JP2002014990A - Question answering system, question answering processing method, transformation rule automatic acquisition processing method and recording medium for these system and method programs - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a question answering system which has high correct answer rates to question statements and also has its easy and flexible construction. SOLUTION: In order to obtain an answer from a data base 11 to a question statement, the question statement is collated with a data base statement at a collation part 14 and resemblance is calculated between both statements. At the same time, a transformation part 16 transforms both question and data base statements by means of a transformation rule that is previously stored in a transformation rule storage part 15. Then the collation and transformation of statements are repeated at the parts 14 and 16 respectively for searching for a data base statement having higher resemblance to the question statement.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a natural language information processing system using a computer, and more particularly to a question answering system using inference based on similarity. It can be used for information retrieval and information extraction.

[0002]

2. Description of the Related Art A question answering system, for example, inputs a question "What signs of the brain are involved in the signs of Parkinson's disease?" It is said that melanocytes in the substantia nigra of the midbrain degenerate and dopamine, a neurotransmitter that is produced in the substantia nigra, causes disease. "
This is a system that searches for such sentences and accurately extracts the "black matter" corresponding to the question from the sentence and outputs it as an answer.

[0003] As such a question answering system, a "question answering system using syntax information" published by the present inventors in Reference 1 below is known. [Reference 1] Masaki Murata and Masao Utiyama and
Hitoshi Isahara, Question Answering System Using S
yntactic Information, 1999, http://xxx.lanl.gov/ab
s / cs.CL / 9911006. This question answering system first extracts keywords from a question sentence, and then outputs the IDF (In
verse Document Frequency) is extracted. Next, the sentence and the extracted sentence are parsed. The query sentence is compared with the extracted sentence using the syntax information of the analysis result, and the similarity between the question sentence and the extracted sentence is calculated in accordance with a predetermined calculation method, while providing solution candidates. From the sentences extracted from the database, the solution is extracted from the one having the highest similarity to the question sentence. The extraction of the solution is performed by using the phrase on the database side associated with the phrase containing the question word in the question sentence as the solution. This system does not consider transforming a query sentence and a sentence obtained from a database so as to increase the similarity and then collating the sentence.

[0004] As a prior art prior to the "question answering system using syntax information", there is known an information retrieval technique using a modification of a sentence described in Reference 2 below. Hereinafter, this is referred to as Katz's method. [Reference 2] Boris Katz, Using English for Indexi
ng and Retrieving, Artificial Intelligence at MIT,
Vol.2, MIT Press, 1990. In general, in deriving an answer by comparing a question sentence and a database sentence in a question answering system, the question sentence and the database sentence are compared to obtain a database sentence that best matches the question sentence. Getting answers from inside is done.

[0005] For example, when there is an input question sentence "Where is the capital of Japan?", A sentence similar to that is extracted from the database. Here, let's say that there is a sentence "The capital of Japan is Tokyo." Then, "where is the capital of Japan" and "the capital of Japan is Tokyo" are collated, and "Tokyo" corresponding to the part of the question word "where" is output as an answer.

However, as described above, the question sentence and the database sentence exactly match, and the answer part cannot always be easily extracted. For example, suppose the database did not have the sentence "The capital of Japan is Tokyo", but only the sentence "Tokyo is the capital of Japan". Then, it cannot be matched with the question sentence and the solution cannot be obtained.

[0007] Katz described in Reference 2 above
Is to convert all sentences to the most common expression (this is called the base expression) and then extract the solution so that the sentence can be matched even if the sentence patterns of the query sentence and the database sentence are different. I do. For example, in the above example, it is not known which base expression is "Japan's capital is Tokyo" or "Tokyo is Japan's capital", but here "Japan's capital is Tokyo" as the base expression. The sentence in the database is transformed into it, and the answer is obtained by collating it with "Where is the capital of Japan?"

[0008]

The "question answering system using syntax information" of the above-mentioned reference 1 mainly uses a semantic constraint for matching a question sentence with a database sentence. However, there are advantages and disadvantages to using semantic constraints and using sentence transformations when matching sentences, and a method using semantic constraints alone cannot always achieve sufficient matching. . A method using sentence transformation is also considered effective for collating a question sentence with a database sentence.

However, the Katz method has the following disadvantages. It is necessary to transform every sentence into the most common representation, the base representation, but it is difficult to define the base representation exactly. For example,
In the case where the base expression is defined as the active sentence and the passive sentence is transformed into the active sentence, a relatively uniform transformation is possible, but in the above example, "Tokyo is the capital of Japan" And the sentence "The capital of Japan is Tokyo" cannot be clearly defined as the base expression. That is, when there are a plurality of mutually deformable sentences, there are many cases where it is ambiguous which one should be used as a base expression. Must be defined as an expression. However, determining such a base representation is actually a very difficult task.

[0010] In addition, in a simple transformation to the base representation, the two sentences "Tokyo is located in the Kanto region" and "Tokyo is located in the Kanto region" indicate that "the Japanese capital is located in the Kanto region". Such a sentence cannot be transformed.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a question answering system which solves the above-mentioned problems, has a high correct answer rate to a question sentence, and is easy and flexible to construct a system. Specifically, when using sentence transformation in a question answering system, it is not necessary to determine a base expression, so that transformation rules can be easily described, and sentence transformation can be performed flexibly. The purpose is to provide a means for automatically acquiring deformation rules.

[0012]

According to the present invention, when a query sentence is compared with a database sentence, both of the query sentence and the database sentence are enhanced by using a transformation rule stored in advance so as to increase the similarity between the query sentence and the database sentence. Rewriting sentences is the most important feature.

The use of the transformation is similar to the above-mentioned Katz's method. In the Katz's method, all sentences are transformed into a base expression. In the present invention, however, the present invention is not limited to the transformation into the base expression. Using the similarity as a scale, the transformation is performed so as to increase the similarity. Thus, the present invention has the following advantages.

In the present invention, there is no need to determine a base expression, so that the description of the deformation rule is facilitated. For example, if "Japan's capital is Tokyo" and "Tokyo is Japan's capital" are mutually deformable, it is unclear which base expression should be used. , It is necessary to force one of them to be defined as the base representation. However, there are many cases where such a definition must be forced, and it is practically difficult to determine a base representation. On the other hand, in the present invention, since it is not always necessary to transform the expression into the base expression, in this case, the problem is solved by writing the following two rules.

Rule 1: "Tokyo is the capital of Japan" is transformed from "Tokyo is the capital of Japan" Rule 2: "Tokyo is the capital of Japan" to "Tokyo is the capital of Japan" Modification As described above, the present invention is characterized in that the right side of the modification rule does not need to be a base expression. However, if something other than the base expression is described on the right side, a mechanism for controlling and managing the deformation is required. In the present invention, the deformation is controlled and managed on the scale of similarity. In other words, by performing deformation so as to increase the similarity, the accuracy in matching the question sentence with the database sentence is improved.

In the present invention, since there is no restriction on the transformation rule for the base expression, the transformation rule can be automatically acquired by using dictionary data such as a Japanese language dictionary or data such as an existing collection of questions and answers. is there. In other words, a description sentence (or definition sentence) of the same word is extracted from the dictionary data read from a plurality of dictionary files, the extracted explanations are matched, and a synonym or synonymous phrase is extracted from the result. To generate another transformation rule for rewriting to another sentence representing the same content.

Alternatively, an existing question sentence and a response sentence thereto are inputted, the input question sentence and the response sentence are matched, and a synonym or a synonym phrase is extracted from the result.
Based on this, a transformation rule is generated.

A program for realizing each of the above processes by a computer can be stored in an appropriate recording medium such as a computer-readable portable medium memory, a semiconductor memory, and a hard disk.

[0019]

FIG. 1 shows an example of a system configuration according to the present invention. In the figure, 1 is a question answering system according to the present invention, 2
Is a transformation rule automatic acquisition system that automatically acquires sentence transformation rules from dictionary data and question answer collections.

The question sentence input section 10 is a means for inputting a question sentence in a natural language. The database 11 is a newspaper,
This is a database that stores digitized text information of papers and other various documents. The keyword extraction / information search unit 12 is a means for extracting a keyword from a question sentence and searching a database. The syntax analysis unit 13 is a means for syntactically analyzing a question sentence and a sentence extracted from the database 11 by a search (this is referred to as a database sentence). The collating unit 14 is a means for collating the input question sentence with the database sentence and calculating their similarity. The transformation rule storage unit 15 stores rules for transforming a sentence into another sentence representing the same content.

The transformation section 16 is means for rewriting a question sentence and a database sentence using the transformation rules stored in the transformation rule storage section 15. The rewritten result is collated again by the collation unit 14, the similarity is calculated, and the processing by the transformation unit 16 and the processing by the collation unit 14 are repeated until the similarity does not improve. The answer output unit 17
This is a means for extracting a solution from the database sentence extracted in the matching having the highest similarity and outputting it as a response sentence.

FIG. 2 shows an example of a transformation rule stored in the transformation rule storage unit 15. FIG. 2A is an example of a transformation rule for a synonym, and this transformation rule can be handled by the Katz method. FIG. 2B is an example of a transformation rule for a synonymous phrase.
Is B "→" B is A "cannot be handled by the Katz method described above.

In addition, not only those dealing with the direct equivalence of the meaning as described above, but also transformation rules related to inference can be used. FIG. 2C shows an example thereof. Here, the left side of the transformation rule is "A" and "if A, B
Are input. This usage example will be described.

For example, it is assumed that there are two sentences, "the weather is fine" and the sentence "if it is fine, an umbrella is unnecessary". When the transformation rule shown in FIG. 2C is applied to these sentences, "sunny" matches A and "umbrella is unnecessary" and B match, and "umbrella is unnecessary" is derived from the result.

In the present invention, the knowledge obtained by the inference as described above can be handled by the transformation rule. The left and right sides of the transformation rule can be anything, and any description can be made, either as part of a sentence or as a union.

Next, the processing of the question answering system 1 shown in FIG. 1 will be described with reference to the flowchart shown in FIG.

First, the question sentence input unit 10 inputs a question sentence (step S1). There is no limitation on the input method such as input from a terminal via a network or input from an application program such as information retrieval.

The keyword extraction / information search unit 12 extracts a keyword from the input question sentence (step S).
2). As a simple method of keyword extraction, for example, there is a method of morphologically analyzing a sentence and leaving only nouns.

Next, the keyword extraction / information search section 12
Is the keyword IDF (Invers
A plurality of sentences having a large sum of (e Document Frequency) are extracted (step S3). As a simple method of calculating the IDF value, for example, when all character strings in the database are N and the frequency of appearance of keywords in the database is n,
og (N / n) can be used. Instead of using the value of IDF, all sentences including keywords appearing in the question sentence may be roughly extracted from the database.

Next, the syntax analyzer 13 analyzes the syntax of the question sentence and all of the sentences (database sentences) extracted from the database 11, and uses them as a set of question sentences and a set of database sentences, respectively (step S4). . In this syntax analysis, for example, a Japanese syntax analysis system described in Reference 3 below can be used. [Reference 3] Sadao Kurohashi, Japanese Dependency /
Case Structure Analyzer KNP version 2.0b6, (Departm
After that, the matching unit 14 initializes a variable S that stores the largest similarity up to the present time to 0 (step S5), and proceeds to step S6. In step S6, each member of the set of question sentences and each member of the set of database sentences are collated in any combination by using the syntax information by the syntax analysis unit 13, and a solution candidate is determined for each combination. Calculate the similarity between a question sentence and a database sentence.

An example of a formula for calculating the similarity will be described. The similarity between the question sentence and the database sentence p is calculated by the following formula Score
given by (p).

Score (p) = BNST1 (p) + α × BNST2 (p) -β ₁
× BNUM _sent (p) −β ₂ × BNUM _all (p) where BNST1 (s) = Σ NEAR (p, b) × JIRITSU (b) (Σ is the sum of clause b of the question sentence) BNST2 (s) = Σ NEAR (p, b) × bnst2 (b1, b2) (Σ is the sum of all dependency relations (b1, b2) in the question.
b1 is related to b2) bnst2 (b1, b2) is JIRITSU (b1) × FUZOKU (b1) × JIRITSU
When (b2) is not 0, bnst2 (b1, b2) = JIRITSU (b1) + FUZOKU (b1) + JIRITSU (b
2) Otherwise, bnst2 (b1, b2) = 0 NEAR (p, b) = NEAR1 (p, b) + NEAR2 (p, b) + NEAR3 (p, b) NEAR1 (p, b) is b If the sentence is the same as the answer, NEAR1 (p,
b) = γ ₁ Otherwise, NEAR1 (p, b) = 1 NEAR2 (p, b) is a case where b is the same sentence as the answer part and the dependency distance between the question word and b and the database sentence correspond to them NEAR2 (p, b) = 1 + γ ₂ / (1 + d) If the greater of the dependency distance between phrases is d, NEAR2 (p, b) = 1 otherwise, b is the answer for NEAR3 (p, b) NEAR3 (p, b) = 1 + γ ₃ / (1 + d ′) If the greater of the phrase distance between the question word and b in the same sentence as the part and the phrase distance between the corresponding phrases in the database sentence is d ′, NEAR3 (p, b) = 1 + γ ₃ / (1 + d ′) Then, NEAR3 (p, b) = 1 All the clauses of the extracted database sentence are
It is associated with any clause on the input side so that the value of e (p) is maximized. JIRITSU (b) is the similarity in the independent word between the phrase b on the input side and the phrase associated with it,
FUZOKU (b) is the degree of similarity in the appendix between the phrase b on the input side and the phrase associated with it.

BNUM _sent (p) is the number of phrases in the database sentence p including the answer part. BNUM _all (p)
Is the number of clauses of the database sentence p. The two BNUMs are
This is a term for ensuring that matching with a sentence without an extra clause is larger if other information is the same.

NEAR raises the value of a phrase close to the answer part, and is used for better matching between the question sentence and the database sentence. The “dependency distance” between two clauses used in the calculation of NEAR means the number of branches between the two clauses in the parse tree, and the “separation distance” between two clauses is the two It means the number obtained by adding 1 between phrases.

Α, β ₁ , β ₂ , γ ₁ , γ ₂ and γ ₃ are constants determined by experiments. In addition, the expression of Score (p) shown here uses only the binary relation with BNST1 and BNST2, but may additionally use a ternary or quaternary relation.

When the similarity between the question sentence and the database sentence is calculated by the above equation, the largest similarity value of this combination is S, and the candidate solution at that time is A (step S6). .

Next, it is determined whether or not the current value of S is larger than the previous value (step S7). If it is larger, the process proceeds to step S8. If not, step S9 is performed.
Proceed to.

In step S8, the transformation unit 16 rewrites each member of the set of question sentences and the set of database sentences by using the transformation rules stored in the transformation rule storage unit 15.
The rewritten sentences are added to a set of question sentences and a set of database sentences, respectively. Thereafter, the process returns to step S6, and the collation by the collation unit 14 is repeated again. In addition, the deformation part 1
6 is repeated, the number of members of the set of question sentences and the set of database sentences becomes enormous, and the calculation cost becomes enormous. At this time, it is preferable to repeat the process while leaving only the members whose similarity between the question sentence and the database sentence increases to some extent and deleting other members during the processing by the matching unit 14.

If it is determined in step S7 that the current value of S is not larger than the previous value of S, the answer output unit 17 extracts a solution from the solution candidate A having the highest similarity and sends it to the question It is output as a response sentence to the sentence (step S9), and the process ends. The generation of a response sentence from a database sentence that is a solution candidate is performed, for example, by using a phrase in a database sentence associated with a phrase including a question word in a question sentence as a solution.

In the above-described processing example, the description has been given assuming that the deformation by the deformation unit 16 is repeated until the improvement of the similarity is not observed. However, due to the calculation time, when the similarity exceeds a certain threshold, the deformation processing is not performed. It is also possible to terminate the repetition or to determine the maximum number of repetitions of the deformation processing in advance, and to terminate the deformation processing after repeating the number of repetitions.

Hereinafter, a specific example of execution by the above processing will be described.

(1) First, it is assumed that the following sentence is input as a question sentence.

Question sentence: "Where is the capital of Japan?" --- a This sentence is a member of a set of question sentences. Keyword extraction /
In the information search unit 12, "Japan" and "capital" are obtained as keywords.

(2) The keyword extraction / information search unit 12
Database 1 with keywords "Japan" and "Capital"
1 and a plurality of sentences having a large IDF
Extract. Here, it is assumed that the following two sentences are obtained.

Database sentence 1: “Tokyo is the capital of Japan” --- b ・ Database sentence 2: “The capital of Japan's neighbor Korea is Seoul” --- c The sentence extracted from these two sentences Become a member of the set.

(3) The syntax analyzer 13 parses all the question sentences and database sentences. That is, a, b, c
Is parsed.

(4) At the start of the collation, the collation unit 14 sets the value of the variable S that stores the maximum similarity to 0.

(5) Thereafter, the collating unit 14 collates the members of the set of question sentences with the members of the set of database sentences in any combination. That is, the comparison between a and b and the comparison between a and c are performed. Here, it is assumed that the similarity between a and b is 22, and the similarity between a and c is 35. At this time, since the maximum similarity of the collation is 35, S is set to 35.

(6) Since the S value 35 is larger than the previous S value 0, the processing of the deforming unit 16 (step S8) is executed. The transformation unit 16 applies a transformation rule to each member of the set of question sentences and the set of database sentences. Here, for the sake of simplicity, it is assumed that only the following deformation rules are used.

Modification rule: A is B → B is A --- d The modification rule of d is applied to each of the sentences a, b, and c, and the following three sentences are newly generated. You.

Deformation of a (use of d): “Where is the capital of Japan” --- e · Deformation of b (use of d): “The capital of Japan is Tokyo” --- f ・Modification of c (use of d): "Seoul is the capital of Korea, the neighbor of Japan." --- g Add these to the set of question sentences and the set of database sentences. Here, e is added to the set of question sentences, and f and g are added to the set of database sentences.

(7) The collating unit 14 again collates the members of the set of question sentences with the members of the set of database sentences in any combination. That is, a and b, a and c, a and f, a and g, e
And b, e and c, e and f, and e and f. Here, it is assumed that the similarity between a and f and e and b is 47, which is the largest. S is set to 47.

(8) The current S value 47 is the previous S value 35
Therefore, the processing of the deforming unit 16 is executed again. At this time, the transformation unit 16 generates only the sentences in the set of question sentences and the set of database sentences this time, so that the set of question sentences and the set of database sentences do not change.

(9) Again, the collating unit 14 collates the members of the set of question sentences with the members of the set of database sentences in any combination. At this time, since the members of the set have not changed from the previous time, the similarity between the matching of a and f and the matching of e and b are the same as in the previous time, and the value is 47. In addition, it is assumed that the solution candidate at this time is “Tokyo” corresponding to the part of the question word at the time of matching.

(10) Since the current S value 47 is the same size as the previous S value, the repetition processing ends here, and “Tokyo”, which is a solution candidate, is output as a solution.

The present invention further has a transformation rule automatic acquisition system 2 for automatically acquiring the transformation rule used by the transformation unit 16. The transformation rule automatic acquisition system 2 can be incorporated in the question answering system 1 as a processing function in the question answering system 1.

In the Katz method described as the prior art, since the transformation rule transforms to a specific base expression, it is necessary to manually describe and create the transformation rule. On the other hand, the transformation rule used in the present invention is as follows:
There is no need to distinguish between a transformation from a sentence B to a sentence B and a transformation from a sentence B to a sentence A. Therefore, as will be described below, it is possible to automatically generate a deformation rule by performing processing by a computer using existing data.

A first method is to prepare a plurality of computer-readable Japanese language dictionaries, obtain knowledge about synonyms and synonymous phrases by matching explanatory sentences (definition sentences) of these dictionaries, and then transform the dictionary. How to get the rules.

The second method is a method of acquiring data on a collection of questions and answers, obtaining knowledge about synonyms and synonyms required to establish the question and answer, and acquiring a transformation rule. .

FIG. 4A is a processing flowchart of the first method for automatically acquiring the deformation rule. First, a plurality of dictionary files such as Japanese language dictionaries are prepared, and an explanatory sentence for the same headword (word) is extracted from the dictionary data read from them (step S11).

Next, the extracted explanatory texts are matched (step S12), and from the result, a part of a different expression excluding the matching part is extracted as a synonym or synonymous phrase (step S13). By assigning the synonyms or synonyms to the left and right sides of the transformation rule so that they can be converted mutually, a transformation rule is generated,
It is stored (step S14). The above processing is repeated for all headwords in the dictionary data, and when all headwords have been processed, the processing is terminated (step S15).

A specific example will be described. For example, assume that there are two dictionaries, dictionary # 1 and dictionary # 2. So "Abekobe"
It is assumed that a transformation rule is generated from the dictionary data of the headword.

Assume that the definition sentence of “Abekobe” in dictionary # 1 is “the relationship between order and position is upside down”, and the definition sentence of “Abekobe” in dictionary # 2 is “ The order, position, and relationship are upside down. " Dictionary # because it is a definition sentence of the same headword
It is considered that the definition sentence 1 and the definition sentence # 2 have the same meaning. When these definition sentences are collated and matched, the part of "the relation of order, position, etc." is very similar to the part of "order, position, relation", so this part is considered to match. From this, it can be seen that the rest of the two definitions correspond to "upside down" and "overturning." From this, the following two deformation rules are obtained.

Deformation rule 1: “turned upside down” → “turned over” Deformation rule 2: “turned over” → “turned upside down” Although the correspondence between the definition sentences of the Japanese language dictionary is used, the above-described first method can be used as long as there are other meaningful correspondences.

FIG. 4B is a processing flowchart of the second method for automatically acquiring the deformation rule. First, a question sentence and a response sentence are read from the text data of the digitized question and answer collection (step S21).

Next, the read question sentence is compared with the response sentence (step S22), and the result is extracted as a synonym or synonymous phrase (step S23).
By assigning the synonym or synonymous phrase to the left and right sides of the transformation rule so that they can be mutually converted, a transformation rule is generated and stored (step S24). The above process is repeated for all the question-and-answer sentences, and when all the question-and-answer sentences have been processed, the process ends (step S25).

A specific example will be described. Now, suppose that there are the following sentences as a question sentence and a response sentence obtained from the question and answer collection.

Question sentence: "Where is the capital of Japan?" Response sentence: "Tokyo is the capital of Japan." It can be seen that they correspond. In addition, the phrase "the capital of Japan" is easily understood. From the above knowledge, the following deformation rules can be obtained. The question particle "ka" is omitted.

Modification rule: “A is B” → “B is A” With this modification rule, in the case of the above example, it can be modified until the question sentence and the database sentence completely match. . In this way, a transformation rule that maximizes the similarity between the question sentence and the database sentence is obtained from the correspondence between phrases and the like.

The question answering system of the present invention is also useful as an information extraction technique. The information extraction is a technology for automatically extracting information such as the name of the prefecture, the location of the prefectural office, the area, the population, the main product, etc. from the existing database if the information is related to the prefecture.

In general, it is said that the current information extraction technology largely depends on the knowledge specific to the target field, and there is a problem that it is costly to transfer the system to another field.

On the other hand, a question answering system such as the present invention is a system which asks, for example, "Where is the prefectural office of Iwate prefecture?" And answers "Morioka". In the case of question answering, it is possible to ask questions in various natural languages, and there is an advantage that there is no dependence on fields and various information can be obtained freely. To use this question-answering system to extract comprehensive information about the prefectures mentioned above, ask "What kind of prefectures are there?" And then ask "Where is the prefectural office located?" , "How much is the area?" If this sequential question is programmed, the question extraction system can solve the information extraction problem without depending on the field.

As a result, for example, the following information can be automatically extracted.

(Aomori prefecture, Aomori, Xkm ² , a person, apple, ...) (Akita prefecture, Akita, Ykm ² , b person, rice, ...) (Iwate prefecture, Morioka, Zkm ² , c person, ...) , ……) ・ ・ ・ ・ ・

[0075]

As described above, according to the present invention,
Since the query sentence and the database sentence are transformed so as to increase the similarity of the matching at the time of matching the query sentence and the database sentence, the accuracy of the matching between the question sentence and the database sentence can be improved. Further, in the present invention, since the transformation operation is controlled on the basis of the similarity, any transformation rule can be used as the transformation rule, not limited to the base expression, and the transformation rule can be easily described or generated. is there.

[Brief description of the drawings]

FIG. 1 is a diagram showing a system configuration example of the present invention.

FIG. 2 is a diagram illustrating an example of a deformation rule.

FIG. 3 is a flowchart of a question answering system.

FIG. 4 is a flowchart of a transformation rule automatic acquisition system.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Question answering system 2 Modification rule automatic acquisition system 10 Question sentence input part 11 Database 12 Keyword extraction / information search part 13 Syntax analysis part 14 Matching part 15 Deformation rule storage part 16 Deformation part 17 Answer output part

────────────────────────────────────────────────── ───

[Procedure amendment]

[Submission Date] June 21, 2001 (2001.6.2)
1)

[Procedure amendment 1]

[Document name to be amended] Statement

[Correction target item name] Claims

[Correction method] Change

[Correction contents]

[Claims]

 ────────────────────────────────────────────────── ─── Continuing on the front page (72) Inventor Hitoshi Isahara 588-2 Iwaoka, Iwaoka-cho, Nishi-ku, Kobe-shi, Hyogo F-term in the Kansai Advanced Research Center, Telecommunications Research Institute, Ministry of Posts and Telecommunications 5B075 ND03 NK02 NK31 NK35 PP24 PR06 QM08 QP03

Claims (7)

[Claims] 1. A question answering system that inputs a question sentence in a natural language, generates a response sentence by comparing it with a sentence in a database, and outputs the answer sentence, storing rules for transforming the sentence into another sentence having the same content. A matching rule storage unit, a matching unit that matches an input question sentence with a sentence extracted from a database, and calculates a similarity between them; and a deformation rule storage unit based on a calculation result of the similarity by the matching unit. A rewriting unit that rewrites the question sentence and the sentence extracted from the database using the reshaping rules stored in the database and a collation unit that maximizes the similarity after repeating the processing by the collation unit and the remodeling unit. And a solution output unit that outputs the solution extracted in (1) as a response sentence. 2. A question-and-response processing method for inputting a question sentence in a natural language and generating and outputting a response sentence by comparing the sentence in a database with a sentence extracted from the database. Calculating the similarity thereof, and rewriting the question sentence and the sentence extracted from the database until the similarity becomes the highest, using a pre-stored sentence transformation rule. Outputting a solution extracted in the matching having the highest similarity as a response sentence. 3. A method for generating, using a computer, a transformation rule for transforming a sentence described in a natural language into another sentence representing the same contents, wherein the same rule is used for the same word from dictionary data read from a plurality of dictionary files. Extracting the description,
The process of matching multiple extracted explanatory sentences and extracting synonyms or synonyms from the results, and generating transformation rules from the extracted synonyms or synonym phrases to rewrite one sentence to another sentence that represents the same content Performing a transformation rule automatic acquisition process. 4. A method for using a computer to generate a transformation rule for transforming a sentence described in a natural language into another sentence representing the same content, comprising the steps of inputting a question sentence and a response sentence thereto. , Matching the input question sentence with the answer sentence, extracting synonyms or synonymous phrases from the results, and transforming the extracted synonyms or synonymous phrases into another sentence representing the same content Generating a rule. 5. A recording medium for recording a program for inputting a question sentence in a natural language and generating and outputting a response sentence by comparing it with a sentence in a database, the program being extracted from the input question sentence and the database. With the sentence
A process of calculating the similarity thereof, a process of rewriting the question sentence and the sentence extracted from the database using a pre-stored sentence transformation rule until the similarity becomes the highest, A question answer processing program recording medium characterized by recording a program for causing a computer to execute, as a response sentence, a solution extracted in the matching having the highest similarity. 6. A recording medium recording a program for generating, using a computer, a transformation rule for transforming a sentence described in a natural language into another sentence having the same content,
A process of extracting a description sentence of the same word from dictionary data read from a plurality of dictionary files, a process of matching a plurality of extracted description sentences and extracting a synonym or synonymous phrase from the result, and a process of extracting a synonym or synonym extracted A transformation rule automatic acquisition processing program recording medium characterized by recording a program for causing a computer to execute a process of generating a transformation rule for rewriting a sentence from a phrase to another sentence having the same content. 7. A recording medium recording a program for generating, using a computer, a transformation rule for transforming a sentence described in a natural language into another sentence representing the same content,
A process of inputting a question sentence and a response sentence thereto, a process of matching the input question sentence and a response sentence, and extracting a synonym or synonymous phrase from the result, and a process of extracting a synonym or synonymous phrase from the extracted synonym or synonymous phrase Processing for generating a transformation rule for rewriting the text into another sentence representing the same content,
A recording medium for automatically acquiring a deformation rule, wherein a program for causing a computer to execute the program is recorded.