JP2012079161A - Natural language text generation device and computer program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a query answering system capable of highly accurately and automatically generating a response text to a query by a natural language text of a broad range of fields.SOLUTION: A query answering device includes: a template assembly storage section 46 for storing a template of a response text to a query text; a template estimation section 48, upon receiving a query text, estimating a template having a prescribed relation with the query; a pattern extension processing section 52 generating an extension template in which a structure of words or a text is corrected by applying a word class and a template extension regulation to the estimated template; a matching section 60 for matching the extension template with a text of a Web corpus 32 and outputting response candidates to the query; a scoring selection section 68 calculating a score of the candidates indicating an eligibility as a replay to the query and outputting them in order of a high score. The extension template has two variables, and a word class which the word should have is specified to each variable.

Description

  The present invention relates to a system that receives a natural language question sentence and generates an answer, and more particularly, to a question answering apparatus that automatically generates a highly accurate answer to an arbitrary question issued in a natural language.

  Development of a question answering system that gives answers to various questions is in progress. For example, Non-Patent Document 1 described later discloses a technique for extracting word pairs having a certain relationship. In the technique disclosed in Non-Patent Document 1, a small amount of related words to be extracted is prepared, and a pattern is learned from the word pairs. At the time of answering, this pattern is used to extract the word pair of the relationship.

  However, in the technique described in Non-Patent Document 1, there is a problem in the extraction accuracy of the target word pair.

  On the other hand, Non-Patent Document 2 discloses an attempt to further expand the above-described pattern by paraphrasing and use it in the question answering system.

Patrick Pantell et al., Espresso: Expanding general patterns to automatically acquire semantic relationships, the 21st International Conference on Computer Language, and the 44th Annual Conference of the Computer Language Society (Colling-ACL-) 06), 113-120, 2006 (Patric Pantel et al., Espresso: Leveraging generic patterns forautomatically harvesting semantic relations. In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the Association for Computational Linguistics (Coling-ACL- 06), pages 113-120, 2006.) Deccan Ring et al., Discovery of Reasoning about Question Answering, Natural Language Engineering, 7 (4): 343-360, 2001 (Dekang Ling et al, Discovery of Inference Rules for QuestionAnswering. Natural Language Engineering, 7 (4): 343-360, 2001.)

  However, although the technique disclosed in Patent Document 2 described above can estimate the answer to a question sentence in a sufficiently learned field, there is a problem that the field that can be handled is limited.

  Limiting questions in advance to a specific field is useful in a system that is used only for a specific purpose, but it is desirable to have a question answering system that can automatically answer a wider range of questions if possible.

  SUMMARY OF THE INVENTION Therefore, an object of the present invention is to provide a question answering system capable of automatically generating an answer sentence with high accuracy in response to a question in a natural language sentence in a wide range of fields.

  A question answering apparatus according to a first aspect of the present invention comprises a template storage means for storing a template of an answer sentence for a question sentence, the template having first and second variable parts, and a question sentence. In response to receiving the input, one or a plurality of templates serving as a prototype of the answer sentence for the question sentence is estimated from the templates stored in the template storage unit, and the first of each template is determined among the question sentences. The template estimation means for outputting together with the constraint condition indicating the word corresponding to the variable part of the template, and the template output by the template estimation means is selected based on the word class of the word constituting the constraint condition of the template By applying the template extension rules prepared in advance to the template, the structure of the word or sentence And a template extension means for generating one or more extension templates were Tadashisa. Each of the extension templates includes first and second variables indicating positions of answer candidates for the question sentence, and the two variables are the first and second variables of the template stored in the template storage unit. The two variable parts are associated with each other, and the attribute to be satisfied by the variable is designated for each. The template expansion means selects an expansion template in which the attribute of the word constituting the constraint condition has a predetermined relationship with the first variable. The question answering device further matches each of the extension templates output by the template extension means with a number of sentences prepared in advance so as to be exchangeable with the second variable of the extension template. Matching means for outputting one or a plurality of words, and the candidate output from the matching means until the candidate is obtained, or the word included in the extended template that matches the candidate and the candidate A scoring and selection means for calculating a score indicating eligibility as an answer to the question sentence based on the occurrence frequency and outputting candidates in order of the highest score is included.

  The attribute of the variable may be a word class of a word assigned to the variable.

  Preferably, the question answering apparatus further includes a question sentence receiving means for receiving the question sentence as a voice signal by communication, converting it into a text string by voice recognition, and inputting the text string to the template estimation means.

  More preferably, the question answering apparatus further includes speech synthesis means for converting candidates output by the scoring and selection means into speech by speech synthesis.

  The question answering apparatus further receives the question sentence as a voice signal by communication, converts it into a text string by voice recognition, and outputs it to the template estimation means, and outputs it by the scoring and selection means. Speech candidates may be included for converting the candidate into a speech signal by speech synthesis and returning it to the terminal that sent the question sentence.

  A computer program according to a second aspect of the present invention comprises a template storage means for storing a computer as a template of an answer sentence for a question sentence, the template having first and second variable parts; In response to receiving the input of the question sentence, one or more templates that are the prototypes of the answer sentences for the question sentence are estimated from the templates stored in the template storage means, A template estimation unit for outputting together with a constraint condition indicating a word corresponding to the first variable part, a word class of a word constituting the constraint condition of the template for each of the templates output by the template estimation unit, and By applying pre-prepared template expansion rules, the structure of the word or sentence is corrected. A template extension means for generating one or a plurality of extended templates, each of the extension templates including first and second variables indicating positions of answer candidates for the question sentence, A variable is associated with each of the first and second variable parts of the template stored in the template storage means, and can be obtained in advance with each of the extension templates output by the template extension means. The matching means for outputting one or more candidates that can be exchanged with the second variable of the extension template by matching a large number of sentences prepared in the above, and the candidate output by the matching means Progress until a candidate is obtained, or a word included in an extended template that matches the candidate Based on the co-occurrence frequency with complement, it calculates a score indicating the eligibility of the answer to the question message, to function as scoring and selection means for outputting the candidate in the order of having a high the score.

1 is a block diagram of a question answering system according to an embodiment of the present invention. It is a figure which shows the example of a seed pattern set. It is a figure which shows the example of an extended pattern set. It is a flowchart which shows the control structure of the program for extending a seed pattern. It is a flowchart which shows the control structure of the program for extracting the word which matches a pattern from a web corpus. 1 is an external view of a computer system that realizes a question answering system according to an embodiment of the present invention. It is a block diagram which shows the hardware constitutions of the computer shown in FIG.

  In the following description and drawings, the same parts are denoted by the same reference numerals. Therefore, detailed description thereof will not be repeated.

[Constitution]
Referring to FIG. 1, a question answering apparatus 30 according to an embodiment of the present invention uses natural language sentences included in a Web corpus 32 composed of sentences collected from the Web, and inputs natural language sentences related to a wide range of fields. This is to generate an appropriate answer sentence 34 for the sentence 20 with high accuracy. It will be apparent to those skilled in the art that there is no problem even if a normal corpus is used instead of the Web corpus 32.

  The question answering device 30 stores a template set storage unit 46 that stores a large number of answer sentence templates for a question sentence, and a template stored in the template set storage part 46 as a template that best matches the pattern as an answer sentence for the input question sentence 20. And a template estimation unit 48 for estimation and extraction. There are usually a plurality of templates extracted by the template estimation unit 48. Each of these is called a seed template, and they are collectively called a seed template set. In the template estimation by the template estimation unit 48, an estimation device in which the relationship between the input sentence and the corresponding template is learned in advance by supervised machine learning can be used. The template estimation unit 48 identifies words that are included in any of the variables included in the seed template among the words obtained from the input question sentence, and outputs them together with the seed template as a constraint condition for the seed template. Therefore, what value should be put in any of the seed templates is determined.

  The question answering device 30 further includes a seed template set storage unit 50 for storing the seed template set output from the template estimation unit 48, a word class of words included in the seed template, a semantic relationship between words, and the like. A thesaurus 58 for storing, and a template expansion rule storage unit for storing template expansion rules prepared in advance to form each of a plurality of templates by expanding each of the seed templates in the seed template set storage unit 50 54, a template extension rule stored in the template extension rule storage unit 54, and a thesaurus 58, and a large number of extension templates different from the seed template are generated from each of the seed templates stored in the seed template set storage unit 50. Ten to do The rate extension processing unit 52, and a extension template set storage unit 56 for storing an extension template set output by the template extension processing unit 52.

  The template stored in the template set storage unit 46 is a sentence pattern for answering a question. Various sentences are collected for learning, question shapes are created from them, and these are prepared as templates in the template set storage unit 46.

  The template estimation unit 48 analyzes the input question sentence 20 and extracts a template that answers it from the template set storage unit 46. The template estimation unit 48 extracts a seed template, for example, by specifying the closest template from the templates stored in the template set storage unit 46 based on the similarity to the character strings constituting the input question sentence 20. A plurality of estimation results can be used. That is, a plurality of seed templates can be stored in the seed template set storage unit 50.

  FIG. 2 shows an example of a seed template. In the example of FIG. 2, it is assumed that the input question sentence 20 is “What is the enemy of Ultraman?”. As a template that matches the question sentence, templates such as “Y enemy is X”, “X is Y enemy”, etc. are obtained using X as a variable. In extracting the template from the input question sentence 20, “Ultraman” corresponds to the variable Y, and the variable X indicates a portion that can be an answer to the input question sentence 20.

  Templates can also be automatically generated from natural language sentences. In this embodiment, it is assumed that a human prepares manually. In this embodiment, the template is described by a so-called regular expression. Various regular expressions are known, but any of them may be used here.

  The thesaurus 58 associates a word class as an attribute for each word.

  The template expansion rule stored in the template expansion rule storage unit 54 is a rule for expanding the seed template stored in the seed template set storage unit 50 and generating an expansion template. Template extension rules can also be described using regular expressions. In this embodiment, each extended rule includes parsing information (typically a parsing tree), and a set of words or word strings (hereinafter referred to as “words”) arranged in the parsing tree. Including. These sets of words have two variables. A variable represents a place that can be replaced with another word. In the present embodiment, the attribute (word class) is specified as a class restriction for each variable. The template expansion rule may include word replacement, sentence structure correction, and the like. Various rules can be used for word replacement, such as replacement with another word in the same word class, replacement with a word of a lower concept of the word, replacement with a synonym of the word, and the like.

  When applying a template expansion rule to a seed template, out of the words included in the seed template, the word class of the word obtained from the question sentence and inserted in the portion corresponding to the variable in the template, and the template expansion rule Matching with the word class attached to the variable part is performed. If they match, the template expansion rule is applied to the seed template. The extension template obtained as a result has two variables, and the attribute (word class) is given as a class restriction to these variables.

  Referring to FIG. 3, as an example of the contents of the extended template set, by applying an extended rule to the seed template shown in FIG. 2, “X <Monster / Yokai>” under the constraint of variable Y = “Ultraman”. Y <hero> rival "," X <monster / youkai> is Y <hero> nemesis "," Y <hero> enemy was X <monster / youkai> ", etc. . X and Y represent variables. The extension template also has two variables, each with a word class. Here, it is assumed that a word class <hero> is obtained from a word “Ultraman” included in an input question sentence by using a thesaurus.

  Referring again to FIG. 1, the question answering apparatus 30 further includes a word class obtained from the thesaurus 58 for matching each of the extended templates stored in the extended template set storage unit 56 with the sentence included in the Web corpus 32. And a matching unit 60 that extracts a word string (referred to as a “match word string”) having a portion that matches any of the extended templates from the Web corpus 32, and is extracted from the Web corpus 32 by the matching unit 60. In order to collect a match word string set storage device 62 that stores a match word string set made up of matched match word strings and a match word string stored in the match word string set storage device 62, a portion that is likely to be an answer to a question For each of the answer candidate collection unit 66 and the match word string output by the answer candidate collection unit 66 Template applied by the rate estimation unit 48, template expansion rules and word classes used for template expansion in the template expansion processing unit 52, and the co-occurrence frequency of the word corresponding to the answer and the words included in the surrounding area Based on the above, a scoring / selecting unit 68 is provided for assigning a score to each match word string, and outputting the match sentence string in descending order of the score.

  In the matching in the matching unit 60, a parsing tree included in the extended template is used to match each subtree in the parsing tree with a sentence included in the Web corpus 32. For example, if a part in which a node in a sentence parse tree of a sentence from the Web corpus 32 is omitted matches the parse tree of the extension template, other conditions (matching word class in variable part, other words As long as (matching of sequence) is satisfied, the matching portion is extracted as a match word sequence. Therefore, a word string that cannot be obtained only by looking at the one-dimensional word string array (a word string that is not actually included in the Web corpus 32) can be extracted as a match word string.

  Referring to FIG. 4, a program for realizing the template extension processing unit 52 of FIG. 1 by a computer has the following control structure. This program assigns the word class of the word of the part supplemented by the input question sentence among the variable parts of all the seed templates with reference to the thesaurus 58, and the template extension stored in the template extension rule storage unit 54. Step 100 for reading all the rules into the main memory of the computer, Step 102 for adding the extension template to the set of extension templates by executing Step 104 described below for each rule read in Step 100, Step After the process of 102 is completed, the process includes a step 106 of outputting the extended template set obtained by the process of step 102 and ending the process.

  In step 104, for all seed templates stored in the seed template set storage unit 50, among the word classes of variables, those supplemented by the input question sentence match with any word class of variables of the extended rule. The step 110 of determining whether there exists is included. If the determination in step 110 is negative, the process for this extended rule is terminated and the process proceeds to the next extended rule.

  Step 104 further includes a step 112 of performing the following step 114 on the seed template that satisfies the condition when the determination of step 110 is affirmative.

  In step 112, the extended rule that is the current processing target is applied to the target seed template to generate a new template (extended template), and the seed that is based on the newly created extended template The product of the template weight and the weight of the applied extension rule is calculated and attached as a weight step 122. The extension template created in step 122 is merged into the extension template set together with the weight calculated in step 122. Step 124. If the extension template created in step 120 has already been merged into the extension template set, the template is not added to the extension template set.

  Referring to FIG. 5, the program for realizing matching unit 60 shown in FIG. 1 includes step 130 for executing the following step 132 for each sentence stored in Web corpus 32.

  Step 130 receives step 140 for performing morphological analysis on the sentence to be processed, and a word string (morpheme string) to which a tag indicating a word class, a utilization form, etc. is attached by the morphological analysis process, Step 141 for parsing the sequence and outputting a word string pattern composed of a parse tree and step 142 for executing the following step 144 for each template stored in the extended template set storage unit 56 are included. Since the target language is Japanese here, step 140 performs morphological analysis. If the target language is a language that separates words with spaces such as English, analysis processing such as part-of-speech analysis may be executed here instead of morphological analysis. For the morphological analysis, an existing morphological analysis program may be used. As a morphological analysis program, for example, JUMAN (URL = http: //nlp.kuee.Kyoto-u.ac.jp/nl-resource/juman.html) or ChaSen (URL = http: //chasen-legacy.sourceforge .jp /) can be used.

  In step 144, step 150 for determining whether or not there is a matching portion including the tree structure in the template to be processed in the word string pattern to be processed, and the determination in step 150 is affirmative. , Step 152 for executing the processing of the following step 154 for each of the matching locations.

  Step 154 includes a step 160 of assigning the weight of the template to be processed to the variable portion (variable X or the like) of the processing target among the matched parts, a word appearing in the obtained template, etc. And a step 162 of further multiplying the template weight by the weight based on the co-occurrence frequency and attaching it to the template for output, and moving the process to the next match location. If the determination in step 150 is negative, nothing is done and the process moves to the next template.

  As described above, in the determination in step 150, the entire sub-tree of the target sentence parse tree is compared with the parse tree of the template, and if they match, they are processed in step 152 and subsequent steps. And For example, if the parse tree of the input sentence with some nodes omitted matches the parse tree of the extension template, and if other conditions are met, the matched part Is extracted as a match. Therefore, a word string that does not actually exist in the Web corpus 32 can be extracted as a match character string.

[Operation]
The question answering apparatus 30 shown in FIGS. 1 to 5 operates as follows. A large number of sentences are collected in advance from the Web and stored in the Web corpus 32 in advance. In the template set storage unit 46, a template for an answer sentence is prepared in advance by manual operation or automatic processing. A thesaurus 58 is also prepared in advance. In the template expansion rule storage unit 54, template expansion rules are prepared in advance by manual operation or automatic processing. These are all stored on a hard disk or the like in a machine-readable form.

  When the input question sentence 20 is given to the question answering device 30, the template estimation unit 48 extracts one or a plurality of the most appropriate answer sentences for the input question sentence 20 from the templates stored in the template set storage unit 46. The seed template is output to the seed template set storage unit 50 (see FIG. 2). For this extraction, a discriminator based on the machine learning result is used.

  When the seed template is stored in the seed template set storage unit 50, the template expansion processing unit 52 operates, and each of the seed templates stored in the seed template set storage unit 50 is stored in the template expansion rule storage unit 54. Extend the seed template by applying template expansion rules. A number of extension templates are generated by this extension and stored in the extension template set storage unit 56.

  The extension template stored in the extension template set storage unit 56 is read by the matching unit 60 and stored in a main storage unit (not shown). The matching unit 60 sequentially reads a large number of sentences stored in the web corpus 32 and performs morphological analysis and syntax analysis on each of them (steps 140 and 141 in FIG. 5). Furthermore, the matching unit 60 expands the parse tree to which the word sequence (morpheme sequence with tags such as word class and semantic class) obtained by morphological analysis and syntax analysis is stored in the main storage unit. It is determined whether or not there is a part that matches the template (step 150). If there is a part that matches any of the extended templates (Yes in step 150), the matching unit 60 uses the part corresponding to the answer (variable X) in the word string represented by the structure as the matched template. Along with the assigned weight, it is output to the match word string set storage device (step 152). The match word string set storage device 62 stores these word strings together with tags and weights attached to the words. The matching unit 60 repeats this for all sentences stored in the web corpus 32.

  For each of the match word strings stored in the match word string set storage device 62, the answer candidate collection unit 66 includes information such as a template, an extended rule, and a word class used for obtaining the answer, the word, Based on the co-occurrence frequency with surrounding words, a score is assigned to each answer candidate and given to the scoring / selection unit 68.

  The scoring / selection unit 68 holds the answer candidates from the answer candidate collection unit 66 in the ascending order of the scores. When all the output of the answer candidates from the answer candidate collection unit 66 is completed, the scoring / selection unit 68 outputs a predetermined number as the answer sentences 34 in order from the highest score.

  In this way, the answer sentence 34 is created. The answer sentence 34 is selected from a large number of answer candidates generated from sentences that match the seed template selected from the template prepared first and the extended template expanded from the seed template. The seed template has the highest possibility as an answer sentence to the input question sentence 20. The extension template includes replacement of a word included in the seed template with a synonym, paraphrase of expression of the seed template, and the like. Accordingly, the answer sentence 34 is selected from a large number of answer candidates for the input question sentence 20. Moreover, since the seed template is expanded by the template expansion rule, a very large number of templates are stored in the expanded template set storage unit 56. Moreover, since a regular expression is used for this template, a very large number of answer candidates are extracted from the web corpus 32 by matching the template and the sentence included in the web corpus 32. Further, in this matching, a matching portion is searched including a subtree of the parse tree. Therefore, although the word “extraction” is used here, expressions that are not included in the Web corpus 32 are also extracted (generated) by the processing of the matching unit 60.

  The Web corpus 32 is considered to contain the largest number of expressions as an available corpus. However, the expressions included in the Web corpus 32 are created by humans, and therefore the number is inevitably limited. On the other hand, in this embodiment, by expanding the template and performing matching with the Web corpus 32 using various extended templates, a much broader expression than that created manually is stored as the match word string. It will be. Accordingly, the answer sentence 34 selected by the scoring / selection unit 68 from these expressions has a high probability of being suitable as an answer to the input question sentence 20. As a result, an answer sentence can be automatically generated and output with high accuracy for various input question sentences 20. However, as described above, it goes without saying that another corpus may be used instead of the Web corpus 32.

  In the above-described embodiment, the variables of the seed template are not attached with those indicating word attributes such as word tags. However, the present invention is not limited to such an embodiment, and a variable with a word tag attached may be used as a seed template. Further, although the number of variables in the seed template is two, the seed template may include three or more variables.

  In the above-described embodiment, the template expansion by the template expansion processing unit 52 is only for the seed template. However, the present invention is not limited to such. You may make it increase the number of templates further by applying a template expansion rule to the expansion template obtained by applying a template expansion rule with respect to a seed template. In this case, the template expansion process may be performed a predetermined number of times, or the template expansion process may be recursively repeated until no new expansion template appears.

  In the above embodiment, various weights are given to the seed template in advance. However, the present invention is not limited to such an embodiment. The weight assigned to the seed template may be constant, and the template weight may be determined only by which template expansion rule is used. Alternatively, when there are a plurality of applicable templates for the word string included in the web corpus 32, a weight that increases according to the number of templates may be given. When a template is created by applying the template expansion rule not only to the seed template but also to the expansion template, it is desirable that the weight of the template is reduced each time the template expansion rule is applied.

  In the above embodiment, each rule is described using regular expressions. However, the present invention is not limited to such an embodiment. As long as the rules can be accurately described according to the purpose, any description method may be used.

  In the above embodiment, sentences collected from the Web are used as the Web corpus 32. At present, since there are a large number of sentences on the Web, it is desirable to use the Web itself as the Web corpus 32. However, of course, a corpus other than the web can be used as the web corpus 32.

[Operation example]
A specific operation example according to this embodiment will be described. The question answering apparatus 30 is an apparatus on the server side of a question answering system that can answer various questions by WWW. In this example, the similar relationship pattern extraction method described in Non-Patent Document 1 is applied to hundreds of millions of WWW pages, and a large number of relationship patterns are extracted in advance. Similarly, data for extracting a seed template in the template estimation unit 48 is also prepared in advance.

  As a specific example, it is assumed that the input question sentence 20 is “who is the enemy of Ultraman”. In response to this input, it is assumed that the template estimation unit 48 estimates the templates “X is an enemy of Y” and “X is an enemy of Y” (Y = Ultraman) from the template set storage unit 46. The template estimation unit 48 stores these two templates as seed templates in the seed template set storage unit 50.

  The template expansion processing unit 52 expands these two seed templates stored in the seed template set storage unit 50. As a result, templates such as “X is Y's rival”, “X is Y's nemesis”, and “Y's enemy was X” are obtained. Furthermore, based on the constraint (Y = Ultraman), a word class <hero> is assigned to Y, and an extension template having a variable that matches this word class is applied to create an extension template. The extension template obtained in this way is stored in the extension template set storage unit 56.

  The matching unit 60 matches between a large number of sentences in the Web corpus 32 and the extended template stored in the extended template set storage unit 56 in consideration of the subtree of the structural analysis tree, and corresponds to the variable X. Ask for expression. These are answer candidates. That is, a set of answer candidates = “Pigmon, Jamira, Baltan, Zetton ...”. From this set of answer candidates, the original template, extended template, relationship between the original word and the replaced word, co-occurrence relationship between the answer candidate word and a word other than that word in the template, etc. The scoring / selection unit 68 gives a score to each candidate, and outputs a predetermined number of high scores, for example, “Pigmon” and “Bartan Alien” as the answer sentence 34.

  The answer sentence 34 may be output as speech by a speech synthesizer (not shown).

[Realization by computer]
The question answering apparatus 30 according to this embodiment can be realized by computer hardware, a program executed by the computer hardware, and data stored in the computer hardware.

  Referring to FIG. 6, the question answering system including question answering apparatus 30 has a computer system 330 functioning as question answering apparatus 30 described above and voices a question sentence (input question sentence 20 shown in FIG. 1) to computer system 330. The mobile phone 300 includes a mobile phone 300 that transmits as a signal and receives an answer sentence (the answer sentence 34 shown in FIG. 1) from the computer system 330 by an audio signal. In the present embodiment, of the communication between the mobile phone 300 and the computer system 330, the question sentence and the answer sentence are transmitted by voice. Therefore, the mobile phone 300 only needs to be capable of voice communication with a normal carrier. On the other hand, the computer system 330 has a function of receiving a voice signal from a telephone, a function of converting the voice signal into a text string by voice recognition, and giving the input question sentence 20 to the template estimation unit 48, and a scoring / selection unit. It is necessary to have a function of converting the answer sentence 34 output by 68 into a voice signal by voice synthesis and returning it to the mobile phone 300 by telephone. Since these are all existing functions, their details are not described here. Of course, if the mobile phone 300 has a data communication function, the response text from the computer system 330 is sent to the mobile phone 300 in text format, or the HTML (combined from the answer and the URL related to the answer) (HyperText Markup Language) document may be returned to the mobile phone 300 and a browser may be activated on the mobile phone 300.

  Referring to FIG. 6, the computer system 330 includes a computer 340 having an FD (flexible disk) drive 352 and a CD-ROM (compact disk read only memory) drive 350, a keyboard 346, a mouse 348, and a monitor 342. including.

  Referring to FIG. 7, in addition to FD drive 352 and CD-ROM drive 350, computer 340 includes CPU (central processing unit) 356 and bus 366 connected to CPU 356, FD drive 352 and CD-ROM drive 350. And a read only memory (ROM) 358 for storing a boot-up program and the like, and a random access memory (RAM) 360 connected to the bus 366 for storing a program command, a system program, work data, and the like. The computer system 330 further includes a network interface (I / F) 344 that provides a connection to the Internet. Although not shown, the computer 340 is connected to the mobile phone network via the network I / F 344 and can perform data communication with the mobile phone 300.

  A computer program for causing the computer system 330 to operate as the question answering apparatus 30 is stored in the CD-ROM 362 or FD 364 inserted in the CD-ROM drive 350 or FD drive 352 and further transferred to the hard disk 354. Alternatively, the program may be transmitted to the computer 340 through a network (not shown) and stored in the hard disk 354. The program is loaded into the RAM 360 when executed. The program may be loaded directly into the RAM 360 from the CD-ROM 362, from the FD 364, or via a network.

  This program includes a plurality of instructions for causing the computer 340 to operate as the question answering apparatus 30 of this embodiment. Some of the basic functions required to perform this operation are provided by operating system (OS) or third party programs running on the computer 340 or various toolkit modules installed on the computer 340. Therefore, this program does not necessarily include all functions necessary for realizing the system and method of this embodiment. This program executes an operation as the above-mentioned question answering apparatus 30 by calling an appropriate function or “tool” (program library) in a controlled manner so as to obtain a desired result. Only need to be included.

  The web corpus 32, template set storage unit 46, seed template set storage unit 50, template extended rule storage unit 54, extended template set storage unit 56, match word string set storage device 62, etc. shown in FIG. 7 is realized by the hard disk 354 or the RAM 360 shown in FIG. In particular, for example, the Web corpus 32 and the template expansion rule storage unit 54 are normally stored in the hard disk 354, and are loaded into the RAM 360 as necessary when the program is executed. The seed template set storage unit 50, the extended template set storage unit 56, the match word string set storage device 62, and the like have work file characteristics. Therefore, these are generated in the RAM 360 at the time of generation, and are stored in the hard disk 354 if necessary to be stored.

  The operation of computer system 330 is well known and will not be repeated here.

  The embodiment disclosed herein is merely an example, and the present invention is not limited to the above-described embodiment. The scope of the present invention is indicated by each claim of the claims after taking into account the description of the detailed description of the invention, and all modifications within the meaning and scope equivalent to the wording described therein are included. Including.

DESCRIPTION OF SYMBOLS 20 Input question sentence 30 Question answering device 32 Web corpus 34 Answer sentence 46 Template set storage part 48 Template estimation part 50 Seed template set storage part 52 Template expansion process part 54 Template extended rule storage part 56 Extended template set storage part 60 Matching part 62 Match word string set storage device 66 Answer candidate collection unit 68 Scoring / selection unit

Claims (6)

A template storage means for storing a template of an answer sentence for a question sentence, the template having first and second variable parts;
In response to receiving an input of a question sentence, one or a plurality of templates as a prototype of an answer sentence for the question sentence is estimated from the templates stored in the template storage unit, and each template in the question sentence is estimated. Template estimation means for outputting together with a constraint indicating a word corresponding to the first variable part of
By applying a template expansion rule prepared in advance to a template selected based on the word class of the words that constitute the constraint condition of the template among the templates output by the template estimation means, a word or sentence Template extension means for generating one or a plurality of extension templates whose structures are modified,
Each of the extension templates includes first and second variables indicating positions of answer candidates for the question sentence, and the two variables are the first and second variables of the template stored in the template storage unit. 2 is associated with each of the two variable parts, and each of them specifies an attribute to be satisfied by the variable,
The template extension means selects an extension template in which the attribute of the word constituting the constraint condition has a predetermined relationship with the attribute of the first variable,
For outputting one or more candidates that can be exchanged with the second variable of the extension template by matching each of the extension templates output by the template extension means and a plurality of sentences prepared in advance. Matching means;
As an answer to the question sentence, based on the process until the candidate is obtained or the co-occurrence frequency of the word included in the extended template that matches the candidate and the candidate for the candidate output by the matching means And a scoring and selection means for calculating a score indicating eligibility of the score and outputting the candidates in order of the highest score. The question answering device according to claim 1, wherein the attribute is a word class of a word assigned to a corresponding variable. The question according to claim 1 or 2, further comprising: a question sentence receiving means for receiving the question sentence as a voice signal by communication, converting it into a text string by voice recognition and inputting the text string to the template estimating means. Answer device. The question answering apparatus according to any one of claims 1 to 3, further comprising speech synthesis means for converting a candidate output by the scoring and selection means into speech by speech synthesis. further,
A question sentence receiving means for receiving the question sentence as a voice signal by communication, converting it into a text string by voice recognition and inputting it to the template estimation means;
The speech synthesis means for converting the candidate output by the scoring and selection means into a speech signal by speech synthesis and sending it back to the terminal that has transmitted the question sentence. The question answering device described. Computer
A template storage means for storing a template of an answer sentence for a question sentence, the template having first and second variable parts;
In response to receiving an input of a question sentence, one or a plurality of templates as a prototype of an answer sentence for the question sentence is estimated from the templates stored in the template storage unit, and each template in the question sentence is estimated. Template estimation means for outputting together with a constraint indicating a word corresponding to the first variable part of
The structure of the word or sentence is corrected by applying the word class of the word constituting the constraint condition of the template and the template expansion rule prepared in advance to each of the templates output by the template estimation means. Template extension means for generating one or more extension templates;
Each of the extension templates includes first and second variables indicating positions of answer candidates for the question sentence, and the two variables are the first and second variables of the template stored in the template storage unit. 2 is associated with each of the two variable parts, and each of them specifies an attribute to be satisfied by the variable,
One or a plurality of candidates that can be exchanged with the second variable of the extension template by matching each of the extension templates output by the template extension means with a number of sentences prepared to be available in advance. Matching means for outputting the number of pieces,
Answers to the question sentence based on the process until the candidate is obtained or the co-occurrence frequency of the word included in the extended template that matches the candidate and the candidate for the candidate output by the matching means A computer program that calculates a score indicating eligibility as a score and functions as scoring and selection means for outputting the candidates in order of the highest score.

Images