Question answering system, data search method, and computer program

The present invention relates to a question answering system, a data retrieval method, and a computer program. More specifically, in a system that inputs a question text and provides an answer corresponding to the question, the question text is divided into a plurality of partial questions, a search for answer candidates based on the partial question is performed, and the obtained answer candidates are obtained. It relates to a question answering system, a data retrieval method, and a computer program that can provide an accurate answer by performing a process of selecting the final answer from.

Recently, network communication via the Internet and the like has become popular, and various services are provided via the network. There is a search service as one of the services via the network. In the search service, for example, a search server receives a search request from a user terminal such as a personal computer or a mobile terminal connected to a network, and the search server executes processing according to the search request and sends the processing result to the user terminal. It is a service to do.

For example, when executing a search process via the Internet, access a website that provides a search service, enter search conditions such as keywords and categories according to the presentation menu of the website, and send it to the server. The server executes the process according to these search conditions and displays the process result on the user terminal.

There are various aspects of the data retrieval process. For example, a keyword-based search method in which a user inputs a keyword and presents a list information of documents containing the input keyword to the user, or a so-called question answering system in which a user inputs a question sentence and provides an answer to a question. is there. The question answering system is widely used because it does not require the user to select a keyword and can receive only the answer to the question.

For example, Patent Document 1 discloses a typical question answering system. In Patent Document 1, the search term set and the question type are determined from the question sentence, the related document set is searched from the document set stored in the document set storage device according to the determined search term set and the question type, and the related document is obtained. The answer to the question text is extracted from, and the extracted answer and the document information from which the answer is extracted are provided as the response result to the question text.

In a general question answering system, the question text provided by the user is input, and the answer corresponding to the question is output instead of outputting the entire hit document. Web information is often used as a source of knowledge to obtain answers. However, at present, it is difficult to say that the question answering system has sufficient answering accuracy, and the degree of spread is behind that of general search systems.

In the current question answering system, first, an independent word is extracted from the question sentence, the obtained independent word is used as a search word, a search is performed for a knowledge source such as a Web page, and an answer is obtained from the search result. The process is executed in the procedure of extracting. For example, in the case of the question "What is the time difference between Japan and Brazil?", "Japan", "Brazil", and "Time difference" are extracted as independent words, and the search is performed using them as search terms. Although "how many hours" is an independent word, it is usually not included in the search term because it is a word that includes interrogative words. By this search, texts such as "The time difference between Japan and Brazil is 12 hours" can be obtained from knowledge sources such as Web pages, and "12 hours" can be extracted as an answer. The search method for question answering is described in, for example, Non-Patent Document 1.

However, the above method has a problem that when the question sentence becomes long and complicated, the search result containing as many independent words as possible in the question sentence is preferentially searched, so that an appropriate search result cannot be obtained. There is a point.

For example "What is the wooden roller coaster in Yomiuri Land?" (The correct answer is "White Canyon") In the case of the question, "Yomiuri Land", "Wooden", and "Roller coaster" are obtained as search terms, and a search process that prioritizes search results containing as many of these search terms as possible is executed.

Therefore, "White Canyon is a roller coaster in Yomiuri Land." "White Canyon is a wooden roller coaster." Even if the description exists in different parts of the knowledge source, "The bandit at Yomiuri Land is similar to the wooden roller coaster Hirakata Park Elves." The description that happens to include all the search terms such as is searched preferentially, and as a result, Wrong answers such as "banded" and "elf" will be extracted.

Thus, in traditional question answering systems, the longer the question text, the less likely it is that the knowledge source will have an appropriate description that includes all the independent words that exist in it, and as a result. There is a problem that there is a high possibility that an incorrect answer will be extracted.<patcit num="1"><text>Japanese Patent Application Laid-Open No. 2002-132811</text></patcit><nplcit num="1"><text>Isozaki, H. "NTT's Question Answering System for NTCIR QAC2", Working Notes of NTCIR-4 Workshop, pp. 326-332 (2004)</text></nplcit>

<p> The present invention has been made in view of the above-mentioned problems, and provides a question answering system capable of extracting an accurate answer even when a question sentence becomes long, a data search method, and a computer program. The purpose is to do.</p><p> Further, in detail, the present invention divides a question sentence into a plurality of partial questions, searches for answer candidates based on the partial questions, and selects a final answer from the obtained answer candidates. The purpose is to provide a question answering system, a data retrieval method, and a computer program capable of providing accurate answers.</p>

<p> The first aspect of the present invention is Parsing means for executing parsing processing of input questions, A question dividing means for dividing the input question into a plurality of partial questions by applying a predetermined question dividing rule to the syntactic semantic analysis result for the input question generated by the syntactic semantic analysis means. A question answering means for acquiring answer candidates corresponding to a plurality of partial questions generated by the question dividing means together with a certainty value as a reliability index. An answer determining means for selecting an answer to the input question from the answer candidates acquired by the question answering means, and It is in a question answering system characterized by having.</p><p> Further, in one embodiment of the question answering system of the present invention, the answer determining means answers the input question with the answer having the largest total certainty value from the answer candidates acquired for the plurality of partial questions. It is characterized in that it is configured to execute the process selected as.</p><p> Further, in one embodiment of the question answering system of the present invention, the answer determination means is a certainty value among the answer candidates that appear in common to each answer candidate from the answer candidates acquired for the plurality of partial questions. The answer is characterized in that the process of selecting the answer having the largest total of is selected as the answer to the input question is executed.</p><p> Further, in one embodiment of the question answering system of the present invention, the question dividing means applies a predetermined question dividing rule to the syntactic semantic analysis result for the input question, and divides the input question into a plurality of search patterns. The question answering means is configured to execute a search process based on the plurality of search patterns.</p><p> Further, in one embodiment of the question answering system of the present invention, the question answering system is further subjected to the answer candidate acquired for one partial question among the plurality of partial questions generated by the question dividing means. It is characterized by having a configuration in which a converted partial question in which the elements of the partial question are replaced is generated and an answer to the converted partial question is obtained.</p><p> Further, in one embodiment of the question answering system of the present invention, the question dividing means refers to the result of the answer candidate acquired by the question answering means, and when no answer candidate having a certainty value exceeding a certain threshold value is found. The feature is that the structure is such that the process of dividing the question sentence into smaller pieces is executed.</p><p> Further, the second aspect of the present invention is The parsing step that executes the parsing process of the input question, A question division step that applies a predetermined question division rule to the syntactic semantic analysis result for the input question generated in the syntactic semantic analysis step and divides the input question into a plurality of partial questions. A question answering step in which answer candidates corresponding to a plurality of partial questions generated in the question division step are acquired together with a certainty value as a reliability index, and a question answering step. An answer determination step of selecting an answer to the input question from the answer candidates acquired in the question answering step, and an answer determination step. It is in the data search method characterized by having.</p><p> Further, in one embodiment of the data retrieval method of the present invention, in the answer determination step, from the answer candidates acquired for the plurality of partial questions, the answer having the largest total confidence value is the answer to the input question. It is characterized by executing the process of selecting as.</p><p> Further, in one embodiment of the data retrieval method of the present invention, the answer determination step is a certainty value among the answer candidates that appear in common to each answer candidate from the answer candidates acquired for the plurality of partial questions. It is characterized in that the process of selecting the answer having the largest total of is selected as the answer to the input question is executed.</p><p> Further, in one embodiment of the data search method of the present invention, the question division step applies a predetermined question division rule to the syntactic semantic analysis result for the input question, and divides the input question into a plurality of search patterns. The question answering step is characterized by executing a search process based on the plurality of search patterns.</p><p> Further, in one embodiment of the data retrieval method of the present invention, the data retrieval method is further based on the answer candidates acquired for one of the plurality of partial questions generated in the question division step. It is characterized by having a step of generating a transformed subquestion in which the elements of the subquestion are replaced and obtaining an answer to the transformed subquestion.</p><p> Further, in one embodiment of the data search method of the present invention, the data search method further refers to the result of the answer candidate acquired in the question answering step, and finds an answer candidate having a certainty value exceeding a certain threshold. If not, it is characterized by having a step of executing a process of dividing the question sentence into smaller pieces.</p><p> Further, the third aspect of the present invention is A computer program that executes data retrieval processing on a computer. The parsing step that executes the parsing process of the input question, A question division step that applies a predetermined question division rule to the syntactic semantic analysis result for the input question generated in the syntactic semantic analysis step and divides the input question into a plurality of partial questions. A question answering step in which answer candidates corresponding to a plurality of partial questions generated in the question division step are acquired together with a certainty value as a reliability index, and a question answering step. An answer determination step of selecting an answer to the input question from the answer candidates acquired in the question answering step, and an answer determination step. It is in a computer program characterized by having.</p><p> The computer program of the present invention provides, for example, a storage medium, a communication medium, such as a CD, FD, MO, etc., provided in a computer-readable format to a computer system capable of executing various program codes. A computer program that can be provided by a medium or a communication medium such as a network. By providing such a program in a computer-readable format, processing according to the program is realized on the computer system.</p><p> Still other objects, features and advantages of the present invention will be clarified by more detailed description based on the examples of the present invention described later and the accompanying drawings. In the present specification, the system is a logical set configuration of a plurality of devices, and the devices having each configuration are not limited to those in the same housing.</p>

<p> According to the configuration of the present invention, the syntactic semantic analysis process of the input question is executed, a predetermined question division rule is applied to the syntactic semantic analysis result, the input question is divided into a plurality of partial questions, and a plurality of generated questions are generated. Since the answer candidates corresponding to the partial question of are acquired together with the conviction value and the final answer to the input question is selected from a plurality of answer candidates, for example, based on the conviction value, when the input question is a long sentence. However, it is possible to provide the best answer.</p><p> Further, according to the configuration of one embodiment of the present invention, the duplicated answer is preferentially selected as the answer of each partial question from the answer candidates obtained for each of the partial questions. It is possible to select a more appropriate answer as the answer to the original question of.</p>

Hereinafter, the question answering system according to the embodiment of the present invention, the data retrieval method, and the details of the computer program will be described with reference to the drawings.

First, an example of a usage pattern of the question answering system of the present invention will be described with reference to FIG. FIG. 1 is a diagram showing a network configuration in which the question answering system 200 of the present invention is connected to a network. The network 100 shown in FIG. 1 is a network such as the Internet or an intranet, and the network 100 is an answer to clients 101-1 to 101-1 to n and clients 101-1 to n as user terminals for transmitting a question to the question answering system 200. Various Web page providing servers 102A to N that provide Web pages as materials for acquiring the above are connected.

The question answering system 200 inputs various question sentences generated by the user from the clients 101-1 to n, and provides the answers to the input questions to the clients 101-1 to n. The answers to the questions are obtained from the Web pages provided by the Web page providing servers 102A to N.

Web page providing servers 102A to N provide Web pages as public pages by the WWW (World Wide Web) system. A Web page is a set of data displayed on a Web browser, and is composed of text data, HTML layout information, images, sounds, and moving images embedded in a document. A collection of web pages is a website, and a website consists of a top page (home page) and other web pages linked from the top page.

The configuration and processing of the question answering system 200 will be described with reference to FIG. The question answering system 200 is connected to the network 100, receives a question from a client connected to the network 100, and searches for an answer using the web page provided by the web page providing server connected to the network 100 as an information source. , Execute the process provided to the client.

As shown in FIG. 2, the question answering system 200 includes a question input means 201, a syntactic semantic analysis means 202, a question division means 203, a question answering means 204, an answer determination means 205, and an answer output means 206.

Hereinafter, the processes executed by the question input means 201, the syntactic semantic analysis means 202, the question division means 203, the question answering means 204, the answer determination means 205, and the answer output means 206 in the question answering system 200 will be described.

[Question input method] The question input means 201 inputs a question text (input question) from the client via the network 100. Below, as a concrete question example, (Input question) "What is the wooden roller coaster in Yomiuri Land?" The processing of each means in the question answering system 200 will be described on the assumption that the question is input from the client.

[Syntactic Semantic Analysis Means] The syntactic semantic analysis means 202 executes the parsing process of the input question. First, the parsing process will be described. Natural language, which is written in various languages such as Japanese and English, is inherently abstract and highly ambiguous, but it can be processed by computer by handling sentences mathematically. As a result, various applications / services related to natural language will be realized by automated processing such as machine translation, dialogue system, search system, question answering system, etc. Such natural language processing is generally divided into processing phases of morphological analysis, parsing, semantic analysis, and context analysis.

In morphological analysis, a sentence is segmented into morphemes, which are the smallest semantic units, and part of speech is recognized. In syntactic analysis, sentence structures such as phrase structures are analyzed based on grammatical rules. Since the grammatical rules are tree-structured, the parsing result is generally a tree-structure in which individual morphemes are joined based on dependency relationships. In semantic analysis, a semantic structure that expresses the meaning conveyed by a sentence is obtained based on the meaning (concept) of a word in a sentence and the semantic relationship between words, and the semantic structure is synthesized. Further, in context analysis, a sentence (discourse), which is a series of sentences, is regarded as a basic unit of analysis, and a semantic unity between sentences is obtained to construct a discourse structure.

Parsing and semantic analysis are considered to be indispensable technologies for realizing applications such as dialogue systems, machine translation, document proofreading support, and document summarization in the field of natural language processing.

In parsing, a natural language sentence is received and a process of determining the dependency relationship between words (phrases) is performed based on grammatical rules. The parsing result can be expressed in the form of a tree structure (dependency tree) called a dependency structure. Further, in the semantic analysis, it is possible to perform a process of determining the case relationship in a sentence based on the dependency relationship between words (phrases). The case relationship here refers to the grammatical role of each element that composes a sentence, such as the subject (SUBJ) and object (OBJ). Semantic analysis may also include processing to determine the tense, aspect, and narration of a sentence.

For examples of syntactic semantic analysis systems, for example, "Masuichi, Okuma," Construction of a practical Japanese analysis system based on Lexical Functional Grammar ", Natural Language Processing, Vol.10, No.2, pp.79-109. (2003) describes the details of the natural language processing system based on LFG. As the parsing means 202 of the question answering system of the present invention, for example, a natural language processing system based on this LFG can be applied.

Figure 3 shows the configuration of the syntactic semantic analysis system 300 that executes natural language processing based on the Lexical Functional Grammar (LFG) LFG. The morphological analysis unit 302 has a morpheme rule 302A and a morpheme dictionary 302B related to a specific language such as Japanese, divides an input sentence into morphemes, which are the smallest semantic units, and performs part-speech recognition processing. For example, if the sentence "My daughter speaks English" is entered, the morphological analysis result will be "My {Noun}'s {up} daughter {Noun} will {up} English {Noun} {up". } Speaking {Verb1} {tr} Masu {jp} .{pt} "is output.

The result of such morphological analysis is then input to the syntax / semantic analysis unit 303. The syntax / meaning analysis unit 303 has dictionaries such as grammar rule 303A and valency dictionary 303B, and analyzes the phrase structure based on grammar rules, etc., and based on the meaning of words in sentences and the semantic relationship between words. The valency dictionary describes the relationship between a verb and other components in a sentence, such as the subject, and describes the semantic relationship between the predicate and the word related to it. Can be extracted). Then, as a result of parsing, the input sentence is questioned based on the "c-structure (constituent structure)" that expresses the phrase structure of the sentence consisting of words and morphological elements as a tree structure, and the case structure such as the subject and the object. Outputs "f-structure (functional structure)" as a result of semantic and functional analysis of sentences, past forms, polite sentences, etc.

That is, c-structure expresses the structure of a natural language sentence as a tree structure by grouping the morphological elements of the sentence into higher-level phrases, and f-structure expresses the sentence tense based on the concept of grammatical function. Semantic information such as structure, tense, aspect, and speech is expressed by an attribute-attribute value matrix structure.

For example, the input question entered by the client (Input question) "What is the wooden roller coaster in Yomiuri Land?" The result of the syntactic semantic analysis is as shown in Fig. 4. FIG. 4 shows data showing the relationship between the modification source, the modification destination, and the modification, and the result of the parsing processing shown in FIG. 4 has the structure shown in FIG. 5 when shown as the tree structure as the above-mentioned c-structure. .. The modification source "Yomiuri Land" has "Aru" as the modification destination, and the modification relationship is second grade. The modification source "Aru" has "roller coaster" as the modification destination, and the modification relationship is arbitrary modification (relative clause), The modification source "wooden" has "roller coaster" as the modification destination, and the modification relationship is arbitrary modification (no case), The modification source "roller coaster" has the "connecting verb" as the modification destination, and the modification relationship is the nominative case. The modification source "interrogative word" has the "concatenated verb" as the modification destination, and the modification relationship is complementary. The above analysis results are obtained. The "connecting verb" is a word such as "desu" or "da".

The syntactic semantic analysis means 202 shown in FIG. 2 executes the syntactic semantic analysis of the input question and generates the syntactic semantic analysis result as shown in FIG.

[Question division method] Next, the processing of the question dividing means 203 will be described. The question dividing means 203 inputs the syntactic semantic analysis result generated by the syntactic semantic analysis means 202 based on the input question, and executes the question dividing process based on the syntactic semantic analysis result.

An example of the question sentence division rule is shown below. [Question split rule a] If the nominative element of the "concatenated verb" ("desu", "da", etc.) is an "interrogative word" and the nominative element of the "concatenated verb" is modified by two or more arbitrary modifiers, then each Divide into multiple question sentences with only one modifier left.

For example (Input question) "What is the wooden roller coaster in Yomiuri Land?" By applying this division rule to the analysis result shown in FIG. 4 corresponding to, the following two partial questions are generated. (Partial Question 1) "What is the roller coaster in Yomiuri Land?" (Partial Question 2) "What is a wooden roller coaster?"

With reference to FIG. 6, the correspondence between the input question and the partial question in the configuration shown as a tree structure as a c-structure will be described. (Input question) "What is the wooden roller coaster in Yomiuri Land?" The c-structure for is the entire configuration shown in Figure 6, where the above question splitting rule, ie, [Question split rule a] If the nominative element of the "concatenated verb" ("desu", "da", etc.) is an "interrogative word" and the nominative element of the "concatenated verb" is modified by two or more arbitrary modifiers, then each Divide into multiple question sentences with only one modifier left. Is applied to divide the question.

In the configuration of FIG. 6, the concatenated verb (desu, da, etc.) is the [concatenated verb] 401 shown in the figure, and its complement is the [question word] 402. The nominative element of this [connecting verb] 401 is [roller coaster] 403.

Furthermore, the modifier for this nominative element [roller coaster] 403 is Modifier 1: [Yomiuri Land] + [Yes], Modifier 2: [Wooden] It is modified with two modifier elements.

If you select each of these two modifiers and generate a question, (Partial Question 1) "What is the roller coaster in Yomiuri Land?" (Partial Question 2) "What is a wooden roller coaster?" It is divided into these two questions. The c-structures of these two partial questions are the subtree structure 411 and the subtree structure 412 shown in FIG.

In this way, the question dividing means 203 inputs the syntactic semantic analysis result generated by the syntactic semantic analysis means 202 based on the input question, and executes the question dividing process based on the syntactic semantic analysis result. The above-mentioned [question division rule a] described as the question division rule is an example, and other rules may be applied.

The following is an example of [Question division rule b] as a rule different from the above-mentioned question division rule. [Question split rule b] If a clause (section 1) ending in a conjunctive form of a "concatenated verb" (concatenated verb 1) is followed by a clause (section 2) whose complement of the "concatenated verb" (concatenated verb 2) is a "question" Add "What?" To the string before the concatenated verb 1 of 1 to generate the first partial question, and then add the main element of the concatenated verb 2 to the concatenated verb 1 for clause 2. Generate a second partial question by replacing it with a complementary element.

An application example of this question division rule b will be described. For example, let the input question be the following question. (Input question) "What is the roller coaster that the announcer of FUJI TV has reported on in Yomiuri Land?" Such a question is divided by applying the above [question division rule b].

The following results can be obtained from the syntactic semantic analysis results of the above question. Section 1: "On a roller coaster that the announcer of FUJI TV has reported on" Section 2: "What is in Yomiuri Land?" Concatenated verb 1: "de", Concatenated verb 2: "is" Nominative element of concatenated verb 2: "thing", Complementary element of concatenated verb 1: "roller coaster"

For this result [Question split rule b] "When a clause (section 1) ending with a conjunctive form of" concatenated verb "(concatenated verb 1) is followed by a clause (section 2) whose complement of" concatenated verb "(concatenated verb 2) is" question verb " Add "What?" To the string before the concatenated verb 1 in clause 1 to generate the first partial question, and then add the main element of the concatenated verb 2 to the concatenated verb 1 for clause 2. Generate a second partial question by replacing it with the complementary element of. " When the rule is applied,

(Partial Question 1) "What is the roller coaster that the announcer of FUJI TV has reported?" (Partial Question 2) "What is the roller coaster in Yomiuri Land?" Can generate two partial questions.

In this way, the question dividing means 203 inputs the syntactic semantic analysis result generated by the syntactic semantic analysis means 202 based on the input question, and executes the question dividing process according to the preset question dividing rule.

[Question answering method] Next, the processing of the question answering means 204 will be described. The question answering means 204 acquires answer candidates based on the partial question generated by the question dividing means 203 by applying an existing question answering system. For example, a conventional question answering system proposed in "Japanese Patent Laid-Open No. 2002-132811" or the like is applied to obtain answer candidates for a plurality of partial questions. That is, for example, a search process such as text pattern matching using a Web page provided by a Web page providing server connected to a network as an information source is executed to acquire answer candidates corresponding to each partial question.

For example, the following input question, (Input question) "What is the wooden roller coaster in Yomiuri Land?" Two partial questions corresponding to (Partial Question 1) "What is the roller coaster in Yomiuri Land?" (Partial Question 2) "What is a wooden roller coaster?" For each of the above, the answer acquisition process based on the question answering system is executed to acquire the answer candidates corresponding to each partial question. It is assumed that each answer candidate is given a certainty value corresponding to the reliability index indicating the reliability of the answer. As for the confidence value, the consistency with the question type, for example, "who" is a highly reliable answer to the question "who", and the confidence value is set high. For the question "Where are you?", The confidence value for the answer of "Place name" is set high. Also, the higher the frequency of appearance, the higher the confidence value is set. These confidence value setting algorithms depend on the question answering system used, but the common configuration is that the confidence value of a more reliable answer is set high.

Question answering means 204 For example Answers corresponding to partial question a: [Answer A: Confidence value = 0.90], [Answer B: Confidence value = 0.62], [Answer C: Confidence value = 0.30] ... Answers corresponding to partial question b: [Answer P: Confidence value = 0.80], [Answer Q: Confidence value = 0.55], [Answer R: Confidence value = 0.27] In this way, the answer group for each partial question is acquired together with the conviction value and output to the answer determination means 205.

[Means for determining answers] Next, the processing of the answer determination means 205 will be described. The answer determination means 205 selects the answer having the largest total confidence value from the answer candidates obtained for the plurality of partial questions as the answer to the input question. Alternatively, from the answer candidates acquired for a plurality of partial questions, among the answer candidates that appear in common to each answer candidate, the answer with the largest total certainty value is selected as the answer to the input question. ..

For example (Input question) "What is the wooden roller coaster in Yomiuri Land?" Partial question (Partial Question 1) "What is the roller coaster in Yomiuri Land?" As an "answer (confidence value)" to "Banded (0.95)" "White Canyon (0.95)" "Elf (0.21)" Is obtained.

on the other hand, (Input question) "What is the wooden roller coaster in Yomiuri Land?" Partial question (Partial Question 2) "What is a wooden roller coaster?" As an "answer (confidence value)" to "ASKA (0.90)" "Jupiter (0.90)" "Elf (0.90)" "Regina (0.90)" "White Canyon (0.90)" "White Cyclone (0.90)" Is obtained.

The result is shown in Fig. 7. The answer determination means 205 selects the answer with the largest total confidence value as the answer to the input question. Alternatively, the answer that appears in common in the answers to each partial question and has the highest total confidence value is selected as the answer to the input question.

In this example, it is assumed that the answer that appears in common in the answers to each partial question and has the highest total confidence value is selected as the answer to the input question. In this case, the answer that appears in common in the answers to each partial question and has the highest total confidence value is "White Canyon", and "White Canyon" is selected as the final answer.

[Answer output means] The answer output means 206 outputs the answer determined by the answer determination means 205 to the client. The answer selected based on the certainty value in the answer determination means 205 described above is output to the client.

Next, the processing sequence executed by the question answering system of the present invention will be described with reference to the flowchart of FIG.

When a question from the client is input in step S101, a syntactic semantic analysis process based on the input question from the client is executed in step S102. This process is a process executed by the syntactic semantic analysis means 202 shown in FIG.

Specifically, for example (Input question) "What is the wooden roller coaster in Yomiuri Land?" If, the syntactic semantic analysis result shown in Fig. 4 is generated.

Next, in step S103, a question division rule is applied to the syntactic semantic analysis result for the input question to generate a partial question. This process is a process executed by the question dividing means 203 shown in FIG. 2, and as described above, for example, [Question split rule a] If the nominative element of the "concatenated verb" ("desu", "da", etc.) is an "interrogative word" and the nominative element of the "concatenated verb" is modified by two or more arbitrary modifiers, then each Divide into multiple question sentences with only one modifier left. [Question split rule b] If a clause (section 1) ending in a conjunctive form of a "concatenated verb" (concatenated verb 1) is followed by a clause (section 2) whose complement of the "concatenated verb" (concatenated verb 2) is a "question" Add "What?" To the string before the concatenated verb 1 of 1 to generate the first partial question, and then add the main element of the concatenated verb 2 to the concatenated verb 1 for clause 2. Generate a second partial question by replacing it with a complementary element. These preset question division rules are applied to execute the question division process and generate a partial question.

Specifically, for example (Input question) "What is the wooden roller coaster in Yomiuri Land?" The following two partial questions corresponding to are generated. (Partial Question 1) "What is the roller coaster in Yomiuri Land?" (Partial Question 2) "What is a wooden roller coaster?"

Next, in step S104, an answer candidate based on the partial question generated by the question dividing means 203 is acquired by applying the existing question answering system, and is executed by the question answering means 204 shown in FIG. is there. A plurality of answer candidates are acquired together with a certainty value by, for example, a search process by text matching for a Web page provided by a Web page providing server connected to a network. For example, the answer candidates shown in FIG. 7 can be obtained.

In step S105, output answers are selected based on the confidence values of a plurality of answer candidates. This is the process executed by the answer determination means 205 shown in FIG. The answer determination means 205 selects the answer with the largest total confidence value as the answer to the input question. Alternatively, a process is executed in which the answer that appears in common in the answers to each partial question and has the highest total confidence value is selected as the answer to the input question.

In step S106, the answer obtained in step S105 is output to the client via the answer output means 206.

As a result of these processes, for example (Input question) "What is the wooden roller coaster in Yomiuri Land?" As the answer corresponding to, "White Canyon" is finally selected and provided to the client.

[Other Examples] Hereinafter, a configuration example different from the above-described embodiment will be described.

Change example 1, In the above embodiment, the question dividing means 203 inputs the syntactic semantic analysis result generated by the syntactic semantic analysis means 202 based on the input question, applies the preset question dividing rule, and executes the question dividing process. Explained as an example.

The question dividing means 203 further applies the search pattern generation rule described below to generate a search pattern for the input question, and based on the generated search pattern, executes a character string search of the input question to divide the question. It may be configured to perform.

[Search pattern generation rule] If the nominative element of the "concatenated verb" ("desu", "da", etc.) is an "interrogative word" and the nominative element of the "concatenated verb" is modified by two or more arbitrary modifiers, then each Divide into multiple search patterns with only one modifier left. However, if the above nominative element is modified in the relationship of "arbitrary modification (nominative)", "" * "{Was | is}" Nominative element "of" Nominative element modifier "{is | is | is | de}" "Nominative element" {| is} "*" of "Nominative element modifier" Generate a search pattern called If the above nominative element is modified by the relation of "arbitrary modification (relative clause)", "*" {Ha | is} "Nominative element modifier clause" "Nominative element" "" Nominative element modifier clause "" Nominative element "" * "" Generates a search pattern called.

A specific example will be described. For example, the input question is "What is the roller coaster in Yomiuri Land?" If it is, "" * "{Is | is} a wooden roller coaster {is | is | is | in}" "Wooden roller coaster {| is}" * "" When "" * "{Ha | ga} roller coaster in Yomiuri Land" "Roller coaster {| is}" * "in Yomiuri Land Generate two types of search patterns.

These search patterns are provided to the question answering means 204, and the question answering means 204 executes a search process based on these a plurality of search patterns. Character strings that match the "*" obtained as a result of this search are used as answer candidates. However, all certainty values are set to "1".

By configuring the search according to this search pattern, it becomes possible to replace the question answering means with a means for performing a search based on pattern matching, and the versatility is enhanced.

b. Change example 2 In the above-described embodiment, the description is based on the premise that answer candidates can be obtained by the search process based on the partial question generated by the question answering means 204, but the answer candidates are not always obtained. In some cases, only answer candidates with low certainty values are acquired.

In such a case, the question sentence dividing means 203 refers to the result of the answer candidate acquired by the question answering means 204, and if no answer candidate having a certainty value exceeding a certain threshold value is found, the question sentence is further divided. ..

For example, it has a subdivision rule for the following search patterns and subdivides. [Subdivision rule] If the nominative element of the "concatenated verb" ("desu", "da", etc.) is an "interrogative word" and the nominative element of the "concatenated verb" is modified by two or more arbitrary modifiers, then each Divide into multiple search patterns with only one modifier left. However, if the above nominative element is modified in the relationship of "arbitrary modification (nominative)", "" * "{Was | is}" Nominative element "of" Nominative element modifier "{is | is | is | de}" "Nominative element" {| is} "*" of "Nominative element modifier" If the search pattern is generated and the search result is not obtained, "" * "{Is | is}" Nominative element "{is | is | is | in}" "" Nominative element "{| is}" * "" When, "" * "{Is | is}" The modifier of the nominative element "{is | is | is | in}" "" Nominative element modifier "{| is}" * "" Generates a search pattern called.

The above search pattern, i.e. "" * "{Is | is} a wooden roller coaster {is | is | is | in}" "Wooden roller coaster {| is}" * "" If you do not get search results with these search patterns, In addition, a new Kensaku pattern, ie "" * "{Is | is} roller coaster {is | is | is}" "Roller coaster {| is}" * "" When, "" * "{Is | is} wooden {is | is | is}" "Wooden {| is}" * "" Two types of search patterns are generated, and the character string that matches "*" is used as each answer candidate.

By subdividing the partial question condition in this way, it is possible to increase the possibility of obtaining answer candidates and to obtain a highly reliable answer with a high conviction value.

c. Modification 3 In the above-described embodiment, the answer determination means 205 executes a process of selecting the answer having the largest total certainty value among the answer candidates commonly appearing in the answer candidates of each partial question as the answer to the question from the client. Explained as what to do.

The answer determination means 205 may apply another answer selection algorithm instead of such a certainty value-dependent answer determination process. For example, a rule is retained in which the answer of a partial question (partial question A) replaces a part of another partial question (partial question B), and the final answer is obtained by the changed partial question B.

A specific example will be described. For example, it is assumed that the question sentence dividing means 203 has the following question dividing rule. [Question split rule] When the nominative element of the "concatenated verb" ("desu", "da", etc.) is "question", the nominative element of the "concatenated verb" is modified by an arbitrary modification (relative clause), and the relative clause If the nominative element or objective element (element 1) in is further modified by an arbitrary modifier element (element 2), add "what?" To the character string corresponding to element 2 + element 1. The first partial question is generated, and the question text with element 2 deleted is used as the second partial question. It is a condition that element 1 is a common noun (not a proper noun).

According to this question division rule, for example (Input question) "Who directed the movie with the highest box office revenue in 2004?" about, Element 1: "Movie" Element 2: "The highest box office revenue in 2004" Is obtained, and the following partial question is obtained. (Partial Question 1) "What was the movie with the highest box office revenue in 2004?" (Partial Question 2) "Who directed the movie?"

Further, the answer determination means 205 replaces the element 1 of the partial question 2 with the answer candidate obtained by the partial question 1.

For example, the following sub-question, ie (Partial Question 1) "What was the movie with the highest box office revenue in 2004?" If there is "Spirited Away" as the answer obtained from (Partial Question 2) "Who directed the movie?" Replace element 1 "movie" with the answer "Spirited Away" of partial question 1 to generate converted partial question 2 which is a conversion of partial question 2. As a result, the following conversion part question 2 is generated. (Conversion part Question 2) "Who oversaw Spirited Away?" The question answering means 204 executes a search process for obtaining an answer based on the converted partial question, and obtains an answer to the converted partial question 2.

The element replacement of the question sentence of the partial question 2 may be performed only when the certainty value of the answer obtained from the partial question 1 is larger than the preset threshold value. Alternatively, a method may be taken in which a plurality of answer candidates for the question text of the partial question 1 are output and the user is made to select the answer that seems to be correct.

Finally, a hardware configuration example of the information processing apparatus constituting the question answering system that executes the above-described processing will be described with reference to FIG. The CPU (Central Processing Unit) 501 executes processing corresponding to the OS (Operating System), syntactic semantic analysis processing based on the input question described in the above embodiment, question division processing, question answering processing, answer determination processing, and the like. To do. These processes are executed according to a computer program stored in a data storage unit such as a ROM or a hard disk of each information processing device.

The ROM (Read Only Memory) 502 stores programs and calculation parameters used by the CPU 501. The RAM (Random Access Memory) 503 stores a program used in the execution of the CPU 501, parameters that change appropriately in the execution, and the like. These are connected to each other by a host bus 504 composed of a CPU bus and the like.

The host bus 504 is connected to an external bus 506 such as a PCI (Peripheral Component Interconnect / Interface) bus via a bridge 505.

The keyboard 508 and the pointing device 509 are input devices operated by the user. The display 510 is composed of a liquid crystal display device, a CRT (Cathode Ray Tube), or the like, and displays various information in text or images.

The HDD (Hard Disk Drive) 511 has a built-in hard disk, drives the hard disk, and records or reproduces programs and information executed by the CPU 501. The hard disk is used, for example, for storing question division rules, storing syntactic semantic analysis results, storing answer candidates, and various other computer programs such as data processing programs.

The drive 512 reads the data or program recorded on the removable recording medium 521 such as the mounted magnetic disk, optical disk, magneto-optical disk, or semiconductor memory, and transfers the data or program to the interface 507 and the external bus 506. , Bridge 505, and RAM 503 connected via host bus 504.

The connection port 514 is a port for connecting the external connection device 522, and has a connection portion such as USB or IEEE1394. The connection port 514 is connected to the CPU 501 or the like via the interface 507, the external bus 506, the bridge 505, the host bus 504, or the like. The communication unit 515 is connected to the network and executes communication with the client and the network connection server.

The hardware configuration example of the information processing device applied as the question answering system shown in FIG. 9 is an example of a device configured by applying a PC, and the question answering system of the present invention has the configuration shown in FIG. Not limited to this, any configuration may be used as long as the processing described in the above-described embodiment can be executed.

The present invention has been described in detail with reference to the specific examples. However, it is self-evident that a person skilled in the art can modify or substitute the embodiment without departing from the gist of the present invention. That is, the present invention has been disclosed in the form of an example, and should not be construed in a limited manner. In order to judge the gist of the present invention, the column of claims should be taken into consideration.

The series of processes described in the specification can be executed by hardware, software, or a composite configuration of both. When executing processing by software, install the program that records the processing sequence in the memory in the computer built in the dedicated hardware and execute it , or execute the program on a general-purpose computer that can execute various processing. It can be installed and run.

For example, the program can be pre-recorded on a hard disk or ROM (Read Only Memory) as a recording medium. Alternatively, the program may be temporarily or permanently placed on a removable recording medium such as a flexible disc, CD-ROM (Compact Disc Read Only Memory), MO (Magneto optical) disc, DVD (Digital Versatile Disc), magnetic disc, or semiconductor memory. It can be stored (recorded). Such a removable recording medium can be provided as so-called package software.

In addition to installing the program on the computer from the removable recording medium as described above, the program can be transferred wirelessly from the download site to the computer, or transferred to the computer by wire via a network such as LAN (Local Area Network) or the Internet. , The computer can receive the program transferred in this way and install it on a recording medium such as a built-in hard disk.

The various processes described in the specification are not only executed in chronological order according to the description, but may also be executed in parallel or individually as required by the processing capacity of the device that executes the processes. Further, in the present specification, the system is a logical set configuration of a plurality of devices, and the devices having each configuration are not limited to those in the same housing.

As described above, according to the configuration of the present invention, the syntactic semantic analysis process of the input question is executed, the predetermined question division rule is applied to the syntactic semantic analysis result, and the input question is asked as a plurality of partial questions. The answer candidates corresponding to the generated multiple partial questions are obtained together with the conviction values, and the final answer to the input question is selected from the multiple answer candidates, for example, based on the conviction values. Even if the input question is a long sentence, it is possible to provide the optimum answer.

Further, according to the configuration of one embodiment of the present invention, the duplicated answer is preferentially selected as the answer of each partial question from the answer candidates obtained for each of the partial questions. It is possible to select a more appropriate answer as the answer to the original question of.

<figref num="1">It is a network block diagram which shows the application example of the question answering system of this invention.</figref><figref num="2">It is a figure explaining the structure of the question answering system which concerns on one Embodiment of this invention.</figref><figref num="3">It is a figure explaining the structural example of the parsing means in the question answering system which concerns on one Embodiment of this invention.</figref><figref num="4">It is a figure which shows the data example of the syntactic semantic analysis result generated by the processing of the syntactic semantic analysis means.</figref><figref num="5">It is the figure which showed the data example of the parsing result generated by the processing of the parsing means as a tree structure.</figref><figref num="6">It is a figure explaining the generation process of a partial question by division based on a syntactic semantic analysis result.</figref><figref num="7">It is a figure explaining the list of the answer and the conviction value to a partial question, and the answer determination process executed by the answer determination means.</figref><figref num="8">It is a figure which shows the flowchart explaining the processing sequence executed by the question answering system which concerns on one Embodiment of this invention.</figref><figref num="9">It is a figure explaining the hardware configuration example of the question answering system which concerns on one Embodiment of this invention.</figref>

Code description

100 networks 101 client 102 Web page providing server 200 Question answering system 201 Question input method 202 Parsing means 203 Question splitting means 204 Question answering means 205 Answer determination method 206 Answer output means 300 Syntax Semantic Analysis System 302 Morphological analysis department 302A Morpheme rule 302B Morpheme dictionary 303 Syntax Semantic Analysis Department 303A Grammar rules 303B Valency dictionary 401 Concatenated verb 402 Interrogative words 403 roller coaster 411,412 Partial tree structure 501 CPU (Central Processing Unit) 502 ROM (Read-Only-Memory) 503 RAM (Random Access Memory) 504 host bus 505 bridge 506 External bus 507 interface 508 keyboard 509 pointing device 510 display 511 HDD (Hard Disk Drive) 512 drives 514 connection port 515 Communication Department 521 Removable recording medium 522 External device