JP4145776B2 - Question answering apparatus and question answering method - Google Patents

Description

  The present invention relates to a question answering device that outputs an answer to a question input by a user, and more particularly to a question answering device that finds an answer from a natural language document.

  As represented by Internet search engines, document search techniques for searching and ranking documents that match user search requests are widely used. However, the document search can satisfy search requests such as “I want to read newspaper articles about XX” and “I want to see a web page about XX”, but “Who is the president of XX?”, “ Can't answer directly to questions like "How tall is Mt. Fuji?" And "Is whales going extinct?" Since document retrieval only outputs a document or a passage in the document, the user must search for an answer from the output result.

  There is a question answering device that directly outputs an answer to the above question. For example, in response to the question “Who is the president of XX company?”, The question answering device outputs the name of the president of XX company instead of outputting a document about XX company, such as the website of XX company. For questions such as “How tall is Mt. Fuji?”, The answer to the question is output as “Mt. Fuji is 3776m above sea level”.

  In recent years, the question answering apparatus has been attracting attention as an advanced form of research such as information retrieval and information extraction, and it has become possible to change a direct answer to a user's question to some extent. For example, a question answering device that outputs an answer to a user's question text and a document (hereinafter referred to as a rationale document) that allows the user to confirm that the answer is an answer to the question Is known (see, for example, Patent Document 1).

Here, the basis document is, for example, a document of an extraction source from which an answer to a question is extracted by a question answering device.
JP 2002-132812 A

  However, since the question text to be entered is written in a natural language, it is not necessarily a simple question. For example, a complicated question such as “What is a confectionery manufacturing company headquartered in Niigata?” Can also be entered. . For such a complicated question, not all information indicating that a certain answer candidate is an answer to the question is included in one document. May be described.

  For example, it is assumed that information that company A's head office is in Niigata and information that company A is a confectionery manufacturing company are described in separate documents. In that case, simply searching for the answer candidate using the input question text as it is may cause the answer candidate (Company A in the above example) to leak from the top of the answer candidate list, and as a result, the correct answer is output. There was a problem that it was impossible. To solve such a problem that there are a plurality of information storage locations, in the field of so-called database search, a technique of dividing an inquiry (query) into a plurality of search conditions is known as a solution.

  However, in question answering devices that find answers from natural language documents, unlike database searches where the format / relationship of the data to be searched is determined in advance, the format / relationship of the data to be searched is unknown, It is not always clear when the question should be divided. Also, unlike database searches that have a low risk of incorrect answers being searched, there is a risk that correct answer candidates will be leaked from the candidate list because incorrect answer candidates are searched, and depending on the expressions included in the search target, There are cases where an inquiry with a question sentence that is not divided can extract an appropriate answer, and an inquiry with an inquiry sentence divided in the opposite direction can extract an appropriate answer.

  As described above, in the conventional question answering apparatus, there is a problem that correct answer candidates are omitted from the candidate list when the information to be answered is described in a plurality of documents.

  In view of the above problems, an object of the present invention is to provide a question answering apparatus capable of outputting appropriate answer candidates even when information to be answered is described in a distributed manner in a plurality of documents.

  In order to achieve the above object, in the present invention, in a question answering apparatus that extracts and outputs answers to an inputted question sentence from a plurality of documents, the inputted question sentence is divided into a plurality of divided question sentences. A question sentence dividing means for dividing into a divided question sentence; and analyzing the question sentence and the divided question sentence to obtain an answer candidate as the answer candidate from the plurality of documents; and the obtained answer candidate; Answer candidate storage means for storing the question sentence or the divided question sentence as a set, and answer candidates stored in the answer candidate storage means are evaluated, and an evaluation score obtained as a result of this evaluation is assigned to each set Answer candidate evaluation means, merging means for merging each answer candidate for the set obtained from the divided question sentence, and a predetermined number of the above-mentioned sets in the order of the set with the highest evaluation score of the set Answer candidates Providing question-answering apparatus characterized by having an answer output means for outputting as an answer.

  According to the present invention, it is possible to provide a question answering apparatus capable of outputting appropriate answer candidates even when information to be answered is described in a distributed manner in a plurality of documents.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
FIG. 1 is a block diagram showing a question answering apparatus according to an embodiment of the present invention.
This question answering apparatus includes an input unit 101 for inputting a question sentence, a control part 102 for controlling the entire question answering apparatus, an output part 103 for outputting an answer result, a divided question sentence generating part 111 for dividing a question sentence, and an input question sentence And question sentence list 112 for managing divided question sentences, answer candidate and answer candidate evaluation points and answer candidate holding part 113 for holding information on rationale documents, question pattern database 121, answer type determination for determining answer type of question sentence Unit 122, document search unit 123 for searching by a sentence and scoring a document, document database 124 in which natural language documents are stored, and an expression that may be a response from a search target document and assigning an answer category Answer expression extraction unit 125, expression category database 126, answer expression database 127, search result, answer type and answer table Select an answer candidate consisting answer candidate generating unit 128 that applies the evaluation point from.

  The question sentence list 112, the answer candidate holding unit 113, the question pattern database 121, the document database 124, the expression category database 126, and the answer expression database 127 include storage media such as a memory and a hard disk drive.

  In the question answering apparatus, the hardware is a computer composed of a processor, ROM, RAM, and the like, and the above functions are realized by the processor executing computer programs read from the storage medium by the computer. Good.

Hereinafter, the overall processing flow of the question answering apparatus will be described with reference to FIG.
The answer expression extraction unit 125 compares the search target document registered in the document database 124 in advance with the expression registered in the expression category database 126 and adds an answer expression that may become an answer with answer type information. The response expression character string, the appearance document ID, the appearance position, and the answer type are stored in the answer expression database 127.

  Regarding the processing of the answer expression extraction unit 125, so-called proper noun extraction and ontology tagger techniques are already known, and these techniques may be applied. Since it is not the characteristic part of this invention, detailed description is abbreviate | omitted here. In response expression extraction, processing may be performed on the surface layer representation of the target document, processing may be performed on the result of morphological analysis, or processing may be performed on the result of syntax / dependency analysis. Also good.

The user inputs a question sentence including a keyword or the like via the input unit 101.
The control unit 102 registers the question text input from the input unit 101 in the first line of the question text list 112 and outputs the question text ID to the divided question text generation unit 111, the answer type determination unit 122, and the document search unit 123. To do.

  The answer type determination unit 122 compares the question text input from the control unit 102 with the expressions registered in the question pattern database 121 to determine the answer type, and sets the answer type information as the determination result to the control unit 102. Output to. For the processing of the answer type determination unit 122, for example, Japanese Patent Laid-Open No. 2002-132812 may be used, and since it is not a main part of the present invention, detailed description thereof is omitted.

  The document search unit 123 extracts a search keyword from the question text input from the control unit 102, performs a keyword search on the document database 124 that is a search target document, and obtains a keyword search result. Then, for the document obtained by this keyword search, the document ID that has been determined in descending order of the document score, the document score of the document ID, and the matched search keyword and the information on the appearance position of the search keyword are output to the control unit 102. To do. As a method for obtaining a document score by this keyword search, various conventionally known techniques may be applied, for example, a document having a higher appearance frequency is set to a higher score. For example, Japanese Patent Laid-Open No. 2002-13281, literature (Takenobu Tokunaga; information retrieval and language processing, University of Tokyo Press, 1994), etc. may be used as the keyword search for obtaining the document score.

  The control unit 102 determines the processing result (response type information) of the answer type determination unit 122 and the processing result of the document search unit 123 (document ID, document score of the document ID, search keyword, and information on the appearance position of the search keyword. Two pieces of information (document ID and document score of the document ID) are output to the answer candidate generating unit 128.

  The answer candidate generating unit 128 searches the answer expression database 127 using the answer type information, the document ID, and the document score of the document ID input from the control unit 102 as keys, and the answer candidate character string (answer candidate expression) and the answer. Information on the appearance position of the candidate character string (answer candidate expression) is obtained.

  The answer candidate generating unit 128 calculates the distance between the answer candidate character string (answer candidate expression) and each keyword matched when searching the answer expression database 127 to obtain the answer candidate character string (answer candidate expression), An evaluation score as an answer is assigned to each answer candidate character string (answer candidate expression) by a predetermined calculation formula. For example, an evaluation score can be given by a calculation formula such as adding the reciprocal of the distance to each matched keyword. Various methods are known in the field of question answering devices for giving a high evaluation score to more likely answer candidates. Here, since it is not the main part of this invention, detailed description is abbreviate | omitted.

  As a question answering technique, for example, a technique known from Japanese Patent Application Laid-Open No. 2003-108583, Japanese Patent Application Laid-Open No. 7-93351, Japanese Patent Application Laid-Open No. 7-192020, or the like may be used.

Next, the answer candidate generating unit 128 leaves the obtained answer candidate character strings (answer candidate expressions) having the same evaluation score from the same surface character strings, and deletes others.
Next, the answer candidate generating unit 128 outputs the answer candidate character string (answer candidate expression), the question sentence ID (number of the question sentence list 112), the evaluation score, and the document ID list to the control unit 102.

  The control unit 102 receives information input from the answer candidate generation unit 128 (answer candidate character string (answer candidate expression) obtained by the answer candidate generation unit 128, its question sentence ID (question sentence ID of the question sentence list 112), (Evaluation point, document ID list) is output to the answer candidate holding unit 113. The data format stored in the answer candidate holding unit 113 will be described in detail later.

The answer candidate holding unit 113 holds this information input from the control unit 102.
The divided question sentence generation unit 111 determines whether or not the question sentence input from the control unit 102 can be divided. If the question sentence can be divided, the divided question sentence generation unit 111 generates a divided question sentence from the question sentence and generates a control part 102. Output to.

  For example, the divided question sentence generation unit 111 parses the input question sentence, determines whether or not it can be divided on the basis of the syntax tree structure, and generates a divided question sentence. Alternatively, a divided question sentence may be generated by determining whether or not division is possible using a regular expression. For example, when the result of the syntax analysis is that the subject of the question sentence has a plurality of linkage modifiers, the question sentence may be divided for each linkage modifier. In addition, when a punctuation mark is included in the middle of a question sentence, the question sentence may be divided before and after the punctuation mark to generate a question.

  2 and 3 are diagrams schematically showing an example of processing for generating a divided question sentence from the question sentence. FIG. 2 is a diagram illustrating an example in which an input question sentence is parsed and divided using a syntax tree structure as a condition. FIG. 3 is a diagram illustrating an example in which a question is generated by dividing a question sentence before and after a punctuation mark when a punctuation mark is included in the middle of the question sentence.

  The divided question sentence generation unit 111 divides the question sentence “Are you a confectionery manufacturing company in Niigata?” Input from the control unit 102, and “Which company has a head office in Niigata?” Two divided question sentences “What is the manufacturing company?” Are generated and output to the control unit 102.

Next, the flow of processing performed by the control unit 102 after the divided question text is input from the divided question text generation unit 111 will be described.
The control unit 102 registers all the divided question sentences input from the divided question sentence generation unit 111 in the question sentence list 112.
Next, the control unit 102 sends the divided question sentences registered in the question sentence list 112 to the answer type determination unit 122 and the document search unit 123 in order one by one.
The processing of the answer type determination unit 122 and the document search unit 123 is the same as that of the input question sentence in the case of the divided question sentence, and outputs each processing result to the control unit 102.
Next, the control unit 102 outputs the processing result of the answer type determination unit 122 and the processing result of the document search unit 123 to the answer candidate generation unit 128.
About the process of the answer candidate production | generation part 128, the case of a divided question sentence is the same as that of the case of an input question sentence, and a process result is output to the control part 102. FIG.
Next, the control unit 102 outputs the information input from the answer candidate generating unit 128 to the answer candidate holding unit 113.
The answer candidate holding unit 113 holds this information input from the control unit 102.
The control unit 102 repeats the same process for all divided question sentences.
When the processing of all the divided question sentences is completed, the control unit 102 reads out the answer candidate character strings (answer candidate expressions) stored in the answer candidate holding unit 113 one by one, and is common to a plurality of divided question sentences. Find the answer candidate and update its evaluation score. This evaluation point update procedure will be described in detail later.

  Next, the control unit 102 reads the answer candidate character strings (answer candidate expressions) held in the answer candidate holding unit 113 and the evaluation points of the answer candidate character strings (answer candidate expressions), and the answer candidates in descending order of the evaluation points. Sort by.

  Finally, the control unit 102 outputs, to the output unit 103, predetermined top n answer candidates among the answer candidates arranged in the order of evaluation points. In this example, the top n cases are determined in advance, but this n need not be fixed and may be variable. Further, the output may be limited to a number of evaluation points.

  The output unit 103 includes a display, and displays answer candidates input from the control unit 102 on the display. An example of this display method will be described later in detail.

The above is the outline of the processing flow of the question answering apparatus.
Hereinafter, the operation of the question answering apparatus will be described in detail by taking as an example the case where the question sentence in FIG. 2 and FIG. 3 is “What is a candy manufacturing company in Niigata?”.
FIG. 4 is a diagram illustrating a data format of the question sentence list 112.
The data of the question sentence list 112 is composed of three items of question sentence ID, type, and text. In the first line, the question sentence ID “001” is entered as “input” indicating the input question sentence, and the text is “What is the candy manufacturing company in Niigata?”. In the second line, a question sentence ID “002” is entered as “division” indicating a division question sentence, and one division question sentence “Where is the company located in Niigata?” Is entered as text. In the third line, “partition” indicating a divided question sentence is inputted as a question sentence ID “003”, and “what is the candy manufacturing company?” Which is the other divided question is inputted as text.

  The answer candidate holding unit 113 has five pieces of information including an answer candidate character string (answer candidate expression), a question sentence ID (question sentence ID of the question sentence list 112 shown in FIG. 4), an evaluation score, a document ID, and a flag (flag). Consists of. The flag is a flag attached to identify data that has already been merged with other data and becomes unnecessary.

  First, the control unit 102 stores, in the answer candidate holding unit 113, answer candidates for the input question sentence (that is, question sentence ID = 001) “What is the candy manufacturing company in Niigata?”. At this time, flag = 1.

Data of the answer candidate holding unit 113 at this time is shown in FIG. 5A.
Next, the control unit 102 stores answer candidates for one of the divided question sentences (that is, question sentence ID = 002) “Which company has the head office in Niigata?” In the answer candidate holding unit 113 (line 6 in FIG. 5B). And the answer candidate for the other divided question sentence (namely, question sentence ID = 003) “Which company is the confectionery manufacturer?” Is stored in the answer candidate holding unit 113 (line 9 in FIG. 5B). Eyes to 11th line). At this time, flag = 1.

  Data of the answer candidate holding unit 113 at this time is shown in FIG. 5B. FIG. 5B is a diagram showing a data format of the answer candidate holding unit 113, and is different from FIG. 5A in that data related to a divided question sentence (question sentence ID = 002, 003) is added.

As described above, answer candidates for all question sentences (input question sentences, divided question sentences) are stored in the answer candidate holding unit 113.
Next, the processing flow of the control unit 102 after answer candidates for all question sentences (input question sentences, divided question sentences) are stored in the answer candidate holding part 113 will be described with reference to the flowchart of FIG.

First, the number of answer candidates stored at the start of processing is substituted into a variable cnt (step S601).
The initial value 1 is substituted for the variable n and the initial value 1 is substituted for the variable i (n = 1, i = 1) (step S602).
It is determined whether or not the variable n is larger than the variable cnt (n> cnt?) (Step S611). If the determination result in this step S611 is Yes, for all data in the answer candidate holding unit 113, a plurality of question sentence IDs are arranged with the same answer candidate character string (when there are a plurality of question sentence IDs, a plurality of question sentence IDs are arranged. The presence or absence of data having (question sentence ID list) is checked, and if found, the flag of the data with the smaller evaluation score is changed to 0 (step S691), and the process ends. On the other hand, if the determination result of step S611 is No, the process proceeds to the next step S612.

  Next, with respect to all the data in the answer candidate holding unit 113, the presence or absence of data having different question sentence IDs or question sentence ID lists with the same answer candidate character string is checked. The flag of the data is changed to 0.

  If No in step S611 (if all answer candidates have not been processed yet), the answer candidate in the nth row has a question sentence ID value of 2 or more (that is, a divided question sentence), and flag = It is determined whether it is 1 (that is, data that has not been merged) (question ID value ≧ 2 and flag = 1) (step S612). If the determination result in this step S612 is Yes, it is a new answer candidate character string that has not yet been merged, so that the variable p is initialized by substituting 0 into the variable p (p = 0) (step S651). On the other hand, if No, the variable n is incremented by 1 (n = n + 1) in step S613, and the process returns to step S611.

Following step S651, the answer candidate character string in the nth row of the answer candidate holding unit 113 is substituted for the variable a (step S652).
Subsequent to step S652, the variable n + 1 is substituted into the variable m and initialized (m = n + 1) (step S653). This variable m is used as a variable indicating the comparison between the data in the nth row and the data in the answer candidate holding unit 113, and is given an initial value n + 1 for comparison with the n + 1th and subsequent rows. ing.

  Following step S653, it is determined whether the variable m is greater than the variable cnt (m> cnt?) (Step S654). That is, it is determined whether or not merging has been completed for all answer candidate character strings in the nth row. If the determination result in step S654 is Yes (all merges have been completed for the answer candidate character string on the nth row), the variable n is incremented by 1 (n = n + 1) in step S613, and the process returns to step S611. On the other hand, if the determination result in step S654 is No (no merge has been completed for the answer candidate character string in the nth row), the answer candidate character in the mth row of the answer candidate holding unit 113 is stored in the variable b. A column is substituted (step S655).

  After step S655, it is determined whether the variable a is equal to the variable b (a = b?) (Step S656). If the determination result in step S656 is Yes (a = b), the process proceeds to the next step S657. If No (a ≠ b), the variable m is incremented by 1 in step S662 (m = m + 1). The process returns to step S654.

  Next, it is determined whether or not the variable p is 0 (step S657). If the determination result in step S657 is Yes, the information on the answer candidate in the nth line of the answer candidate holding unit 113 is copied to the variable cnt + i line, and the flag in the answer candidate nth line is set to 0 (step S658). If No, the process proceeds to step S660.

After step S658, the variable i is incremented by 1 (i = i + 1) (step S659), and the process proceeds to the next step S660.
In the next step S660, the following three processes are performed. First, the evaluation score of the answer candidate in the m-th row is added to the evaluation score of the answer candidate in the variable cnt + i-th row. Second, the question sentence ID and the document ID are added to the variable cnt + i line. The third is flag = 0 in the m-th row.

After step S660, the variable p is incremented by 1 (p = p + 1) (step S661), and the process proceeds to step S662.
In short, the merge process is sequentially performed on answer candidates that have not yet been merged in order from n = 1.
Data having the same answer candidate character string for answer candidates having a question sentence ID value of 2 or more (that is, answer candidates based on divided questions) and a flag value of 1 (that is, data that has not yet been merged) Check for the presence or absence.

  If found, the top data (data in the nth row) among the duplicate data, that is, the answer candidate character string, question sentence ID, evaluation score, document of the answer candidate in the nth row ID, flag (flag = 1) is stored in the variable cnt + i line of the answer candidate holding unit 113, and the flag of the answer candidate in the nth line is set to 0.

  Then, the data appearing second and later from the top among the overlapping data is sequentially merged into the variable cnt + i line. More specifically, the evaluation score of the answer candidate in the m-th line that matches the answer candidate character string is added to the evaluation score of the answer candidate in the variable cnt + i line, the question sentence ID and the document ID are added, and the m-th line The flag of the answer candidate is set to 0.

As a result of the above processing, the answer candidate data based on the divided question is merged.
For example, when n = 6, in step S658, the question sentence ID “002”, the evaluation score “0.2”, and the document ID “0021” on the sixth answer candidate line in FIG. 5B are the 12th answer candidate line in FIG. 5C. Each is copied to the first line (variable cnt + 1 line, that is, 11 + 1 = 12th line, which is the number of pieces of information held by the answer candidate holding unit 113 at this time). The flag of the answer candidate 6 (n) line is 0, and the flag of the variable cnt + 1 line as the copy destination is 1.

  Further, for the duplication of the divided question text, a search is performed by changing the variable m. When m = 9, the evaluation score “0.2” on the answer candidate 9th line in FIG. 5B is shown in step S660. The value added to the 12th answer candidate line of 5C is “0.4”, and the question sentence ID “003” and the document ID “0311” are added to the answer candidate 12th line of FIG. 5C. . The flag in the answer candidate 9 (m) line is 0. Data of the answer candidate holding unit 113 at this time is shown in FIG. 5C.

  In short, among the duplicated data, the data in the top row is copied to an empty area (step S658), and the second and subsequent data are added to the copied row (step S660). ) Data of the answer candidate holding unit 113 at this time is shown in FIG. 5D. Thus, there is only one data with flag = 1 for each answer candidate character string.

  Next, for all the data in the answer candidate holding unit 113, data having different question sentence IDs (a question sentence ID list in which a plurality of question sentence IDs are arranged when there are a plurality of question sentence IDs) in the same answer candidate character string. The flag of the data with the smaller evaluation score is changed to 0 if it is found.

Data of the answer candidate holding unit 113 at this time is shown in FIG. 5D.
The above is the flow of the update process of the evaluation score of the answer candidate holding unit 113 by the control unit 102.
In this embodiment, the answer candidate obtained from the input question sentence and the answer candidate obtained from the divided question sentence are separately calculated without being added. However, the same effect can be obtained even if the total is calculated without distinguishing the answer candidates obtained from the input question sentence / divided question sentence.

Next, an output form to the output unit 103 will be described in detail with reference to the drawings.
FIG. 7A shows a case in which the question sentence is not divided and only the answer candidate for the input question sentence (data with question sentence ID 001 in FIG. 5D, ie, data in FIG. 5A) is output to the output unit 103. Is a display example. This example is an example in which the top three answers are output from the first place to the third place, and answer candidate character strings in the first to third answer candidates are displayed. In this way, the images are sorted and displayed in the order of evaluation points.

  FIG. 7B is a display example when an answer is output from the whole answer candidates for the input question sentence and the divided question sentence. In this example, the evaluation score on the first line of the answer candidate and the 12th line of the answer candidate is 0.4, but the answer candidate for the input question sentence is regarded as correct and is set as the first answer candidate. In the third candidate answer, the second candidate answer line with the second highest evaluation score of 0.32 is displayed. In this way, the images are sorted and displayed in the order of evaluation points.

Next, a display example in the case of displaying information indicating that it is an answer output obtained from a divided question sentence will be described using FIG.
In FIG. 8, the second-ranked answer “XX Confectionery” is the answer output obtained from the divided question text (FIG. 5D answer candidate 12th line), so that another color is displayed next to the answer output on the screen. -Or, the control unit 102 controls and displays "question division" with the shape character, so that the user can know that the second-ranked answer output is the answer output obtained from the divided question text.

  In this example, the character is displayed in a different color or shape next to the answer output. For example, for the answer output obtained from the divided question text, the character of the answer output is changed by changing the color of the answer output character. It may be possible to identify the font by changing the font and character modification from the answer output obtained from the input question sentence.

  As described above, when the answer output is obtained from the divided question sentence, the user may want to know what kind of divided question sentence is obtained. In order to cope with this, as shown in FIG. 9, in addition to the answer output of FIG. 8, the control unit 102 may control and display the divided question text so as to display it. In FIG. 9, the second-ranked answer output “XX Confectionery” is divided into two divided question sentences (question sentence ID = 002 “Which company has the head office in Niigata?”, Question sentence ID = 003 “confectionery manufacturing The answer output obtained from "What is the company?"), So that the second answer output is displayed next to and below the second answer output, so that the display corresponds to the second answer output Display in association with. At this time, the control unit 102 obtains a divided question sentence by referring to the question sentence list 112 shown in FIG. 4, and outputs and displays it on the screen corresponding to the second-ranked answer output.

  Further, there are cases where the user wants to know what document is the basis document for answer output. In order to cope with this, as shown in FIG. 10, in addition to the response output of FIG. 9, the document that is the basis of the response output is displayed in addition to the response output. Actually, the control unit 102 obtains the document ID of the document (foundation document) that is the basis of the answer output from the data of the answer candidate holding unit 113 shown in FIG. 5D, and the document database 124 via the document search unit 123. The basis document is obtained from the above, and output and displayed on the screen corresponding to the second-ranked answer output.

  As described above, in the present embodiment, even when the basis for the answer to the input question sentence is distributed among a plurality of documents, the answer candidate based on the input question sentence and the answer candidate based on the divided question sentence are acquired. These can be evaluated and answer candidates can be presented appropriately.

  In the present invention, the sum of the evaluation points is used when the answer candidates are merged. For example, the coefficients may be multiplied by a total of 0.9.

In addition, the present invention can be applied in various ways without departing from the spirit of the present invention.
Note that the present invention is not limited to the above-described embodiment as it is, and can be embodied by modifying the components without departing from the scope of the invention in the implementation stage. In addition, various inventions can be formed by appropriately combining a plurality of components disclosed in the embodiment. For example, some components may be deleted from all the components shown in the embodiment. Furthermore, constituent elements over different embodiments may be appropriately combined.

The block diagram which shows the structure of the question answering apparatus of this invention. The figure which shows an example of a question sentence division | segmentation rule (example which divided | segmented on condition of syntax tree structure). The figure which shows an example of a question sentence division | segmentation rule (example which divided | segmented the question sentence before and after the punctuation mark, and produced | generated the question). It is a figure which shows the format of the data of the question sentence memorize | stored in the question sentence list. The figure which shows the format of the data of the answer candidate which the answer candidate holding part 113 hold | maintains (example input about the input question sentence). The figure which shows the format of the data of the answer candidate which the answer candidate holding part 113 hold | maintains (example input about the input question sentence and the division | segmentation question sentence). The figure which shows the format of the data of the answer candidate which the answer candidate holding part 113 hold | maintains (example for demonstrating merge). The figure which shows the format of the data of the answer candidate which the answer candidate holding part 113 hold | maintains (example for demonstrating merge). 5 is a flowchart showing a flow of merge processing of the control unit 102. Display example of output unit 103 ((a): Display example when answer is output only from answer candidate for input question sentence. (B) When answer is output from all of answer questions for input question sentence and divided question sentence. Display example). The example of a display of the output part 103 (The example of a display in the case of displaying so that it is an answer output obtained from the division | segmentation question sentence so that identification is possible). Display example of output unit 103 (display example in the case where answer candidates obtained from divided question sentences are displayed in an identifiable manner) A display example of the output unit 103 (a display example in which a document that is the basis for answer output is displayed in addition to the answer output).

Explanation of symbols

  DESCRIPTION OF SYMBOLS 101 ... Input part, 102 ... Control part, 103 ... Output part, 111 ... Divided question sentence generation part, 112 ... Question sentence list, 113 ... Answer candidate holding part, 121 ... Question pattern database, 122 ... Answer type determination part, 123 Document retrieval unit 124 Document database 125 Answer expression extraction unit 126 Expression category database 127 Answer expression database 128 Answer candidate generation unit

Claims (4)

In a question answering apparatus that extracts and outputs answers to an inputted question sentence from a plurality of documents registered in a document database ,
A question sentence dividing means for dividing the inputted question sentence into divided question sentences which are a plurality of divided question sentences;
Retrieve a search keyword from the question sentence and the divided question sentence , perform a keyword search for the plurality of documents , search the answer expression database using the document ID and document score of the document obtained by the keyword search as a key, Answer candidate storage means for obtaining an answer candidate character string and storing the obtained answer candidate character string and the question sentence or the divided question sentence as a set;
An answer candidate evaluation unit that evaluates and calculates a distance between the answer candidate character string stored in the answer candidate storage unit and the search keyword, and assigns an evaluation score obtained as a result of the evaluation to each set;
For the set obtained from the divided question text , a merge means for merging when the answer candidate character strings are duplicated , and adding the evaluation score ;
A question answering apparatus comprising: answer output means for outputting a predetermined number of answer candidates as answers in the order of the highest evaluation score of the set obtained finally. The answer output means includes
The question answering apparatus according to claim 1, further comprising means for attaching and outputting information indicating that the answer is obtained from the divided question sentence to the answer obtained from the divided question sentence. The answer output means includes
The question answering apparatus according to claim 1, further comprising means for outputting the divided question sentence as a pair with the answer to the answer obtained from the divided question sentence. In the question answering method that extracts and outputs the answers to the input question text from multiple documents registered in the document database ,
The question sentence dividing means divides the inputted question sentence into a plurality of divided question sentences, which are divided question sentences,
A search keyword is extracted from the question sentence and the divided question sentence by the answer candidate storage means, the keyword search is performed on the plurality of documents , and the answer is obtained using the document ID and document score of the document obtained by the keyword search as keys. Search the expression database, obtain an answer candidate character string , store the obtained answer candidate character string and the question sentence or the divided question sentence as a set,
The answer candidate evaluation means evaluates and calculates the distance between the answer candidate character string stored in the answer candidate storage means and the search keyword, and assigns an evaluation score obtained as a result of this evaluation for each set,
For the set obtained from the divided question sentence by merging means, if the answer candidate character strings are duplicated , merge , add the evaluation score,
A question answering method characterized in that the answer output means outputs a predetermined number of the answer candidates as answers in the order of the highest evaluation score of the set obtained finally.

Images