JP2006091993A - Question/answering device and method and question/answering program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To quickly answer a question including relative expressions which should be answered from descriptive contents in a plurality of text data by retrieving a database. <P>SOLUTION: A database generating part 107 extracts relevant information from text data in a text information database 104 from such a point of view that the sequencing of extraction results is possible, and generates an extraction information database 108 from the information extracted from the same point of view. Thus, it is possible to acquire an answer by searching a point or range shown by relative expressions included in the question from the generated extraction information database 108. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a question answering apparatus and method and a question answering program that search a database for a question in a natural language from a user and output an answer.

  A question answering system that outputs an answer to a question in a natural language from a user usually holds an answer database, searches for an answer from the database, and outputs it to the user. Particularly in recent years, a question answering system for searching for an answer from a database storing a large amount of text data has been realized by using a technology developed in the field of text information search and text information extraction (for example, Non-patent document 1).

  In such a question answering system, for example, if there is a question “who is the president of XX company”, the database is first searched using the question. For example, if the search result includes text data such as “X Corporation has announced the appointment of Mr. △ President as President”, the text string “Mr. XX” is included in this text data. Is extracted and output as an answer to the question.

  In such a question answering apparatus, if there is a difference between the expressions described in the question and the text data to be searched, the search cannot be performed well. For example, the word “USA” is used in the question, but “US” is described in the text data to be searched.

In order to solve such a problem and the system obtains an answer, it is necessary to eliminate the difference in expression between the question and the text data. As a method for solving this problem, for example, it is considered to perform a search using synonyms and related words, or to change a question. As such a system, for example, a system is known that obtains an answer after performing transformation so that the similarity between the question and the text data to be searched is increased using a transformation rule prepared in advance (for example, JP 2002-14990 A).
JP 2002-14990 A E.Voorhees, "The TREC-8 question answering track report. In Proceedings of TREC-8, 1999.", [online], [searched on September 21, 2004], Internet <URL: http: // trec. nist.gov/data/qa/t8_qadata.html>

  Here, consider a case where a question answering system is given a question that includes relative expressions such as “before XX”, “next to XX”, “before XX”, etc. . For example, it is assumed that the question “who is the president of XX Co., Ltd. before Mr. XX” is given, and the search target database includes the following text data A, B, and C.

A: “1995, Monday, Sunday, XX Co., Ltd. announced that Mr. ◇ was appointed as President ...”
B: “January, 1997, XX Co., Ltd. announced that Mr. △ took office as President ...”
C: “Monday, Sunday, 1999, XX Co., Ltd. announced that Mr. □ was appointed president.”
In order to obtain correct answers to questions that use such relative expressions, it is necessary to compare the dates described in each text data. In the case of the above example, by comparing the dates described in the text data A, B, and C, it can be understood that the answer is “Mr. Here, it is assumed that text data when the president of XX Co., Ltd. changes is included in the database.

  However, in the conventional search using synonyms and related words and the modification of the question using the transformation rule, the processing according to the contents of the multiple text data to be searched is not performed. There is a problem that an appropriate process for obtaining an answer cannot be performed on a question including a relative expression on the axis (in the above example, on the time axis) from the viewpoint.

  In order to achieve the above object, according to the present invention, a plurality of patterns prepared in advance, axis data that can be rearranged on a certain axis, and related data associated with the data, a text information database A database generating means for generating an extracted information database composed of a plurality of extracted information data corresponding to the plurality of patterns by combining the axis data and the related data, and an input for inputting a question in a natural language from a user Means, analyzing the question and extracting a response type and a keyword of an answer to the question, a relative expression analyzing means for analyzing a relative expression with respect to the axis from the question input from the input means, Using the answer type and the keyword, a plurality of extractions in the extraction information database The extracted information data determining means for determining which extracted information data is used in the report data, the extracted information data determined by the extracted information data determining means and the analysis result by the relative expression analyzing means, There is provided a question answering device comprising answer search means for searching for an answer to the question from an extracted information database, and output means for outputting an answer searched by the answer search means.

The present invention relating to the apparatus is also established as an invention relating to a method, and the present invention relating to a method is also established as an invention relating to an apparatus.
Further, the present invention relating to an apparatus or a method has a function for causing a computer to execute a procedure corresponding to the invention (or for causing a computer to function as a means corresponding to the invention, or for a computer to have a function corresponding to the invention. It can also be realized as a program (for realizing the program), and can also be realized as a computer-readable recording medium on which the program is recorded.

  In the question answering apparatus according to the present invention, after extracting relevant information from text data from a viewpoint that can be ordered, a database is generated based on the extracted contents of information extracted from the same viewpoint. An answer is obtained by obtaining the point or range indicated by the relative expression included in the question. Accordingly, it is possible to obtain an answer by searching a database for a question including a relative expression that requires an answer to be obtained from descriptions in a plurality of text data.

  This embodiment is typically realized by a computer controlled by software. The software in this case includes programs and data, and realizes the operational effects of the present invention by physically utilizing the computer hardware. Applied. Furthermore, the specific types and configurations of hardware and software that implement the present invention, the scope of processing by software, and the like can be freely changed. Therefore, in the following description, a virtual function block diagram illustrated in a block form for each function constituting the present invention is used. Note that a program for operating a computer to implement the present invention is also an embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
(First embodiment)
FIG. 1 is a diagram showing a system configuration of a question answering apparatus according to the first embodiment of the present invention.
The question answering apparatus 100 of the present invention includes a question input unit 101 for inputting a user's question, a question processing unit 102 for processing a question to perform keyword extraction and answer type analysis, and a relative for analyzing a relative expression included in the question. An expression analysis unit 103, a text information database 104, a related information extraction unit 105 that extracts related information from the text information database 104 from the viewpoint that the extraction results can be ordered, and a plurality of pieces of extracted information extracted from the same viewpoint From the extracted information organizing unit 106 for organizing, the database generating unit 107 for generating a database from the extracted / organized information, the extracted information database 108 generated by the database generating unit 107, and the text information database 104 or the extracted information database 108 Answer information search unit that searches for information that includes answers And 09, and answer extraction unit 110 for extracting an answer to be presented to the user from the search results of answer information retrieval unit 109, and a reply output section 111 for outputting an answer to the user.

  Here, the text information database 104 is a database in which text data to be searched for answers is registered, and is generated by an index generation technique such as an inverted index method used in an existing text information search technique.

Next, the operation of the entire apparatus and the operation of each unit will be described.
The operation of the entire apparatus is divided into a database generation phase for generating an extraction information database for answers in advance and a question answer phase for responding to questions.
In the following, first, the operation in the database generation phase will be described with reference to the flowchart of FIG. 2, and then the operation in the question answering phase will be described with reference to the flowchart of FIG.

(Database generation phase)
In the database generation phase, first, in step S201, the related information extraction unit 105 uses a pattern prepared in advance to search the text data registered in the text information database 104 for relevant information that matches the pattern. Extract. The pattern used for extraction is variously prepared assuming either advance what questions to be. For example, the following pattern is prepared.
Pattern P1: “@PRODUCT released on @DATE”
Pattern P2: “@PRODUCT (@DATE release)”
Here, consider a case where the text data to be extracted has the following description.
Text T1: “... ABC was released on XX month x day”
Text T2: "... CDE (released on the month and day) ..."
The related information extraction unit 105 first extracts a unique name from the input text data. The unique name from the text data is extracted by an existing technique using an extraction rule based on a unique name dictionary or pattern matching. For example, a dictionary is used for names of people and places, and an extraction rule is used for dates. From the text of the above example, it is possible to extract the product name and date as follows. Here, “@PRODUCT” represents the product name, and “@DATE” represents the date.
Text T1 ': "... @ PRODUCT has been released to @DATE"
Text T2 ': "... @ PRODUCT (@DATE released) ..."
Based on this result and the previous patterns P1 and P2, the pattern P1 and the text T1 ′ and the pattern P2 and the text T2 ′ are successfully verified.
Here, referring to the text T1 which is the original text data for the matched unique name portion for the text T1 ′,
“(○ month x day, ABC)”
Is required. Similarly, for the text T2 ′, referring to the text T2 that is the original text data for the matched unique name portion,
“(○ month △ day, CDE)”
Is required.
For example,
Pattern P3: “@COMPANY. * @ PERSON becomes the president of @COMPANY at @DATE”
(Date, company name, president name) is extracted using a regular expression pattern such as In this pattern, the name extracted by @PERSON is the president name.
For example, it is assumed that the following text data T3 and T4 exist.
Text T3: “○ × Company ... □ Mr. X □ assumed the post of President on the day of the month △ month × ...”
Text T4: “○ × company is ... △ Mr. △ appointed as president on □ month △ month × day ...”
The related information extraction unit 105 can extract a unique name from the text data T3 and T4 and extract a company name, a person name, and a date as follows. Here, “@COMPANY” represents the company name, “@PERSON” represents the name of the president, and “@DATE” represents the date.
Text T3 ': “When @COMPANY ... @PERSON became president of @DATE ..."
Text T4 ': “When @COMPANY ... @PERSON became president of @DATE ..."
(Date, company name, president name)
“(○ year △ month × day, ○ × company, Mr. ○ □)”
“(□ year △ month × day, ○ × company, ○ △ Mr.)”
Can be extracted.
In step S202, the extracted information organizing unit 106 collects a plurality of pieces of information extracted from the same viewpoint, and arranges them based on the extracted contents.
For example, the result of rearranging information on (date, product name) in the above extraction example is as shown in FIG. In FIG. 4, each extracted information is rearranged in order of time.
In addition, although information on time such as (date, product name), (date, company name, president name) is described here, for example, (price, product name), (location, building name) The same processing can be performed for other information such as. When such information is extracted, the product names can be rearranged in order of price, or the building names can be rearranged in the north-south (or east-west) direction.

  In this embodiment, a method of rearranging data and detecting a response from the rearranged data is considered to be the most natural and efficient method. Therefore, this method will be described. However, if comparison of data is possible, the method is not necessarily described. You may make it detect an answer, without rearranging.

  In step S <b> 203, the database generation unit 107 generates an extracted information database 108 for searching the extracted / organized information. Here, generate an index that searches the extracted information such as (date, product name), (date, company name, president name) etc. from each element (date, product name, company name, president name, etc.) . This can be realized by an existing database generation technique such as generating a hash table with each element as a key.

  In the above embodiment, the database generation unit 107 generates the extraction information database 108 before receiving a question. However, it is not always necessary to generate the extraction information database 108 before receiving a question. For example, the database generation unit 107 uses the keyword and the answer type obtained by analyzing the question by the question processing unit 102 after the question is input from the user in the question answering phase to be described later. The extracted information database 108 may be generated from the text information database 104.

  If it takes time to generate the extracted information database 108 due to the large amount of data in the text information database 104 at this time, the text information database 104 is searched once using the keyword extracted from the question, For example, the top tens of cases that are closely related to keywords are extracted. Then, by generating the extraction information database 108 from the extracted data, the accuracy of the answer is slightly lowered, but the time required to generate the extraction information database 108 can be shortened. As a result, the time required from receiving a question to outputting an answer can be reduced.

(Question answer phase)
Subsequently, processing in the question answering phase will be described.
(Step S301)
First, in step S301, the question input unit 101 inputs a question from a user.
(Step S302)
Next, in step S302, the question processing unit 102 extracts keywords from the question input from the question input unit 101 and analyzes the answer type. This keyword extraction can be realized using existing technology. For example, it can be realized by extracting a morpheme (for example, a general noun, proper noun, etc.) having specific part-of-speech information from a morpheme analysis result of a question using a morpheme analysis technique. Response type analysis can also be realized with existing technology. For example, a rule corresponding to each answer type can be prepared and analyzed based on a result of matching with a question. The analysis rules are as follows, for example.
Rule R1: “. * Who” → @ PERSON
Rule R2: “When is *.” → @ DATE
When the question matches the pattern “. * Who”, the rule R1 sets the matched answer type to @PERSON.
When the question matches the pattern “when is *.”, The rule R2 sets the answer type to @DATE.
(Step S303)
Next, in step S303, the relative expression analysis unit 103 analyzes the relative expressions (such as “previous” and “next”) included in the question. In this method, a relative expression pattern to be analyzed is prepared in advance, and a relative expression included in the question is analyzed based on a result of matching with the question.

For example, the following character string pattern is prepared.
Pattern B1: “Before @PERSON”: A = −1
Pattern B2: “Before the @PERSON”: A = −1
Pattern B3: “next to @PERSON”: A = + 1
Pattern B4: “Before @PERSON”: A = −2
Pattern B5: “Two generations before @PERSON”: A = −2
Pattern B6: “Before @PERSON”: A <@PERSON
Here, the right side of “:” in each pattern indicates a point or range represented by the relative expression. A = −1 indicates the previous one, that is, one generation before the extracted @PERSON.
Here, it is assumed that the next question has been input in step S301.
Q1: “Who is the previous president of Mr. △?”
First, the relative expression analysis unit 103 extracts a unique name from the question sentence. As described in the related information extraction unit 105, the extraction of the proper name from the question can be realized by an existing technique. Here, the following results are obtained by extracting the proper name for the question.
Q1 ': “Who is the previous president of @PERSON”
When pattern matching is performed on this result, matching with the pattern B2 succeeds. As a result, it is understood that the question Q1 includes a relative expression, and the original unique name can be obtained from the question Q1.
As a result, the following unique name and the point or range indicated by the relative expression are obtained (step S303).
("Mr. △", A = -1)
Next, the answer information search unit 109 searches the text information database 104 for data including the answer to the question Q1. First, the answer information search unit 109 obtains an analyzed keyword and answer type from the question processing unit 102, and obtains an analysis result from the relative expression analysis unit 103.

(No in step S304)
Here, when the relative expression is not analyzed from the question Q1 in the relative expression analysis unit 103 (No in step S304), it is determined that the question Q1 is a question that does not include the relative expression, and is analyzed from the question Q1. The answer information is searched for the text information database 104 using only the keyword and the answer type (step S306), and, for example, the top five answer candidates are presented to the user. The reason why the top five answer candidates are extracted in this way is that it is difficult to obtain correct answers from the text information database 104 using only keywords and answer types. These can be realized by using existing text information retrieval technology.

(Yes in step S304)
On the other hand, when the relative expression is analyzed from the question Q1 in the relative expression analysis unit 103 (Yes in step S304), the answer information search unit 109 searches the extracted information database 108 for the answer information based on the analysis result. (Step S305).

First, the answer information search unit 109 searches the extracted information database 108 using the unique name in the analysis result.
Next, information on the point or range indicated by the relative expression is extracted from the extracted information database 108 as answer information for this search result. For example, when the analysis result of the relative expression analysis unit 108 becomes (“Mr. △△”, A = −1), first, the extraction information database 108 is searched using the unique name “Mr. △△”. .
Here, it is assumed that information (date and time, company name, president name) as shown in FIG. 5 is registered in the extracted information database 108. In addition, in FIG. 5, it is assumed that various extracted information data such as extracted information data regarding when the product was released are registered.

  The answer information search unit 109 searches the extracted information database 108 for data using the keyword and answer type input from the question processing unit 102. For example, in the above example, the keywords: “Mr. XX”, “XX Company”, “President”, and answer type: @PERSON are used.

At this time, the response information search unit 109 searches for the extracted information data including the president name (PERSON) from the response type @PERSON and the keyword “President”. Further, from this extracted information data, the keyword “XX company” is searched. , By searching for “Mr.

“(April 1, 2003, ○ × company, Mr. ○ △)”,
…,
“(March 28, 2004, ○ × company, Mr. ○ △)”,
The search result is obtained.
Further, in the extracted information database 108, since A = −1, information immediately before this information, that is, “(March 31, 2003, XX company, Mr. □ △)”
Is searched, and "Mr.
Specifically, from the information registered in the extracted information database 108, the date and time of the information is past data from April 1, 2003, and the keywords: “XX Company”, “President”, answer By extracting the data of type: @PERSON and sorting by date and time, “(October 11, 1999, ○ × company, ◇ ×)” in FIG.
…,
“(March 29, 2003, XX company, Mr. □ △)”,
“(March 31, 2003, ○ × company, Mr. □ △)”
Get. The search is performed from the newest data among the sorted data, and is the data of the president's name that changes for the first time (first time) from “Mr.
“(March 31, 2003, ○ × company, Mr. □ △)”
Is searched, and "Mr.
When the content of the question is not the previous generation but the second generation, the data of the president name that changes the second time from “Mr.
As a modified example of the search method, data on all dates and times is searched from the extracted information database 108 using the keywords: “XX Company”, “President”, and the answer type: @PERSON. The same answer can be obtained even if the answer is the data of the president's name that changes for the first time (first time) from “Mr.

  If the search fails and no answer information is obtained from the extracted information database 108, using the same method as in step S304, where the relative expression information is not analyzed from the question, The response information is searched again for the text information database 104.

  Next, in step S307, the answer extraction unit 110 extracts an answer from the search result of the answer information search unit 109. This is possible with the technology used in existing question answering systems. The same type of information as the answer type is extracted from the search result by proper name extraction, and an answer that is considered appropriate can be selected based on the distance between the extracted information and each keyword in the question Q1. The unique name can be extracted in the same manner as the processing in the related information extraction unit 105.

Finally, in step S308, the answer output unit 111 outputs the answer obtained by the answer extraction unit 110 to the user.
In this embodiment, as an example, a case where a relative expression related to the date (@DATE) is included in the question has been described. However, the present invention is not limited to relative expressions on the time axis such as date. For example, you can extract information such as (price, product name) and answer the question “What is the cheapest ○ ◇ next to ○ □”, or extract information such as (location, building name) It is also possible to answer the question “What is △ in the north of XX”. If the information extracted in this way can be rearranged in order, it can be similarly applied to information extracted from an arbitrary viewpoint.

  As described above, according to the present invention, it is possible to obtain an answer by searching a database for a question including a relative expression that requires an answer to be obtained from descriptions in a plurality of text data.

  It should be noted that the present invention is an answer within a range that can be understood from the range stored in the text information database 104. For example, when the text information related to the questioned content is not stored at all in the text information database 104, the correct answer is given. Is not derived. However, this is within the limits of the error in the present invention.

(Second Embodiment)
FIG. 6 is a diagram showing a system configuration of a question answering apparatus according to the second embodiment of the present invention.
The question answering apparatus according to this embodiment can further include data input means 601 for inputting external data, and can use external data 602 such as an external database for generating the extraction information database 108. In this system, external data input from the data input unit 601 is sent to the related information extraction unit 105 and used to generate the extraction information database 108. Other configurations and operations are the same as those in the first embodiment, and a description thereof will be omitted.

  As described above, according to the present invention, it is possible to obtain an answer by searching a database for a question including a relative expression that requires an answer to be obtained from descriptions in a plurality of text data.

The figure which shows the system structural example of a question answering apparatus. The figure which shows the process in a database production | generation phase. The figure which shows the process in a question response phase. The figure which shows the arrangement | sequence result of extraction information. The figure which shows the search of the extraction information database. The figure which shows the other system configuration example of a question answering apparatus.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 100 ... Question answering apparatus, 101 ... Question input part, 102 ... Question processing part, 103 ... Relative expression analysis part, 104 ... Text information database, 105 ... Related information extraction part, 106 ... Extraction information arrangement part, 107 ... Database generation part , 108 ... extraction information database, 109 ... answer information search unit, 110 ... answer extraction unit, 111 ... answer output unit.

Claims (4)

Using a plurality of patterns prepared in advance, axis data that can be rearranged on a certain axis and related data associated with the data are extracted from a text information database, and the axis data and the related data are grouped Database generating means for generating an extraction information database comprising a plurality of extraction information data corresponding to the plurality of patterns;
An input means for inputting natural language questions from users;
A question analysis means for analyzing the question and extracting an answer type and a keyword of an answer to the question;
Relative expression analysis means for analyzing a relative expression for the axis from the question input from the input means;
Extraction information data determination means for determining which extraction information data to use among a plurality of extraction information data in the extraction information database using the answer type and the keyword;
Using the extracted information data determined by the extracted information data determining means and the analysis result by the relative expression analyzing means, an answer searching means for searching for an answer to the question from the extracted information database;
A question answering apparatus comprising output means for outputting an answer searched by the answer searching means. Furthermore, an external data input unit for inputting text information from the outside is provided.
The question answering apparatus according to claim 1, wherein the database generation unit extracts the axis data and the related data from text information input by the external data input unit. Using a plurality of patterns prepared in advance, axis data that can be rearranged on a certain axis and related data associated with the data are extracted from a text information database, and the axis data and the related data are grouped Generating an extraction information database comprising a plurality of extraction information data corresponding to the plurality of patterns;
Enter natural language questions from users,
Analyzing the question and extracting answer types and keywords for answers to the question,
Analyzing the relative expression for the axis from the question,
Using the answer type and the keyword, it is determined which extracted information data to use among a plurality of extracted information data in the extracted information database,
Using the determined extraction information data and the analysis result of the relative expression, search for an answer to the question from the extraction information database,
A question answering method characterized by outputting the retrieved answer. In a program for causing a computer to function as a question answering device,
The program is stored in the computer.
Using a plurality of patterns prepared in advance, axis data that can be rearranged on a certain axis and related data associated with the data are extracted from a text information database, and the axis data and the related data are grouped Generating an extraction information database including a plurality of extraction information data corresponding to the plurality of patterns;
Let users enter questions in natural language,
Analyzing the question and extracting the answer type and keyword of the answer to the question,
Analyzing a relative expression for the axis from the question input from the input means;
Using the answer type and the keyword, it is determined which extracted information data to use among a plurality of extracted information data in the extracted information database,
Using the determined extraction information data and the analysis result of the relative expression, the answer to the question is searched from the extraction information database,
A question answering program for outputting the retrieved answer.

Images