i * f _i / n _i . The advantage of this method is that the certainty factor can be changed by the value of f _i / n _i even if p _i is the same. However, n _i is the number of input keywords, and in the case of Method 3, in the case of KR, 1 is replaced with f _i . In the evaluation, keywords were not included in the extracted results. In addition to the character type method, there are also the following methods.

(B) Method based on part of speech In the method based on part of speech, for example, if there is only a noun in the input expression, the expression other than the noun is omitted at the time of output, and if the input expression only has an adjective, the expression other than the adjective is omitted at the time of output. That's it. Furthermore, when the expression is composed of a plurality of words, the part of speech information of the last word (morpheme) can be used.

(Description 1 by example)
If the input keyword is:
“Fun” “sad” “happy” “very happy” “very sad”
If the following is obtained as an extract:
"Very""New""Beautiful""Verybeautiful""Verydifficult"
Estimate the part of speech of the last word in the expression of the extract, and in the above input keyword, the part of speech of the last word in the extract is not “adjective” because the last word has only “adjective” , Excluding adverbs ("very").

(Description 2 by example)
If the input keyword is:
"Fun""joy""sadness""sad"
In the above input keyword, when there are plural types such as “adjective” and “noun”, those parts of speech are output, and expressions other than those parts of speech are not output.

  In order to obtain part-of-speech information such as the estimation of the part-of-speech of the last word (morpheme) as described above, the following morpheme analysis system (morpheme analysis means) is required.

Description of the morphological analysis system In order to divide Japanese into words, it is necessary to use the morphological analysis system in the keyword extraction unit 122. Here, ChaSen will be explained (the morphological analysis system tea bowl developed at Nara Institute of Technology. Http://chasen.aist-nara.ac.jp/index.html.jp).

  This splits the Japanese sentence and also estimates the part of speech of each word. For example, if “go to school” is entered, the following results can be obtained.

School Gacco School Noun-General To He To particle-Case particle-General Go Iku Go Verb-independence
In this way, each line is divided so that one word is included, and reading and part-of-speech information are given to each word.

(C) Method based on common partial character string For example, when input expressions all have the same common end expression “Shi”, an expression that does not have “Shi” is omitted at the time of output. This can be done not only at the end but also at the top character string.

(Description by example)
If the input keyword is:
"Sad""fun""happy"
If the following is extracted:
"Joy""Sorrow""Beautiful""New"
Since the common partial character string of the input keyword is “Shi”, “Joy” and “Sadness” that do not have “Shi” are deleted and output.

(D) Specification of constraint by user In the above description, the method of automatically obtaining the constraint from the input expression has been described. However, this constraint can be made to be allowed by the user. For example, when the user selects the option of “Kanji only”, the output using a character type other than Kanji can not be output. In addition, when the user selects the option “Shi” at the end, it is possible to prevent the output not having “Shi” at the end in the output. Furthermore, when the user selects the option that the part of speech is a noun, the output is made so that expressions other than the noun are not output.

Next, a question created by the question creation unit 13 or a question configured by the output keyword output by the keyword increase unit 12 when the question answering apparatus 1 does not include the question creation unit 13 (“X” The process of extracting answer candidates for “Y?”) Will be described. The answer candidate extraction unit 14 extracts answer candidates using a machine learning technique.
(Machine learning method)
The machine learning method prepares many sets of question-answer pairs, learns with it, learns what kind of answer is the problem, and uses the learning result to create a new problem. This is a method that makes it possible to guess the answer (for example, see the following document (7)).

Reference (7): Masaki 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
In order to convey the problem situation to the machine, what kind of problem is required, features (elements constituting the problem with information used for analysis) are 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 answer “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-answer pairs, learns with it, learns what type of feature set the answer will be, and uses the learning results. This is a method of extracting a set of features from a new problem and guessing an answer in the case of the feature.

  First, general machine learning techniques will be described. As a method of machine learning, generally, methods such as a k-nearest neighbor method, a simple Bayes method, a decision list method, a maximum entropy method, and a support vector machine method are used.

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

  The Simple Bayes method is a method of estimating the probability of each classification based on Bayes' theorem and determining the classification having the highest probability value as a classification destination.

  In the simple Bayes method, the probability of outputting the classification a in the context b is given by the following equation (13).

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 the 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 . When the value obtained by performing maximum likelihood estimation as P (f _i | a) is used, the value often becomes zero, and it may be difficult to determine the classification destination because the value of Expression (14) is zero. Therefore, smoothing is performed. Here, a smoothed image using the following equation (15) is used.

Here, freq (f _i , a) means the number of cases having the feature f _i and the classification a, and freq (a) means the number of cases having the classification a.

  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

p (a | b) = p (a | fmax) Equation (16)
However, fmax is given by the following equation.

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

In the maximum entropy method, when a set of preset features f _j (1 ≦ j ≦ k) is F, an expression (entropy) that satisfies a predetermined conditional expression (equation (18)) below ( 19) This is a method of obtaining a probability distribution p (a, b) when maximizing 19) and determining a classification having the largest probability value among the classification probabilities obtained according to the probability distribution.

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

  Expression (18) is to obtain the expected value of the frequency of the output and feature pair by multiplying the probability p and the function g meaning the appearance of the pair of output and feature, and the expected value in the known data on the right side And the expected value calculated based on the probability distribution calculated on the left side is the constraint, and entropy maximization (smoothing of the probability distribution) is performed to determine the probability distribution of the output and the context. Details of the maximum entropy method are described in the following documents (8) and (9).

Reference (8): Eric Sven Ristad, Maximum Entropy Modeling for Natural Language, (ACL / EACL Tutorial Program, Madrid, 1997
(9): Eric Sven Ristad, Maximum Entropy Modeling Toolkit, Release 1.6beta, (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. 27 shows the concept of margin maximization in the support vector machine method. In FIG. 27, 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. FIG. 27A is a conceptual diagram when the interval between the positive example and the negative example is small (small margin), and FIG. 27B 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 and negative examples, the larger the interval (margin) between the positive and negative examples in the learning data, the less likely it is to make an incorrect classification with open data. As shown in FIG. 27B, a hyperplane that maximizes this margin is obtained, and classification is performed using it.

  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 (kernel) The one that has been introduced) is used.

  This extended method is equivalent to classification using the following discriminant function, 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 ∈ {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 (21) under the constraints of Expression (22) and Expression (23).

  The function K is called a kernel function, and various functions are used. In this embodiment, the following polynomial is used.

K (x, y) = (x · y + 1) d Equation (24)
C and d are constants set experimentally. In a specific example to be described later, C is fixed to 1 through all the processes. 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 Expression (20) is normally calculated using only this case. That is, only actual cases called support vectors are used for actual analysis.

  The details of the extended support vector machine method are described in the following literature (10) and literature (11).

Reference (10): Nello Cristianini and John Shawe-Taylor, An Introduction to Support Vector Machines and other kernel-based learning methods, (Cambridge University Press, 2000)
Reference (11): Taku Kudoh, Tinysvm: Support Vector machines, (http://cl.aist-nara.ac.jp/taku-ku//software/Tiny SVM / index.html, 2000)
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 pairwise method, in the case of data having n classifications, every pair (n (n-1) / 2) of two different classification destinations is generated, and a binary classifier indicates which is better for each pair. That is, it is obtained by the support vector machine method processing module and finally obtains the classification destination by majority decision of the classification destination by n (n−1) / 2 binary classification.

  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. Look at how the three support vector machines estimate the binomial candidates to be estimated. Of the three support vector machines, the other classification destination and the support vector This is a method for obtaining a classification destination of the one farthest from the separation plane of the machine. For example, when a candidate is farthest from the separation plane in the support vector machine created by the learning process of the “classification destination a and other” group, the candidate classification destination is estimated as a.

Next, a specific method for extracting answer candidates in the embodiment of the present invention will be described.
(Answer candidate extraction method 1)
<Problem structure>
In advance,
The question “Answer“ Z1 ”in the question“ What is Y1 of X1? ”” --- Answer “Correct”
Answer “Question“ What is Y2 of X2? ”“ Z2 ”” --- Answer “Correct”
Answer “Z3” with the question “What is Y3 of X3?” --- Answer “Incorrect”
Create many pairs of questions and answers.

Also, for example, the following features expressing the above problem are prepared.
-Xi, Yi, Zi words themselves-Semantic class of Xi, Yi, Zi words-Number of articles searched by Xi, Yi-Number of articles where Zi exists in articles searched by Xi, Yi-Xi, Yi are close The number of articles that appeared (within a certain number of words) · The number of articles that appeared in close proximity (within the range of a certain number of words) · Appeared most frequently in articles searched with Xi, Yi Whether or not Zi matches the word that has been searched for · whether or not Zi matches the word that appears jth most frequently in the article searched by Xi and Yi • For example, the answer candidate extraction unit 14 uses the newspaper as a keyword Take out an article group that seems to be written an answer from the document data group (not shown) such as article data / encyclopedia data, take out the language expression of the taken out article group as an answer candidate, and take out the answer candidate The Previous order (e.g., Score _near (c)) when rearranged in, whether the highest candidate and Zi of the order match, also, or above if j th candidate and Zi of the order match By processing, a large number of sets of feature sets and answer sets are prepared.

Here, Score _near (c) used as the priority order is described in Non-Patent Document 1 described above, and indicates a score based on the answer candidates and the proximity of the keywords.

  Next, semantic classes will be described. In general, there is a table describing what semantic class each word has, and by using the table, the semantic class of the word can be obtained. For example, there is a classification vocabulary table. In the classification vocabulary table, a word is expressed by a 10-digit number called a classification number, and a word having a similar similarity is a similar word. For example, the first 3 or 5 digits of this number are used as a word semantic class. For example, the classification number of “villager” is 1230102050, which indicates that the word belongs to 123 (semantic class such as race, citizen, social hierarchy) and 12301 (semantic class such as citizen, resident).

  Once you define the problem structure and features, you can use machine learning techniques. That is, the answer candidate extraction unit 14 learns what kind of answer is obtained when using a set of prepared features and answers, and uses the learning result to create a new Answers estimated in the case of a set of features for a simple problem are extracted as answer candidates.

<Specific examples of problems and features>
Examples of problems:
Answer "Question" What is the capital of Japan? "" Tokyo "--- Answer" Correct "
The question “Answer“ Osaka ”in the question“ What is the capital of Japan? ”” --- Answer “Incorrect”
The question “Answer“ Pan ”in the question“ What is the capital of Japan? ”——— Answer“ Incorrect ”
Specific examples of features:
The question “Question“ What is the capital of Japan? ”Answers“ Tokyo ”” --- If the answer is “correct” • Xi word itself: Japan • Yi word itself: capital • Zi word itself: Tokyo • Meaning of Xi Class: 12590 (class of place name)
・ Yi meaning class: 12540 (city village class)
・ Zi semantic class: 12590 (class of place name)
(The first five digits of the classification vocabulary table are used as the semantic class.)
-Number of articles searched by Xi, Yi: Number of articles including Japan and the capital. For example 1000
-Number of articles in which Zi exists in articles searched by Xi, Yi: Number of articles including Japan, the capital, and Tokyo. 100 for example
Number of articles in which Xi and Yi appear close to each other (within a certain number of words): For example, the number of articles in which Japan and the capital are within 10 words. For example 500
-Number of articles in which Xi, Yi, Zi appear close to each other (within a certain number of words): For example, the number of articles in which Japan, the capital, and Tokyo are within 10 words. For example 50
・ Whether or not Zi matches the word that appears most frequently in the articles searched by Xi and Yi: For example, here, the word is particularly limited to a noun, and as a noun, “ko” is the most frequent. ”And“ Tokyo ”do not match, so this feature is“ No ”.
-Whether the j-th most frequently appearing word in articles searched with Xi, Yi matches Zi: For example, here, the word is particularly limited to a noun, and as a noun, "Tokyo" is the second most frequent Then, the feature in the case of j = 2 is “Yes”.
Using Xi and Yi as keywords, for example, the answer candidate extraction unit 14 extracts a group of articles likely to be written from a group of document data such as newspaper article data / encyclopedia data, and language representation of the extracted article group And when the retrieved answer candidates are rearranged in the priority order (for example, Score _near (c)), whether or not Zi matches the candidate with the highest order, and Whether the j-th candidate in the order matches Zi. For example, if the first candidate is “that” and the second candidate is “Tokyo,” “whether Zi matches the highest candidate in the order. “No”, and “whether Zi matches the j-th candidate in that order” is “Yes” when j = 2.

When learning with more cases, for example, the answer candidate extraction unit 14
Yi word itself: Semantic class of capital Zi: 12590 (place name class)
"If the Zi matches the highest candidate in that order" is "Yes"
Or, if j = 2 of “whether Zi matches the j-th candidate in that order” is “yes”,
Answer “Correct”
Learn to be
Yi word itself: Semantic class of capital Zi: except 12590 (class of place name)
Answer "Incorrect"
Yi's word itself: the capital “whether Zi matches the highest candidate in that order” is “No” and
If j = 2 to 10 in “whether Zi matches the jth candidate in that order” or not,
Answer "Incorrect"
To learn.

The learning results are accumulated in the learning DB 17 by the answer candidate extraction unit 14. And the answer candidate extraction part 14 uses the learning result information in learning DB17, for example, a new question (Namely, the question comprised by the output keyword output by the keyword increase part 12):
Question "What is the capital of France?"-For the answer "Paris"
Yi word itself: Semantic class of capital Zi: 12590 (class of place name)
“Whether Zi matches the highest candidate in that order” is “Yes” or
“Yes” when j = 2 of “whether Zi matches the j-th candidate in that order”
Therefore, it is judged as “correct”.

Also new issues:
Question "What is the capital of France?"-For the answer "Signal", using the above learning results,
Since the word Yi itself is other than the semantic class of the capital Zi: 12590 (class of place name), it is determined as “incorrect”.

  Here, not only the answer is obtained by the machine learning method, but also the degree of the correct answer and the degree of the incorrect answer can be obtained at the same time.

  That is, the answer candidate extraction unit 14 first uses what set of feature sets and answer sets to determine what kind of answer (correct answer or incorrect answer) at what kind of feature set. That is, machine learning is performed, and information indicating what kind of answer is obtained in what feature set is stored in the learning DB 17 as learning result information. Then, the answer candidate extraction unit 14 uses a new question (a question composed of the output keywords output by the keyword increase unit 12) to extract a set of features from a question newly created by the answer candidate extraction unit 14). Based on the learning result information stored in the learning DB 17, the level of what kind of answer is likely to be obtained in the case of the extracted feature set, that is, “it is likely to be correct” is obtained.

  Then, for example, the answer candidate extraction unit 14 sends the “answer” in the question (question-answer pair) to the answer table output unit 15 as an answer candidate when the degree of “probably correct” is the highest. Output. The answer table output unit 15 outputs a table in which each answer candidate is associated with a question as an answer table. In the embodiment of the present invention, for example, a predetermined number of questions may be selected in descending order of “prone to be correct”, and “answers” in the selected questions may be set as answer candidates. Further, for example, it is possible to select a question whose degree of “it is likely to be correct” is a predetermined threshold or more and set “answer” in the selected question as a candidate answer. In addition, for example, a predetermined ratio (for example, 90%) of the degree of the problem having the greatest degree of “prone to be correct” is set as a threshold, and the problem in which the degree of “prone to be correct” is equal to or greater than this threshold is set. The “answer” in the selected question may be selected and set as the answer candidate.

  The method of obtaining the degree of “prone to be correct” is different depending on various methods used by the answer candidate extraction unit 14 as a machine learning technique.

  For example, in the embodiment of the present invention, when the answer candidate extraction unit 14 uses the k-nearest neighbor method as the machine learning method, the answer candidate extraction unit 14 sets a large number of sets of feature sets and answer sets prepared above. Is used to define the degree of similarity between problems based on the proportion of overlapping features (the proportion of how many of the same features) in the feature set. Then, the answer candidate extraction unit 14 stores the defined similarity and problem (and answer) in the learning DB 17 as learning result information.

  And the answer candidate extraction part 14 is comprised by the output keyword output by the keyword increase part 12 when the question produced by the question creation part 13 or when the question answering apparatus 1 takes the structure which is not provided with the question creation part 13 is taken. When a new problem is created using the question to be asked, the similarity and the problem stored in the learning DB 17 are referred to, and the k questions (and answers) in descending order of the similarity to the newly created problem. Are selected from the questions (and answers) stored in the learning DB 17, and the classification destination (correct answer or incorrect answer) determined by the majority decision in the selected k questions is set as the answer to the newly created problem. In the case of using the k-nearest neighbor method, the degree of “it is likely to be a correct answer” is the number of votes of the majority vote in the selected k problems, that is, the number of votes obtained by the classification “correct answer”.

  The answer candidate extraction unit 14 outputs the “answer” in the question (question-answer pair) to the answer table output unit 15 as an answer candidate when the degree of “it is likely to be correct” is the highest. In the embodiment of the present invention, for example, a predetermined number of questions may be selected in descending order of “prone to be correct”, and “answers” in the selected questions may be set as answer candidates. Further, for example, it is possible to select a question whose degree of “it is likely to be correct” is a predetermined threshold or more and set “answer” in the selected question as a candidate answer. In addition, for example, a predetermined ratio (for example, 90%) of the degree of the problem having the greatest degree of “prone to be correct” is set as a threshold, and the problem in which the degree of “prone to be correct” is equal to or greater than this threshold is set. The “answer” in the selected question may be selected and set as the answer candidate.

  For example, in the embodiment of the present invention, when the answer candidate extraction unit 14 uses the Simple Bayes method as a machine learning method, for example, the answer candidate extraction unit 14 includes the set of the prepared feature and the answer. Are stored in the learning DB 17 as learning result information.

  The answer candidate extraction unit 14 includes a question created by the question creation unit 13 or an output keyword output by the keyword increase unit 12 when the question answering apparatus 1 does not include the question creation unit 13. When a new problem is created using a question, a set of features is extracted from the newly created problem. Then, the answer candidate extracting unit 14 is a case of a feature set extracted from a newly created problem based on a Bayes' theorem based on a set of answers and feature sets stored in the learning DB 17. The probability of becoming each of the categories is calculated, and the category having the largest probability value is set as the answer to the problem. When the Simple Bayes method is used, the degree of “prone to be correct” is the probability of being classified as “correct”.

  The answer candidate extraction unit 14 outputs the “answer” in the question (question-answer pair) to the answer table output unit 15 as an answer candidate when the degree of “it is likely to be correct” is the highest. In the embodiment of the present invention, for example, a predetermined number of questions may be selected in descending order of “prone to be correct”, and “answers” in the selected questions may be set as answer candidates. Further, for example, it is possible to select a question whose degree of “it is likely to be correct” is a predetermined threshold or more and set “answer” in the selected question as a candidate answer. In addition, for example, a predetermined ratio (for example, 90%) of the degree of the problem having the greatest degree of “prone to be correct” is set as a threshold, and the problem in which the degree of “prone to be correct” is equal to or greater than this threshold is set. The “answer” in the selected question may be selected and set as the answer candidate.

  Further, for example, in the embodiment of the present invention, when the answer candidate extraction unit 14 uses a decision list method as a machine learning method, for example, the answer candidate extraction unit 14 has a feature and a classification destination for a prepared problem. Is stored in the learning DB 17 in a predetermined priority order. The answer candidate extraction unit 14 includes a question created by the question creation unit 13 or an output keyword output by the keyword increase unit 12 when the question answering apparatus 1 does not include the question creation unit 13. When a new problem is created using a question, a set of features is extracted from the newly created problem.

  Then, the answer candidate extraction unit 14 compares the feature extracted from the newly created problem with the rule feature in descending order of the priority stored in the list stored in the learning DB 17, and finds the rule with the matching feature. The classification destination is the answer to the question. When the decision list method is used, the degree of “it is likely to be correct” is a predetermined priority or a numerical value or scale corresponding to it.

  The answer candidate extraction unit 14 outputs the “answer” in the question (question-answer pair) to the answer table output unit 15 as an answer candidate when the degree of “it is likely to be correct” is the highest. In the embodiment of the present invention, for example, a predetermined number of questions may be selected in descending order of “prone to be correct”, and “answers” in the selected questions may be set as answer candidates. Further, for example, it is possible to select a question whose degree of “it is likely to be correct” is a predetermined threshold or more and set “answer” in the selected question as a candidate answer. In addition, for example, a predetermined ratio (for example, 90%) of the degree of the problem having the greatest degree of “prone to be correct” is set as a threshold, and the problem in which the degree of “prone to be correct” is equal to or greater than this threshold is set. The “answer” in the selected question may be selected and set as the answer candidate.

  Also, for example, in the embodiment of the present invention, when the answer candidate extraction unit 14 uses the maximum entropy method as a machine learning method, for example, the answer candidate extraction unit 14 selects a classification that can be an answer to a prepared question. A probability distribution consisting of a set of features when satisfying a predetermined conditional expression and maximizing an expression showing entropy and a binary term of a classification that can be an answer is obtained and stored in the learning DB 17.

  And the answer candidate extraction part 14 is comprised by the output keyword output by the keyword increase part 12 when the question produced by the question creation part 13 or when the question answering apparatus 1 takes the structure which is not provided with the question creation part 13 is taken. When a new question is created using the question to be asked, the probability probability of the classification that can be the answer to the set of features of the new problem is obtained using the probability distribution stored in the learning DB 17, and the largest probability value A category that can be an answer having the following is specified, and the specified category is set as an answer to the problem.

  That is, when the maximum entropy method is used, the degree of “prone to be correct” is the probability of being classified as “correct”.

  The answer candidate extraction unit 14 outputs the “answer” in the question (question-answer pair) to the answer table output unit 15 as an answer candidate when the degree of “it is likely to be correct” is the highest. In the embodiment of the present invention, for example, a predetermined number of questions may be selected in descending order of “prone to be correct”, and “answers” in the selected questions may be set as answer candidates. Further, for example, it is possible to select a question whose degree of “it is likely to be correct” is a predetermined threshold or more and set “answer” in the selected question as a candidate answer. In addition, for example, a predetermined ratio (for example, 90%) of the degree of the problem having the greatest degree of “prone to be correct” is set as a threshold, and the problem in which the degree of “prone to be correct” is equal to or greater than this threshold is set. The “answer” in the selected question may be selected and set as the answer candidate.

  Further, for example, in the embodiment of the present invention, when the answer candidate extracting unit 14 uses the support vector machine method as a machine learning method, for example, the answer candidate extracting unit 14 may classify as a possible answer to a prepared question. Is divided into positive and negative examples, and the space between the positive and negative examples of the problem is defined on a space whose dimension is a set of problem features according to a predetermined execution function using a kernel function. The hyperplane for maximizing and dividing the positive example and the negative example by the hyperplane is obtained and stored in the learning DB 17.

  And the answer candidate extraction part 14 is comprised by the output keyword output by the keyword increase part 12 when the question produced by the question creation part 13 or when the question answering apparatus 1 takes the structure which is not provided with the question creation part 13 is taken. When a new problem is created using the question to be asked, the hyperplane in the learning DB 17 is used, and either the positive example side or the negative example side in the space in which the feature set of the new problem is divided by the hyperplane is used. A classification determined based on the identified result is determined as the answer to the question.

  That is, when the support vector machine method is used, the degree of “it is likely to be correct” is the size of the distance from the separation plane to the positive example space. More specifically, when a problem with a correct answer is a positive example and a problem with an incorrect answer is a negative example, a problem that is located in the space on the positive side with respect to the separation plane is a correct answer. The problem is judged to be a problem, and the problem that the distance from the separation plane is large is more likely to be “correct answer”.

  The answer candidate extraction unit 14 outputs the “answer” in the question (question-answer pair) to the answer table output unit 15 as an answer candidate when the degree of “it is likely to be correct” is the highest. In the embodiment of the present invention, for example, a predetermined number of questions may be selected in descending order of “prone to be correct”, and “answers” in the selected questions may be set as answer candidates. Further, for example, it is possible to select a question whose degree of “it is likely to be correct” is a predetermined threshold or more and set “answer” in the selected question as a candidate answer. In addition, for example, a predetermined ratio (for example, 90%) of the degree of the problem having the greatest degree of “prone to be correct” is set as a threshold, and the problem in which the degree of “prone to be correct” is equal to or greater than this threshold is set. The “answer” in the selected question may be selected and set as the answer candidate.

<Specific example>
For example, a question created by the question creating unit 13 or a question “French capital” composed of an output keyword output by the keyword increasing unit 12 when the question answering apparatus 1 does not include the question creating unit 13. Will be explained.

  First, the answer candidate extraction unit 14 acquires a document including the output keywords “France” and “capital” from a document data group such as newspaper article data / encyclopedia data. To extract the keywords "France" and "Capital" from the question "What is the French capital?", Use morphological analysis technology. When the question answering apparatus 1 omits the question creating unit 13 and the answer candidate extracting unit 14 extracts the answer candidates for the question configured by the output keyword output by the keyword increasing unit 12, the above question is used. It is not necessary to extract the keywords “France” and “Capital” from “What is the French capital?” Using the morphological analysis technology, and the answer candidate extraction unit 14 outputs the output keywords “France”, “ Documents including them are acquired from the document data group using the “capital” as it is.

  The answer candidate extraction unit 14 extracts a linguistic expression in the document including the keywords “France” and “capital” as a candidate for an answer expression of the question. For extraction of this expression, for example, the solution extraction process described in Non-Patent Document 1 is used.

The extracted answer expression candidates are rearranged in the descending order of, for example, the value of Score _near (c), and the top few candidates of the value are extracted, and the candidates (candidate 1, candidate 2,...) Are extracted. about,
The question “Question“ What is the French capital? ”Answered“ Candidate 1 ”]
The question “Question“ What is the French capital? ”Answered“ Candidate 2 ”]
...
Create Here, “candidate 1” and “candidate 2” indicate candidates for the expression “answer” of the above question.

  Apply the machine learning method described above to the created question (question-answer pair), and find the “answer” in the question (question-answer pair) when the degree of “prone to be correct” is the greatest. Then, it is output to the answer table output unit 15 as an answer candidate.

  The answer table output unit 15 stores the answer to the question “What is the French capital?” In the answer table (for example, a row corresponding to the data item “France” and a column corresponding to the data item “Capital”). The corresponding answer candidate is stored in the grid where the crosses.

(Answer candidate extraction method 2)
<Problem structure>
The answer candidate extraction unit 14
Question “Question“ What is Y1 of X1? ”” --- Answer “Location Name”
Problem “Question“ What is Y2 of X2? ”——— Answer“ Place Name ”
Problem “Question“ What's Y3 of X3? ”” --- Answer “Person Name”
Question "Question" What is Y4 of X4? "" --- Answer "Number"
. . .
Create many pairs of questions and answers.
As features,
The meaning classes of the words Xi and Yi themselves may be considered.
Once you define the problem structure and features, you can use machine learning techniques.

<Specific examples of problems and features>
Examples of problems:
Question "Question" What is the capital of Japan? "--- Answer" Location "
Problem “Question“ What is the Japanese Prime Minister? ”——— Answer“ Person Name ”
Problem “Question“ What is the area of Japan? ”——— Answer“ Number ”
Specific examples of features:
Question "Question" What is the capital of Japan? "--- If the answer is" place name "
-Xi word itself: Japan-Yi word itself: capital-Semantic class of Xi: 12590 (class of place name)
・ Yi meaning class: 12540 (city village class)
(Use the first 5 digits of the classification lexicon as the semantic class)
Learning more examples, for example,
Yi's word itself:
Answer “place name”
Learn to be
Yi ’s word itself:
Answer "person name",
Yi ’s word itself:
Answer “number”,
That is, the answer candidate extraction unit 14 learns and accumulates the learning result in the learning DB 17.

And the answer candidate extraction part 14 judges an answer using the learning result accumulate | stored in learning DB17.
For example, a new problem:
The answer to the question "What is the French capital?"
Yi word itself: Since it is a capital, it is determined as “place name”.

<Specific example>
For example, a question created by the question creating unit 13 or a question “French capital” composed of an output keyword output by the keyword increasing unit 12 when the question answering apparatus 1 does not include the question creating unit 13. Will be explained.

First, the answer candidate extraction unit 14 uses the machine learning technique described above,
Regarding the question "What is the French capital?"
The result that the answer is “place name” is acquired.

  The obtained answer “place name” is used as an answer type when extracting answer candidates to be stored in the cells of the answer table.

  That is, the answer candidate extraction unit 14 uses keywords (“France”, “Capital”) that constitute the question “What is the French capital?” Created by the question creation unit 13 from a document data group such as newspaper article data / encyclopedia data. ) And the language expression included in the extracted document that matches the answer type is output as an answer candidate to the answer table output unit 15. When the question answering apparatus 1 does not include the question creating unit 13, the answer candidate extracting unit 14 selects a document including the output keywords (“France” and “Capital”) output by the keyword increasing unit 12 as the document data. Candidate answers to the question “What is the French capital?” Composed of the output keywords “France” and “Capital” for the language expressions that are extracted from the group and included in the extracted document and that match the above answer type Is output to the answer table output unit 15.

  The answer table output unit 15 stores the answer to the question “What is the French capital?” In the answer table (for example, a row corresponding to the data item “France” and a column corresponding to the data item “Capital”). The corresponding answer candidate is stored in the grid where the crosses.

  In the present invention, for example, when extracting an answer candidate for the question “What is Y of X?”, For example, the answer candidate extracting unit 14 does not use a machine learning technique, An article group including the keyword “X” and the keyword “Y” is extracted from a large amount of document data group (not shown) such as encyclopedia data, and appears in the document data group among the linguistic expressions of the extracted article group. It is also possible to adopt a configuration in which the answering frequency is output as answer candidates. Further, in the embodiment of the present invention, it is possible to adopt a configuration in which a predetermined number of linguistic expressions of the extracted article group are extracted in descending order of appearance in the document data group and output as answer candidates. .

  Here, when the answer candidate extraction unit 14 outputs answer candidates using the answer type, the specific expression extraction technique described in the description of Non-Patent Document 1 is used. A specific expression is a linguistic expression that means a specific thing / quantity, such as a numerical name such as a name, a place name, or an organization name, and a numerical expression such as a monetary amount. This is a technology for automatic extraction by a computer. For example, if a specific expression is extracted for a sentence “the Japanese prime minister is Junichiro Koizumi”, the specific expressions “Japan” and “Junichiro Koizumi” are extracted as place names and personal names. In the embodiment of the present invention, the answer candidate extraction unit 14 determines whether or not the extracted unique expression matches the answer type, and outputs the matching specific expression as an answer candidate.

Hereinafter, an example of a general technique for extracting a specific expression will be described.
(1) A method using machine learning There is a method of extracting a specific expression using machine learning (for example, see the following document (12)).

Reference (12): Masayuki Asahara, Yuji Matsumoto, Utilization of Redundant Morphological Analysis in Japanese Named Expression Extraction Information Processing Society of Japan Natural Language Processing Study Group NL153-7 2002
First, for example, the sentence “Japan's prime minister is Mr. Koizumi” is divided into each character, and the correct characters such as B-LOCATION and I-LOCATION are assigned to the divided characters as follows. Set the correct answer. The first column below is each divided character, and the correct tag of each character is the second column.
Sun B-LOCATION
I-LOCATION
O
Neck O
Phase O
Is O
Small B-PERSON
Izumi I-PERSON
O
N
At O
O
. O
In the above, B-? ? ? Is a tag that signifies the start of the type of proper expression below the hyphen. For example, B-LOCATION means the beginning of a unique expression called place name, and B-PERSON means the beginning of a unique expression called person name. I-? ? ? Is a tag that means something other than the beginning of the type of proper expression below the hyphen, and O is something else. Therefore, for example, the character “day” is a character that corresponds to the beginning of the unique name “place name”, and the character “book” is the unique name “place name”.

  In this way, the correct answer of each character is set, learned from such data, this correct answer is estimated with new data, and from this correct answer tag, the beginning of each proper expression and how far it is. Is recognized and the proper expression is estimated.

When learning from the correct data set for each character, the system uses various information in the form of features. For example,
Sun B-LOCATION
Part of
Japan-B Noun-B
Such information is used. Japan-B means the beginning of the word Japan, and noun-B means the beginning of the noun. For recognition of words and parts of speech, for example, morphological analysis by ChaSen described above is used. If ChaSen is used, the input Japanese can be divided into words. For example, ChaSen divides a Japanese sentence and estimates the part of speech of each word as described above. For example, if “go to school” is entered, the following results can be obtained.

School Gacco School Noun-General To He To particle-Case particle-General Go Iku Go Verb-independence
In this way, each line is divided so that one word is included, and reading and part-of-speech information are given to each word.

  For example, in the above-mentioned document (12), as features, characters themselves (for example, “small”), character types (for example, hiragana and katakana), part-of-speech information, tags, Information (for example, “B-PERSON” etc.) is used.

  Learning using these features. Investigate what features appear in the characters that estimate the tag and the surrounding characters, learn what features are likely to appear when the features appear, and use the learning results to create new data Perform tag estimation. For machine learning, for example, a support vector machine is used.

  In addition to the above-described method, there are various methods for extracting the proper expression. For example, there is a technique of extracting a specific expression using a maximum entropy model and a rewrite rule (see Document (13)).

  Reference (13): Kiyotaka Uchimoto, Ma Aoi, Maki Murata, Hiromi Osaku, Masao Uchiyama, Hitoshi Isahara, Named Expression Extraction Based on Maximum Entropy Model and Rewrite Rules, Journal of the Language Processing Society, Vol.7, No.2, 2000).

  Also, for example, the following document (14) describes a technique for extracting Japanese proper expressions using a support vector machine.

Reference (14): Hiroyasu Yamada, Taku Kudo, Yuji Matsumoto, Japanese Named Expression Extraction using Support Vector Machine, Journal of Information Processing Society of Japan, Vol.43, No.1 ", 2002
(2) Method using morpheme analysis If a morpheme analysis system (for example, ChaSen described above) is used, the input Japanese can be divided into words.

  For example, ChaSen divides a Japanese sentence and estimates the part of speech of each word as described above. For example, if “go to school” is entered, the following results can be obtained.

School Gacco School Noun-General To He To particle-Case particle-General Go Iku Go Verb-independence
In this way, each line is divided so that one word is included, and reading and part-of-speech information are given to each word.

In particular,
input:
The capital of Japan is Tokyo Output:
Japanese Japanese nouns-proprietary nouns-region-country nouns-incorporated capitals sta capital nouns-general is ha particles-corresponders tokyo tokyo nouns-proprietary nouns-regions-generals
EOS
Is the output of chasen, and the part-of-speech of noun-proper noun-region is output.
Using this system, it is possible to extract, for example, a specific expression of a place name.

Also,
input:
The output that Prime Minister Murayama said:
Murayama Murayama Murayama nouns-proper nouns-personal names-surname prime minister Shusha prime nouns-general ga is a particle-case particle-general saying ii verb-self-supporting five-stage, wa-gakuin
EOS
Is the output of chasen, but in this case, the part of speech of noun-proper noun-person name is output. Using this system, for example, it is possible to extract a specific expression of a person name.
(3) A method using a created rule There is also a method of manually creating a rule to extract a specific expression.

For example,
Noun + “san” is the name of the person + “prime” is the name of the person + “town” is the name of the place + “city” is the place of name.

  FIG. 5 is a diagram showing an example of a question response process flow according to the first embodiment of the present invention. The first keyword and the second keyword are input to the keyword input unit 11 as input keywords (step S1). For example, the first keyword “Japan” and the second keyword “area” are input.

The pattern extracting unit 121 of the keyword increasing unit 12 performs a full text search for the input keyword in the keyword extracting DB 16, and extracts a pattern that appears around the input keyword as c _i (step S2). The pattern appearing in the vicinity is appropriately defined. Extraction of the pattern c _i are performed for each of the first keyword and the second keyword.

In keyword extraction section 122 of the keyword increasing portion 12, and full-text search for a pattern c _i extracted by the pattern extraction unit 121 by the keyword extraction DB 16, and at the same time extracts the representation exp extracted by the pattern c _i, the extracted expression exp Sorting is performed in descending order of the Score value, and the result is output as a keyword (step S3). By the process of step S3, for example, the first keyword is increased to three third keywords “Japan”, “America”, and “Germany”, and the second keyword is “area”, “population”, Increase to three fourth keywords of “latitude”.

  Next, the question creation unit 13 creates a question composed of the output keywords (step S4). In step S4, a question composed of the third keyword and the fourth keyword is created. For example, the question creating unit 13 creates a question “What is the American population?” Composed of the third keyword “USA” and the fourth keyword “population”. When the question answering apparatus 1 has a configuration that does not include the question creating unit 13, the process of step S4 is omitted.

  Next, the answer candidate extraction unit 14 extracts the answer candidates for each created question using the above-described machine learning technique (step S5). When the question answering apparatus 1 does not include the question creating unit 13, in step S5, the answer candidate extracting unit 14 selects an answer candidate for the question configured by the keyword output by the keyword increasing unit 12 as a machine. Extract using learning techniques. And the answer table output part 15 outputs an answer table (step S6).

  FIG. 6 is a diagram illustrating another example of the configuration of the question answering apparatus according to the first embodiment of the present invention. The question answering device 10 is a device that increases an input keyword and outputs an answer to a question constituted by the increased keyword. Among the constituent elements included in the question answering apparatus 10 shown in FIG. 6, the same reference numerals as those included in the question answering apparatus 1 shown in FIG. 1 are the same as the constituent elements included in the question answering apparatus 1. It has the function of.

  In the embodiment of the present invention, the question creating unit 13 is omitted from the configuration shown in FIG. 6, and the answer candidate extracting unit 14 is configured by the third keyword and the fourth keyword output by the keyword increasing unit 60. A configuration may be adopted in which answer candidates for a question to be asked are extracted and output.

  The keyword increasing unit 60 of the question answering apparatus 10 increases the keyword input to the keyword input unit 11. That is, the keyword increasing unit 60 outputs, for example, a third keyword larger than the number of the first keywords based on the first keyword input to the keyword input unit 11. Moreover, the keyword increase part 60 outputs the 4th keyword more than the number of 2nd keywords based on the 2nd keyword input into the keyword input part 11, for example.

  The word data database (DB) 61 stores correspondence information between words and word fields. For example, correspondence information between words and word fields as shown in FIG. 7 is stored. For example, words such as Japan, the United States, Germany,... Are stored as words corresponding to the field “country name”.

  The thesaurus database (DB) 62 stores thesaurus data, which is word classification information based on semantic similarity. For example, in the thesaurus DB 62, correspondence information between words and 10-digit numbers (classification numbers) assigned to the words as shown in FIG. 8 is stored as thesaurus data. In the example shown in FIG. 8, thesaurus data is shown in the form of a classification vocabulary table.

  The classification vocabulary table is generally a table in which words are arranged based on meaning, and a number called a classification number is assigned to each word. This 10-digit classification number indicates a 7-level hierarchical structure, with the top five levels being represented by the first five digits of the classification number, the sixth level is the next two digits, and the lowest level is the last three levels. It is expressed in digits.

  The similarity calculation unit 100 calculates the similarity between the keyword input to the keyword input unit 11 and the word in the thesaurus data based on the thesaurus data in the thesaurus DB 62. The keyword extraction unit 101 extracts, for example, a word whose calculated similarity is equal to or greater than a predetermined threshold as a keyword and outputs the keyword. In addition, the keyword extraction unit 101 can take a configuration in which a predetermined number of words are extracted from the thesaurus data in the descending order of the calculated similarity and are output as keywords.

  In the embodiment of the present invention, the keyword extraction unit 101 has the same field as the keyword input to the keyword input unit 11 based on the correspondence information between the word and the field of the word stored in the word data DB 61. It is also possible to adopt a configuration in which words are extracted as keywords and output.

  The question answering process flow when the above question answering apparatus 10 is used is the same as the question answering process flow shown in FIG. 5 except that the processes in steps S2 and S3 are different. In an example of the question answering process flow when the question answering apparatus 10 is used, the keyword input to the keyword input unit 11 by the keyword extracting unit 101 of the keyword increasing unit 60 instead of steps S2 and S3 in FIG. Are extracted from the word data DB 61 and output as keywords.

  For example, if the first keyword “Japan” is input to the keyword input unit 11, the keyword extraction unit 101 belongs to the field “country name” to which the word “Japan” corresponds from the word data DB 61 shown in FIG. Corresponding) words “Japan”, “America”, “Germany”,... Are extracted and output as a third keyword. For example, if the second keyword “area” is input to the keyword input unit 11, the keyword extraction unit 101 reads the field “numerical expression” corresponding to the word “area” from the word data DB 61 illustrated in FIG. 7. “Area”, “population”, “latitude”,... That belong to (corresponding to) are extracted and output as a fourth keyword.

  In another example of the question answering process flow when the question answering device 10 is used, for example, the similarity calculating unit 100 of the keyword increasing unit 60 performs keyword input instead of step S2 and step S3 in FIG. The similarity between the keyword input to the unit 11 and the word in the thesaurus DB 62 is calculated, and the keyword extraction unit 101 of the keyword increasing unit 60 outputs a word whose calculated similarity is equal to or greater than a predetermined threshold as a keyword. .

  For example, the keyword extraction unit 101 can take a configuration in which a predetermined number of words are extracted from the thesaurus data in descending order of the calculated similarity and are output as keywords.

  The similarity calculation unit 100 calculates the similarity between the input keyword and the word in the thesaurus DB 62, for example, as follows. The similarity is obtained by using the proportion of the digits of each digit in the 10-digit classification number assigned to each word in the thesaurus data (classification vocabulary table) stored in the thesaurus DB 62 shown in FIG. That is, for example, with respect to the classification number assigned to each word in the classification vocabulary table, the digit number matches between the keyword input to the keyword input unit 11 and the classification number assigned to the same word. The ratio is calculated, and the calculated value is set as the similarity. For example, the sixth and seventh digits of the classification number and the eighth, ninth and tenth digits are considered as one continuous number.

  For example, if the keyword input as the first keyword in the keyword input unit 11 is “Japan”, the words “Japan” and “America” in the classification vocabulary table shown in FIG. The number is given. In the following, the upper five levels, the sixth level, and the lowest level of the classification number are shown separated by blanks.

Japan: 12590 01 012
United States: 12590 04 192
For example, since the first five digits match in the top five levels of the classification numbers of both words, the similarity between the calculated keyword “Japan” and the word “America” in the classification lexicon is 5 .

  For example, when the keyword input as the second keyword in the keyword input unit 11 is “area”, the words “area” and “population” in the classification vocabulary table have the following classification numbers, respectively. It is shaken.

Area: 12630 13 015
Population: 12630 10 012
For example, since the first five digits match at the top five levels of the classification numbers of both words, the similarity between the calculated keyword “area” and the word “population” in the classification lexicon is similarity 5. .

  For example, when the keyword input as the second keyword in the keyword input unit 11 is “population”, the words “population” and “latitude” in the classification vocabulary table have the following classification numbers, respectively. It is shaken.

Population: 12630 10 012
Latitude: 12630 10 015
For example, in the top five levels of the classification numbers of both words, the first five digits match and the two-digit number “10” at the sixth level matches, so the calculated keyword “population” and the classification vocabulary table The similarity with the word “latitude” in the middle is 7.

  For example, when the keyword input as the second keyword in the keyword input unit 11 is “population”, the words “population” and “America” in the classification vocabulary table have the following classification numbers, respectively. It is shaken.

Population: 12630 10 012
United States: 12590 04 192
For example, since the first two digits match at the top five levels of the classification numbers of both words, the similarity between the calculated keyword “population” and the word “America” in the classification vocabulary table is similarity 2 .

  FIG. 9 is a diagram illustrating an example of the configuration of the question answering apparatus according to the second embodiment of the present invention. In the second embodiment, for example, when the first keyword “Japan” and the second keyword “capital” + question pronoun “where is?” Are entered, the first keyword “Japan” Based on this, the first keyword is increased to, for example, “Japan”, “America”, and “Germany”. And the first keyword after the increase and the second keyword “capital” + interrogative pronoun “Where is?” For example, “Where is the capital of Japan?”, “Where is the capital of America?” ? "And" Where is the German capital? "

  The question answering device 2 is a device that increases the number of input keywords and outputs an answer to a question composed of the increased keywords. Among the constituent elements of the question answering device 2 shown in FIG. 9, the keyword input unit 11, answer table output unit 15, keyword extraction DB 16, pattern extraction unit 121, and keyword extraction unit 122 are the question answering device shown in FIG. 1 are the same as those of the same reference numerals. In the embodiment of the present invention, the question creation unit 23 to be described later is omitted from the configuration shown in FIG. 9, and the answer candidate extraction unit 24 selects answer candidates for the question configured by the keywords output by the keyword increase unit 18. A configuration of extracting and outputting may be adopted.

  A keyword is input to the keyword input unit 11. For example, the first keyword “Japan” and the second keyword “capital” are input. The interrogative pronoun input unit 21 receives the interrogative pronoun associated with the second keyword input to the keyword input unit 11. For example, the question pronoun “Where is?” Is input. In addition, examples of the question pronouns input to the question pronoun input unit 21 include “What time is it?” And “Who is?”. The interrogative pronoun input to the interrogative pronoun input unit 21 may be input based on a user's designated input, or may be input by a computer different from the question answering device 2. There may be.

  The answer type estimation unit 22 has a configuration in which a question created by a question creation unit 23 described later based on the question pronoun input to the question pronoun input unit 21 or the question answering device 2 does not include the question creation unit 23. When adopting, an answer type that is a type of language expression of an answer candidate for a question configured by a keyword output by the keyword increasing unit 18 described later is estimated. For example, when the input question pronoun is “Where?”, It is estimated that the answer type is “proprietary noun (place name)”. In the embodiment of the present invention, the answer type estimation unit 22 may estimate the answer type based on a predetermined question pronoun instead of the question pronoun input to the question pronoun input unit 21.

  The keyword increase unit 18 uses the keyword extraction technique to extract keywords in the same field as the input first keyword from the keyword extraction DB 16, increases the first keyword, and outputs it as a third keyword. . In the second embodiment, the keyword increasing unit 18 outputs the second keyword (for example, “capital”) to the question creating unit 23 without increasing it. When the question answering device 2 is configured not to include the question creating unit 23, the keyword increasing unit 18 outputs the third keyword and the second keyword to the answer candidate extracting unit 24.

  The question creating unit 23 outputs the third keyword output by the processing of the keyword increasing unit 18, the second keyword, and the question pronoun (or predetermined question pronoun) input to the question pronoun input unit 21. Create multiple questions based on it.

  In the knowledge database (DB) 25, document data groups to be searched for answer candidates are accumulated. Examples of the document data group to be accumulated include document data groups such as newspaper article data and encyclopedia data.

  The answer candidate extracting unit 24 includes document data including keywords (or the third keyword and the second keyword output from the keyword increasing unit 18) constituting each question created by the question creating unit 23 from the knowledge DB 25. , And from the document data extracted by this search process, a linguistic expression that matches the answer type estimated by the answer type estimation unit 22 is extracted as an answer candidate.

  The answer table output unit 15 outputs a table in which each extracted answer candidate is associated with a question as an answer table. For example, an answer table as shown in FIG. 10 is output.

  In the answer table shown in FIG. 10, for example, as an answer to the question “Where is the capital of Japan?”, A row corresponding to the data item “Japan” and a column corresponding to the data item “capital” intersect. In the eyes, "Tokyo" is stored, and the answer to the question "Where is the US capital?" Intersects the row corresponding to the data item "USA" and the column corresponding to the data item "capital" To the intersection of the row corresponding to the data item “USA” and the column corresponding to the data item “Capital” as an answer to the question “Where is the German capital?” , “Berlin” is stored.

  FIG. 11 is a diagram illustrating an example of a question response process flow according to the second embodiment of the present invention. The first keyword and the second keyword are input to the keyword input unit 11 as input keywords (step S11). For example, the first keyword “Japan” and the second keyword “capital” are input. In addition, the question pronoun associated with the second keyword is input to the question pronoun input unit 21 (step S12). For example, the question pronoun “where is?” Associated with the second keyword “capital” is input.

A pattern extraction unit 121 of the keyword increasing portion 18, the first keyword and full-text search on the keyword extraction DB 16, extracts the emerging pattern around the first keyword as c _i (step S13). The pattern appearing in the vicinity is appropriately defined.

In keyword extraction section 122 of the keyword increasing portion 18, and full-text search for a pattern c _i extracted by the pattern extraction unit 121 by the keyword extraction DB 16, and at the same time extracts the representation exp extracted by the pattern c _i, the extracted expression exp Sorting is performed in descending order of the Score value, and the result is output as the third keyword (step S14). By the process of step S14, for example, the first keyword increases to three third keywords “Japan”, “America”, and “Germany”.

  The answer type estimation unit 22 estimates the answer type based on the question pronoun input to the question pronoun input unit 21 (step S15). For example, when the input question pronoun is “where is?”, The answer type estimation unit 22 estimates that the answer type is “proper noun (place name)”.

  The question creating unit 23 creates a question composed of the third keyword and the second keyword using the question pronoun input to the question pronoun input unit 21 (step S16). For example, the question creating unit 23 creates a question “Where is the US capital?” Composed of the third keyword “USA” and the second keyword “capital”. When the question answering apparatus 2 adopts a configuration that does not include the question creating unit 23, the process of step S16 is omitted.

  Next, the answer candidate extraction unit 24 extracts answer candidates for each created question (step S17). That is, the answer candidate extraction unit 24 searches the knowledge DB 25 for document data including the keyword that constitutes each question created by the question creation unit 23, and uses the answer type estimation unit from the document data extracted by this search process. A linguistic expression that matches the answer type estimated by 22 is extracted as an answer candidate. When the question answering apparatus 2 does not include the question creating unit 23, in step S17, the answer candidate extracting unit 24 searches the knowledge DB 25 for document data including the keyword output by the keyword increasing unit 18. Then, from the document data extracted by this search processing, a linguistic expression that matches the answer type estimated by the answer type estimating unit 22 is extracted as an answer candidate for the question constituted by the keyword output by the keyword increasing unit 18. . Then, the answer table output unit 15 outputs the answer table (step S18). For example, the answer table as shown in FIG. 10 described above is output.

  FIG. 12 is a diagram illustrating a configuration example of Modification 1 of the second embodiment of the present invention. The question answering device 20 is a device that increases the number of input keywords and outputs an answer to a question composed of the increased keywords. Among the constituent elements included in the question answering apparatus 10 shown in FIG. 12, the constituent elements included in the question answering apparatus 10 shown in FIG. 6 or the constituent elements provided in the question answering apparatus 2 shown in FIG. Has the same function as the constituent elements of the question answering device 10 or the question answering device 2.

  The keyword increasing unit 63 of the question answering apparatus 20 increases the first keyword input to the keyword input unit 11 and outputs it as a third keyword. The second keyword input to the keyword input unit 11 is output to the question creating unit 23 without being increased. In the embodiment of the present invention, the question creation unit 23 is omitted from the configuration shown in FIG. 12, and the answer candidate extraction unit 24 is configured by the third keyword and the second keyword output by the keyword increase unit 63. A configuration may be adopted in which answer candidates for a question to be asked are extracted and output.

  The question answering process flow when the above question answering apparatus 20 is used is the same as the question answering process flow shown in FIG. 11 except that the processes in steps S13 and S14 are different. In an example of the question answering process flow when the question answering device 10 is used, the keyword extracting unit 101 of the keyword increasing unit 63 inputs the keyword input unit 11 to the keyword input unit 11 instead of steps S13 and S14 of FIG. A word in the same field as the one keyword is extracted from the word data DB 61 and output as a third keyword.

  Further, in another example of the question answering process flow when the question answering device 10 is used, the similarity calculating unit 100 of the keyword increasing unit 63 is replaced with the keyword inputting unit 11 instead of step S13 and step S14 of FIG. The degree of similarity between the first keyword input to the word and the word in the thesaurus DB 62 is calculated, and the keyword extraction unit 101 of the keyword increase unit 63 determines that the calculated similarity is equal to or greater than a predetermined threshold. Output as a keyword.

  In addition, the keyword extraction unit 101 can take a configuration in which a predetermined number of words are extracted from the thesaurus data in descending order of the calculated similarity and are output as the third keyword.

  In the second modification of the second embodiment of the present invention, in the question answering device 2 shown in FIG. 9 or the question answering device 20 shown in FIG. 12, the question pronoun input unit 21 is input to the keyword input unit 11. A questionable pronoun that is not associated with a keyword is entered. The keyword increasing unit 18 of the question answering device 2 (or the keyword increasing unit 63 of the question answering device 20) outputs the third keyword as an output keyword based on the first keyword input to the keyword input unit 11, Based on the second keyword input to the keyword input unit 11, the fourth keyword is output as an output keyword. The answer type estimation unit 22 estimates the answer type based on the question pronoun input to the question pronoun input unit 21. Then, the answer candidate extraction unit 24 searches the knowledge DB 25 for document data including the keyword output by the keyword increase unit 18, and is estimated by the answer type estimation unit 22 from the document data extracted by this search process. A linguistic expression suitable for the answer type is extracted as an answer candidate for the question constituted by the keyword output by the keyword increasing unit 18. Then, the answer table output unit 15 outputs the answer table. In the second modification of the second embodiment of the present invention, the answer type estimation unit 22 is not based on the question pronoun input to the question pronoun input unit 21 but based on a predetermined question pronoun. You may take the structure which estimates an answer type.

  FIG. 13 is a diagram illustrating an example of the configuration of the question answering apparatus according to the third embodiment of the present invention. In the third embodiment, instead of using the machine learning method as in the first embodiment, answer candidates are extracted using the input answer type (or a predetermined answer type).

  In the third embodiment, for example, when the first keyword “Japan”, the second keyword “capital”, and the answer type “proper noun (place name)” are input, the first keyword “Japan”. Based on the above, the first keyword is increased to, for example, three third keywords “Japan”, “America”, and “Germany”. Further, based on the second keyword “capital”, the second keyword is increased to, for example, three fourth keywords “capital”, “old capital”, and “southmost city”.

  And the combination of the increased third keyword and fourth keyword, for example, “What is the capital of Japan?”, “What is the old capital of America?”, “What is the southernmost city of Germany?”・ An answer to each question is output. More specifically, as will be described later, the third keyword “Japan” and the fourth keyword “capital” that make up the question “What is the capital of Japan?” Are document data groups to be searched for answer candidates. And the linguistic expressions in the document including both keywords are extracted as answer candidates, and the extracted answer candidates that match the answer type “proper noun (place name)” are output as answers.

  The question answering device 3 is a device that increases the number of input keywords and outputs an answer to a question composed of the increased keywords. Among the components of the question answering device 3 shown in FIG. 13, the keyword input unit 11, the keyword increase unit 12, the question creation unit 13, the answer table output unit 15, the keyword extraction DB 16, the pattern extraction unit 121, and the keyword extraction unit 122 are 1 are the same as those of the question answering apparatus 1 shown in FIG. 1, and the answer candidate extraction unit 24 and the knowledge DB 25 are the same as those of the question answering apparatus 2 shown in FIG. It is. In the embodiment of the present invention, the question creating unit 13 is omitted from the configuration shown in FIG. 13, and the answer candidate extracting unit 24 extracts answer candidates for the question constituted by the keywords output by the keyword increasing unit 12. A configuration of outputting may be adopted.

  A keyword is input to the keyword input unit 11. For example, the first keyword “Japan” and the second keyword “capital” are input. The answer type input unit 31 includes a question created by the question creating unit 13 or a keyword output by the keyword increasing unit 12 when the question answering device 3 omits the question creating unit 13. The answer type of the answer candidate for the question is entered. For example, an answer type “proper noun (place name)” is input.

  In addition, as answer types input to the answer type input unit 31, for example, “proprietary noun (numerical value)”, “proprietary noun (person name)”, “katakana expression” (represented only by katakana), “noun” ”,“ Verb ”and the like. Note that the answer type input to the answer type input unit 31 may be input based on a user's designated input, or may be input by a computer different from the question answering device 3. There may be.

  As described with reference to FIG. 1, the keyword increase unit 12 uses the keyword extraction technique to extract keywords in the same field as the input keywords from the keyword extraction DB 16 and increase the keywords.

  For example, the keyword increasing unit 12 outputs a third keyword that is larger than the number of first keywords based on the input first keyword. In addition, for example, the keyword increasing unit 12 outputs a fourth keyword that is larger than the number of second keywords based on the input second keyword.

  The question creating unit 13 creates a plurality of questions configured by the third keyword and the fourth keyword. The answer candidate extraction unit 24 stores a document including the keywords (or the third keyword and the fourth keyword output by the keyword increase unit 12) constituting each question created by the question creation unit 13 from the knowledge DB 25. A language expression that matches the answer type (or predetermined answer type) input to the answer type input unit 31 is extracted as an answer candidate from the language expressions included in the document extracted by the search process. To do.

  The answer table output unit 15 outputs a table in which each extracted answer candidate is associated with a question as an answer table.

  FIG. 14 is a diagram illustrating an example of a question response process flow according to the third embodiment of the present invention. The first keyword and the second keyword are input to the keyword input unit 11 as input keywords (step S21). For example, the first keyword “Japan” and the second keyword “capital” are input. Moreover, the answer type of the answer candidate with respect to the question created by the question creation part 13 is input into the answer type input part 31 (step S22). For example, “proper noun (place name)” is input as the answer type. When the question answering device 3 does not include the question creating unit 13, the answer type input unit 31 receives the answer type of the answer candidate for the question configured by the keyword output by the keyword increasing unit 12. Is done.

A pattern extraction unit 121 of the keyword increasing portion 12, and full-text search input keywords in the keyword extraction DB 16, extracts the emerging pattern around the plurality of input keywords as c _i (step S23). The pattern appearing in the vicinity is appropriately defined. Extraction of the pattern c _i are performed for each of the first keyword and the second keyword.

In keyword extraction section 122 of the keyword increasing portion 12, and full-text search for a pattern c _i extracted by the pattern extraction unit 121 by the keyword extraction DB 16, and at the same time extracts the representation exp extracted by the pattern c _i, the extracted expression exp The values are sorted in descending order of the Score value and output as keywords (step S24). By the process of step S24, for example, the first keyword increases to three third keywords “Japan”, “America”, and “Germany”. In addition, the second keyword increases to three fourth keywords of “capital”, “old capital”, and “southernest city”.

  The question creating unit 23 creates a question composed of the output keywords (step S25). In step S25, a question composed of the output third keyword and fourth keyword is created. For example, the question creating unit 23 creates a question “What is the US capital?” Composed of the third keyword “USA” and the fourth keyword “capital”. When the question answering device 3 is configured not to include the question creating unit 13, the process of step S25 is omitted.

  Next, the answer candidate extraction unit 24 extracts answer candidates for each created question (step S26). That is, the answer candidate extraction unit 24 searches the knowledge DB 25 for document data including keywords that constitute each question created by the question creation unit 13, and uses the answer type input unit from the document data extracted by this search process. A linguistic expression that matches the answer type input in 31 is extracted as an answer candidate. When the question answering device 3 does not include the question creating unit 13, in step S26, the answer candidate extracting unit 24 extracts the keyword (the third keyword and the third keyword) output from the knowledge DB 25 by the keyword increasing unit 12. 4), and the linguistic expression suitable for the answer type input to the answer type input unit 31 is output from the document data extracted by the search process. It extracts as an answer candidate with respect to the question comprised by 3 keywords and a 4th keyword. And the answer table output part 15 outputs an answer table (step S27). For example, an answer table as shown in FIG. 15 is output.

  FIG. 16 is a diagram illustrating a configuration example of Modification 1 of the third embodiment of the present invention. The question answering device 30 is a device that increases the number of input keywords and outputs an answer to a question composed of the increased keywords. Of the constituent elements provided in the question answering apparatus 30 shown in FIG. 16, the constituent elements provided in the question answering apparatus 1 shown in FIG. 1, the constituent elements provided in the question answering apparatus 10 shown in FIG. 6, or the question answering apparatus 3 shown in FIG. The same reference numerals as those of the question answering device 1, the question answering device 10, or the question answering device 3 have the same functions.

  In the embodiment of the present invention, the question creating unit 13 is omitted from the configuration shown in FIG. 16, and the answer candidate extracting unit 24 is configured by the third keyword and the fourth keyword output by the keyword increasing unit 60. A configuration may be adopted in which answer candidates for a question to be asked are extracted and output.

  The question answering process flow when the above question answering apparatus 30 is used is the same as the question answering process flow shown in FIG. 14 except that the processes in steps S23 and S24 are different. In an example of the question answering process flow when the question answering device 30 is used, the keyword extracting unit 101 of the keyword increasing unit 60 inputs the keyword input unit 11 to the keyword input unit 11 instead of steps S23 and S24 in FIG. A word in the same field as the one keyword is extracted from the word data DB 61 and output as a third keyword. Further, the keyword extraction unit 101 extracts words in the same field as the second keyword input to the keyword input unit 11 from the word data DB 61 and outputs them as the fourth keyword.

  In another example of the question answering process flow when the question answering device 30 is used, the similarity calculating unit 100 of the keyword increasing unit 60 replaces the keyword input unit 11 in place of steps S23 and S24 in FIG. The similarity between the first keyword input to the word and the word in the thesaurus DB 62 is calculated, and the keyword extracting unit 101 of the keyword increasing unit 60 calculates a word whose calculated similarity is equal to or greater than a predetermined threshold to the third Output as a keyword. In addition, the similarity calculation unit 100 calculates the similarity between the second keyword input to the keyword input unit 11 and the word in the thesaurus DB 62, and the keyword extraction unit 101 determines the calculated similarity in advance. A word equal to or greater than the threshold is output as the fourth keyword.

  Further, the keyword extraction unit 101 may take a configuration in which, for example, a predetermined number of words are extracted from the thesaurus data in descending order of the calculated similarity and are output as the third keyword and the fourth keyword. it can.

  In the second modification of the third embodiment of the present invention, for example, in the question answering device 3 shown in FIG. 13 or the question answering device 30 shown in FIG. 16, the answer type input unit 31 includes the keyword increasing unit 12 (or An answer type which is a type of linguistic expression of an answer candidate for a question configured by an output keyword output by the keyword increasing unit 60), and is associated with the second keyword input to the keyword input unit 11 The answer type is entered. The keyword increasing unit 12 (or the keyword increasing unit 60) outputs the third keyword as an output keyword based on the first keyword input to the keyword input unit 11, and the second keyword input to the keyword input unit 11 Are output as output keywords. The answer candidate extraction unit 24 searches the knowledge DB 25 for document data including the keywords (the third keyword and the second keyword) output by the keyword increase unit 12 (or the keyword increase unit 60). From the extracted document data, a linguistic expression that matches the answer type input to the answer type input unit 31 corresponding to the second keyword, and the third keyword output by the keyword increasing unit 12 and the second keyword are output. It is extracted as an answer candidate for a question composed of keywords. Then, the answer table output unit 15 outputs the answer table.

  In the third modification of the third embodiment of the present invention, the answer candidate extraction unit 24 uses the keyword (the third keyword and the keyword) output from the knowledge DB 25 by the keyword increasing unit 12 (or the keyword increasing unit 60). The document data including the second keyword) is searched, and from the document data extracted by the search processing, a predetermined language expression that matches the answer type associated with the second keyword is increased. A configuration may be adopted in which it is extracted as an answer candidate for a question configured by the third keyword output by the unit 12 and the second keyword.

  In the third modification of the third embodiment of the present invention, for example, in the question answering device 3 shown in FIG. 13 or the question answering device 30 shown in FIG. 16, the keyword input unit 11 includes a first keyword and a plurality of keywords. And the second keyword grouped by the group. For example, the first keyword “Japan”, the second keyword “Prime Minister” and “mayor” belonging to the personal name group, and the second keyword “capital” and “old capital” belonging to the place name group are input. . The answer type input unit 31 receives an answer type associated with each group to which the second keyword input to the keyword input unit 11 belongs. For example, the answer type “person name” is input as the answer type corresponding to the group of person names, and the answer type “place name” is input as the answer type corresponding to the group of place names.

  The keyword increasing unit 12 (or the keyword increasing unit 60) outputs the third keyword as an output keyword based on the first keyword. Further, based on the second keyword, the fourth keyword is output as an output keyword for each group to which the second keyword belongs. For example, “Japan”, “Germany”, and “USA” are output as the third keyword. Also, for example, “Prime Minister”, “Mayor”, “Nobel Prize Winner” are output as the fourth keyword belonging to the personal name group, and “Capital”, “Old” are the fourth keywords belonging to the place name group, for example. "Capital" and "Southernmost city" are output.

  The answer candidate extraction unit 24 selects an answer candidate for a question composed of the third keyword output by the keyword increasing unit 12 (or the keyword increasing unit 60) and the fourth keyword belonging to the group of person names as an answer type. The answer type “person name” input to the input unit 31 is used for extraction. For example, answer candidates for the question “Who is the German Nobel Prize winner?” Are extracted using the answer type “person name”. In addition, the answer candidate extraction unit 24 selects answer candidates for a question configured by the third keyword output by the keyword increasing unit 12 (or the keyword increasing unit 60) and the fourth keyword belonging to the place name group. The answer type “location name” input to the answer type input unit 31 is used for extraction. For example, answer candidates for the question “What is the US capital?” Are extracted using the answer type “place name”. Then, the answer table output unit 15 outputs the answer table.

  FIG. 17 is a diagram illustrating an example of a configuration of a question answering apparatus according to the fourth embodiment of the present invention. In the fourth embodiment, answer candidates are extracted using the similarity relationship of keywords.

  In the fourth embodiment, for example, the first keyword “Japan”, “America”,... And the second keyword “area”, “capital”,. ) ”,“ Proper noun (place name) ”, and so on. The input answer type is associated with each input second keyword. For example, the answer type associated with the second keyword “area” is “proper noun (numerical value)”, and the answer type associated with the second keyword “capital” is “proper noun (place name)”. is there.

  In the fourth embodiment, for example, based on the input first keywords “Japan”, “America”,..., The first keyword is converted into a number of third keywords (for example, “Japan”, “ "America", "Germany", ...). Further, the keyword increasing unit 12 selects the second keyword as a number of fourth keywords (for example, “area”, “capital”, “old capital” based on the second keywords “area”, “capital”,. ", ...).

  Next, among the second keywords (the same fourth keyword), a keyword similar to the fourth keyword is determined as a similar keyword. For example, the second keyword (the same fourth keyword) “capital” similar to the fourth keyword “old capital” is determined as a similar keyword.

  And the candidate of the answer with respect to the question comprised by the combination of a 3rd keyword and a 4th keyword is used for the answer type matched with the similar keyword similar to the 4th keyword which comprises the said question. Extract and output the answer sheet. For example, answer candidates for the question “What is the old capital of Japan?” Are extracted using the answer type “proper noun (place name)” associated with the similar keyword “capital”, and an answer table is output.

  The question answering device 4 is a device that increases the number of input keywords and outputs an answer to a question composed of the increased keywords. Among the components of the question answering device 4 shown in FIG. 17, the keyword input unit 11, the keyword increase unit 12, the answer table output unit 15, the keyword extraction DB 16, the pattern extraction unit 121, and the keyword extraction unit 122 are respectively shown in FIG. The knowledge DB 25 is the same as the knowledge DB 25 provided in the question answering apparatus 2 shown in FIG. 9, and the answer type input unit 31 is shown in FIG. This is the same as the answer type input unit 31 provided in the question answering device 3.

  A keyword is input to the keyword input unit 11. For example, the first keyword “Japan”, “America”,... And the second keyword “area”, “capital”,. The answer type input unit 31 includes a question created by the question creating unit 42 or a keyword output by the keyword increasing unit 12 when the question answering device 4 omits the question creating unit 42. The answer type of the answer candidate for the question is entered. The input answer type is particularly associated with the second keyword.

  For example, an answer type “proprietary noun (numerical value)” corresponding to the second keyword “area” is input to the answer type input unit 31, and “specific” corresponding to the second keyword “capital” is input. The answer type “noun (place name)” is input.

  As described with reference to FIG. 1, the keyword increase unit 12 uses the keyword extraction technique to extract keywords in the same field as the input keywords from the keyword extraction DB 16 and increase the keywords. By the processing of the keyword increasing unit 12, the third keyword is output from the first keyword, and the fourth keyword is output from the second keyword.

  The similar keyword determination unit 41 determines a second keyword (the same fourth keyword) input to the keyword input unit 11 that is similar to each fourth keyword as a similar keyword. A method for determining similar keywords will be described below.

(Method using co-occurrence vector (1))
For each fourth keyword, the number of appearances in the keyword extraction DB 16 co-occurring with the pattern c _i extracted by the keyword increasing unit 12 is calculated, and a vector (hereinafter, “co-occurrence vector” having the calculated number of times as an element is calculated. ").

For example, in the keyword extraction process in the keyword increasing unit 12, the number of times the fourth keyword (1) appears co-occurring with the pattern c ₁ is 0, the number of times the fourth keyword (1) appears co-occurring with the pattern c ₂ is 1,. If the number of occurrences co-occurring with the pattern c _n is 1, the co-occurrence vector for the fourth keyword (1) is obtained as (0, 1,... 1). Similarly, co-occurrence vectors for other fourth keywords (second keyword (2), second keyword (3),...) Are obtained.

The degree of similarity between the co-occurrence vector for the fourth keyword identical to the second keyword input to the keyword input unit 11 and the co-occurrence vector for the fourth keyword for which a corresponding similar keyword is desired is obtained. For example, the co-occurrence vector (a ₁ , a ₂ , a ₃ ,..., _An ) for the fourth keyword that is the same as the second keyword input to the keyword input unit 11 is obtained for the corresponding similar keyword. If the co-occurrence vector for the fourth keyword is (b ₁ , b ₂ , b ₃ ,... B _n ), (a ₁ −b ₁ ) ² + (a ₂ −b ₂ ) ² + (a ₃₋ b ₃ ) ² +... ( _An −b _n ) ² is calculated. The calculated value indicates the degree of similarity between the two co-occurrence vectors. The lower the calculated value, the higher the degree of similarity.

  The fourth keyword that is the same as the second keyword input to the keyword input unit 11 when the calculated value is the lowest is set as a similar keyword similar to the fourth keyword for which a corresponding similar keyword is desired.

(Method using co-occurrence vector (2))
The similar keyword determination unit 41 first searches the knowledge DB 25 using the fourth keyword, and extracts words (co-occurrence words) that appear along with each fourth keyword. Then, a vector whose element is the number of times each fourth keyword co-occurs with the extracted co-occurrence word and appears in the knowledge DB 25 is obtained as a co-occurrence vector for each fourth keyword.

For example, the number of occurrences of the fourth keyword (1) co-occurring with the co-occurrence word w ₁ is 2, the number of occurrences of co-occurrence with the co-occurrence word w ₂ is 0, and the co-occurrence word w ₃ is co-occurring to number of appearances is 1 ..., when the number of emerged and co-occur with pattern c _n is 1, the co-occurrence vector for the fourth keyword (1), (2,0,1, ...・ 1) is obtained. Similarly, co-occurrence vectors for other fourth keywords (second keyword (2), second keyword (3),...) Are obtained.

The degree of similarity between the co-occurrence vector for the fourth keyword identical to the second keyword input to the keyword input unit 11 and the co-occurrence vector for the fourth keyword for which a corresponding similar keyword is desired is obtained. For example, the co-occurrence vector (a ₁ , a ₂ , a ₃ ,..., _An ) for the fourth keyword that is the same as the second keyword input to the keyword input unit 11 is obtained for the corresponding similar keyword. If the co-occurrence vector for the fourth keyword is (b ₁ , b ₂ , b ₃ ,... B _n ), (a ₁ −b ₁ ) ² + (a ₂ −b ₂ ) ² + (a ₃₋ b ₃ ) ² +... ( _An −b _n ) ² is calculated. The calculated value indicates the degree of similarity between the two co-occurrence vectors. The lower the calculated value, the higher the degree of similarity.

  The fourth keyword that is the same as the second keyword input to the keyword input unit 11 when the calculated value is the lowest is set as a similar keyword similar to the fourth keyword for which a corresponding similar keyword is desired.

  In the embodiment of the present invention, the similar keyword determination unit 41 may be configured to obtain the co-occurrence vector using other document data instead of the knowledge DB 25. For example, it is possible to adopt a configuration in which the co-occurrence vector is obtained using a large-scale corpus (not shown) in which a large amount of document data is stored.

(Method using thesaurus data)
A thesaurus database (not shown in FIG. 17) in which thesaurus data is stored in the form of a classification vocabulary table is prepared. The similar keyword determination unit 41 uses the fourth keyword, using the ratio of the digit of each digit in the 10-digit classification number assigned to each word in the thesaurus data stored in the thesaurus database, The similarity to the second keyword (the same fourth keyword) input to the keyword input unit 11 is obtained.

  That is, for example, for the classification number assigned to the same word as the fourth keyword for which the corresponding similar keyword is to be obtained in the classification vocabulary table, the second keyword (the same as the second keyword input to the keyword input unit 11) is assigned. 4) and the classification number assigned to the same word are calculated, and the percentage of the numbers in each digit is calculated, and the magnitude of the calculated value is used as the similarity. Then, the second keyword (the same fourth keyword) as that when the calculated value is the largest is determined as a similar keyword similar to the fourth keyword for which a corresponding similar keyword is desired.

  The question creating unit 42 creates a question configured by a combination of the third keyword and the fourth keyword output by the processing of the keyword increasing unit 12.

  In the embodiment of the present invention, the question creating unit 42 is omitted from the configuration shown in FIG. 17, and the answer candidate extracting unit 43 is configured by the third keyword and the fourth keyword output by the keyword increasing unit 12. A configuration may be adopted in which answer candidates for a question to be asked are extracted and output.

  The answer candidate extraction unit 43 searches the knowledge DB 25 for documents including the third keyword and the fourth keyword that constitute each question created by the question creation unit 42, and includes them in the document extracted by this search process. Language expressions that match the answer type input to the answer type input unit 31 in association with similar keywords similar to the fourth keyword that constitutes each question are extracted as answer candidates. The answer candidate extraction unit 43 searches the knowledge DB 25 for documents including the third keyword and the fourth keyword that constitute each question created by the question creation unit 42, and includes them in the document extracted by this search process. The language expression that matches the answer type previously associated with the similar keyword similar to the fourth keyword constituting each question may be extracted as answer candidates.

  The answer candidate extraction unit 43 includes the third keyword and the fourth keyword output from the knowledge DB 25 by the keyword increasing unit 12 when the question answering device 4 omits the question creating unit 42. A document is searched, and it is input to the answer type input unit 31 in association with a similar keyword similar to the fourth keyword among the language expressions included in the document extracted by this search processing (or as a similar keyword in advance). A linguistic expression that matches the associated answer type is extracted as an answer candidate for each question configured by the third keyword and the fourth keyword.

  That is, an answer type associated with a similar keyword similar to the fourth keyword constituting each question is used to extract answer candidates for each question.

  For example, the answer type “proprietary noun (place name)” associated with the similar keyword “capital” similar to the fourth keyword “old capital” is selected as the answer candidate for the question “What is Japan's old capital?” Use to extract.

  The answer table output unit 15 outputs a table in which each extracted answer candidate is associated with a question as an answer table.

  FIG. 18 is a diagram showing an example of a question response process flow according to the fourth embodiment of the present invention. The first keyword and the second keyword are input as input keywords to the keyword input unit 11 (step S31). For example, the first keyword “Japan”, “America”,... And the second keyword “area”, “capital”,. In addition, the answer type input unit 31 inputs the answer type associated with the second keyword (step S32). For example, an answer type “proper noun (numerical value)” associated with the second keyword “area” and an answer type “proper noun (place name)” associated with the second keyword “capital” are input.

A pattern extraction unit 121 of the keyword increasing portion 12, and full-text search input keywords in the keyword extraction DB 16, extracts the emerging pattern around the plurality of input keywords as c _i (step S33). The pattern appearing in the vicinity is appropriately defined. Incidentally, extraction of the pattern c _i are performed for each of the first keyword and the second keyword.

In keyword extraction section 122 of the keyword increasing portion 12, and full-text search for a pattern c _i extracted by the pattern extraction unit 121 by the keyword extraction DB 16, and at the same time extracts the representation exp extracted by the pattern c _i, the extracted expression exp The values are sorted in descending order of the Score value and output as keywords (step S34).

  By the processing in step S34, for example, the first keyword is converted into a number of third keywords (for example, “Japan”, “America”, “Germany”, “Italy”, “France”, “United Kingdom”,...). To increase. In addition, the second keyword increases to a large number of fourth keywords (for example, “area”, “population”, “latitude”, “capital”, “old capital”, “southmost city”, etc.). .

  The similar keyword determination unit 41 determines a similar keyword similar to the fourth keyword (step S35). For example, as the similar keyword similar to the fourth keyword “Old Capital”, the “capital” that is the second keyword (the same fourth keyword) input to the keyword input unit 11 is determined.

  While the keyword is input to the keyword input unit 11 (step S36), the above-described processing of step S31 to step S35 is repeated.

  In step S36, when there is no input keyword input to the keyword input unit 11, the question creating unit 42 creates a question composed of the third keyword and the fourth keyword (step S37). For example, questions such as "What is the old capital of Japan?", "What is the area of America?", "What is the latitude of Germany?" When the question answering apparatus 4 does not include the question creating unit 42, the process of step S37 is omitted.

  The answer candidate extraction unit 43 extracts answer candidates for each created question (step S38). That is, the answer candidate extraction unit 43 searches the knowledge DB 25 for documents including the third keyword and the fourth keyword that constitute each question created by the question creation unit 42, and extracts the document extracted by this search process. Among the included language expressions, a language expression that matches an answer type associated with a similar keyword in which the fourth keyword constituting each question is similar is extracted as an answer candidate.

  In step S38, the answer candidate extraction unit 43 uses the third keyword and the fourth keyword output from the knowledge DB 25 by the keyword increase unit 12 when the question answering device 4 omits the question creation unit 42. And the answer type input to the answer type input unit 31 in association with a similar keyword similar to the fourth keyword among the language expressions included in the document extracted by the search process. The linguistic expression to be extracted is extracted as an answer candidate for each question composed of the third keyword and the fourth keyword.

  For example, for extraction of answer candidates for a question composed of the fourth keyword “latitude”, the answer type “proper noun (numerical value) associated with the similar keyword“ area ”similar to the fourth keyword“ latitude ”is used. Is used.

  For example, for extraction of answer candidates for a question composed of the fourth keyword “Old Capital”, the answer type “proper noun (place name) associated with a similar keyword similar to the fourth keyword“ Old Capital ”is used. ) ”.

  And the answer table output part 15 outputs an answer table (step S39). For example, an answer table as shown in FIG. 19 is output.

  FIG. 20 is a diagram illustrating another example of the configuration of the question answering apparatus according to the fourth embodiment of the present invention. The question answering device 40 is a device that increases the number of input keywords and outputs an answer to a question composed of the increased keywords. Among the constituent elements of the question answering apparatus 40 shown in FIG. 20, the constituent elements of the question answering apparatus 1 shown in FIG. 1, the constituent elements of the question answering apparatus 10 shown in FIG. 6, or the question answering apparatus 4 shown in FIG. The same reference numerals as those of the question answering device 1, the question answering device 10, or the question answering device 4 have the same functions. In the embodiment of the present invention, the question creating unit 42 is omitted from the configuration shown in FIG. 20, and the answer candidate extracting unit 43 is configured by the third keyword and the fourth keyword output by the keyword increasing unit 60. A configuration may be adopted in which answer candidates for a question to be asked are extracted and output.

  The question answering process flow when the above question answering device 40 is used is the same as the question answering process flow shown in FIG. 18 except that the processes in steps S33 and S34 are different. In an example of the question answering process flow when the question answering device 40 is used, the keyword extracting unit 101 of the keyword increasing unit 60 inputs the keyword input unit 11 to the keyword input unit 11 instead of steps S33 and S34 in FIG. A word in the same field as the one keyword is extracted from the word data DB 61 and output as a third keyword. Further, the keyword extraction unit 101 extracts words in the same field as the second keyword input to the keyword input unit 11 from the word data DB 61 and outputs them as the fourth keyword.

  In another example of the question answering process flow when the question answering device 40 is used, the similarity calculating unit 100 of the keyword increasing unit 60 replaces the keyword input unit 11 in place of steps S33 and S34 in FIG. The similarity between the first keyword input to the word and the word in the thesaurus DB 62 is calculated, and the keyword extracting unit 101 of the keyword increasing unit 60 calculates a word whose calculated similarity is equal to or greater than a predetermined threshold to the third Output as a keyword. In addition, the similarity calculation unit 100 calculates the similarity between the second keyword input to the keyword input unit 11 and the word in the thesaurus DB 62, and the keyword extraction unit 101 determines the calculated similarity in advance. A word equal to or greater than the threshold is output as the fourth keyword.

  In addition, the keyword extraction unit 101 takes, for example, a configuration in which a predetermined number of words are extracted from the thesaurus data in descending order of the calculated similarity and are output as the third keyword and the fourth keyword. You can also.

  Further, in the fourth embodiment of the present invention, the question answering device 4 shown in FIG. 17 or the question answering device 40 shown in FIG. 20 is input to the keyword input unit 11 instead of the answer type input unit 31, for example. A question pronoun input unit (not shown) for inputting a question pronoun associated with the second keyword that has been entered, and a keyword increase unit 12 (or keyword) based on the question pronoun input to the question pronoun input unit A configuration may be adopted that includes an answer type estimation unit (not shown) that estimates an answer type that is a type of a linguistic expression of an answer candidate for a question configured by an output keyword output by the increase unit 60).

  In the above configuration, the answer candidate extraction unit 43 searches the knowledge DB 25 for documents including the third keyword and the fourth keyword output by the keyword increasing unit 12 (or the keyword increasing unit 60), and this search. Of the linguistic expressions included in the document extracted by the processing, the answer type estimation unit estimates based on the question pronoun input corresponding to the similar keyword similar to the fourth keyword and input to the question pronoun input unit. A linguistic expression suitable for the answer type may be extracted as an answer candidate for each question configured by the third keyword and the fourth keyword. Further, the answer candidate extraction unit 43 searches the knowledge DB 25 for documents including the third keyword and the fourth keyword output by the keyword increase unit 12 (or the keyword increase unit 60), and is extracted by this search process. Among the linguistic expressions contained in the document, the pronoun pronoun associated with a similar keyword similar to the fourth keyword (that is, predetermined as being associated with the second keyword input to the keyword input unit 11) A linguistic expression suitable for the answer type estimated by the answer type estimation unit based on the question pronoun is extracted as an answer candidate for each question composed of the third keyword and the fourth keyword. Good.

  The present invention can also be implemented as a program that is read and executed by a computer. The program for realizing the present invention can be stored in an appropriate recording medium such as a portable medium memory, a semiconductor memory, or a hard disk, which can be read by a computer, provided by being recorded on these recording media, or communication. It is provided by transmission / reception using a network via an interface.

It is a figure which shows an example of a structure of the question answering apparatus in the 1st Embodiment of this invention. It is an example of an answer table. It is a figure which shows an example of the relevance rate and recall rate with respect to the extraction result of a keyword. It is a figure which shows an example of correct data. It is a figure which shows an example of the question response process flow in the 1st Embodiment of this invention. It is a figure which shows another example of a structure of the question answering apparatus in the 1st Embodiment of this invention. It is a figure which shows the data structural example of word data DB. It is a figure which shows the data structural example of a thesaurus DB. It is a figure which shows an example of a structure of the question answering apparatus in the 2nd Embodiment of this invention. It is an example of an answer table. It is a figure which shows an example of the question response process flow in the 2nd Embodiment of this invention. It is a figure which shows the structural example of the modification 1 of the 2nd Embodiment of this invention. It is a figure which shows an example of a structure of the question answering apparatus in the 3rd Embodiment of this invention. It is a figure which shows an example of the question response process flow in the 3rd Embodiment of this invention. It is an example of an answer table. It is a figure which shows the structural example of the modification 1 of the 3rd Embodiment of this invention. It is a figure which shows an example of a structure of the question answering apparatus in the 4th Embodiment of this invention. It is a figure which shows an example of the question response process flow in the 4th Embodiment of this invention. It is an example of an answer table. It is a figure which shows another example of a structure of the question answering apparatus in the 4th Embodiment of this invention. It is an example of a list of candidate solutions and scores. It is an example of the output result using the method of adding the score of a solution candidate simply. It is an example of the output result with respect to a question. It is an example of the output result with respect to a question. It is an example of the output result with respect to a question. It is an example of the output result with respect to a question. It is a figure which shows the concept of margin maximization of a support vector machine method.

Explanation of symbols

1, 2, 3, 4, 10, 20, 30, 40 Question answering device 11 Keyword input unit 12, 18, 60, 63 Keyword increasing unit 13, 23, 42 Question creating unit 14, 24, 43 Answer candidate extracting unit 15 Answer table output section 16 Keyword extraction DB
17 Learning DB
21 Question pronoun input part 22 Answer type estimation part 25 Knowledge DB
31 Answer type input part 41 Similar keyword determination part 61 Word data DB
62 Thesaurus DB
100 similarity calculation unit 101, 122 keyword extraction unit 121 pattern extraction unit

Claims (10)

The first keyword itself is a plurality of keywords, the second keyword itself is a plurality of keywords, and the first keyword and the second keyword are input. The first keyword and the second keyword Along with answers to question data expressed in natural language composed of keywords, a third keyword obtained by increasing the first keyword and a fourth keyword obtained by increasing the second keyword A question answering device that outputs an answer to question data expressed in a natural language,
A keyword input means for inputting a plurality of keywords as input keywords;
Keyword increasing means for extracting more keywords than the number of input keywords based on the input keywords and outputting them as output keywords;
The third keyword, which is an output keyword obtained by increasing the first keyword as an input keyword by the keyword increasing means, and the second keyword as an input keyword obtained by the keyword increasing means. Answer candidate extraction means for extracting an answer candidate that is an answer candidate for a question constituted by the fourth keyword that is an output keyword from a document data group that is a search target of answer candidates stored in advance;
An answer table output means for outputting, as an answer table, a table in which each extracted answer candidate is associated with a question;
The keyword increasing means is:
The input keyword is subjected to a full text search in a keyword extraction database in which document data for keyword extraction is stored, and character strings appearing before and after the plurality of keywords in a search result of the plurality of keywords of the input keyword are used as a pattern. Pattern extracting means for extracting;
A keyword extraction unit that performs a full text search in the keyword extraction database for the pattern extracted by the pattern extraction unit, extracts an expression surrounded by the pattern, and outputs the extracted expression as an output keyword;
A question answering apparatus characterized by that. The question answering device according to claim 1,
The keyword increasing means is:
Based on the input first keyword, a third keyword is output as an output keyword. Based on the input second keyword, a fourth keyword is output as an output keyword.
The answer candidate extraction means learns what kind of answer the problem will be by using a large number of sets of questions prepared in advance and answers to the problem, and outputs the output based on the learning result. A question answering apparatus, wherein an answer candidate that is a candidate for an answer to a question composed of the third keyword and the fourth keyword is extracted. The question answering device according to claim 1,
The keyword increasing means is:
Based on the input first keyword, a third keyword is output as an output keyword. Based on the input second keyword, a fourth keyword is output as an output keyword.
The answer candidate extracting unit extracts the output third document data and the fourth keyword data including the fourth keyword from a large amount of document data stored in the storage unit in advance, and language representation of the extracted document data From the above, using the frequency of appearance in the large amount of document data group, an answer candidate for the question constituted by the output third keyword and fourth keyword is extracted. . The question answering device according to claim 1,
Interrogative pronoun input means for inputting interrogative pronouns associated with the second keyword;
Based on the question pronoun input by the question pronoun input means, an answer type estimation means for estimating an answer type which is a type of a linguistic expression of an answer candidate for a question configured by an output keyword output by the keyword increasing means And
The keyword increasing means outputs a third keyword as an output keyword based on the input first keyword, and outputs the input second keyword as an output keyword.
The answer candidate extraction means searches the document data group including the third keyword and the second keyword output by the keyword increasing means from the document data group that is the search target of the answer candidate. Extracting from the extracted document data a linguistic expression that matches the answer type estimated by the answer type estimating means as an answer candidate of a question composed of the third keyword and the second keyword. A question answering device. The question answering device according to claim 1,
Based on a question pronoun associated with the second keyword determined in advance, an answer type that is a type of a linguistic expression of an answer candidate for a question constituted by an output keyword output by the keyword increasing means is estimated Answer type estimation means for
The keyword increasing means outputs a third keyword as an output keyword based on the input first keyword, and outputs the input second keyword as an output keyword.
The answer candidate extraction means searches the document data group including the third keyword and the second keyword output by the keyword increasing means from the document data group that is the search target of the answer candidate. Extracting from the extracted document data a linguistic expression that matches the answer type estimated by the answer type estimating means as an answer candidate of a question composed of the third keyword and the second keyword. A question answering device. The question answering device according to claim 1,
The keyword increasing means is:
Based on the input first keyword, a third keyword is output as an output keyword. Based on the input second keyword, a fourth keyword is output as an output keyword.
The answer candidate extraction means searches the document data group including the third keyword and the fourth keyword output by the keyword increasing means from the document data group that is the search target of the answer candidate. A linguistic expression suitable for a predetermined answer type is extracted from the extracted document data as an answer candidate of a question composed of the output third keyword and fourth keyword. Question answering device. The question answering device according to claim 1,
An answer type that is a type of linguistic expression of answer candidates for a question configured by the output keyword output by the keyword increasing means, and that is associated with the second keyword input to the keyword input means An answer type input means for inputting the type is provided.
The keyword increasing means outputs a third keyword as an output keyword based on the input first keyword, and outputs a fourth keyword as an output keyword based on the input second keyword. And
Similar keyword determining means for determining a similar keyword to the output fourth keyword among the second keywords as a similar keyword for each of the fourth keywords,
The answer candidate extraction means searches the document data group including the third keyword and the fourth keyword output by the keyword increasing means from the document data group that is the search target of the answer candidate, and extracts by this search processing A linguistic expression that matches the answer type input to the answer type input means by associating the output fourth keyword with a similar keyword that is similar to the output fourth keyword. And a fourth keyword as a candidate for answering a question. The question answering device according to claim 7,
The similar keyword determination means includes
A co-occurrence word that is a word that co-occurs with the fourth keyword output by the keyword extraction means is extracted from a large amount of document data stored in advance in the storage means, and the fourth For each keyword, find a co-occurrence vector that is a vector whose elements are the number of occurrences in the document data group co-occurring with each of the extracted co-occurrence words,
The degree of similarity between the co-occurrence vector for each fourth keyword and the co-occurrence vector for the same fourth keyword as the second keyword input to the keyword input means is obtained, and the obtained degree of similarity is obtained. A question answering apparatus characterized in that a fourth keyword identical to a second keyword similar to each of the fourth keywords is determined as the similar keyword. The first keyword itself is a plurality of keywords, the second keyword itself is a plurality of keywords, and the first keyword and the second keyword are input. The first keyword and the second keyword Along with answers to question data expressed in natural language composed of keywords, a third keyword obtained by increasing the first keyword and a fourth keyword obtained by increasing the second keyword A question answering method for outputting an answer to question data expressed in a natural language,
Inputting a first keyword composed of a plurality of keywords and a second keyword composed of a plurality of keywords;
Based on the input first keyword and second keyword, the third keyword and the fourth keyword are extracted by increasing the number of the respective keywords constituting the first keyword and the second keyword. Output step,
An answer candidate that is a candidate for an answer to a question constituted by a plurality of increased output keywords including the third keyword and the fourth keyword is extracted from a document data group that is a search target for answer candidates stored in advance. Steps,
Outputting a table in which each extracted answer candidate and a question are associated with each other as an answer table,
Extracting more keywords than the number of input keywords includes:
The input keyword is subjected to a full text search in a keyword extraction database in which document data for keyword extraction is stored, and character strings appearing before and after the plurality of keywords in a search result of the plurality of keywords of the input keyword are used as a pattern. Extracting steps;
Searching the full text of the extracted pattern in the keyword extraction database, extracting an expression surrounded by the pattern, and outputting the extracted expression as an output keyword.
A question answering method characterized by that. The first keyword itself is a plurality of keywords, the second keyword itself is a plurality of keywords, and the first keyword and the second keyword are input. The first keyword and the second keyword Along with answers to question data expressed in natural language composed of keywords, a third keyword obtained by increasing the first keyword and a fourth keyword obtained by increasing the second keyword A question answering program for causing a computer provided in a question answering device to output an answer to question data expressed in a natural language,
In the computer,
A process of inputting a first keyword composed of a plurality of keywords and a second keyword composed of a plurality of keywords;
Based on the input first keyword and second keyword, the third keyword and the fourth keyword are extracted by increasing the number of the respective keywords constituting the first keyword and the second keyword. Processing to output
An answer candidate that is a candidate for an answer to a question constituted by a plurality of increased output keywords including the third keyword and the fourth keyword is extracted from a document data group that is a search target for answer candidates stored in advance. Processing,
And a process of outputting a table in which each extracted answer candidate and a question are associated as an answer table ,
The process of extracting more keywords than the number of input keywords is as follows:
The input keyword is subjected to a full text search in a keyword extraction database in which document data for keyword extraction is stored, and character strings appearing before and after the plurality of keywords in a search result of the plurality of keywords of the input keyword are used as a pattern. Processing to extract,
A question answering program, comprising: performing a full text search on the extracted pattern in the keyword extraction database, extracting an expression surrounded by the pattern, and outputting the extracted expression as an output keyword.

Images