JP5882241B2 - Method and apparatus for generating search keyword for question answering, and program - Google Patents

Description

  The present invention relates to a query response search keyword generation method, apparatus, and program, and more particularly to a question response search keyword generation method, apparatus, and program for generating a search keyword for searching for an answer to a question sentence.

  In a question answering system using a computer, a user's question is input in a natural language. For example, a factoid-type question that asks about simple facts and events is entered in a form such as "Who did you tell me about glasses?" On the other hand, it is desirable that the question answering system returns an answer “Francisco Xavier”.

  In addition, non-factoid type questions that ask reasons, causes, opinions, methods, etc. are entered in a form such as “I was bitten by a dog. What should I do?” Or cut and display the method part on the homepage. The following is an example in which the question answering system returns “Francisco Xavier” to the user question “Who brought the glasses to Japan?” User question “The dog has been bitten. An example will be described in which the question answering system returns “When a dog bites you, wash your wounds well and go to the hospital and receive treatment immediately”.

  FIG. 8 shows a configuration example of the question answering system. This question answering system operates as follows in response to an input user question. First, in the question analysis, a user question is analyzed, an answer type is identified, and converted into a search keyword for information search. Next, by performing an information search, an Internet search is performed based on the search keyword passed from the question analysis, and the top n snippets containing a large number of search keywords are obtained from the document including the search keyword or its summary sentence (snippet). To do.

  In the answer candidate extraction, answer candidates are created based on the n snippets obtained by the information search and the answer type obtained by the question analysis. If the answer type is a factoid question type, a specific expression is extracted from the snippet obtained by the information search. Nonpatent literature 1 etc. exist in the answer candidate extraction method in a non fact fact type question. Then, the answer candidate evaluation selects and outputs the most suitable answer among the plurality of answer candidates obtained by answer candidate extraction.

  In addition, as a method for concisely rephrasing the input sentence while maintaining the factuality of the event, there is known a method for converting the predicate of the sentence into the simplest form without changing its meaning as much as possible (for example, , See Patent Document 1).

JP 2011-145844 A

Murata M, Tsukawaki S, Kanamaru T, Ma Q, and Isahara H, `` A System for Answering Non-Factoid Japanese Questions by Using Passage Retrieval Weighted Based on Type of Answer. '', In Proceedings of NTCIR-6 Workshop Meeting, Tokyo, Japan, 2007, pp.477-482

  One of the important elements in the above series of processes is that it is necessary to set an appropriate search keyword and obtain a snippet (document) containing many true answers by information search. If an inappropriate search keyword is set, a large number of documents that do not include a true answer are searched, and a correct answer cannot be output.

  For example, for the user question of Example 1 above, when the query analysis generates a search keyword with only the content word “Japanese glasses tell”, snippets such as document 1-1 and document 1-2 shown in FIG. (When the number of snippets obtained by information retrieval is n = 2). However, since neither the document 1-1 nor the document 1-2 obtained by the search keyword “Japanese glasses tell” contains “francis Xavier” which should be the correct answer, a correct answer cannot be output.

  Further, in the case of the user question of Example 2 above, when the query analysis generates a search keyword having only the content word “Dog Kam”, snippets such as Document 1-1 and Document 1-2 shown in FIG. (In Internet search, it matches the search keyword after performing morphological analysis of the document. For the search keyword “Kam”, the appearance form of the document “Kama (Kam's blank form)” “Kami ( The document 1-1 is a document related to dog discipline, and the document 1-2 is an advertisement for dog gum, so output a countermeasure when a dog is bitten. I can't.

  As described above, when only the content word of the user question is used as a search keyword, a document that is different from the intention of the user question cannot be obtained because there is no negative “no”, “being (being)” indicating the state, tense, and modality information. Many searches.

  As another conventional method, a method in which a user question is decomposed into phrases, and for phrases including verbs and adjectives, not only content words but also function words can be used as search keywords. In this method, from the user question of Example 1 above, “Japanese glasses transmitted” is obtained as a search keyword, and snippets shown in document 2-1 and document 2-2 in FIG. 9 are obtained. Since the snippet includes “Francisco Xavier” which is the correct answer of Example 1, the question answering system can also output a correct answer.

  However, with respect to Example 2 above, the search keyword “dog was bitten” was obtained, and documents matching “Chaita” had priority, such as Document 3-1 and Document 3-2 in FIG. It is not possible to obtain a countermeasure when a dog is bitten by a search.

  In Japanese, don't say "when you got bitten by a dog". When you use the combination modification, the polite expression "mas" is missing or "chau" and "ta" mean completion. It is difficult to appear at the same time. Therefore, if an unnecessary functional expression is included in the search keyword, there is a problem that many documents obtained by information search do not include the answer to the user question.

  The present invention has been made to solve the above problem, and provides a query response search keyword generation method, apparatus, and program capable of obtaining a search keyword for accurately searching for an answer to a question sentence. With the goal.

In order to achieve the above object, a query keyword search keyword generation method according to the present invention includes a predicate function expression normalization unit, a content word extraction unit, and a query in a query keyword search keyword generation device including a search keyword generation unit. A method for generating a search keyword for response, which includes a content word included in the question sentence and the content word following the content word based on the morphological analysis result of the question sentence input by the pre-description unit function expression normalization means Predetermined function words that represent combinations of function expressions that are character strings that include at least one function word, function words that represent negation, function words that represent states, and functions that represent tense word, and leave the function words representing the modality, function words representing the negative, function words representing the state, function words representing the tense, and other than function words representing the modality of the function words, to remove Normalizing the previous description part, and extracting the content word from the part other than the predicate part of the question sentence based on the morphological analysis result of the question sentence by the content word extracting means, A search for an answer to the question sentence by a combination of the predicate normalized by the pre-description part functional expression normalizing means and the content word extracted by the content word extracting means by the search keyword generating means Generating as a search keyword.

The search keyword generation device for question response according to the present invention is a content word included in the question sentence based on a morphological analysis result of the inputted question sentence, and a character string following the content word, and at least 1 For predicates composed of combinations of function expressions that are character strings including two function words, the function words representing negation, function words representing state, function words representing tense, and function representing modalities are predetermined. Predicate function expression that normalizes the predescription part by deleting the function words other than the function word that leaves the word, the function word that represents negation, the function word that represents the state, the function word that represents the tense, and the modality Normalization means, content word extraction means for extracting content words from a part other than the predicate of the question sentence based on the morphological analysis result of the question sentence, and normalization by the pre-description part functional expression normalization means Predicates and The combination of content words extracted by the serial contents word extraction means is configured to include a search keyword generating means for generating a search keyword for searching an answer to the question message.

  According to the search keyword generation method for question response and the search keyword generation device for question response according to the present invention, the predicate function expression normalization means includes the question sentence based on the morphological analysis result of the inputted question sentence. In the predicate consisting of a combination of a content word and a functional expression that is a character string that follows the content word and includes at least one functional word so as not to affect the meaning of the predicate Normalize predicates by converting functional expressions into simple forms.

  Then, the content word extraction unit extracts the content word from the portion other than the predicate of the question sentence based on the morphological analysis result of the question sentence.

  Then, the search keyword for searching for the answer to the question sentence by the combination of the predicate normalized by the predicate function expression normalizing means and the content word extracted by the content word extracting means by the search keyword generating means Generate as

  In this way, in order not to affect the meaning of the predicate of the question sentence, the function expression is converted into a simple form to normalize the predicate, and the normalized predicate and the content word included in the question sentence Is generated as a search keyword for searching for an answer to a question sentence, a search keyword for accurately searching for an answer to the question sentence can be obtained.

  In addition, the pre-description part functional expression normalization means is based on the morphological analysis result of the input question sentence, and the predicate composed of a combination of the content word and the functional expression included in the question sentence The previous description part can be normalized by deleting function words that do not affect the meaning of the previous description part and redundant function words and converting the function expression into a simple form.

  Further, the pre-description part function expression normalizing means is configured to predetermine a function word other than a function word representing negation, a function word representing a state, a function word representing tense, and a function word representing a modality. It can be deleted as a function word that does not affect the meaning of.

  A program according to the present invention is a program for causing a computer to execute each step of the above-described question answer search keyword generation method.

  As described above, according to the query keyword search keyword generation method, apparatus, and program of the present invention, the function expression is converted into a simple form so as not to affect the meaning of the predicate of the question sentence. By normalizing predicates and generating a combination of the normalized predicates and content words contained in the question text as search keywords for searching for answers to the question text, the answers to the question text can be accurately obtained. The effect that the search keyword for searching can be obtained is acquired.

It is a block diagram which shows the example of 1 structure of the question answering apparatus which concerns on embodiment of this invention. It is a figure which shows the example of the morphological analysis result about a question sentence "The personal computer has broken." It is a figure which shows the example of the predicate extraction result about a question sentence "a personal computer has broken." It is a figure which shows the example of the meaning label provision result about a question sentence "a personal computer has broken." It is a figure which shows the example of the NULL deletion result about a question sentence "a personal computer has broken." It is a figure which shows the example of the redundant label deletion result about a question sentence "a personal computer has broken." It is a flowchart which shows the content of the question answer process routine in the question answering apparatus which concerns on embodiment of this invention. It is explanatory drawing for demonstrating a prior art. It is a figure which shows the example of the document obtained by the search keyword with respect to the question sentence "Who did you tell glasses to Japan?", And an information search. It is a figure which shows the example of the document obtained by the search keyword and information search with respect to the question sentence "The dog has bitten. What should I do?"

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

  In this embodiment, a query including a correct answer to be answered by the system in information retrieval is generated by generating a search keyword after briefly rephrasing a question sentence created by a user while maintaining the factuality of an event. Alternatively, an object is to obtain many summary sentences (hereinafter referred to as snippets) and to extract answers to questions from the obtained snippets.

  As a method for succinctly rephrasing the question sentence while maintaining the fact of the event, a method of converting the predicate of the question sentence into the simplest form without changing its meaning as much as possible is used (see Patent Document 1 above). ). For predicates, the entire converted predicate is used as a search keyword, and for other parts, the content word is used as a search keyword.

  In the following example, the question answering device responds to the question “Who brought the glasses to Japan?” With the question answering device “Francisco Xavier”. An example will be described as an example 2 in which the question answering device responds “When a dog bites you, wash your wounds well and go to the hospital immediately for treatment”.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

<System configuration of question answering device>
FIG. 1 is a block diagram showing a question answering apparatus 100 according to an embodiment of the present invention. This question answering apparatus 100 is composed of a computer having a CPU, a RAM, and a ROM storing a program for executing a question answering processing routine described later, and is functionally constructed as follows. Yes.

  As shown in FIG. 1, the question answering apparatus 100 according to the present embodiment includes an input unit 1, a calculation unit 2, and an output unit 3.

  The input unit 1 receives a question sentence created by the user. The question sentence is a sentence created in a natural language.

  The calculation unit 2 includes a question analysis unit 20, an information search unit 21, an answer candidate extraction unit 22, and an answer candidate evaluation unit 23.

  The question analysis unit 20 analyzes the question sentence, identifies an answer type for the question sentence, and generates a search keyword for information search.

  The question analysis unit 20 includes a morpheme analysis unit 201, a predicate function expression normalization unit 202, a content word extraction unit 203, a search keyword generation unit 204, and an answer type determination unit 205.

  The morpheme analysis unit 201 performs morphological analysis on the question sentence input by the input unit 1 to decompose it into words, and gives parts of speech to each word.

  The predicate function expression normalization unit 202 simplifies the function expression of the predicate of the question sentence based on the morpheme analysis result analyzed by the morpheme analysis unit 201 so as not to affect the meaning of the predicate. Convert. Here, the predicate is composed of a combination of “content word” and “function expression” which is a character string subsequent to “content word” and including at least one function word. “Content word” refers to a word having a general meaning such as a verb, noun, adjective, adjective verb, or adverb, and “function word” refers to a word having a grammatical role such as a particle or an auxiliary verb.

  Specifically, the predicate function expression normalization unit 202 extracts the predicate of the question sentence, and then performs normalization that paraphrases the function expression of the predicate into a simple expression. The normalization method of the predicate function expression is as disclosed in Patent Document 1 described above, and does not affect the meaning of the predicate for a predicate composed of a combination of a content word and a function expression. Remove function words and redundant function words to convert the function expression into a simple form.

  As a method of deleting a function word that does not affect the meaning of the predicate, it does not correspond to a predetermined function word representing negation, a function word representing a state, a function word representing tense, and a function word representing a modality. The function word is deleted as a function word that does not affect the meaning of the predicate. In other words, function words that represent negation, state, tense, and modality that affect the meaning of the event remain, and other function words are deleted, so that the functional expression of the predicate can be expressed simply without changing the meaning of the event. In other words.

  Here, the operation of the predicate function expression normalization unit 202 will be described in detail by taking as an example a case where the question sentence “PC has been broken” is input to the question answering apparatus 100.

  In order to normalize the predicate function expression, a function expression meaning label dictionary is used. Here, the functional expression semantic label dictionary is a semantic label that represents the meaning of the functional expression that affects the meaning of the predicate, a list of character strings in the standard form of the functional expression corresponding to each semantic label, and a single description. When the function expression corresponding to the part includes a plurality of function words having the same semantic label, the dictionary is formed by combining redundant rules for each semantic label indicating how to retain the function words. The function expression meaning label dictionary stores the following meaning labels and character lists (1) to (4).

(1) A meaning label of “Negation” is stored as a meaning label for distinguishing negation or affirmation of the event represented by the predicate, and “None” as a character list corresponding to the meaning label of “Negation”, “Hey” is stored. Note that the function word included in the character list corresponding to the meaning label “deny” is an example of a function word representing negation.
(2) Semantic labels for “passive” and “usage” are stored as semantic labels for distinguishing predicate states. The following character lists are stored as character lists corresponding to the respective semantic labels.
"Passive": "Be""Be"
“Utility”: “Make” “To be”
Note that the function words included in the character list corresponding to the meaning labels “passive” and “usage” are examples of function words representing the state.
(3) A semantic label of “complete” is stored as a semantic label characterizing the tense of the event represented by the predicate, and “ta”, “chau”, “ “Chimau”, “Te / Eku”, and “Te / Oku” are stored. Note that the function word included in the character list corresponding to the meaning label “complete” is an example of a function word representing tense.
(4) “Question”, “Invitement / will”, “Aspiration” as a semantic label for distinguishing whether or not the event represented by the predicate includes the subjectivity (ie, modality expression) of the speaker (user) "," Request "," Recommendation "," Necessary "," Allow "," Inference ", and" Possible "are stored. The following character lists are stored as character lists corresponding to the respective semantic labels.
"Question": "ka"
“Invite / Will”: “U”
"Wish": "I want""Gana""I want / I want"
"Request": "Te / Me""Te / I want"
“Recommended”: “Ta / One / Bad / Good” “To / Good”
"Necessary": "Should""None / Te / Has / Can / No"
"Allow": "Te / mo / good"
“Inference”: “Maybe / Knowable / None” “Like” “Yo” “So” “Well”
"Possible": "Let""Koto / can / can"
Note that the function words included in the character list corresponding to these semantic labels are examples of function words representing modalities.

  When the morphological analysis is performed by the morphological analysis unit 201 for the question sentence “PC is broken”, and the morphological analysis result as shown in FIG. 2 is input to the predicate function expression normalization unit 202, first, “Cha” and “Ta” that match the list (entry) of character strings in the standard form of the functional expression in the functional expression meaning label dictionary (not shown) are recognized as function words that affect the meaning of the predicate. The

  Next, “yo” belonging to the part of speech “particle” is recognized as a function word. Finally, “broken” belonging to the part of speech “verb—independence” is recognized as a content word. As a result, “It's broken” is extracted as a predicate. FIG. 3 shows an example of the predicate extraction result.

  Next, the extracted predicate “has been broken” is first analyzed by “longest matching method from the back”. However, since there is no “yo” entry in the functional expression semantic label dictionary, an empty semantic label “NULL” is given. Next, “ta” and “chau” are analyzed, and a “complete” semantic label is assigned to each. This semantic label assignment operation is terminated immediately before the last morpheme when the morpheme constituting the predicate is processed by the longest match method from the back, that is, immediately before the content word “break”. FIG. 4 shows an example of the meaning label assignment result. Note that the method of assigning a semantic label is not limited to the longest matching method from the back, and other methods may be used.

  Next, with respect to the functional expression “Cha-Tyo” in the upper description part, “Yo” with the empty semantic label “NULL” is deleted, and the functional expression “Cha-ta” is simplified. FIG. 5 shows an example of a NULL deletion result.

  Next, using the redundancy rule described in the function expression meaning label dictionary, all entries (word information) of redundant function words are deleted. The redundancy rule is based on the knowledge that “when there are a plurality of function words representing the same meaning in the same predicate, the meaning represented by the predicate can be retained by leaving only one of them”. If the predicate after semantic label assignment and NULL deletion is “broken”, it is the same in the word information for each morpheme corresponding to the predicate after provision of the semantic label and after NULL deletion. It is checked whether or not there are two or more semantic labels. If there are two or more semantic labels, deletion is performed according to the redundancy rule. Since the word information for each morpheme corresponding to the upper description part “has been broken” has two semantic labels “completed”, it is subject to a redundancy rule. The redundancy rule “completed” is “if the surface shape is the same, leave the first (First), otherwise leave the last (Last)”. Accordingly, the function word “Cha” is deleted, and the function word “ta” of the meaning label “complete” is left as a normalized function expression. FIG. 6 shows an example of the redundant label deletion result.

  Next, the content word “broken” is connected to the normalized function expression “ta”. For the morpheme connection process including word utilization, a utilization generator based on a language model can be used. A language model utilization generator learns in advance from the correct answer data “what connection is likely” using the surface form, part of speech, and utilization type of the front word as well as the surface form and part of speech of the back word. It is a generator using a model. Therefore, in the utilization generator based on the language model, “notation; part of speech; utilization type” of the content word “break” (ru); “break; verb—independence; one step” and “ta; auxiliary verb of the function expression“ ta ” By inputting “special / ta” and “.; Symbol-punctuation” of the morpheme “.” Indicating the end of the sentence, it is possible to generate “broken” that is a correctly connected predicate. . In addition, even when a predicate that does not include a functional expression (a predicate that does not require normalization) is included in the question sentence, the predicate can be included in the search keyword.

  When the predicate function expression normalization unit 202 performs normalization of the predicate function expression with respect to the example 1, the question sentence of the above example 1 is the predicate, and the normalized predicate is also transmitted. " This is a form in which the function word “ta” with the meaning of completion is given to the predicate “Tell”, but “completion” belongs to the tense, and deleting this changes the event to the present tense. , Normalized by adding “ta” to “tell”. Further, in the case of the above example 2, “it has been bitten” is determined as a predicate in the predicate function expression normalization, and normalization is performed. In this predicate, "Kamu" is a predicate, "Ru" is a function word with a semantic label "Passive", "Cha" is a function word with a semantic label "Complete", "Mas" is a function word that represents politeness, "Ta" Is a function word of the semantic label “complete”. Of these, “mas” does not affect the meaning of the event, so it is deleted, and “chau” and “ta” appear redundantly as function words of the meaning label of “complete”, so “chau” is judged unnecessary And deleted. As a result, “bite”, “re” and “ta” remain as normalization results, and “bite” is output as a normalized predicate.

  The content word extraction unit 203 extracts a content word from a part other than the predicate of the question sentence based on the morphological analysis result by the morpheme analysis unit 201. For example, in Example 1 above, the two words “Japan” and “Glasses” are content words other than the predicate. In Example 2 above, “dog” is extracted as the content word in the portion other than the predicate.

  The search keyword generation unit 204 searches for a combination of the predicate normalized by the predicate function expression normalization unit 202 and the content word extracted by the content word extraction unit 203 to search for an answer to the question sentence. Generate as a keyword. In the case of the above example 1, as a result, “Japanese glasses transmitted” is generated as a search keyword. In the case of Example 2 above, “dog bitten” is generated as a search keyword.

  The answer type determination unit 205 determines and outputs an answer type corresponding to the question sentence based on the morphological analysis result by the morpheme analysis unit 201. Answer types are determined in advance, and in this embodiment, there are eight types of answer types of factoid type questions: person name, place name, organization name, artifact name, date, time, amount of money, ratio, non-factoid type There are two types of question answer types: reason and method. The answer type is determined by a clue word (word or expression) included in the question sentence or a classifier using machine learning. For example, in Example 1 above, since the clue word “who” is included, it is determined that the person's name is being asked. In the above example 2, since the clue word “how to” is included, it is determined that the method is answered.

  The information search unit 21 performs an Internet search based on the search keyword generated by the search keyword generation unit 204. From the obtained search results, the top n snippets including many search keywords among the snippets including the search keywords. To get. Here, n is usually several tens. For example, in the above example 1, the snippet of the document 2-1 and the document 2-2 shown in FIG. 9 is obtained with respect to the search keyword “Japanese glasses transmitted”. In the above example 2, snippets such as the document 2-1 and the document 2-2 shown in FIG. 10 are obtained for the search keyword “dog bite”.

The answer candidate extraction unit 22 extracts answer candidates based on the n snippets acquired by the information search unit 21 and the answer type obtained by the answer type determination unit 205.
If the answer type determined by the answer type determination unit 205 is the answer type of the factoid question, the specific expression is extracted from the snippet obtained by the information search unit 21. In the extraction of the specific expression, the morpheme sequence representing each of the person name, the place name, the organization name, the artifact name, the date, the time, the amount, and the ratio corresponding to the eight answer types of the factoid type question is used as the specific expression from the snippet. Extract. If the snippet was “Glasses was introduced to Japan by Francisco Xavier in 1549”, “1549” is the date, “Frances Xavier” is the name of the person, and “Japan” is the place name. The answer candidate extraction unit 22 performs the extraction process of the unique expression on all snippets obtained by the Internet search, and the unique expression that matches the determined answer type (person name in the above example 1) from the extracted unique expression. All expressions are extracted as answer candidates.
If the answer type determined by the answer type determination unit 205 is a non-factoid question type, the snippet (document) is decomposed into paragraphs and sentences, and then processing according to the answer type is performed. (See Non-Patent Document 1 above). When the answer type is a method, for example, paragraphs and sentences including the clue expressions “method”, “procedure”, “by”, “to”, “to tomo” are extracted as answer candidates (the above non-patent document) Reference 1).

  The answer candidate evaluation unit 23 selects and outputs the most suitable answer candidate from among the plurality of answer candidates obtained by the answer candidate extraction unit 22. If the answer type determined by the answer type determination unit 205 is a factoid question type, the answer candidate that appears in the most snippet is selected. For example, in the case of the above-described example 1, when the personal name “Francisco Xavier” appears in four snippets and “Yoshitaka Ouchi” appears in two snippets, “Francisco Xavier” is output as an answer. On the other hand, if the answer type determined by the answer type determining unit 205 is a non-factoid question type, a score is calculated for each of the answer candidates extracted by the answer candidate extracting unit 22, and the score and An answer is output based on a predetermined threshold (see Non-Patent Document 1 above).

  The output unit 3 outputs the answer output by the answer candidate evaluation unit 23 as a result. Since the document 2-1 and the document 2-2 shown in FIG. 9 include “francis Xavier” which is the correct answer, the question answering apparatus 100 outputs “francis xavier” as a correct answer. be able to. In addition, in the document 2-1 and the document 2-2 shown in FIG. 10 described above, the countermeasures when the dog is bitten are written, so the question answering apparatus 100 matches the countermeasure part with the question sentence. Can be output as an answer.

<Operation of question answering device>
Next, the operation of the question answering apparatus 100 according to the present embodiment will be described. First, when a question sentence created by the user is input to the question answering apparatus 100, the question answering process routine shown in FIG.

  First, in step S100, an input of a question sentence is accepted. Next, in step S102, a morphological analysis process is performed on the question sentence received in step S100.

  Next, in step S104, the predicate function expression normalization unit 202 identifies and extracts the predicate of the question sentence based on the morpheme analysis result analyzed in step S102.

  In step S106, the predicate function expression normalization unit 202 converts the predicate of the question sentence extracted in step S104 to the meaning of the predicate based on the morpheme analysis result analyzed by the morpheme analysis unit 201. Convert the functional representation into a simple form and normalize it so as not to affect it.

  In step S108, the content word extraction unit 203 extracts content words from portions other than the predicate of the question sentence based on the morphological analysis result in step S102.

  In step S110, the search keyword generation unit 204 generates a combination of the predicate normalized in step S106 and the content word extracted in step S108 as a search keyword for searching for an answer to the question sentence. To do.

  Next, in step S112, the answer type determination unit 205 determines the answer type corresponding to the question sentence based on the morphological analysis result in step S102.

  In step S114, the information search unit 21 performs an Internet search based on the search keyword generated in step S110. From the obtained search results, the snippet including the search keyword includes a high number of search keywords. Get n snippets.

  In step S116, the answer candidate extraction unit 22 extracts answer candidates based on the n snippets acquired in step S114 and the answer type determined in step S112.

  Next, in step S118, the answer candidate evaluation unit 23 selects the answer candidate most suitable for the answer from among the plurality of answer candidates obtained in step S116.

  In step S120, the output unit 3 outputs the answer candidate selected in step S118 as a result, and the question answering process routine is terminated.

  As described above, according to the question answering apparatus 100 according to the present embodiment, the function expression is converted into a simple form and the predicate is normalized so as not to affect the meaning of the predicate of the question sentence. , Search for accurately searching for answers to a question sentence by generating a combination of normalized predicates and content words contained in the question sentence as search keywords for searching the answer to the question sentence Keywords can be obtained.

  In addition, by generating a search keyword that deletes the function words of negation, state, tense, and modality related to the event and deletes other function words from the question sentence, the snippet that includes the word necessary for the event is given priority. To be searched.

  In addition, since the functional expression that does not affect the event is deleted, the risk of searching for a snippet that does not become an answer to the question answer is reduced, and as a result, the answer accuracy of the question answer is improved.

  In addition, by using the search keyword generated by the question analysis unit 20, in the information search, many snippets including necessary function expressions and not including unnecessary function expressions are matched, and the question answering apparatus returns an appropriate answer. It becomes like this.

  Note that the present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  For example, the predicate function expression normalization unit 202, the content word extraction unit 203, and the answer type determination unit 205 in the question answering apparatus 100 have been described as an example in which processing is performed based on the morpheme analysis result by the morpheme analysis unit 201. However, the present invention is not limited to this. For example, the question text input by the input unit 1 may have been subjected to morphological analysis.

  Although the above-described question answering apparatus has a computer system therein, the “computer system” includes a homepage providing environment (or display environment) if the WWW system is used.

  In the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium.

2 Calculation unit 20 Question analysis unit 100 Question answering device 202 Predicate function expression normalization unit 203 Content word extraction unit 204 Search keyword generation unit

Claims (7)

A query response search keyword generation method in a query response search keyword generation device including a predicate function expression normalization means, a content word extraction means, and a search keyword generation means,
Based on the morphological analysis result of the question sentence input by the pre-description unit function expression normalization means, the content word included in the question sentence and a character string following the content word and at least one function word For predicates consisting of combinations of functional expressions that are character strings containing , leave the predetermined functional words that represent negation, functional words that represent states, functional words that represent tense, and functional words that represent modalities. A function word representing negation, a function word representing a state, a function word representing tense, and a function word other than a function word representing a modality, thereby normalizing the previous description part; and
A step of extracting a content word from a portion other than a predicate of the question sentence based on a morphological analysis result of the question sentence by the content word extraction means;
A search for an answer to the question sentence by a combination of the predicate normalized by the pre-description part functional expression normalizing means and the content word extracted by the content word extracting means by the search keyword generating means Generating as search keywords,
For generating query keywords for question answering including   The step of normalizing by the pre-description part functional expression normalizing means is composed of a combination of the content word and the functional expression included in the question sentence based on the morphological analysis result of the inputted question sentence The predicate part is normalized by deleting function words that do not affect the meaning of the previous description part and redundant function words and converting the function expression into a simple form. Search keyword generation method for question answering.   The step of normalizing by the pre-description part functional expression normalizing means includes a predetermined functional word other than the functional word representing the negative, the functional word representing the state, the functional word representing the tense, and the functional word representing the modality. The method for generating a search keyword for question answering according to claim 2, wherein the keyword is deleted as a function word that does not affect the meaning of the previous description part. Based on the morphological analysis result of the input question sentence, the content word included in the question sentence, and a functional expression that is a character string that follows the content word and includes at least one functional word For predicates that are composed of combinations, function words that represent negation, function words that represent negation, function words that represent states, function words that represent tense, and function words that represent modalities remain, and represent function words and states that represent negation. A predicate function expression normalizing means for normalizing the previous description part by deleting function words other than function words, function words representing tense, and function words representing modality ;
Based on the result of the morphological analysis of the question sentence, content word extracting means for extracting a content word from a part other than the predicate of the question sentence;
A search keyword for generating a combination of the predicate normalized by the predescription part functional expression normalizing means and the content word extracted by the content word extracting means as a search keyword for searching for an answer to the question sentence Generating means;
Search keyword generating device for question answering including   The pre-description part functional expression normalization means performs pre-description on a predicate composed of a combination of the content word and the functional expression included in the question sentence based on the morphological analysis result of the input question sentence. The query keyword for question answering according to claim 4, wherein the previous description part is normalized by deleting function words that do not affect the meaning of the description part and redundant function words and converting the function expression into a simple form. Generator.   The pre-description part function expression normalization means predetermines the function words other than the function word representing negation, the function word representing the state, the function word representing the tense, and the function word representing the modality. The search keyword generation device for question response according to claim 5, wherein the keyword is deleted as a function word that does not affect the function.   The program for making a computer perform each step which comprises the search keyword production | generation method for question answering in any one of Claims 1-3.