JP2010009471A - Query reply retrieval system, and method and program therefor - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a query reply retrieval system exhibiting a proper narrowing keyword, when incapable of narrowing a reply from an input query sentence, and capable of supporting specification of the reply, and a method and a program therefor. <P>SOLUTION: This query reply retrieval system includes a case database storing a case including a pair of the query sentence and a reply sentence, and a narrowing keyword selecting means for calculating a significance of the keyword included in the query sentence of the case, using the similarity between the paired reply sentences of the query sentences, and for selecting the narrowing keyword for narrowing a reply for an inquiry sentence from the keywords, based on the significance. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a question answer search system, method and program thereof.

  An operator of a company contact center or the like is required to respond appropriately to inquiries from customers by telephone or e-mail. The contact center operator record is stored as a case database in which pairs of questions from customers and operator responses are recorded. If there is a question answering system that can search for a reference answer from such a case database, it can be expected to improve the quality of the answer contents of the operator and shorten the response time.

Conventionally, a question answering system using a case database has been proposed. When a user inputs a question sentence, these detect the question sentence stored in the case database having a high similarity to the question sentence, and present the answer sentence to the user as an answer candidate. For example, Patent Document 1 is characterized in that, when calculating the similarity, a question sentence input using precise language analysis is compared with a question sentence in a case database.
JP 2006-244262 A

  The problem with conventional question answering systems is that the answers cannot always be specified sufficiently by the input question text alone. For example, as shown in FIG. 1, as a result of searching the failure case database using “paper jam” as a question sentence, answer sentences of five cases similar to the question sentence (including “paper jam”) are searched as answer candidates. . Here, pairs of failure phenomena and causes are stored in the failure case database, and a question sentence is searched for a phenomenon sentence, and a cause sentence is presented as an answer candidate. From this example, the phenomenon of “paper jam” alone does not reveal whether the cause of the failure is in the “LD unit” or the “pick roller”, and the answer to the question cannot be specified sufficiently.

  Therefore, the present invention has been invented in view of the above problems, and in the case where answers cannot be narrowed down based on an inputted question sentence, an appropriate narrowing keyword is presented, and a question answer search that supports the identification of answers It is to provide a system and method and program thereof.

  The present invention for solving the above-mentioned problems is a case database in which a case including a pair of a question sentence and an answer sentence is stored, and the importance of a keyword included in the question sentence of the case is determined between the answer sentences of the question sentence pair. And a narrowed keyword selection means for selecting a narrow keyword for narrowing down answers to a query sentence from the keywords based on the degree of importance.

  The present invention for solving the above problem is to calculate the importance of a keyword included in a question sentence of a case including a pair of a question sentence and an answer sentence using the similarity between the answer sentences of the question sentence pair, This is a question answer search method in which a narrow-down keyword for narrowing down answers to an inquiry sentence is selected from the keywords based on importance.

  The present invention for solving the above-described problem is a process of calculating the importance of a keyword included in a question sentence of a case including a pair of a question sentence and an answer sentence using the similarity between the answer sentences of the question sentence pair; A program for causing the information processing apparatus to execute a process of selecting, from the keywords, a narrowing keyword for narrowing down answers to the inquiry sentence based on the importance.

  The present invention can support the identification of an answer by presenting a keyword for narrowing down an answer when the answer cannot be sufficiently identified from the input question text (cannot be narrowed down).

  The best mode for carrying out the present invention will be described in detail with reference to the drawings.

  First, an outline of an embodiment of the present invention will be described.

  For example, it is assumed that “paper jam” is input as a question sentence. The case search means searches for cases including “paper jam” in the question sentence as shown in FIG. As a result, five cases of ID1 to ID5 are obtained.

  Next, the narrow-down keyword selection unit calculates the importance of the keyword included in the question sentence of the case obtained by the case search unit using the similarity between the answer sentence paired with the question sentence. Here, if the independent word included in the question sentence is a keyword, “paper jam”, “frequent occurrence”, “printing”, “light”, “recent”, “think”, and the like correspond to the five cases. The importance of all these keywords is obtained by the average value of the similarities between the answer sentences of the cases including the keyword. For example, as shown in FIG. 2, when calculating the importance of the keyword “light”, the similarity between the answer sentences ID1 to ID3 including “light” sim (1,2), sim (1,3), Find the average value of sim (2,3). Similarly, when calculating the importance of “printing” as shown in FIG. 3, the average value of the similarity between the answer sentences ID1 to ID5 including “printing” is obtained. A method of calculating the similarity will be described later. Here, it is assumed that the average value (importance) of the similarity between the answer sentences is 0.78 for “light” and 0.43 for “print”.

  The narrowed keyword selection means selects a number specified in advance in the order of importance or a keyword having an importance level equal to or higher than a threshold as a narrow keyword and presents it to the user.

  The user looks at the presented refinement keywords and confirms that there is a failure corresponding to the refinement keywords such as “(Light)” and “Abnormal sound” in addition to “Paper jam”. If it has occurred, it is possible to identify (narrow down) the answers by searching again with the question sentence including the narrowed keyword. 4 and 5 show the narrowing keywords and the corresponding cases, and FIG. 4 shows the narrowing keyword “(printed) is light” and the corresponding cases, and FIG. 5 shows the narrowing. The keyword “abnormal sound” and a corresponding example are displayed.

  A method for calculating the importance of a keyword based on the similarity between documents including the keyword is called Term Strength, which is referred to in Reference 1 (Wilbur, JW and Sirotkin, K., The automatic identification of stop words, Journal of Information Science. , 18, pp.45-55, 1992.). This is an index that aims at the ease of becoming the subject of a keyword (≒ importance of the keyword role in a document ≒ less likely to become an unnecessary word) .The closer the keyword is to the subject of the document, the more similar the content of the target document containing the keyword. Based on the nature of being easy. The difference between the above document 1 and the present embodiment is that if document 1 is applied as it is, the importance of the keyword included in the question sentence is calculated based on the similarity between the question sentences. In this form, the importance of the keyword included in the question sentence is calculated based on the similarity between the answer sentences. This difference makes it possible to present a refined keyword that makes it easier to identify answers (high similarity between answers) when answers cannot be sufficiently identified (cannot be narrowed down) from the entered question text. Has an effect.

  Intuitively, there is a correlation between the contents of the question sentences in the case and the contents of the answer sentences. Therefore, if the question sentences in the case are similar, the answer sentences are also similar. It is considered that the same effect can be obtained. However, as in the example of FIG. 1, the phenomenon of failure is described very finely in the question sentence, and there are many parts (independent words) that are irrelevant to the identification of the answer. Therefore, there are cases where the answer sentences are similar but the question sentences are not similar. Even if the question sentences are similar to each other, the answer sentences are not necessarily similar if the question sentences in these cases have very abstract contents. This is effective when there is no correlation between the content of the question text and the content of the answer text.

  Next, a specific first embodiment will be described.

  Referring to FIG. 6, the question answer search system according to the present embodiment includes a data processing device 1 that operates under program control, a storage device 2 that stores information, an input device 3 such as a keyboard, and an output device 4 such as a display. Including.

  In the question answer search system, when the above four devices are mounted on one hardware, the data processing device 1 and the storage device 2 exist on the server, and the input device 3 and the output device 4 are the clients. May be present in another device above. In the latter, information input from the input device 3 on the client is transmitted to the data processing device 1 on the server via the network, and information output from the data processing device 1 on the server is transmitted to the client via the network. It is transmitted to the output device 4.

  The data processing device 1 includes a case search unit 10, a refined keyword selection unit 11, and a display unit 12.

  The case search unit 10 searches the case database 20 for cases where the question sentence input through the input device 3 and the question sentence in the case database 20 are similar, and stores the result in the case ID storage unit 21.

  The refinement keyword selection unit 11 refers to the case ID storage unit 21 and the case database 20 to determine the importance of the keyword included in the question text of the case obtained by the case search unit 10 and the similarity between the answer sentences. Calculate using. Then, the number of keywords specified in advance in the order of importance or the keyword whose importance is greater than or equal to the threshold value is selected as a refinement keyword and stored in the refinement keyword storage unit 22.

  The display unit 12 refers to the refined keyword storage unit 22 and the case database 20, and displays the refined keyword obtained by the refined keyword selection unit 11 and the example obtained by the case search unit 10 including the refined keyword in the output device 4. Send.

  The storage device 2 includes a case database 20, a case ID storage unit 21, and a narrowed keyword storage unit 22. The storage device 2 is usually realized by an auxiliary storage device such as an HDD, but may be a memory. Further, the case database 20, the case ID storage unit 21, and the narrowed keyword storage unit 22 may not be all included in the storage device 2. May be in different places.

  In the case database 20, one case includes a pair of a question sentence and an answer sentence, and each case is given a case ID for identifying the case.

  The case ID storage unit 21 stores the case ID obtained by the case search unit 10.

  The refinement keyword storage unit 22 stores refinement keywords selected by the refinement keyword selection unit 11.

  Next, the operation of the present embodiment will be described in detail with reference to FIG. Note that failure case data is stored in the case database 10, and each case is composed of a pair of “failure phenomenon” corresponding to a question sentence and “cause of failure” corresponding to an answer sentence. An example will be described. Also, the question sentence and the answer sentence in the case database 10 are divided into words by morphological analysis in advance, and only independent words are extracted and stored in the storage device 2.

  First, the case search unit 10 searches the case database 20 for cases where the question sentence input through the input device 3 is similar to the question sentence in the case database 20, and stores the result in the case ID storage unit 21 (FIG. 7 step S1).

  The question sentence includes a word input such as “paper jam” and a sentence input such as “paper jam occurs”. In the case of word input, a case that includes the word in the question sentence is searched. When a plurality of words are input as in the case of “paper jam occurrence” (space is a word separator), a case including all words is searched. In the case of sentence input, only independent words are extracted using morphological analysis, and the words are the same as word input. Here, “paper jam” is input from the input device 3, ID1 to ID5 of FIG. 8 are searched from the case database 20 as examples including “paper jam” in the question sentence, and stored in the case ID storage unit 21. Shall.

  Next, the refinement keyword unit 11 refers to the case ID storage unit 21 and the case database 20 to determine the importance of the keyword included in the question sentence of the case obtained by the case search unit 10 between the answer sentences. The calculation is performed using the similarity, and the number of keywords specified in advance in the order of importance or keywords whose importance is greater than or equal to the threshold value are stored in the refined keyword storage unit 22 as refined keywords (step S2 in FIG. 7).

  Here, the keyword included in the question sentence represents an independent word included in the question sentence. FIG. 9 shows independent words included in the question sentence and the answer sentence of the cases ID1 to ID5 obtained by the case search unit 10. “Abnormal sound”, “light”, “printing”, “paper jam”, “occurrence”, “think”, “middle”, “power”, “frequent”, “large”, “UNIT”, “recent”, “Okuru” and “Throw in” are keywords for importance calculation.

  The importance of the keyword is calculated using the similarity between the answer sentences. Specifically, as shown in FIG. 10 (Equation 1), the importance score Score (N) of the keyword N is selected from the cases D_N that are examples obtained by the case search means and include N in the question sentence. All combinations of similarity sim (d_i, d_j) of the two selected answer sentences are calculated, and the average value is obtained. The number of combinations is | D_N | x (| D_N | -1) / 2 with respect to | D_N |. There are various methods for calculating the importance of the keyword, as will be described later.

  FIG. 11 (Formula 5) shows a calculation formula of sim (d_i, d_j). sim (d_i, d_j) is a calculation method generally called cosine similarity. In (Expression 5), the weight w (d, t) of the word (independent word) t included in the case d is calculated by the method of (Expression 6) or (Expression 7). In (Equation 6), the weight of every word is always 1 and (Equation 7) is the product of the number of occurrences tf (d, t) of t in d and the inverse of the number of occurrences df (t) of t. Ask. Note that the calculation method of sim (d_i, d_j) can be applied to any method for calculating the similarity between two documents other than the cosine similarity. For example, a Jaccard coefficient or a dice coefficient may be used.

  An example of calculating the importance of the keyword N is shown in FIG. FIG. 12A calculates the score Score (thin) of the keyword “thin” included in the question sentences of the examples ID1 to ID5 in FIG. 8 obtained by the document search unit 10. “Thin” is included in the question texts of the three cases ID1-ID3, so there are three similarities, sim (1,2), sim (1,3), and sim (2,3). Ask for. (B) of FIG. 12 calculates sim (1, 2) as a calculation example of sim (d_i, d_j). Each value of w (d, t) is calculated as shown in FIG. For example, in the case of ID1, “LD” appears once and “LD” appears in 3 out of 5 cases, so 1 * log (5/3) = 1.42.

  FIG. 13 shows the result of calculating the importance of the keyword included in the question sentence of the case of FIG. 8 obtained by the case search unit 10. (Equation 1) in FIG. 10 cannot be calculated because the denominator is 0 when the number of cases including keywords | D_N | is 1. Therefore, the keyword of | D_N | = 1 is 0. The narrowed keyword selection unit 11 stores the number of keywords specified in advance in the order of importance or the keywords whose importance is equal to or greater than the threshold as narrowed keywords in the narrowed keyword storage unit 22. For example, when a keyword having an importance of 0.7 or more is selected as a narrowing keyword, two of “abnormal sound” and “light” are stored in the narrowing keyword storage unit 22 as narrowing keywords. At this time, the narrowed keyword storage unit 22 registers the narrowed keyword and the case ID including the narrowed keyword as shown in FIG.

  Finally, the display unit 12 refers to the narrowed keyword storage unit 21 and the case database 20, and transmits the narrowed keyword obtained by the narrowed keyword selection unit 11 and the case obtained by the case search unit 10 to the output device 4. (Step S3 in FIG. 7).

  As shown in FIG. 14, the narrowed keyword storage unit 22 stores correspondences of case IDs obtained by the case search unit 10 including “abnormal sound” and “light”, and the contents are output to the output device 4. To send to.

  The operation of the present embodiment has been described above. The importance of the keyword included in the question text of the case obtained by the case search unit 10 in step S2 of FIG. 7 was calculated using the similarity between the answer sentences. In the above, (Equation 1) in FIG. 10 is used as a specific implementation method, but there are other implementation methods.

  In the realization method of (Equation 2) in FIG. 10, the importance Score (N) of the keyword N is the case obtained by the case search unit 10 and the center of the answer sentence of the case D_N including the keyword in the question sentence is obtained. After that, the average value of the similarity sim (d_i, d_DN) of each answer sentence and the center of the answer sentence is obtained. The center of the answer sentence is obtained by concatenating all the answer sentences of D_N. For example, the center of the response from ID1 to ID5 in Fig. 8 is "LD unit was broken / LD unit replaced./LD unit failed. Replacement / pick roller failure./ Apparently pick roller failure. ("/" Is a delimiter of the answer sentence). In (Equation 1), it is necessary to find all combinations of the similarity between the two answer sentences D_N, but (Equation 2) only needs to calculate the similarity between the center of each answer sentence and the answer sentence. , (Equation 1) can be realized at higher speed.

  In the realization method of (Equation 3) in FIG. 10, the importance score Score (N) of the keyword N is changed from the value of (Equation 1) to the expected value E [Score (N)] of the importance of the keyword having the number of appearance cases. By subtracting the value of For example, in the case of FIG. 8 obtained by the document search means, “thin” appears in three question sentences ID1 to ID3. The “thin” score is obtained by the method of (Equation 1), the expected value of the score of (Equation 1) of the keyword appearing in the three question sentences is obtained, and the difference between the two is obtained. Due to the nature of (Equation 1), there is a general tendency that a keyword with a smaller number of appearance cases has a higher score and a keyword with a higher number of appearance cases has a lower score. Therefore, if (Equation 1) is applied as it is, a keyword with a small number of appearance cases will be overestimated. As shown in (Expression 3), this problem can be solved by subtracting the expected value.

  The implementation method of (Expression 4) in FIG. 10 is the same as (Expression 3) except that (Expression 2) is used instead of (Expression 1).

  (Equation 9) to (Equation 12) in FIG. 15 show the importance of the keywords included in the question sentence of the case obtained by the case search unit 10 in addition to the similarity between the answer sentences, It differs from (Equation 1) to (Equation 4) in that the number of cases is used. With this implementation method, it is possible to present keywords that are easy to identify answers and at the same time have a high possibility of realizing a user's question as a response keyword. The reason why the logarithm log is given to the term of the number of appearance cases | D_N | in (Expression 9) to (Expression 12) is to reduce the influence of the number of appearance cases. Of course, it is not necessary to add a log.

  The question answer search system presents the narrowed keyword to the user, and if the user can materialize the question text that was initially input by the narrowed keyword, the search is newly performed with the question sentence including the narrowed keyword. . For example, when the initial question sentence “paper jam” is presented as a keyword that narrows “light” and “abnormal sound”, the user can add “(paper) light” and “abnormal sound” in addition to “paper jam”. ”Is investigated whether a phenomenon such as“ As a result, if “(printing) is thin” occurs, a new case is searched for “thin” and “paper jam”. Here, if there are 3 cases including “thin” and 1 case including “abnormal sound”, it can be said that “thin” is more likely to actually occur than “abnormal sound”. . In this way, by considering the number of appearance cases, it is possible to present a keyword with a high possibility that the user can materialize the question as a response keyword.

  (Equation 13) and (Equation 14) in FIG. 16 show the importance of keywords included in the question sentence of the case obtained by the case search unit 10 in addition to the similarity between the answer sentences. It differs from (Expression 1) to (Expression 4) and (Expression 9) to (Expression 12) in that the similarity is used. In QScore (N) in (Expression 13) and (Expression 14), Score (N) of (Expression 1) to (Expression 4) and (Expression 9) to (Expression 12) is calculated for the question sentence. That is, | D_N | may be calculated by replacing | D_N | with a case set including N in the question sentence, and sim (d_1, d_2) with the similarity between the question sentence d_1 and the question sentence d_2. AScore (N) is the same as Score (N) in (Formula 1)-(Formula 4) and (Formula 9)-(Formula 12).

  When there are a plurality of correct answers to a certain question, if the similarity between the answer sentences is used, the importance of the keyword for identifying those answers does not necessarily increase. On the other hand, there is a certain correlation between the content of the question text and the content of the answer text, so if the similarity between the question texts is high, it may be a keyword that identifies the answer even if the similarity between the answer texts is not high Becomes higher. Therefore, the similarity between question sentences is also used.

  Another embodiment will be described.

  Up to this point, we have used cases that consist of pairs of one question sentence and one answer sentence, but some cases consist of pairs of one question sentence and multiple answer sentences. There are also things. For example, in a QA site on the Web, as shown in FIG. 17, when a user registers a question sentence through the Web, an answer sentence to that is registered from another user. One of the registered answer sentences is selected as the best answer sentence. Currently known as QA sites are Yahoo Wisdom Bag (http://chiebukuro.yahoo.co.jp/), Teach Goo (http://oshiete.goo.ne.jp/), OKWave "(http://okwave.jp/) exists.

  A method of applying the present invention to an example including a pair of such one question sentence and a plurality of answer sentences will be described. The simplest method is to execute the narrow-down keyword selection unit 11 using the best answer sentence selected in advance among a plurality of answer sentences. In this case, it can be considered that it is substantially composed of a pair of one question sentence and one answer sentence.

  As another method, a narrowed keyword portion 11 is obtained after expanding a case including a pair of n response sentences for one question sentence into n cases including one answer sentence for one question sentence. Is to apply. For example, a case consisting of {Q1, {A1, A2}} is expanded to {Q1, A1} {Q1, A2}. Here, {Q1, {A1, A2}} means one example consisting of question sentence Q1 and two answer sentences A1, A2, question sentence Q1 and answer sentence A1, and question sentence Q1 and answer sentence A2. This represents an expansion to two cases.

  As yet another method, the narrowed-down keyword unit 11 is executed after obtaining the centers of a plurality of answer sentences for one question sentence. In this case, it can be considered that it is substantially composed of a pair of one question sentence and one answer sentence. The method for obtaining the center may be obtained by concatenating a plurality of answer sentences as in the case of (Formula 2) in FIG.

  As another method, when the refinement keyword unit 11 calculates the importance of the keyword included in the question sentence of the case obtained by the case search unit 10 using the similarity between the answer sentences of the cases, There is a method in which the similarity of all combinations of the plurality of answer sentences and the plurality of case answer sentences is obtained, and the minimum value, maximum value, or average value thereof is used as the degree of similarity between the case answer sentences. For example, in case A {Q1, {A1, A2}, case B {Q2, {A3, A4}}, the similarity between the answer sentences of case A and case B is expressed as sim (A1, A3), sim (A1, A4), sim (A2, A3), and sim (A2, A4) are obtained from the minimum value, maximum value, or average value.

  The effect of this embodiment is to support the identification of answers by presenting response keywords that make it easy to identify answers when the answers cannot be sufficiently identified from the input question text (cannot be narrowed down). It is possible. The reason is that the importance of keywords included in the question text of the case obtained by the case search unit 10 is calculated using the similarity between the answer sentences, and the number or importance specified in advance in the order of importance. This is because the narrowed keyword part 11 is used as a narrowed keyword.

  The question answer search system according to the present invention can be used when searching for a reference answer from an accumulated case database in order to quickly and accurately answer a customer inquiry received by an operator of a contact center. It can also be used to search for QA for users to reference on the Web QA site.

FIG. 1 is a diagram for explaining the problems of the prior art. FIG. 2 is a diagram for explaining the outline of the present invention. FIG. 3 is a diagram for explaining the outline of the present invention. FIG. 4 is a diagram for explaining the outline of the present invention. FIG. 5 is a diagram for explaining the outline of the present invention. FIG. 6 is a diagram showing the configuration of the embodiment of the present invention. FIG. 7 is a flowchart showing the operation of the embodiment of the present invention. FIG. 8 is a diagram for explaining the operation of the embodiment of the present invention. FIG. 9 is a diagram for explaining the operation of the refinement keyword unit 11 according to the embodiment of this invention. FIG. 10 is a diagram for explaining the operation of the refinement keyword unit 11 according to the embodiment of this invention. FIG. 11 is a diagram for explaining the operation of the refinement keyword unit 11 according to the embodiment of this invention. FIG. 12 is a diagram for explaining the operation of the refinement keyword unit 11 according to the embodiment of this invention. FIG. 13 is a diagram for explaining the operation of the refinement keyword unit 11 according to the embodiment of this invention. FIG. 14 shows a storage example of the narrowed keyword storage unit 22 according to the embodiment of this invention. FIG. 15 is a diagram for explaining the operation of the refinement keyword unit 11 according to the embodiment of this invention. FIG. 16 is a diagram for explaining the operation of the refinement keyword unit 11 according to the embodiment of this invention. FIG. 17 is a diagram for explaining the operation of the refinement keyword unit 11 according to another embodiment of the present invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Data processing device 2 Storage device 3 Input device 4 Output device 10 Case search part 11 Refinement keyword selection part 12 Display part 20 Case database 21 Case ID storage part 22 Refinement keyword storage part

Claims (33)

A case database that stores cases containing pairs of question and answer sentences;
The importance of the keyword included in the question sentence of the case is calculated using the similarity between the answer sentences of the question sentence pairs, and the narrow-down keyword for narrowing down the answer to the inquiry sentence based on the importance is A question answer search system having narrowed keyword selection means for selecting from keywords.   The question answer search system according to claim 1, further comprising case search means for searching the case database for cases where the inputted inquiry sentence is similar to the question sentence in the case database.   The narrowed keyword selection means calculates the importance of the keyword included in the question sentence of the case searched by the case search means using the similarity between the answer sentences of the question sentence pairs, and based on the importance The question answer search system according to claim 2, wherein a narrow-down keyword for narrowing down answers to the inquiry sentence is selected from the keywords.   The question answer search system according to claim 1, wherein the narrowing keyword selection unit selects a predetermined number of keywords as narrowing keywords in the order of importance.   The question answer search system according to claim 1, wherein the narrowed keyword selection unit selects a keyword having the importance level equal to or higher than a threshold as a narrow keyword.   The question answer search system according to claim 1, further comprising display means for transmitting the narrowed keyword and an example including the narrowed keyword to an output device.   The narrowed-down keyword selection unit calculates the importance of the keyword included in the question sentence of the case using the similarity between the answer sentences of the case and the number of appearance cases. Question answer search system described in.   The narrowed keyword selection means calculates the importance of a keyword included in the question sentence of the case using the similarity between the answer sentences of the case and the similarity between the question sentences of the case. The question answer search system according to claim 6.   The narrowed-down keyword selection means calculates the similarity between the answer sentences of the case examples including the keyword in the question sentence for all combinations of the answer sentences in the case examples that include the keyword in the question sentence, and calculates the average value of the importance of the keyword The question answer search system according to any one of claims 1 to 6.   The narrowed-down keyword selection means obtains the center of the answer sentence of the case including the keyword in the question sentence, and calculates the average value of the similarity between each answer sentence of the case including the keyword in the question sentence and the center of the answer sentence, The question answer search system according to any one of claims 1 to 6, wherein the importance level of the keyword is set.   7. The narrowed-down keyword selection means uses a value obtained by subtracting an expected value of the importance of a keyword having the number of appearance cases of the keyword from the obtained importance of the keyword as the importance of the keyword. Question answer search system described in any one.   The narrow-down keyword selection means, when one case is composed of a pair of one question sentence and a plurality of answer sentences, one example is regarded as a pair of one question sentence and one answer sentence. The question answer search system according to any one of claims 1 to 11.   The narrow-down keyword selection means, when one case is composed of a pair of one question sentence and a plurality of answer sentences, selects the best answer sentence from a plurality of answer sentences, and selects the case as the question sentence. The question and answer search system according to claim 12, which is regarded as a pair of the selected answer sentence and the selected answer sentence.   The narrow-down keyword selection means expands into n cases including one answer sentence for one question sentence when one case is composed of a pair of one question sentence and n answer sentences. The question and answer search system according to claim 12.   The narrow-down keyword selection means obtains the center of a plurality of answer sentences for one question sentence when one example is composed of a pair of one question sentence and a plurality of answer sentences, The question answer search system according to claim 12, which is regarded as a pair of the answer sentence and the center of the answer sentence.   The narrow-down keyword selection means, when one case is composed of a pair of one question sentence and a plurality of answer sentences, obtains the similarity of all combinations of a plurality of answer sentences of the target case, The question answer search system according to claim 12, wherein the minimum value, the maximum value, or the average value is set as the similarity between the answer sentences in the cases.   The importance of the keyword included in the question sentence of the case including the pair of the question sentence and the answer sentence is calculated using the similarity between the answer sentences of the question sentence pair, and the query sentence is calculated based on the importance degree. A question answer search method for selecting a narrow-down keyword for narrowing down an answer from the keyword.   The question answer search method according to claim 17, wherein a case where a query sentence and a question sentence are similar is searched from a case database storing cases including pairs of question sentences and answer sentences.   The importance of the keyword included in the question sentence of the searched case is calculated using the similarity between the answer sentences of the question sentence pairs, and the answers to the inquiry sentence are narrowed down based on the importance. The question answer search method according to claim 18, wherein a narrow-down keyword is selected from the keywords.   18. The question answer search method according to claim 17, wherein a predetermined number of keywords are selected as narrowing keywords in order of importance.   The question answer search method according to claim 17, wherein keywords having the importance level equal to or higher than a threshold are selected as narrowing keywords.   The question answer search method according to claim 17, wherein the narrowed keyword and an example including the narrowed keyword are displayed.   The question answer search method according to any one of claims 17 to 22, wherein the importance of a keyword included in the question sentence of the case is calculated using a similarity between the answer sentences of the case and the number of appearance cases. .   The importance of the keyword included in the question sentence of the case is calculated using the similarity between the answer sentences of the case and the similarity between the question sentences of the case. The method for searching for answers to questions.   18. The degree of similarity between the answer sentences of cases including a keyword in the question sentence is calculated for all combinations of the answer sentences of cases containing the keyword in the question sentence, and the average value is set as the importance of the keyword. Item 23. The question answer search method according to Item 22.   The center of the answer sentence of the case including the keyword in the question sentence is obtained, and the average value of the similarity between each answer sentence of the case including the keyword in the question sentence and the center of the answer sentence is set as the importance of the keyword. The question answer search method according to any one of claims 17 to 22.   The question answer according to any one of claims 17 to 22, wherein a value obtained by subtracting an expected value of the importance of a keyword having the number of appearance cases of the keyword from the obtained importance of the keyword is set as the importance of the keyword. retrieval method.   Claims 17 to 17 in which one example is regarded as a pair of one question sentence and one answer sentence when one example is composed of a pair of one question sentence and a plurality of answer sentences. Item 28. The question answer search method according to any one of Items 27.   When one example is composed of a pair of one question sentence and a plurality of answer sentences, the best answer sentence is selected from among a plurality of answer sentences, and the example sentence is selected as the question sentence and the selected answer sentence. The method for retrieving a question answer according to claim 28, which is regarded as a pair with.   30. The method according to claim 28, wherein when one example is composed of a pair of one question sentence and n answer sentences, the one question sentence is expanded into n cases including one answer sentence. Question answer search method.   When one example is composed of a pair of one question sentence and a plurality of answer sentences, the center of a plurality of answer sentences for one question sentence is obtained, and the example is defined as the center of the question sentence and the answer sentence. The method for searching for answers to a question according to claim 28, which is regarded as a pair.   When one case is composed of a pair of one question sentence and multiple answer sentences, the similarity of all combinations of multiple answer sentences of the target case is obtained, and the minimum value, maximum value, or average 29. The question answer search method according to claim 28, wherein the value is a similarity between the answer sentences in the cases. A process of calculating the importance of a keyword included in a question sentence of a case including a pair of a question sentence and an answer sentence using the similarity between the answer sentences of the question sentence pair;
A program for causing an information processing apparatus to execute a process of selecting a narrow-down keyword for narrowing down answers to an inquiry sentence from the keyword based on the importance.

Images