JP4461738B2 - Question answering apparatus and method - Google Patents

Description

  The present invention relates to a question answering technique for outputting a text including an answer to an inputted question.

  Bulletin boards and mailing lists have some common purpose, and are often used by asking questions related to that purpose and obtaining answers. For example, a bulletin board or mailing list is used for the purpose of an SE in charge of a product in the company to ask questions regarding the installation of products to customers, the contents of inquiries from customers, and to share the results. May be used. The accumulation of messages on such bulletin boards and mailing lists includes questions on how to deal with various troubles that have occurred so far, questions on how to respond to inquiries from customers, and answers to them. In the case of an SE (system engineer) who is in charge of the same product but with a different customer or an SE in charge of a similar product, similar troubles often occur. However, even if a new question is asked each time, it is cumbersome to repeatedly give the same answer as in the past, so it may not be possible to obtain an appropriate answer from experienced people or people with knowledge. Many. If you can find the same or similar questions and answers from past messages appropriately, you can use the past accumulation to solve the problem or ask only the different points to make the respondents take time. Without getting answers, and finding the right person who seems to know the answers. However, it takes a lot of work to look through all past messages and search for questions. In addition, even if a trouble has never occurred in that product, it is possible that another product with a similar function or operating environment might have the same trouble. In such a case, you are usually involved. Finding similar questions from previous messages about other products that don't have much more work.

  A method has been proposed in which message accumulation is organized, and when typical data describing problems and countermeasures for them is created, an appropriate countermeasure is searched from there (Patent Document 1). However, it is difficult to manually organize a large number of messages, extract common problems, and collect similar problems. Although automatic creation of an FAQ list has been attempted, it is difficult to extract with sufficient accuracy in practice at present. Also, in the case of product troubles such as those mentioned above, similar problems may be concentrated in a short period of time, such as when a product is released, and it is often not possible to make a problem database etc. over time. .

  On the other hand, a search function is provided in a bulletin board system or a mail system, and a message search using a general text search technique such as a keyword search is realized. In such a method, each message is searched as one equivalent text, so that a message that is not a question or an answer or a part that is not related to a question or an answer matches a keyword.

  Also, a question answering system that searches for and outputs answers from a large number of documents in response to a question input by a user has been proposed (Patent Document 2). Even when a message is registered in such a system, regardless of whether the message is a question message or an answer message, text including a sentence similar to the question sentence is searched, and the answer is extracted from the retrieved text. In addition, the current question answering system is intended for those that can answer with a short noun phrase, such as “what”, “who”, “when”, etc. and have a clear answer. On the other hand, in the case of dealing with troubles and inquiries as described above, there are many questions about procedures and methods that cannot be answered with a single word such as `` how '' and `` why ''. Have difficulty.

Furthermore, in the case of such questions, answers are often different depending on the environment of various customers such as operating OS (operating system) and other products used together, and if these conditions are different, it will not be helpful at all. Or a large number of messages with different conditions are searched, and if you do not read all of them, you will not be able to find an appropriate message. However, when a person inputs a question into a search system or a question answering system, there is a tendency that a simple question sentence is first inserted and detailed information is not input. Also, in situations where the questioner does not know what conditions are related to the problem, it is difficult to enter information such as the environment and background. Therefore, even if a detailed search such as a related document search is performed and a message having a content similar to the question sentence is searched, the necessary information is not included because the information necessary for the original question sentence is not sufficiently included. It is not easy for the user to obtain an appropriate answer because it is not searched or a large number of messages are searched.
JP 2002-366556 A JP-A-2002-132911

  The present invention has been made in view of these points, and one of the objects of the present invention is to provide a large amount of unorganized messages by providing appropriate hint information to a user's ambiguous question. It is to provide a question answering technique capable of finding and presenting a question message and a response message suitable for the user's intention from the text.

  According to one aspect of the present invention, in order to solve the above problem, referring to the specific example of FIG. 1, the question answering apparatus associates a message expressed in text with a response message for message information. A message database 1 registered by the user; a question sentence input means 2 for a user to input a question sentence; a text analysis means 3 for analyzing a message in the message database and an input question sentence input from the question sentence input means; and a message database Question message extracting means 5 for extracting a message including a sentence representing a question from the message in 1; Searching for one or more messages similar to the input question sentence from the question messages extracted by the question message extracting means Similar question message search means 8; similarities searched by similar question message search means Hint information extracting means 10 for extracting a dependency relation between a phrase included in the input question sentence and a phrase not included from the response message of the message and the similar question message; and a dependency relation extracted by the hint information extraction means And the hint information feedback means 11 for receiving at least one instruction of a relationship that matches the user's intention and a relationship that does not match the user's intention; and scoring the response message based on the instruction received by the hint information feedback means Response message scoring means 12; and response message output means 13 for outputting a response message based on the scoring result of the response message scoring means are provided.

  In this configuration, since the question message similar to the input question sentence is detected from the already obtained question message group, and the corresponding response message is extracted and scored, an appropriate response message is extracted from the past message. You can narrow down.

  It is possible to provide hint information to the user by paying attention to the dependency relationship between words included in the input question sentence and words not included. Many past question messages and response messages correspond to environments that are slightly different from the input question text. Considering the dependency relationship of words not included in the input question text, for example, messages that match the user's intention Can be extracted.

  In some cases, the relationship between words of only one of the existing question message and response message may be used as hint information. A dependency relationship between words included in the input question sentence may be added to the hint information.

  Moreover, you may score the response message corresponding to the question message obtained by searching using grammatical information or knowledge information.

  The message database used is, for example, an electronic bulletin board or a mailing list, but is not limited thereto. It is sufficient that the response message is associated with the message. Of course, a message other than a question or a response to the question may be included. The association may be anything that can be identified. It can be an indentation representing a thread. The message database is a means for storing and holding messages, and is not limited to the phrase “message database”, but may be specified using other phrases.

  The past question messages are preferably extracted in advance, but some labels may be attached so that they can be easily retrieved from the message database. At the time of searching for similar question messages, the question messages themselves may be extracted.

  According to another aspect of the present invention, a question answering apparatus: a message database in which a message expressed in text and a response message corresponding to the message are registered in association with each other; a question sentence input for a user to input a question sentence A similar question message search means for searching for one or more messages similar to the input question sentence from question messages extracted as messages including a sentence representing a question from the messages in the message database; Means for scoring a response message associated with the similar question message searched by the similar question message search means; and response message output means for outputting a response message based on the score are provided.

  Even in this configuration, since the question message similar to the input question sentence is detected from the already obtained question message group, and the response message corresponding to this is extracted and scored, the appropriate response message is determined from the past message. You can narrow down.

  The present invention can be realized not only as an apparatus or a system but also as a method. Of course, a part of the invention can be configured as software. Of course, software products used to cause a computer to execute such software are also included in the technical scope of the present invention.

  These and other aspects of the invention are set forth in the appended claims and will be described in detail below with reference to examples.

  According to this invention, it is possible to appropriately narrow down the response messages that match the input question text by focusing on the question messages stored in the message database. In particular, if appropriate hint information is given to a user's ambiguous question, it is possible to search and present a question message and its response message that suit the user's intention from a large amount of unorganized message text. .

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

  FIG. 1 shows a configuration of a question answering system according to an embodiment of the present invention. 1, the question answering system includes a message database 1, a question sentence input unit 2, a text analysis unit 3, a message analysis result database 4, a question message extraction unit 5, a search index creation unit 6, a search index holding unit 7, The similar question message search unit 8, the response message extraction unit 9, the hint information extraction unit 10, the hint information feedback unit 11, the response message scoring unit 12, the response message output unit 13, and the like are configured.

  In the message database 1, one or more message groups are recorded. Each message group is a message group such as one mailing list or bulletin board, and is composed of a plurality of messages composed of text and response relationships between the messages. The response relationship between messages indicates the relationship of which message is a reply message of a message and which message is a reply message of a message.

  The question sentence input part 2 receives the input of the question sentence from a user.

  The text analysis unit 3 performs text analysis on each message in the message database 1 and an input question sentence from the user. The text analysis unit 3 first performs morphological analysis on the input text, divides it into words, and gives part-of-speech information. Further, the dependency relationship between words is analyzed as necessary. Various methods can be used for dependency analysis. For example, a method based on syntax analysis or a method based on shallow analysis such as pattern matching may be used. The analysis result of each message analyzed by the text analysis unit 3 is registered in the message analysis result database 4.

  The question message extraction unit 5 determines whether each message includes a question sentence from the analysis result of the text analysis unit 3, and extracts a message including the question sentence. Whether or not a question sentence is included is determined based on a rule from the morphological analysis result of each sentence in the message, for example. This rule includes, but is not limited to, “the sentence ends with the particle“ ka ””, “the sentence ends with the particle“ ne ”+ question mark”, “including the question word and the sentence ends with“ ka ””, etc. Not. For example, a rule that uses not only the morphological analysis result but also the dependency analysis result may be used.

  The search index creation unit 6 creates a question message search index from the analysis result of the message determined to contain the question sentence by the question message extraction unit 5 and registers it in the search index holding unit 7. In the search index holding unit 7, for example, independent words in the question message and their frequencies are registered. The search index is not limited to this, and may be any method depending on the search method used by the similar question message search unit 8. Further, not the entire question message but only the sentence determined as the question sentence may be set as an index creation target.

  The similar question message search unit 8 searches the search index holding unit 7 from the analysis result of the input question sentence analyzed by the text analysis unit 3 and extracts a message including a similar question sentence. The search method, for example, extracts independent words from the input question sentence, scores each question message with the frequency of occurrence of these independent words, tf * IDF (Term Frequency / Inverse Document Frequency), and the like. Extract high messages as similar messages. The search method is not limited to this, and only a word having a specific part of speech may be extracted from the input question sentence, or a relationship between words included in the input question sentence may be used. The number of similar question messages to be extracted may be set in advance, may be set in advance on a user setting screen, or may be specified when a question is input.

  The response message extraction unit 9 extracts the reply message of each similar question message extracted by the similar question message search unit 8 from the response relationship between the messages recorded in the message database 1. Response messages can be extracted by extracting only direct reply messages of similar question messages, extracting by reply including indirect reply messages such as reply replies, and grouping reply messages below direct reply messages. Various methods such as a method of extracting as a unit are conceivable, and any method may be used.

  The hint information extraction unit 10 creates hint information to be presented to the user from the analysis result of each response message extracted by the response message extraction unit 9. First, the hint information extraction unit 10 reads the analysis result of each response message from the message analysis result database 4. Next, a dependency relationship between words included in the input question sentence and words not included is extracted. Here, the words included in the input question sentence are those that satisfy the predetermined conditions among the analysis results of the input question sentence, such as all independent words, or certain parts of speech such as nouns, verbs, adjectives, and adjective verbs. Select a word. For example, words in the input question sentence may be scored by tf * IDF or the like, and a word having a high score may be selected. Similarly, the words that are not included in the question sentence out of the words included in the response message are all self-supporting words, words with a certain part of speech, words with high scores, or the like. In addition, the dependency relationship between the words included in the input question sentence and the words not included may be selected from all the dependency relationships, or a relationship that is determined to be important in advance, for example, a case relationship between a noun and a verb. You may limit to. Furthermore, the importance of words and dependency relationships may be determined, the score of each dependency relationship may be calculated, and a certain number of high scores may be extracted as hint information. Not only when words included in the input question sentence have a dependency relationship with other words alone, but also dependency relationships of compound words including words included in the input question sentence may be extracted.

  The hint information feedback unit 11 presents the hint information extracted by the hint information extraction unit 10 to the user, and specifies the hint information that does not match the intention and the hint information that matches the intention of the user. Only hint information that matches the intention or only hint information that does not match the intention may be specified.

  The response message scoring unit 12 scores the response message based on the hint information specified by the user. For example, the hint information designated by the user as fitting to the intention is given a positive weight, the hint information designated as not meeting the intention is assigned a minus weight, and scoring is performed based on the dependency information included in each response message. . Scoring may be performed only by the user's selection, or weights may be corrected by the user's selection for the score of each dependency relationship based on the importance of the word based on tf * IDF and the importance of the relationship. For example, the score of the response message may be calculated by adding the dependency relationship scores. Further, the appearance frequency of the hint information in the response message may be added to scoring. When the selected hint information is included in the question message, this is reflected in the score of the response message. For example, a score is added to the response message of the question message including the hint information selected as ◯, and the score of the response message of the question message including the hint information selected as X is deducted.

  The response message output unit 13 outputs a response message or a set of a question message and a response message together with the score based on the scoring result of the response message scoring unit 12. As the response message to be output, a certain number having a high score may be selected. When there are a plurality of response messages in one question message, the response relationship between them may be output together. Further, the hint information selected by the user as “◯” included in each question message or response message may be output together.

  The response message display unit 14 displays the response message output from the response message output unit 13 on a display or the like. There are various display methods, such as displaying the response message or question message and response message in descending order of score, displaying only the message title and following the body by the link, displaying the message or title together with the score, etc. A method is conceivable. Further, when there are a plurality of response messages of the same question message, one thread (same question message) may be displayed together, not necessarily in the order of score. Furthermore, the hint information selected by the user included in the message as ◯ may be displayed together with the title of the message. By doing so, it is possible to assist the user in selecting an appropriate response message without reading the text.

  FIG. 2 is a flowchart showing the operation when creating the analysis result data of the message analysis result database 4 and the index of the search index holding unit 7 of the question answering system of the present invention.

  In FIG. 2, first, a message in the message database 1 is sent to the text analysis unit 3 for morphological analysis (SP1) and dependency analysis (SP2). The message subjected to the dependency analysis is registered in the message analysis result database 4 together with, for example, a message ID (SP3). On the other hand, the message subjected to morphological analysis is sent to the question message extraction unit 5 to determine whether each message includes a question sentence (SP5). The morphological analysis result of the message determined to include the question sentence is sent to the search index creation unit 6 as a question message. The search index creation unit 6 creates a search index (SP6) and registers it in the search index holding unit 7 (SP7).

  FIG. 3 is a flowchart showing an operation when a user inputs a question to the question answering system of the present invention.

  In FIG. 3, the message database 1 contains a large number of messages registered on the bulletin board related to the trouble of the software A. The analysis result data of the message analysis result database 4 and the search index of the search index holding unit 7 are as follows. It is assumed that it has already been created.

  First, the user inputs a question to the question input unit 2 (SP11). For example, assume that the input question sentence Q shown in FIG. 4 is input. The input question sentence Q is morphologically analyzed by the text analysis unit 3 (SP12) and sent to the similar question message search unit 8.

  The similar question message search unit 8 first extracts “software A”, “install”, “error”, and “occurs” as keywords from the input question sentence Q (SP13). Next, query messages containing these words are searched from the search index of the search index holding unit 7 (SP14), ranked based on the appearance frequency, etc., and a predetermined number, for example, 100 messages as similar question messages. Extract (SP15) and send to response message extractor 9. For example, the similar question message QM shown in FIG. 5 is extracted as one of the similar question messages.

  The response message extraction unit 9 extracts the response message of the question message searched by the similar question message search unit 8 from the message database 1 (SP16) and sends it to the hint information extraction unit 10. For example, as shown in FIG. 6, a QM response message AM is extracted as one of the response messages.

  First, the hint information extraction unit 10 acquires the dependency analysis result of the similar question message extracted by the similar question message search unit 8 and the response message extracted by the response message extraction unit 9 from the message analysis result database 4. (SP17). Next, from the acquired dependency relationship, the keywords extracted from the input question sentence Q ("soft A", "install", "error", "occurs") and other nouns, verbs, adjectives, adjective verbs The dependency relationship is extracted (SP18). For example, the dependency relationship shown in FIG. 7 is extracted from the similar question message QM, and the dependency relationship shown in FIG. 8 is extracted from the response message AM. In the dependency relationships shown in FIG. 7 and FIG. 8, the first term represents the relationship. When the independent words are connected by particles, etc., the surface expression is used. Shown in combination of part of speech. The second term is the surface representation of the dependent word, and the third term is the surface representation of the receiving word. The hint information extracted by the hint information extraction unit 10 is sent to the hint information feedback unit 11.

  The hint information feedback unit 11 displays the hint information extracted by the hint information extraction unit 10 on, for example, a display and presents it to the user (SP19). FIG. 9A is a display example of hint information presented to the input question sentence Q. Unlike the dependency relationship examples shown in FIGS. 7 and 8, the phrase is displayed in the form of a short phrase that can be read by humans so as to be easily understood by the user. From the displayed hint information, the user determines what suits the intent of his / her question and what does not fit, and feeds back (SP20). FIG. 9B is an example of user feedback. A check box of ○ indicates a thing that suits the intention, and a check box of × indicates a thing that does not fit the intention. The user selects information that is not clearly shown in the input question sentence and that is likely to affect the answer to the question, for example, “install in OS-B”, “error code is E007”. Check the check box of ○ as the information that suits the intention, and check that the information that is not related to your situation, for example, “Install to OS-C”, “Error code is E008”, etc. Check the box. The user feedback information is sent to the response message scoring unit 12.

  The response message scoring unit 12 scores the response message extracted by the response message extraction unit 9 based on the user feedback information (SP21). Here, the importance of the dependency relationship (hint information) extracted from the response message is calculated, the dependency relationship selected by the user as ○ is multiplied by N times, and the relationship selected as × is multiplied by −N. , Add all relationship scores. When the question message includes the relationship selected by the user, a score that is similarly multiplied by N or -N is added to the score of the response message. The response message scored in this way is sent to the response message output unit 13 together with the score.

  The response message output unit 13 outputs the response messages scored by the response message scoring unit 12 for the preset number of messages together with the score (SP22).

  The output response message is displayed on the display by the response message display unit 14 (SP23). FIG. 10 is an example of display. Here, the titles of the question message and the response message are displayed so that the response relationship can be understood, and the user can see the linked text by selecting the message. If the message includes hint information selected by the user as ◯, it is displayed together.

  In the present embodiment, the message in the message database 1 is analyzed in advance by the text analysis unit 3, the question message is extracted by the question message extraction unit 5, the analysis result data in the message analysis result database 4, and the search index holding unit Although the search index of 7 is created and the response message is extracted using these when inputting a question, the present invention is not limited to this, and the text analysis of the message and the extraction of the question sentence after the question is input are described. For example, various configurations can be adopted.

  This is the end of the description of the embodiment.

  The present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the spirit of the invention. For example, although the hint information is used in the above-described example, it may be determined that the response message is an appropriate response message based on grammatical information, knowledge information, statistical information, and the like, and output it. Hint information, grammatical information, knowledge information, statistical information, and the like may be used together. In the latter case, when it is determined that the score is not definite because a sufficiently large score or the like cannot be obtained, the user's instruction may be requested by hint information.

It is a block diagram which shows the structure of the question answering system of the Example of this invention. It is a flowchart which shows the operation | movement of creation of the message analysis result database and search index in the question answering system of the above-mentioned Example. It is a flowchart which shows the operation | movement of the response message output in the question answering system of the said Example. It is a figure which shows the example of the input question sentence of the question answering system of the above-mentioned Example. It is a figure which shows the example of the similar question message of the question answering system of the above-mentioned Example. It is a figure which shows the example of the response message of the question answering system of the above-mentioned Example. It is a figure which shows the example of the dependency relation extracted from the similar question message of the question answering system of the above-mentioned Example. It is a figure which shows the example of the dependency relation extracted from the response message of the question answering system of the above-mentioned Example. The figure which shows the hint example of the question answering system of the above-mentioned Example, and the example of a display with respect to it. It is a figure which shows the example of a display of the response message of the question answering system of the said Example.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Message database 2 Question sentence input part 3 Text analysis part 4 Message analysis result database 5 Question message extraction part 6 Search index creation part 7 Search index holding part 8 Similar question message search part 9 Response message extraction part 10 Hint information extraction part 11 Hint Information Feedback Unit 12 Response Message Scoring Unit 13 Response Message Output Unit 14 Response Message Display Unit

Claims (4)

A message database in which a response request message expressed in text and a response message corresponding to the response request message are registered in association with each other;
A question sentence input means for the user to input a question sentence;
Text analysis means for analyzing the response request message and the response message in the message database and the input question text input from the question text input means;
From the response request message in the message database, the query message extracting means for extracting a query message containing a statement indicating a question,
From among the question message extracted by the query message extraction unit, a similarity query message retrieval means for retrieving one or more similar query message that is similar to the input question,
From the response message of the similarity query message search means the similarity query message and the similarity query messages retrieved, the hint information extraction for extracting a dependency relationship between the phrase does not include the words contained in the input question Means,
Hint information feedback means for displaying the dependency relationship extracted by the hint information extraction means, and accepting an instruction of at least one of a relationship that matches the user's intention and a relationship that does not match the user's intention;
Based on the instruction accepted by the hint information feedback unit, a response message scoring means for scoring the response message,
And a response message output means for outputting the response message based on a scoring result of the response message scoring means. Further comprising a response message display means, the response message display means, together with the response message and the response message or the inquiry message, question answering according to claim 1 for displaying the selected relationship if the user meets the intent apparatus. A message database in which a response request message expressed in text and a response message corresponding to the response request message are registered in association with each other;
A question sentence input means for inputting a question sentence from a user;
Said message extracted from the response request message in the database, the statement indicating the question from the including questions message, one or more similar query message that is similar to the input question sentence inputted by the question content input means Similar question message search means for searching for,
From at least one of the response message of the retrieved similar inquiry message searching section similar question message and the similar query message, hint information extraction means for extracting a dependency relationship between words with respect to words contained in the input question When,
Hint information feedback means for displaying the dependency relationship extracted by the hint information extraction means, and accepting an instruction of at least one of a relationship that matches the user's intention and a relationship that does not match the user's intention;
Based on the instruction accepted by the hint information feedback unit, a response message scoring means for scoring the response message,
And a response message output means for outputting the response message based on a scoring result of the response message scoring means. Computer
A message database in which a response request message expressed in text and a response message for the response request message are associated and registered;
Question sentence input means for the user to input a question sentence,
Text analysis means for analyzing the response request message and the response message in the message database and the input question text input from the question text input means;
A question message extracting means for extracting a question message including a sentence representing a question from the response request message in the message database;
Similar question message search means for searching for one or more similar question messages similar to the input question sentence from the question messages extracted by the question message extraction means,
Hint information extraction for extracting a dependency relationship between a word included in the input question sentence and a word not included from the similar question message searched by the similar question message search means and a response message of the similar question message means,
Hint information feedback means for displaying the dependency relationship extracted by the hint information extraction means and receiving an instruction of at least one of a relationship that matches the user's intention and a relationship that does not match the user's intention;
A response message scoring means for scoring the response message based on the instruction received by the hint information feedback means; and
Response message output means for outputting the response message based on the scoring result of the response message scoring means
Question Answering computer program to function as.

Images