JP2012141910A - Information acquisition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To extract useful information from text data composed of a pair of a question and an answer.
The contents of the question and the answer are analyzed, and useful information is extracted by combining the analysis results. Specifically, an input means for inputting text data composed of a pair of a question and an answer, an information extraction means for extracting information from the text data, and an output means for outputting a result extracted by the information extraction means. The information extraction means comprises: question text analysis means for analyzing the question part text input by the input means; answer text analysis means for analyzing the answer part text input thereto; the question text analysis means; and the answer text analysis. The information acquisition apparatus includes a conforming text determination unit that performs text conformity determination from the analysis result of the unit.
[Selection] Figure 1

Description

  The present invention relates to an analysis technique for text data including a pair of a question and an answer.

  Many companies have a support service department and a call center that accept questions and requests from customers, and record and store the contents of questions and the responses and responses of the companies as text data. Many companies try to capture customer needs and product evaluations and opinions from such data and use them as clues to expand sales and develop new products. In order to obtain effective clues, a technique for analyzing text data accumulated by a support service and extracting necessary information is important.

  Patent Document 1 discloses a technique for extracting the frequency of words contained in a text as customer needs, which is obtained by using an existing text analysis technique such as morphological analysis for text data obtained by collecting customer opinions.

  Further, Patent Literature 2 discloses a technique for classifying customer opinions based on emotional axes such as joy and anger based on expression.

JP 2007-226568 A JP 2003-281161 A

  Using the method of Patent Document 1, word-based information such as words that frequently appear in text data can be obtained. However, most of the words included in inquiries sent to support services are the names of existing products, words related to product functions, or general technical terms in related fields. Therefore, there is a problem that only information that is commonplace such as information equivalent to the number of units sold or information known in the target field can be obtained.

  In addition, inquiries sent to support services and the like have a large number of negative opinions, and using the method of Patent Document 2, there is a problem that only a classification biased toward negative emotions can be performed.

  In the present invention, for text data composed of a pair of a question and an answer, the contents of each text of the question and the answer are analyzed, and the analysis results of both are combined to obtain information and emotion that cannot be obtained on a word basis. Extract information from a specific point of view rather than the axis. Specifically, an input means for inputting text data composed of a pair of a question and an answer, an information extraction means for extracting information from the text data, and an output means for outputting a result extracted by the information extraction means, The information extraction means includes: question content analysis means for analyzing the question part text input by the input means; answer content analysis means for analyzing the answer part text of the input; the question content analysis means; and the answer content The information acquisition device includes a conforming text determination unit that performs text conformity determination from the analysis result of the analysis unit.

  According to the present invention, there is an effect that useful information can be acquired from text data composed of a question and an answer.

It is a figure explaining the structure of this invention. It is a figure explaining a question answer table. It is a figure explaining the processing system of an information extraction program. It is a figure explaining the processing system of a question content analysis process. It is a figure explaining the processing method of an answer content analysis process. It is a figure explaining the processing system of a suitable text determination process. It is a figure explaining the processing system of a required partial text extraction process. It is a figure explaining a conformity text table. It is a figure explaining the example of a display of a processing result. It is a figure explaining the time series of this invention. It is a figure explaining the example of a display of a processing result. It is a figure which shows the block diagram of the process part which acquires statistical information from the accumulate | stored result.

  Embodiments of the present invention will be described below with reference to FIGS.

  FIG. 1 is a diagram for explaining the configuration of this embodiment. An apparatus 101 for realizing the present invention includes a CPU 102 that executes various processes, an input apparatus 103 that is input by a user, an output apparatus 104 that provides various information to the user, an OS 106, and a program such as an information extraction program 107. And a storage device 105 for storing a work memory 111 for storing various data and temporarily storing calculation results being processed.

  The input device 103 can be configured with devices such as a keyboard and a mouse.

  The output device 104 can be configured by a display device such as a display.

  The storage device 105 can be configured by a non-volatile memory such as SRAM or flash memory, and may be configured such that programs and invariant data are divided into ROM and variable data is divided into RAM. It can also be configured by a storage medium such as a disk.

  FIG. 2 is a diagram for explaining the configuration of data handled in this embodiment. The data can be configured as a question answer table 201. The question text 202, the answer text 203, the question flag 204, and the answer flag 205 may be elements of the table, and a necessary number of attributes 206 may be added to the elements as necessary. The question / answer table 201 is stored in the storage device 105.

  The question text 202 is a text that describes questions, opinions, requests, etc. of customers etc. sent to support services, call centers, and the like.

  The answer text 203 is a text that describes an answer, a corresponding content, etc. with respect to the question text 202.

  The question flag 104 has an initial value of 0 and is set to 1 according to the result of question content analysis described later.

  The answer flag 105 has an initial value of 0, and is set to 1 according to the result of answer content analysis described later.

  Attributes relating to questions and answers such as product names and customer types can be stored in the attribute 106 area.

  FIG. 3 is a schematic PAD showing the processing flow of the information extraction program 107. For each element registered in the question / answer data table, a question content analysis process 301, an answer content analysis process 302, and a matching text determination process 303 are performed. Each process (301, 302, 303) corresponds to a process performed by each of the question analysis module 108, the answer analysis module 109, and the matching text determination module 110, respectively. Furthermore, the necessary partial text extraction process 304 may be performed by adding a module to the information extraction program 107.

  FIG. 4 is a schematic PAD of the question content analysis process 301. This process is performed by the question analysis module 108. It is determined whether or not the question text 202 of the processing target element in the question answer table 201 satisfies a predetermined expression rule (401), and when satisfied, the question flag 204 of the target element of the question answer table 201 is set to 1. (402). The above expression rule will be described by taking as an example a case where a questioner selects a question that he / she desires for a product or service. Questions such as “I want to change the lower limit value of a variable” and “I want to execute process A and process B at the same time” are questions that indicate a desire that the questioner wants to do something or wants to do it. A technique for extracting a desired expression from a text by a modality such as “I want to do” is generally known. Using the modality as an extraction rule, it is possible to select a question that includes an expression that expresses the willingness of the questioner to do or want to do something. However, the modality alone includes questions that indicate that you want people to respond to the product, such as "I want to ask a question" or "I want to consult". To do. By using an expression rule that "includes expressions listed as extraction rules, but does not match expressions listed in exclusion rules", the question text containing the content that the questioner wants for the product Extract.

  FIG. 5 is a schematic PAD of the response content analysis process 302. This process is performed by the answer analysis module 109. It is determined whether or not the answer text 203 of the processing target element in the question answer table 201 satisfies a predetermined expression rule (501). If it is satisfied, the answer flag 205 of the target element in the question answer table 201 is set to 1. (502). The above expression rule will be described by taking as an example the case of selecting answer texts that answer that the subject matter is not supported. Expressions indicating that the question target item is not supported, such as “not supported”, “not supported”, “cannot be”, are enumerated as extraction rules. An expression indicating a past state such as “Not supported” or an expression indicating a situation under a limited condition such as “Unsupported in” can be used as an exclusion rule. By using an expression rule that includes the expressions listed as extraction rules but does not match the expressions listed in the exclusion rules, the answer text that does not support the question item is extracted. To do.

  FIG. 6 is a schematic PAD of the matching text determination process 303. This process is performed by the matching text determination module 110. It is determined whether or not both the question flag and the answer flag of the target element are 1 (601). If both are 1, the question answer pointer that is an address indicating the storage position of the target element on the storage device 105 Is stored in the matching text table (described later, FIG. 8). According to the expression rules shown as examples in the description of FIGS. 4 and 5, in response to a question, a questioner requests that the answer is not supported. Therefore, this is important information because it is a question / answer describing matters that can be expected to improve the service through support.

  As a result, information that cannot be obtained by analyzing the question and the answer independently can be obtained by a combination of both analysis results.

  FIG. 7 is a schematic PAD of the necessary part text extraction process 304. This process can be executed by adding a module to the information extraction program 107. The following processing is performed on the elements extracted in the matching text determination processing 303.

  In process 701, the text immediately before the expression constituting the expression rule used in the question content analysis process 302 is extracted from the question text of the target element. In other words, in the expression rule used in the description of FIG. 4, an expression constituting the extraction rule, for example, a text corresponding to “to” of “to want” can be set as the extraction target text. For example, from question texts such as “I want to change the lower limit value of a variable” or “I want to execute process A and process B at the same time”, “Change the lower limit value of variable” and “Execute process A and process B at the same time” respectively. Extract. The required partial text can be extracted by setting the extraction start point immediately after the phrase or conjunction and the extraction end point immediately before the extraction rule expression.

  In the process 702, the text immediately before the expression constituting the expression rule used in the answer content analysis process 303 is extracted from the answer text of the target element. That is, in the expression rule used in the description of FIG. 5, the expression immediately before the expression constituting the extraction rule, “not supported” or “not supported” can be set as the extraction target text. For example, response texts such as “Changing the lower limit value is not supported” and “Simultaneous execution of process A and process B are not supported” indicate “change of the lower limit value”, “process A and process B,” respectively. Is executed simultaneously. The extraction method can be the same as the processing 701.

  Either the process 701 or the process 702 may be performed.

  In process 703, one or both of the texts extracted in processes 701 and 702 are selected. For example, when the extracted text of the process 701 is basically used, and when the extraction result is insufficient, such as when the number of characters of the extracted result of the process 701 is extremely short, such as several characters, the extracted result of the process 702 is used. The process of doing. The accuracy can be improved by complementing the results of both the question and the answer.

  The necessary part text extraction process 304 makes it possible to confirm important matters at a glance without referring to the full text of the question and the answer. For example, when the information collected by the support service is expanded as useful information to a department in charge such as a design department, visibility is improved by presenting only a necessary text portion as a list. An output example will be described later. That is, the necessary part text extraction processing is not an essential component in the information acquisition apparatus, but by executing this processing, there is an effect that it is possible to improve the convenience of confirmation by the user.

  FIG. 8 is a diagram showing a configuration example of data for storing the processing result of FIG. The data can be organized as a matching text table 801. The matching text table 801 is stored in the storage device 105.

  The question answer pointer 804 to the element of the question answer table 201 determined to be applicable in the matching text determination processing 303 is stored. When the necessary part text extraction process 304 is performed, the extracted necessary part text is stored (802). If necessary, an attribute value 803 may be stored.

  Since the question answer table 201 can be referred to by the question answer pointer 804, the question acceptance date / time, the product name, the type of the questioner, etc. are stored in the attribute 206 of the question answer table 201, so that they are stored in the conformance text table 801. The processing results can be classified.

  By storing time information such as the date and time when a question is received or the date and time of answering, it is possible to know a change in time. Further, if the type of the questioner for each product is recorded from the product name, for example, by industry, the result for each industry can be obtained.

  FIG. 12 is a configuration diagram of a processing unit that obtains statistical information from the accumulated results obtained by the matching text determination processing 303. The statistical information acquisition module 1201 includes a similarity calculation unit 1202 and a totaling unit 1203, and can be realized as a module in the information extraction program 107 in FIG.

  The similarity calculation unit 1202 obtains the similarity based on the number of matching words included in the necessary partial text 802, and determines that the case is the same when the similarity exceeds a certain value. By storing the addresses and IDs of the elements that are the same as a list or array as the attribute value 803 of the compatible text table 801, it is possible to store information as to which element each element of the compatible text table 801 is the same as. .

  The totaling unit 1203 calculates the number of cases. A large number of cases indicates that there are many requests, so it can be recognized that the content is more important. More detailed information can be obtained by classifying the content of the case by time or product.

  FIG. 9 shows an example of the processing result display screen of FIG. An example is shown in which the results of each month are displayed by tab selection (901, 902, 903), divided into question acceptance months or answer acceptance months. The target product name 904 and the text content 905 extracted by the necessary part text extraction 304 are displayed. Since only necessary text is written in the contents 905, the display becomes simple and easy to read. As described above, the same number may be obtained and the number 906 may be displayed. In FIG. 9, the information is classified and displayed according to the reception time, but it can be classified using other information stored in the attribute 206 of the question answer table 201 such as the product name and the type of the questioner. It is also possible to increase.

  FIG. 10 is a time series diagram of this embodiment. Data including a question / answer text and other attribute values is input 1001 from the input device 103, and converted into a question / answer table 201 format by the CPU 102 and stored in the storage device 105 1002. In response to the information extraction request 1003 from the input device 103, the CPU 102 acquires data from the storage device 105 1004, performs information extraction processing by the information extraction program 107, and stores the processing result in the storage device 105 1005. When the result display request 1006 is received from the input device 103, the CPU 102 acquires result data from the storage device 105 1007 and displays 1008 on the output device 104 1008. Even if there is no result display request 1006, the result display 1008 may be a method that is performed when the information extraction process is completed upon receipt of the information extraction request 1003.

  In the outline PAD of the question content analysis process 301 in FIG. 4, an example will be described in which the expression rule of the determination process 401 is a rule for extracting an expression asking whether or not it is possible. Questions that ask whether a questioner can execute a certain matter, such as “Can the log output destination of the process T be changed?” Or “Can the output message be controlled?” Are common. Such a question can be extracted by using a modality representing possible or impossible and a modality representing question. For example, expressions such as “Can you do”, “Can you do” or “Can you do” are enumerated, and including them is used as an expression rule.

  In contrast to the expression rule of the determination process 401 of the question content analysis process 301 described above, the expression rule of the outline PAD determination process 501 of the answer content analysis process 302 of FIG. To do. This expression rule can also be created by making an extraction rule using a modality expressing the impossibility. That is, expressions that indicate impossibility such as “cannot be performed” and “impossible” are listed, and the inclusion of these expressions is used as an expression rule.

  From the results of the question content analysis process 301 and the answer content analysis process 302 described above, when the matching text determination process 303 is performed, a question-answer pair that cannot be answered to a question asking about possibility is obtained.

  Further, the necessary partial text extraction process 304 is performed in the same manner as in the first embodiment, the result is stored in the matching text table 801, and the stored data is aggregated to obtain the number of the same case. The matter with many cases is a matter that has not been realized even though there are many needs, so the convenience of the user can be improved by making this matter possible.

  The expression rule of the determination process 401 of the question content analysis process 301 in FIG. 4 is the same as that in the second embodiment, and the expression rule of the determination process 501 is an expression that presents an alternative measure in the schematic PAD of the answer content analysis process 302 in FIG. A description will be given of an example in which the rule is used to extract. For example, “Command A cannot change variable X. However, if you change the value of X in the definition statement of file B, you can change variable X to another value.” If you can't do it the way the questioner asked you, but if you can achieve the goal by using another method, you can use the function G instead. " It happens often. Summarizing the cases where such alternatives are answered, the response time can be shortened because the same type of question can be answered immediately without searching from the beginning. In addition, when there are many similar questions, it is possible to find a possibility that the method of asking questions is more general, and it is also effective as an opportunity to study specification changes that can be realized by that method.

The expression rule of the determination process 501 is to enumerate expressions that present alternatives, such as “However, if you can ...”, “Please, instead ...”, and include the enumerated expressions. By doing so, you can extract answers that present alternatives.
In the necessary part text extraction process 304 of FIG. 7, a process of extracting text representing the contents of the alternative may be performed. It can be extracted by creating an extraction rule such as text between “how” and “if”, or between “instead” and “please”. The extraction result can be stored in the attribute column 803 of the matching text table 801.

  FIG. 11 is a display example of questions and alternative measures. A list of product names 1101, question contents 1102, and alternative measures 1103 is displayed. It is also possible to increase the number of display items according to the information stored in the attribute column 206 of the question / answer table 201. If a question is included that is similar to a new question, the answer time can be reduced because the alternatives are readily known.

  The present invention can be applied to, for example, text data analysis of questions from customers and the like collected by the telephone, e-mail, documents, etc. to support services and call centers, and collected on the Web.

  DESCRIPTION OF SYMBOLS 101 Information extraction device, 102 CPU, 103 Input device, 104 Output device, 105 Storage device, 106 OS, 107 Information extraction program, 108 Question analysis module, 109 Answer analysis module, 110 Conformance text determination module, 111 Work memory, 201 Question Answer table, 202 Question text, 203 Answer text, 204 Question flag, 205 Answer flag, 206 Attribute, 301 Question content analysis processing, 302 Answer content analysis processing, 303 Conformance text determination processing, 304 Necessary partial text extraction processing, 801 Conformance text Table, 802 Required text, 803 attribute, 804 Question answer pointer.

Claims (9)

An input means for inputting text data composed of a pair of question text and answer text;
A question content analysis means for analyzing the question text;
An answer content analyzing means for analyzing the answer text;
A conformance text determination means for performing text conformity determination from the analysis results of the question content analysis means and the answer content analysis means;
And an output means for outputting a result determined by the matching text determination means. Furthermore, it has a necessary partial text extracting means for extracting a necessary partial text from the text data that is adapted by the matching text determination means,
The information acquisition apparatus according to claim 1, wherein the extracted result is output as the determination result.   The information acquisition apparatus according to claim 1, wherein time information is added to the determination result, and the time information is output together with the determination result.   2. The information acquisition apparatus according to claim 1, further comprising storage means for storing a table storing attributes belonging to the question text.   The information acquisition apparatus according to claim 4, wherein the determination result is classified and output based on the attribute.   3. The information acquisition apparatus according to claim 2, wherein the necessary part text extracting unit extracts the necessary part text based on a predetermined extraction rule for the question text and / or the answer text. Furthermore, from a plurality of the necessary partial texts, a similarity calculation unit for obtaining a similarity according to the degree of coincidence of words included in the necessary partial text;
The information acquisition apparatus according to claim 2, further comprising: a totaling unit that tabulates the cases that are recognized as the result of obtaining the similarity.   The information acquisition apparatus according to claim 1, wherein the matching text determination unit extracts a pair of the question text asking for possibility and the answer text indicating impossibility.   The information acquisition apparatus according to claim 1, wherein the matching text determination unit extracts a pair of the question text and the answer text indicating an alternative to the question text.

Images