JP5774459B2 - Discourse summary template creation system and discourse summary template creation program - Google Patents

Description

  The present invention relates to a technology for natural language processing, and in particular, a discourse summary template creation system and a discourse summary template for creating a template for generating a discourse summary from a result of analyzing the structure of discourse data obtained by speech recognition processing. The present invention relates to a technology that is effective when applied to a creation program.

  For example, in a call center operation, an operator or a communicator needs to perform a series of operations such as creating a response record and storing it in a response management system after receiving a call. At this time, there is a problem that it takes time for the operator or the like to create a response record based on his / her memory.

  Further, quality (content, uniformity, completeness, conciseness, etc.) varies according to the skill of each operator. In this way, the quality of the response records varies or the quality is low, so it is not possible to accurately grasp the customer's voice, or from the accumulation of response records, new technologies such as text mining are used. Problems such as inability to obtain knowledge (for example, finding potential problems and needs, selecting FAQ candidates, etc.) also occur.

  In order to solve such problems, for example, the contents of the dialogue at the reception are recorded and recorded as voice data, the data is converted into text by using voice recognition technology, and further, natural language processing technology is applied. It is also considered to automatically generate a response record by summarizing the contents. As a result, it is possible to improve the usability of the service record by ensuring the uniformity and simplicity while reducing the cost for creating the service record.

  As such a technique, for example, in Non-Patent Document 1, in voice conversation at a call center, voice data is converted into text, and rules are recursively obtained by inductive learning from past conversation contents and business daily reports. It describes a technology that automatically generates daily business reports by determining and extracting important points for unknown dialogue contents using rules and performing style conversion.

Junji Yano, Kenji Araki, “Performance evaluation of daily report generation method using inductive learning for voice conversation in call center”, Information Processing Society of Japan Research Report 2007-NL-178, March 28, 2007, p. 21-28

  When using data obtained by converting a person's discourse into text by speech recognition (hereinafter sometimes referred to as “discourse data”), the sentences in the discourse data are usually recorded in spoken language, so follow the correct grammar. Japanese language sentences and existing natural language processing technology based on written language may not be able to cope well. Moreover, the accuracy of the current speech recognition engine cannot eliminate recognition errors, and there are many noises such as typographical errors and omissions due to erroneous recognition. Therefore, it is difficult to analyze and effectively use discourse data using natural language processing.

  On the other hand, in the technology of Non-Patent Document 1, the entire dialogue document is analyzed based on the rules obtained by the inductive learning process, the important sentences are extracted, the style conversion from spoken language to written language is performed, and the daily business report is automatically generated. be able to. However, in the technique of Non-Patent Document 1, since a procedure for extracting an important sentence from an entire document and performing a style conversion is performed, as in a general automatic summary generation technique, a summary of a dialog document can be obtained. However, it is uncertain whether the summary contains the items and contents that the user needs, and the required items are not included, or the contents different from the contents that should be pointed are summarized. Cases can also arise.

  On the other hand, if the information including the structure of the summary and the conditions to which it is applied is defined as a summary template in advance and the pattern of word concatenation specified as the condition in the summary template is included in the discourse data, A method of outputting a summary sentence by applying a corresponding word in discourse data to the summary template is conceivable. As a result, it is possible to obtain a summary including items and contents set in the summary template and not including unnecessary items. In addition, since it is possible to generate a summary by extracting only the necessary parts from the discourse data content and discourse structure analysis results, even if other parts contain noise or incorrect analysis results due to noise, You can generate summaries without being affected by these.

  However, when applying the summary creation method using the summary template as described above to actual work, it is required to create and define a sufficient summary template in advance. There is a problem that it is difficult to manually create a large number of summary templates based on rules of thumb and rules of thumb.

  Therefore, an object of the present invention is to easily and accurately create a summary template used to generate a summary that includes desired items and contents and does not include unnecessary items from the result of analyzing the structure of discourse data. It is another object of the present invention to provide a discourse summary template creation system and a discourse summary template creation program that assists a user in creating such a summary template. The above and other objects and novel features of the present invention will be apparent from the description of this specification and the accompanying drawings.

  Of the inventions disclosed in this application, the outline of typical ones will be briefly described as follows.

  A discourse summary template creation system according to a representative embodiment of the present invention is used when generating a summary about a discourse based on discourse data in which the content of a discourse is converted into text, and a predetermined content in the discourse data. A list of correspondences between word concatenation patterns for identifying the parts to be included in the summary and a template of summary sentences included in the summary is specified, and When the discourse data matches, a system for creating a summary template used to generate the summary sentence from the template of the summary sentence corresponding to the matched connection pattern and to make the summary, It has the following characteristics.

  That is, the discourse summary template creation system includes a clustering unit that obtains the summary records from the summary history in which summary records that are past summaries are accumulated, classifies them into similar ones, and clusters them for each cluster. A candidate sentence selection unit that selects a representative summary record from the summary records included in a cluster and sets it as a candidate sentence; and substitutes a specific expression with a parameter for each candidate sentence as a template of the summary sentence, and the summary And a templating unit for creating a template.

  The present invention can also be applied to a discourse summary template creation program that causes a computer to function as the above-described discourse summary template creation system.

  Among the inventions disclosed in the present application, effects obtained by typical ones will be briefly described as follows.

  That is, according to a typical embodiment of the present invention, a summary used to generate a summary that includes desired items and contents and does not include unnecessary items from the result of analyzing the structure of discourse data. It is possible to easily create a template with high accuracy, or to assist the user in creating such a summary template.

It is the block diagram shown about the outline | summary of the structural example of the discourse summary template creation system which is one embodiment of this invention. It is the block diagram shown about the outline | summary of the structural example of the discourse structure analysis system in one embodiment of this invention. It is the figure shown about the example of the discourse in one embodiment of this invention, and the concept of discourse semantics. It is the figure shown about the flow of discourse data and the example of a context in one embodiment of this invention. It is the figure shown about the example of the flow analysis rule of the discourse structure analysis system in one embodiment of this invention. It is the flowchart which showed the example of the flow of the flow analysis process in the flow analysis part of the discourse structure analysis system in one embodiment of this invention. It is the figure shown about the example of the specific expression analysis rule of the discourse structure analysis system in one embodiment of this invention. It is the flowchart which showed the example of the flow of the specific expression analysis process in the specific expression analysis part of the discourse structure analysis system in one embodiment of this invention. It is the figure shown about the example of cohesion analysis in the discourse structure analysis system in one embodiment of this invention. It is the flowchart which showed the example of the flow of the cohesiveness analysis process in the cohesiveness analysis part of the discourse structure analysis system in one embodiment of this invention. It is the figure shown about the example of the discourse reduction in one embodiment of this invention. It is the flowchart which showed the example of the flow of the discourse reduction process in the discourse reduction part of the discourse structure analysis system in one embodiment of this invention. It is the figure shown about the example of the discourse summary in the discourse summary production | generation system in one embodiment of this invention. It is the figure shown about the specific example of the content of the summary template in one embodiment of this invention. It is the flowchart which showed the example of the flow of the discourse summary process in the discourse summary part of the discourse summary production | generation system in one embodiment of this invention. It is the figure shown about the example of preparation of the summary template in the discourse summary template creation system in one embodiment of this invention. It is the flowchart which showed the example of the flow of the preparation process of the summary template in one embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. Note that components having the same function are denoted by the same reference symbols throughout the drawings for describing the embodiment, and the repetitive description thereof will be omitted.

  The discourse summary template creation system according to one embodiment of the present invention uses discourse semantics (information related to the discourse structure for grasping the semantic content of the entire discourse), which is the analysis result of the discourse structure for discourse data including spoken words. This is a system for creating a summary template used to generate a summary sentence having a predefined structure based on it. As will be described later, when the discourse data includes the word concatenation pattern specified in the summary template, by outputting the summary text specified in the summary template, It is possible to realize a discourse summary generation system that generates a summary.

  Here, discourse semantics means, for example, the meaning content from character strings and morphemes in discourse data, the flow as described later for discourse data, and cohesiveness (discourse segments), proper expressions, discourse reduction It is the analysis result obtained by analyzing the discourse structure such as. This discourse semantics may be generated from discourse data by using a discourse structure analysis system, which will be described later, or may be generated using a sentence structure analysis system using other natural language processing, etc. . It is also possible to manually create a tag by tagging the discourse data. In the present embodiment, description will be made assuming that the generation is performed using a discourse structure analysis system described later.

  The discourse semantics in this embodiment do not have to be the result of strict analysis of the content and structure of Japanese sentences. For example, the discourse semantics have a level of accuracy that does not impede practical use, such as creating a response record at a call center. I just need it. In this specification, “conversation” refers to an utterance action by one or more speakers, for example, a speech at a lecture to be held by one person, an operator in a call center operation as described above, and a customer. Communication, multi-person conference calls and video conferences, and other multi-person conversations and dialogues.

  FIG. 3 is a diagram showing an example of discourse and the concept of discourse semantics in the present embodiment. The example of FIG. 3 shows an example of a discourse consisting of interaction between a call center operator and a customer. The utterance on the left side is the utterance of the operator, and the right side shows the utterance of the customer. In the present embodiment, one sentence of each speaker is called a “statement”. In addition, a group of consecutive statements for each speaker is referred to as a “block” (even if consecutive statements of the same speaker are separated in time, they become different blocks).

  In FIG. 3, “flow” is information that simply indicates the semantic content of each statement. By tracing this information, the flow of discourse can be grasped. For example, in the operator statement of FIG. 3, the first “Yes.” Indicates that the flow is “greeting”, while the next “Yes.” Indicates that the flow is “confusion”. In this way, depending on the discourse situation (context), the flow may be different even for statements of the same wording.

  In FIG. 3, “proprietary expression” is an expression that individually characterizes the content of the discourse. For example, in the discourse of exchange at the call center in FIG. 3, ID, person name, company name, telephone number, location, etc. Applicable to proper nouns. Further, the “segment” is a segmented set of common (united) contents of the discourse for each block. The information about the above “flow”, “specific expression”, and “segment” is a constituent element of “Discourse Semantics” as the analysis result of the discourse structure.

[System configuration (Discourse structure analysis system)]
FIG. 2 is a block diagram showing an outline of a configuration example of the discourse structure analyzing system according to the embodiment of the present invention. The discourse structure analysis system 3 is configured by a computer system such as a server or a PC (Personal Computer). For example, the preprocessing unit 31, the flow analysis unit 32, the morpheme analysis unit 34, and the proper expression analysis unit 35 implemented by a software program. , The cohesiveness analysis unit 33, the discourse reduction unit 36, the flow analysis rule 37, and the specific expression analysis rule 38, and the discourse data 51 in which the content of the discourse is converted into text by the speech recognition engine 4. Is input, the discourse structure is analyzed, and discourse semantics 57 is output.

  Note that the existing voice recognition engine 4 can be used. As described above, the discourse structure analysis system 3 according to the present embodiment has a feature that it is resistant to noise of the discourse data 51 (such as recognition errors by the speech recognition engine 4), but the recognition rate of the speech recognition engine 4 is higher. Needless to say, it is possible to analyze the discourse structure with higher accuracy.

  The pre-processing unit 31 of the discourse structure analysis system 3 has a function of receiving pre-conversation data 51 and performing pre-processing such as data layout conversion so that processing can be performed in other units as necessary. . Here, for example, continuous statements for each speaker are grouped into blocks, sequence numbers (IDs) indicating the order are assigned to each block and each statement, and speaker data of each block is extracted and set. Perform the process. The discourse data 51 is represented as a data structure using, for example, XML (eXtended Markup Language).

  The flow analysis unit 32 has a function of analyzing a rule-based flow based on the flow analysis rule 37 on the discourse data 51 after the preprocessing by the preprocessing unit 31 and outputting flow information 53 as an analysis result. . The flow information 53 holds the ID of each statement in the discourse data 51 and the flow information corresponding thereto, and is represented as a data structure using, for example, XML.

  This flow information 53 is included in the discourse semantics 57 as an output of the discourse structure analysis system 3 and is an essential component. That is, the discourse semantics 57 includes at least flow information 53 as a flow analysis result for the discourse data 51. The details of the flow analysis rule 37 and details of the processing in the flow analysis unit 32 will be described later.

  The morpheme analysis unit 34 performs so-called morpheme analysis for discriminating parts of speech by dividing the discourse data 51 after preprocessing by the preprocessing unit 31 into columns of morphemes (the smallest unit having meaning in the language). As a result, it has a function of outputting morpheme information 55. The morpheme information 55 holds the ID of each statement in the discourse data 51 and the information of the morpheme string corresponding to the ID, and is represented as a data structure using, for example, XML.

  Similar to the flow information 53, the morpheme information 55 is included in the discourse semantics 57 as an output of the discourse structure analysis system 3, and is also input to the specific expression analysis unit 35 and the cohesiveness analysis unit 33 described later. That is, when the specific expression analysis by the specific expression analysis unit 35 and the cohesiveness analysis (segmentation) by the cohesiveness analysis unit 33 are not performed, the morpheme analysis by the morpheme analysis unit 34 may be unnecessary. In the present embodiment, the morpheme analysis by the morpheme analysis unit 34 is indispensable in order to use the specific expression analyzed by the specific expression analysis unit 35 in the discourse summary template creation system and the discourse summary generation system described later. Note that an existing morpheme analysis engine (for example, MeCab (Japanese cloth candy) or ChaSen (tea candy)) can be used for the morpheme analysis unit 34.

  The specific expression analysis unit 35 is based on the discourse data 51 after preprocessing by the preprocessing unit 31 and the morpheme information 55 output from the morpheme analysis unit 34. And the specific expression information 56 is output as an analysis result. That is, the analysis processing in the specific expression analysis unit 35 is premised on that at least morpheme analysis by the morpheme analysis unit 34 is performed. The specific expression information 56 holds information of each statement ID in the discourse data 51 and a list of specific expressions corresponding to the statement ID, and is represented as a data structure using, for example, XML.

  Like the flow information 53, the specific expression information 56 is included in the discourse semantics 57 as an output of the discourse structure analysis system 3. However, the specific expression information 56 is used by the application using the discourse semantics 57 and the required discourse semantics. It may be unnecessary depending on accuracy. Therefore, the output of the unique expression information 56 is not essential, and the user may be able to select whether or not to output. In the present embodiment, since the specific expression information 56 is used in the discourse summary template creation system and the discourse summary generation system, which will be described later, the output of the specific expression information 56 is essential. The details of the specific expression analysis rule 38 and details of the processing in the specific expression analysis unit 35 will be described later.

  The cohesiveness analysis unit 33 uses the discourse data 51 after the preprocessing by the preprocessing unit 31, the flow information 53 output from the flow analysis unit 32, and the morpheme information 55 output from the morpheme analysis unit 34. The cohesiveness is analyzed and segmented, and the cohesiveness information 54 is output as an analysis result. That is, the analysis processing in the cohesiveness analysis unit 33 is based on the premise that at least morpheme analysis by the morpheme analysis unit 34 and flow analysis by the flow analysis unit 32 are performed. The cohesiveness information 54 holds segment ID information corresponding to each block ID in the discourse data 51, and is represented as a data structure using, for example, XML.

  This cohesiveness information 54 is included in the discourse semantics 57 as an output of the discourse structure analysis system 3, as with the unique expression information 56. However, the usage form by the application using the discourse semantics 57 and the required discourse semantics are included. Depending on the accuracy, etc., it may be unnecessary. Therefore, the output of the cohesiveness information 54 is not essential, and the user may be able to select whether or not to output. In the present embodiment, the cohesiveness information 54 may be used in a discourse summary template creation system and a discourse summary generation system, which will be described later, and the cohesiveness information 54 may be output according to the situation. Details of processing in the cohesiveness analysis unit 33 will be described later.

  The discourse reduction unit 36 deletes unnecessary expressions from the discourse data 51 based on the discourse data 51 after the preprocessing by the preprocessing unit 31 and the flow information 53 output from the flow analysis unit 32, and the meaning of the discourse is reduced. It has a function of extracting a certain part (hereinafter, the extracted discourse data 51 may be described as “simplification”) and outputting the reduction information 52 as a processing result. In other words, the reduction process in the discourse reduction unit 36 is based on the premise that at least the flow analysis unit 32 has analyzed the flow. The reduction information 52 holds, for example, the ID of each statement extracted as a reduction from the discourse data 51, the newly configured block, and information on the ID, and is represented as a data structure using, for example, XML. The

  Similar to the flow information 53, the simplified information 52 is included in the discourse semantics 57 as an output of the discourse structure analysis system 3. However, the usage form by the application that uses the discourse semantics 57 after that, and the required discourse semantics 57 are included. Depending on the accuracy, it may not be necessary. Accordingly, the output of the reduction information 52 is not essential, and the user may be able to select whether or not to output. In this embodiment, since the reduction information 52 is used in the discourse summary template creation system and the discourse summary generation system described later, the output of the reduction information 52 is essential. Details of the processing in the discourse reduction unit 36 will be described later.

  As described above, the discourse semantics 57 output as the analysis result by the discourse structure analysis system 3 includes the flow information 53, the morpheme information 55, the unique expression information 56, the cohesiveness information 54, and the reduction information 52. . At this time, a set of these data may simply be handled as discourse semantics 57, or these data may be combined into one data structure that uses XML, for example.

[System configuration (Discussion summary template creation system)]
FIG. 1 is a block diagram showing an outline of a configuration example of a discourse summary template creation system according to an embodiment of the present invention. Similar to the discourse structure analysis system 3, the discourse summary template creation system 1 is configured by a computer system such as a server or a PC. For example, a clustering unit 11, a candidate sentence selection unit 12, and a templating unit implemented by a software program 13 and so on. In addition, a summary history 61 including a database, a file, or the like that has a table or database such as the cluster information 14 and the summary template 15 and in which each summary of past discourse contents is stored, is input. This is a system for selecting a sentence to be summarized as a summary, such as a sentence including frequently appearing as, and generating a summary template 15 based on the sentence.

  Here, for example, in the present embodiment, the summary history 61 is a record in which a response record (summary record) created by each call center operator to summarize and report the response contents for the customer is accumulated. Each reception record is created from, for example, the discourse structure analysis system 3 shown in FIG. 2 and the summary 59 generated by the discourse summary generation system 2 described later in FIG. 1 based on the content of the discourse between the operator and the customer. It is also possible that the operator manually summarizes and records the contents of the response.

  Note that when a response record is created from the summary 59 generated by the discourse structure analysis system 3 and the discourse summary generation system 2, there is discourse data 51 that is the basis of the response record. The structure is analyzed by the structure analysis system 3, and the discourse semantics 57 'is generated. When the operator manually summarizes and records such data, such data does not exist. In this case, from each response record, for example, the morphological analyzer of the discourse structure analyzing system 3 shown in FIG. The discourse semantics 57 ′ such as the morpheme information 55 and the specific expression information 56 may be generated in advance by each unit such as 34 and the specific expression analysis unit 35.

  The clustering unit 11 acquires each summary (response record) included in the summary history 61, classifies it into a similar one based on its expression (content), generates a cluster, and records the information in the cluster information 14 Has the function of The clustering technique is not particularly limited, and a known technique for clustering sentences can be used as appropriate. For example, as will be described later, for each word included in each response record of the summary history 61, the importance of the word in each response record is scored to generate a word vector having the score as an element, and for each response record Clustering is performed by the k-means method or the like based on the distance between the word vectors.

  The candidate sentence selection unit 12 has a function of selecting a representative sentence, that is, a candidate sentence to be templated, from the response records included in each cluster based on the cluster information recorded in the cluster information 14. As a method for selecting candidate sentences, for example, the importance scores calculated for each word for each response record in the summary history 61 in the clustering unit 11 are summed, and the sum is the maximum (considered to be the most important). ) Is selected as a candidate sentence. In addition, taking into account the fact that it is made into a template, for example, the score is weighted based on the shortness of the sentence in the response record and the number of words (content words such as proper nouns) that can be parameterized. May be.

  The template forming unit 13 has a function of creating a summary template 15 as will be described later based on the candidate sentence selected from each cluster by the candidate sentence selecting unit 12. The created summary template 15 may be used as a draft, and a final summary template 15 may be created by a user such as an administrator editing the content, editing the content, or adding / deleting it. The created summary template 15 is, for example, taken into the discourse summary generation system 2 and used to generate a summary (response record) from discourse data.

  The discourse summary generation system 2 is configured by a computer system such as a server or a PC, similar to the discourse summary template creation system 1. For example, the discourse summary generation system 2 is created by the discourse summary unit 21 implemented by a software program and the discourse summary template creation system 1. A summary 59 about the discourse data 58 is input with the discourse data 58 having the summarized template 15 and the discourse data 58 converted into text and the discourse semantics 57 which is the analysis result of the discourse structure for the discourse data 58. Generate and output.

  Note that the discourse data 58 is equivalent to the discourse data 51 preprocessed with respect to the discourse data 51 by the preprocessing unit 31 of the discourse structure analysis system 3 described above, that is, for example, continuous statements for each speaker are collected as a block. The sequence number (ID) indicating the order is assigned to each block and each statement, and the speaker data of each block is set. Also, the discourse semantics 57 is output as a result of analysis of the discourse structure (flow, proper expression, cohesiveness, morpheme, and discourse reduction) for the discourse data 58 by the discourse structure analysis system 3 or the like described above, for example. It is a thing.

  In this embodiment, the analysis result of the discourse structure output by the discourse structure analysis system 3 is input to the discourse summary generation system 2, and the summary template 15 used in the discourse summary generation system 2 is used as the discourse summary template creation system. However, the present invention is not limited to such a configuration. For example, one or more of these systems are configured on one computer system so that the created summary template 15 is automatically used for summarization, or the summary 59 is generated from the discourse structure analysis processing. The processes up to may be continuously executable.

  The discourse summary unit 21 of the discourse summary generation system 2 uses the discourse data 58 to identify the word concatenation pattern specified in the summarization template 15 based on the unique expression information 56, the cohesiveness information 54, and the reduction information 52 of the discourse semantics 57. Is included, the summary text specified in the summary template 15 is generated and output. The summary 59 has a format in which the contents of the discourse data 58 are collected according to the configuration / pattern specified in advance as the summary template 15 and is represented as a data structure using text data, XML, or the like. Details of the summary template 15 and the summary 59 and details of the processing in the discourse summary unit 21 will be described later.

[Flow analysis]
The details of the flow analysis processing by the flow analysis unit 32 of the discourse structure analysis system 3 shown in FIG. 2 will be described below. FIG. 4 is a diagram showing an example of a flow and context in the discourse data 51. In the present embodiment, a context is set for each block of discourse. In the example of FIG. 4, the context of the first block (speaker: operator) is “start”, and the next block (speaker: customer) and the next block (speaker: operator) are both “declaration”. It has become.

  As shown in the example of FIG. 4, in this embodiment, in principle, the flow set in the last statement of the block immediately before a different speaker is set as the context of the target block. Since there is no previous block for the head block of the discourse data 51, “start” is always set as the context. By determining the context in this way, it is possible to analyze the flow along the discourse context even if the statements have the same wording.

  FIG. 5 is a diagram showing an example of the flow analysis rule 37 of the discourse structure analysis system 3. The flow analysis rule 37 is held in a file, database, or the like, and includes an expression for determining a flow in a statement (hereinafter sometimes referred to as “flow expression”) and an expression that matches the flow expression. It is a table that holds a list designated in advance with respect to correspondence with a flow pattern (hereinafter, sometimes referred to as “flow definition”) set in accordance with the context of a statement it has.

  As shown in the example of FIG. 5, the flow expression is specified by a regular expression of a character string (for example, “/Yes./” or “/ ?? /”) that serves as a clue for determining the flow. Therefore, the flow analysis unit 32 or the discourse structure analysis system 3 needs to be able to perform processing necessary for regular expressions by introducing, for example, an existing regular expression library.

  Note that the character string specified as the flow expression should preferably specify, for example, a greeting at the beginning of a sentence, a reply, a particle, a conjunction, a sentence ending expression, etc. Specify in descending order. As for the priority, for example, the complexity of the regular expression may be calculated according to a predetermined condition, and the priority may be set in descending order of the complexity. Note that the flow expression pattern to be specified may differ depending on the type of discourse (for example, exchange at a call center or regular meeting in a company) targeted by the discourse structure analysis system 3.

  For each flow expression, for example, a flow definition is specified by a label “flow”. In the flow definition, set “greeting” as the flow if the context is “start” in the statement that matches the context of the target statement and the flow that is set in the statement (for example, “/Yes./”) Specify). A plurality of flow definitions by the label “flow” can be specified according to the type of context assumed in each flow expression.

  In addition, for example, a default flow (for example, a flow expression of “/Yes./”, which is a flow to be set when the context of the target statement does not correspond to any “flow” label by the label “default”. In the statement that matches, if the context of the statement does not match any of the “flow” labels, the flow is set to “contrast”. Note that one default flow is always set in each flow expression.

  For the flow definition, it is desirable to designate a main flow pattern as a general discourse flow according to the type of discourse targeted by the discourse structure analysis system 3. For example, in the exchange at the call center, the flow definition can be specified based on a pattern such as “apology” if the customer makes a “complaint”.

  In the flow analysis rule 37, as described above, the flow expression to be specified and the content of the corresponding flow definition may differ depending on the type of discourse targeted by the discourse structure analysis system 3. A plurality of flow analysis rules 37 to be applied can be held so that the user can be switched, so that it is possible to flexibly cope with the characteristics of the discourse.

  As described above, in this embodiment, the part of speech is not specified in the flow analysis rule 37. That is, since the flow analysis unit 32 does not determine the part of speech, it is not necessary to perform the morpheme analysis by the morpheme analysis unit 34 in advance in the analysis processing by the flow analysis unit 32. In order to improve accuracy, the morpheme information 55 is used, for example, a regular expression in the flow expression of the flow analysis rule 37 is applied to a morpheme sequence (including part of speech information) of the morpheme information 55 as described later. It is also possible to perform matching under more detailed conditions as applied.

  FIG. 6 is a flowchart showing an example of the flow of the flow analysis process in the flow analysis unit 32 of the discourse structure analysis system 3. When starting the flow analysis process, the flow analysis unit 32 first starts a loop process that repeats the process for each statement of the discourse data 51 after the preprocessing by the preprocessing unit 31 (S101). When loop processing for each statement is started, loop processing for repeating processing for each flow expression of the flow analysis rule 37 is started (S102).

  When the loop processing for each flow expression is started, it is determined whether or not the target statement matches the target flow expression (S103). If there is no match, the process proceeds to the next flow expression process (S108, S102). On the other hand, if there is a match, first, for the target statement, the flow set in the last statement in the block immediately before the block to which the statement belongs is set as a context (S104). At this time, as shown in FIG. 4, when the block to which the statement belongs is the first block of the discourse data 51, “start” is set in the context.

  Next, for the context set in step S104, it is determined whether or not there is a context that matches the context specified by the “flow” label of the flow definition in the target flow expression (S105). If there is a match, the flow corresponding to the context in the matching flow definition is set as the flow of the target statement (S106). On the other hand, if there is no match, the default flow is set as the flow of the target statement (S107). Thereafter, the loop processing for each flow expression is exited.

  When the loop processing for each flow expression is completed, it is determined whether or not a flow is set in the target statement (S109). If the flow is set, nothing is performed. If the flow is not set, the “statement” flow is set in the target statement (S110), and the process proceeds to the next statement (S111). , S101). When all the loop processes for each statement are completed, the ID of each statement and the flow information corresponding thereto are output as flow information 53, and the flow analysis process is terminated.

  In this way, flow analysis that considers spoken language is used to match each statement, and by setting an appropriate flow according to the context of the matched statement, flow analysis that is less susceptible to spoken language and noise for the entire discourse data 51 It can be carried out.

[Specific expression analysis]
In the following, details of the specific expression analysis processing by the specific expression analysis unit 35 of the discourse structure analysis system 3 shown in FIG. 2 will be described. FIG. 7 is a diagram showing an example of the specific expression analysis rule 38 of the discourse structure analysis system 3. The specific expression analysis rule 38 is held in a file, a database, or the like, and is a regular expression for specifying a specific expression in a statement and a rule for specifying and extracting a specific expression from a portion that matches the regular expression. It is a table that holds a list designated in advance for correspondence with (hereinafter may be described as “specific expression rules”).

  As shown in the example of FIG. 7, the specific expression in the statement is specified by a regular expression. Therefore, the specific expression analysis unit 35 or the discourse structure analysis system 3 needs to be able to perform processing necessary for regular expressions by introducing, for example, an existing regular expression library. In the specific expression analysis, it is necessary to determine not only the regular expression for the normal character string but also the expression of the morpheme string of each statement held in the morpheme information 55. It is expanded so that the information can be judged.

  For example, in the example of FIG. 7, the range of the morpheme string is specified by “()” in each regular expression, and this part is matched with the information of the morpheme string instead of the character string. In the first regular expression example, the part of “(<f: organization> +)” indicates one or more morpheme sequences analyzed that the part of speech (f) is “organization”. (?: Of)? "Indicates a part of speech of 0 or more and 1 or less analyzed that the surface shape is" no ". The part of “(<f: person name, surname>)” indicates one morpheme string that has been analyzed with the part of speech being “person name, surname”. This regular expression matches expressions such as “AAA Shoji Suzuki” and “AAA Shoji Suzuki”. Note that the representation of the morpheme string may differ depending on the output specifications of the morpheme analysis engine in the morpheme analysis unit 34.

  At this time, in the corresponding specific expression rule, the first element in the morpheme string (the morpheme string whose part of speech is “organization”, for example, “AAA Shoji”) is specified as a specific expression having the attribute “company name”. The second element in the column (the morpheme string whose part of speech is “person name, surname”, for example “Suzuki”) is specified to be specified as a specific expression having the attribute “person name”. It should be noted that a portion (for example, “no”) enclosed by “(?:)” Is treated as not included in the element of the target morpheme string.

  In the example of the second regular expression of FIG. 7, the part of speech is specified by a normal regular expression without particular consideration, and matches the expression of a telephone number such as “03-1234-5678”. At this time, in the corresponding specific expression rule, the first element in the morpheme string (in this example, there is only one element such as “03-1234-5678”) is defined as a specific expression having the attribute “phone number”. It is specified to be specified.

  It is desirable that the character string of the unique expression specified by the regular expression is, for example, a proper noun that individually characterizes the content of the discourse or a character string made up of numerical values. In addition, for example, a character string composed of expressions (for example, complaint expression or gratitude expression) indicating the emotion of a speaker (customer, etc.) is specified and specified as a specific expression having a corresponding emotion attribute. May be. In addition, as a result of analysis, a certain character string may be included in a plurality of unique expressions.

  In the specific expression analysis rule 38, as described above, a list of correspondence between the regular expression for specifying the specific expression and the specific expression rule is specified in advance. As with the flow analysis rule 37, the normal expression to be specified is specified. As the expression and the corresponding specific expression rule, different ones are set according to the type of discourse targeted by the discourse structure analysis system 3, or the user switches the specific expression analysis rule 38 to be applied according to the type of discourse. It is possible to respond flexibly according to the characteristics of discourse.

  FIG. 8 is a flowchart showing an example of the flow of the specific expression analysis process in the specific expression analysis unit 35 of the discourse structure analysis system 3. When the specific expression analysis unit 35 starts the specific expression analysis process, it starts a loop process that repeats the process for each statement of the discourse data 51 (S201). When the loop processing for each statement is started, loop processing for repeating the processing for each regular expression for specifying the specific expression of the specific expression analysis rule 38 is started (S202).

  When the loop processing for each regular expression is started, it is determined whether or not each morpheme sequence in the morpheme information 55 corresponding to the target statement matches the target regular expression (S203). If there is no match, the processing proceeds to the next regular expression (S205, S202). On the other hand, if a match is found, a specific expression is identified and extracted by applying a specific expression rule corresponding to the target regular expression (S204), and the process proceeds to the next regular expression process (S205, S202).

  When all the loop processes for each regular expression are completed, the process proceeds to the next statement (S206, S201). When all of the loop processing for each statement is completed, the ID of each statement and the information on the list of specific expressions corresponding thereto are output as specific expression information 56, and the specific expression analysis process is ended.

  In this way, by performing matching with a regular expression considering the part of speech of the morpheme string and extracting the morpheme string, it is possible to perform a specific expression analysis considering a grammar peculiar to spoken words such as omission of a particle.

[Cohesion analysis]
The details of the cohesiveness analysis process by the cohesiveness analysis unit 33 of the discourse structure analysis system 3 shown in FIG. 2 will be described below. FIG. 9 is a diagram showing an example of cohesiveness analysis in the discourse structure analysis system 3. In the cohesiveness analysis process, first, as shown in the upper diagram, all blocks in the discourse data 51 are set as one segment.

  Thereafter, segment division processing is performed first. In the dividing process, it is determined for each statement that a flow of a specific type is set in the flow information 53, and a segment is divided in a block to which the statement belongs. Here, it is desirable that the type of the flow to be identified is a flow that is estimated to appear when the discontinuity or content changes.

  For example, in the case of multi-person discourse such as exchanges at a call center, it is often the case that the discourse progresses depending on the response of the question and the response to it, so the content is often switched in the “question” statement of the flow. It is done. Therefore, in the present embodiment, the statement of the flow “question” (hereinafter sometimes referred to as “question statement”) is determined based on the flow information 53, and as shown in the middle diagram of FIG. Each time a question statement appears, the segment shall be split.

  Thereafter, segment integration processing is performed. In the integration process, when there is a question statement corresponding to a predetermined integration condition for each segment, the segment having the question statement is integrated with the immediately preceding segment. Here, the integration condition is designation of an expression that causes the segment having the question statement to be inferred that the content continues from the previous segment (highly cohesive with the previous segment).

  Examples of the integration condition include a case where the question statement has a word (proprietary noun) common to the immediately previous question statement or a case where a directive is included. A segment having a question statement corresponding to these integration conditions is determined to be highly cohesive with the previous segment, and is integrated with the immediately preceding segment as shown in the lower diagram of FIG. It should be noted that the integration process is repeated until there are no segments that can be integrated.

  FIG. 10 is a flowchart showing an example of the flow of cohesiveness analysis processing in the cohesiveness analyzing unit 33 of the discourse structure analyzing system 3. When the cohesiveness analysis unit 33 starts cohesiveness analysis processing, first, all the blocks of the discourse data 51 are set as one segment (S301). Specifically, for example, the initial value of the segment ID is assigned to all the blocks. After that, as a segment division process, a loop process for repeating the process for each block of the discourse data 51 is started (S302).

  When the loop processing for each block is started, it is determined whether or not a question statement is included in the target block (whether or not the flow has a “question” statement) (S303). If a question statement is included, the segment is divided between the target block and the immediately preceding block (S304). Specifically, for example, the next segment ID is assigned to all blocks after the target block. Thereafter, the processing moves to the next block (S305, S302).

  When the loop processing (division processing) for each block is completed, next, loop processing that repeats the processing for each segment divided by the above-described division processing is started as segment combination processing (S306). When the loop processing for each segment is started, it is determined whether or not there is a question statement corresponding to the integration condition in the target segment (S307).

  If there is a question statement corresponding to the integration condition, the target segment and the previous segment are integrated (S308). Specifically, for example, the integration is performed by updating the segment ID of each block included in the target segment to the segment ID of the immediately preceding segment. At this time, the segment ID of each block included in each subsequent segment may be sequentially incremented. Thereafter, the process proceeds to the next segment (S309, S306). When the loop processing (integration processing) for each segment is completed, the information on the list of segment IDs corresponding to the ID of each block (including each block) is output as the cohesiveness information 54 for cohesiveness analysis. End the process.

  Accordingly, each statement in the discourse data 51 is segmented based on the flow information 53 obtained by the flow analysis, and can be used as useful information for various applications using the discourse semantics 57.

[Discussion reduction]
In the following, the details of the discourse reduction process by the discourse reduction unit 36 of the discourse structure analysis system 3 shown in FIG. 2 will be described. FIG. 11 is a diagram illustrating an example of discourse reduction. In the discourse reduction process, unnecessary expressions such as compliments and greetings are deleted from the discourse data 51 based on the flow information of each statement analyzed by the flow analysis unit 32, and other parts are extracted as essential parts in the discourse. To do.

  As shown in the example of FIG. 11, for the discourse data 51 subjected to the upper flow analysis, statements in which the flow is determined to be “greeting”, “conflict”, “standard” are deleted as unnecessary expressions, The blocks shown by the same speaker can be connected and reconstructed into one block, so that the reduction shown in the lower part can be obtained. In this embodiment, the flow that is determined to be unnecessary expression is defined as “greeting”, “conformity”, and “standard (a flow that indicates a typical expression that is used regardless of the content of the discourse)”. However, the definition of the flow that is determined as an unnecessary expression may be appropriately changed according to the type of discourse.

  FIG. 12 is a flowchart showing an example of the flow of discourse reduction processing in the discourse reduction unit 36 of the discourse structure analysis system 3. When the discourse reduction unit 36 starts the discourse reduction process, it first starts a loop process that repeats the process for each statement of the discourse data 51 (S401). When the loop processing for each statement is started, the flow information 53 is referred to for the target statement, and it is determined whether or not the flow corresponds to one of “greeting”, “consonance”, and “standard” (S402).

  If the flow does not correspond to any of them, the target statement is not an unnecessary expression, and a reduction flag is set for the target statement to identify this information (S403). In this embodiment, a reduction flag is set for a statement that is not an unnecessary expression. Conversely, when the target statement corresponds to any flow in step S402, an unnecessary expression flag is set as an unnecessary expression. You may make it set. Thereafter, the processing proceeds to the next statement (S404, S401).

  When the loop processing for each statement is completed, the block is reconfigured by a statement in which the reduction flag is set (the remaining statements from which unnecessary expressions are deleted) (S405). Here, when the blocks of the same speaker are continuous due to the deletion of unnecessary expressions, these are combined into one block (for example, the last block in the discourse data 51 in the lower part of FIG. 11). . At this time, for example, when the combined block becomes longer than a predetermined length, it may be divided into a plurality of blocks having appropriate lengths. After the block is reconfigured, the block ID may be renumbered. Thereafter, the ID of each statement for which the reduction flag is set and the information of the reconfigured block are output as the reduction information 52, and the discourse reduction process is terminated.

  As a result, the content of the discourse data 51 can be compressed as faithfully as possible without leaking the essential part to reduce the amount of data, and the analysis efficiency and utilization in various applications that use the discourse semantics 57 thereafter. Efficiency can be improved.

  Note that the flow information 53, the unique expression information 56, the cohesiveness information 54, and the reduction information 52 included in the discourse semantics 57 output by the above-described flow analysis, proper expression analysis, cohesiveness analysis, and discourse reduction processing are as follows. In the discourse structure analysis system 3, an interface may be provided that allows the user to refer to the contents and make corrections as appropriate.

[Discussion summary]
Hereinafter, the details of the discourse summary processing by the discourse summary unit 21 of the discourse summary generating system 2 shown in FIG. 1 will be described. FIG. 13 is a diagram showing an example of a discourse summary in the discourse summary generation system 2. The upper part of FIG. 13 shows an example of the result of the discourse structure analysis for the discourse data 58. For each statement of the result of simplifying the discourse data 58 based on the simplification information 52, each piece of information ( In the example of FIG. 13, a flow and a unique expression representing “amount” are shown.

  With respect to these pieces of information, the discourse summary generation system 2 according to the present embodiment uses a word concatenation pattern (“cancel” in the example of FIG. A summary template 15 designating "I want to do" is prepared in advance.

  For example, the utterance of an operator at a call center is standardized to some extent by a talk script or the like, and items that need to be communicated to customers are often predetermined. However, these utterance contents are not always exactly the same every time, and there are cases where the words are slightly different in the response, and therefore often cannot be specified by matching with a complete phrase. On the other hand, with a simple matching with a keyword, there is a high possibility that an extra utterance will be matched. Therefore, in this embodiment, an utterance to be included as a summary is specified by matching with a word connection pattern.

  The word concatenation pattern in the present embodiment is assumed to be included in an utterance having a specific content (desired request / request in this example), such as “cancel” or “want to” in the example of FIG. The keyword-like words to be specified are defined together with their appearance order.

 If the discourse content of each block matches the pattern of word concatenation specified in one of the summary templates 15, the corresponding summary text specified in the summary template 15 (in the example of FIG. ”, Etc.) are output to a predetermined position (slot) in time series, so that a summary 59 about the discourse data 58 can be obtained as shown in the lower part of FIG.

  When matching the summary template 15 with respect to each block, a target block (first block) and a subsequent second block (for example, a block pair as shown in the upper part of FIG. 13) are targeted. Match as. This is because the contents to be summarized may be specified depending on the conversational relationship between the speaker and the other party. Therefore, since blocks that contain only greetings and comments need to be excluded because they become noise, here the result of simplifying the discourse data 58 based on the simplified information 52 (blocks after excluding statements such as greetings and comments) Is the target of matching.

  FIG. 14 is a diagram showing a specific example of the contents of the summary template 15. The summary template 15 is held in a file, a database, or the like, and a word concatenation pattern for specifying a portion to be summarized (portion to be included in the summary 59) from the simplified block of discourse data 58, and an output summary. It is a table holding a list specified in advance for correspondence with a template of a sentence.

  As shown in the example of FIG. 14, each template is classified by a “template” label, and a template for a summary sentence is designated for the “template” label. The “from” label indicates that a pattern of word concatenation is matched with respect to the first block (original block) of the pair of blocks to be matched. It is also possible to target the second block (destination block) by specifying the “to” label instead of the “from” label.

  In the “seqs” label, a pattern of word concatenation to be matched in the block is designated. If there is a block containing the specified words in the specified order, the summary text specified by the “template” label is output as the summary 59. At this time, if the “slot” label is designated, the summary text is output to the corresponding slot in the summary 59. In addition, when matching the pattern of the sentence in a block and the connection pattern of a word, various existing algorithms, such as a regular expression, can be utilized, for example.

  Further, as shown in the second template in FIG. 14, it is also possible to specify a variable parameter for the template of the summary sentence specified by the “template” label. In this case, the parameter part is replaced by the corresponding specific expression in the target block at the time of output. In addition, with the “by” label, it is possible to match a block spoken by a designated speaker (for example, an operator or a customer) out of a pair of blocks to be matched.

  A summary 59 is generated by outputting a summary sentence from the template for blocks that match the word connection pattern in each template in the summary template 15. The specification content in the example of FIG. 14 is an example for the purpose of creating a response record, which is a summary 59 about exchanges at a call center, for example, depending on the type of discourse for which the summary 59 is to be created. A plurality of items may be prepared in advance and the user may be switched according to the target business. Further, an interface that allows the user to edit the contents of the summary template 15 may be provided.

  FIG. 15 is a flowchart showing an example of the flow of discourse summary processing in the discourse summary section 21 of the discourse summary generation system 2. When the discourse summarization section 21 starts the discourse summarization process, first, each block (result of reconstructing the block after excluding statements such as greetings and reconciliation) as a result of simplifying the discourse data 58 based on the simplification information 52 A loop process for repeating the process is started (S501). When the loop processing for each block is started, first, the target block is set as the first block (original block), and the next block following is set as the second block (destination block) (S502).

  Next, a loop process for repeating the process for each template in the summary template 15 is started (S503). When the loop processing for each template is started, in the block (first or second block) corresponding to the designation of the “from” label, “by” label, etc. of the target template, the word designated in the “seqs” label It is determined whether or not a connection pattern is included (S504). If the pattern is not included, the process proceeds to the next template process (S506, S503).

  If a word concatenation pattern is included in step S504, a summary sentence is generated from the summary sentence template specified in the “template” label of the target template and output as a summary 59 (S505). At this time, the parameter portion is replaced with the corresponding specific expression in the target block as necessary. Also, when outputting, the generated summary text is added to the end of the summary 59, and if the “slot” label is specified in the target template, it is added to the end of the slot specified by the label And If the target slot is not in the summary 59, the target slot is created and added.

  When the loop process for each template is completed, the process proceeds to the next block (S507, S501). In other words, the block that has been the second block in the loop processing for the immediately preceding block becomes the first block in the next block processing, so that the block pair moves one block at a time. When the loop processing for each block ends, a summary 59 is output and the discourse summary processing ends. In the summary 59, summary sentences related to items and contents designated by the summary template 15 are recorded in time series for each slot.

  In the above-described example, when the summary 59 is generated, the entire discourse data 58 reduced based on the reduction information 52 is targeted, but the segment of the discourse data 58 held in the cohesiveness information 54 of the discourse semantics 57 is used. Based on this information, the above processing may be repeated for each segment to generate a summary 59 subdivided for each segment. Similarly to the contents of the discourse semantics 57, the content of the output summary 59 may be provided with an interface that allows the user to appropriately modify the discourse summary generation system 2 by referring to the contents.

  As described above, according to the discourse summary generation system 2 of the present embodiment, when the word concatenation pattern specified in the preset summarization template 15 is included in the block in the discourse data 58, the summarization template By outputting the summary text specified in 15, it is possible to obtain a summary 59 including items and contents set in the summary template 15 and not including unnecessary items. Further, since only a necessary portion is extracted from the content of the discourse data 58 and the summary 59 is generated, even if noise is included in other unnecessary portions, the summary 59 is not affected by these. Can be generated.

[Discussion summary (template creation)]
Below, the detail of the process which produces the summary template 15 used in the case of the talk summary by the talk summary template creation system 1 shown in FIG. 1 is demonstrated. FIG. 16 is a diagram showing an example of creating the summary template 15 in the discourse summary template creating system 1. In the present embodiment, the summary template 15 is created from the summary history 61 in which the response records created as the summary of the response contents for the customer by each operator of the call center are accumulated.

  In the present embodiment, each response record accumulated in the summary history 61 is shown in, for example, the discourse structure analysis system 3 shown in FIG. 2 and the discourse structure analysis system 3 shown in FIG. 1 based on the contents of the discourse between the operator and the customer. Generated from the summary 59 generated by the discourse summary generation system 2 or manually prepared by the operator by summarizing the discourse contents, and the morphological analysis of the discourse structure analysis system 3 shown in FIG. It is assumed that the analysis processing is performed by each unit such as the unit 34. That is, it is assumed that the discourse structure analysis system 3 has already analyzed the structure of each response record (or the discourse statement that is the basis thereof) and the discourse semantics 57 'has been generated.

  First, the clustering unit 11 classifies the sentences of each response record accumulated in the summary history 61 into similar ones, and obtains cluster information 14. The cluster information 14 includes one or more clusters including one or more response records classified as similar. As described above, the clustering method is not particularly limited. For example, for each word included in each response record accumulated in the summary history 61, the importance of the word in each response record is scored, and each response record is recorded. Then, a word vector having the score as an element is generated, and clustering is performed by the k-means method based on the distance between the word vectors.

  As a method for scoring the importance of words, various methods can be used as long as they can be evaluated as numerical values. In the present embodiment, for example, they are generally used. An index is a TF (Term Frequency) value that is an appearance frequency of a target word in a sentence and an IDF (Inverse Document Frequency) value that is an inverse of the number of sentences in which the target word appears in all sentences. The product TF-IDF value is used. Thereby, it is possible to determine that a word having a large TF-IDF value is a word having high importance in the response recording. In addition, for each response record, a word vector whose element is the score of each word is generated, and the degree of similarity is determined by the algebraic distance (or the angle between the word vectors) of each word vector in the vector space. Each response record can be clustered.

  From each of the obtained clusters, the candidate sentence selection unit 12 selects a response record to be a summary template 15 as a candidate sentence 16. The candidate sentence 16 is, for example, a sentence including what frequently appears as a summary expression or pattern in each reception record included in the target cluster, or a typical sentence that is considered to represent the contents well. Here, as described above, for example, the score of importance calculated for each word for each response record is summed, and the response record having the maximum sum is selected as the candidate sentence 16. You may make it weight a score by the shortness of the sentence of a response record, the number of words which can be parameterized, etc.

  For the selected candidate sentence 16, for example, a template corresponding to a so-called content word such as a specific expression is parameterized (replaced with a parameter) by the template forming unit 13 and set as a template with a “template” label in the summary template 15 To do.

  In addition, for each cluster in the cluster information 14, a word ranking 17 ranking each word is obtained. Here, for example, for each word, the importance scores in each reception record in the cluster are summed, and the words are ranked in order from the highest score. A combination obtained based on a predetermined number of words in the word ranking 17 (three in the example of FIG. 16, “wrapping”, “phone”, and “like”) is displayed in the summary template 15 by the “seqs” label. By setting as the connection pattern, the summary template 15 as shown in FIG. 14 is automatically created. Further, the user can edit (edit) the summary template 15 by enabling the user to edit the contents of the summary template 15 as necessary.

  FIG. 17 is a flowchart showing an example of the flow of processing for creating the summary template 15 in the discourse summary template creation system 1. First, as necessary, the discourse structure analysis processing is performed for each response record accumulated in the summary history 61 using the discourse structure analysis system 3 to obtain discourse semantics 57 '(S601). In the present embodiment, it is assumed that at least analysis by the morpheme analysis unit 33 and the specific expression analysis unit 35 is performed, and the morpheme information 55 and the specific expression information 56 are obtained. As described above, if each response record accumulated in the summary history 61 is created from the summary 59 generated by the discourse structure analysis system 3 and the discourse summary generation system 2, step S601 is not necessary.

  Next, the clustering unit 11 temporarily excludes words corresponding to so-called content words such as specific expressions from the words included in each response record accumulated in the summary history 61 for clustering (S602). Since words such as proper expressions are parameterized when they are templated, the accuracy of clustering can be improved in consideration of templating by not including them in similar judgments when clustering response records. Can be improved.

  Next, for each remaining word, the importance is scored for each reception record, and a word vector for the reception record having the score as an element is generated (S603). For importance scoring, as described above, for example, the TF-IDF value for each word is used. Thereafter, each reception record is clustered by an arbitrary clustering method such as the k-means method based on the distance between the word vectors for each reception record to obtain cluster information 14 (S604).

  Thereafter, a loop process for repeating the process for each cluster in the cluster information 14 is started (S605). When the loop processing for each cluster is started, first, the candidate sentence selection unit 12 calculates a score after performing predetermined weighting as necessary for each response record in the cluster to be processed (S606). The response record with the highest score is selected as the candidate sentence 16 (S607). As described above, the score of each reception record is calculated by, for example, summing the importance scores calculated for each word for each reception record. At that time, weighting is performed by a predetermined coefficient so that the score becomes larger as the length of the response record sentence is shorter or the number of words that can be parameterized (for example, content words such as specific expressions) is larger. May be.

  After that, for the selected candidate sentence 16, the template forming unit 13 parameterizes words corresponding to so-called content words such as specific expressions (S608). On the other hand, word ranking 17 in the cluster is generated for words in the cluster to be processed (S609). Here, as described above, for example, for each word, the score of importance in each reception record in the cluster is summed, and the words are ranked in descending order of score.

  Thereafter, template processing is performed based on the candidate sentence 16 parameterized in step S608 and the word ranking 17 generated in step S609, and a summary template 15 is created (S610), and the process proceeds to the next cluster ( S611, S605). In creating the summary template 15, as described above, for example, the parameterized candidate sentence 16 is set as a template by the “template” label, and further obtained based on a predetermined number of words in the word ranking 17. The combination to be used is set as a connection pattern by the “seqs” label.

  When the loop processing for each cluster is completed, the user edits the contents of the summary template 15 as necessary (S612), completes the summary template 15 and ends the processing. By using this summary template 15 in the discourse summary generation system 2, for example, it is possible to obtain a summary 59 that includes items and contents set in the summary template 15 and does not include unnecessary items.

  As described above, according to the discourse summary template creation system 1 of the present embodiment, the response record information accumulated in the summary history 61 is clustered into similar ones, and the candidate sentence 16 is selected for each cluster. Thus, it is possible to parameterize the candidate sentence 16 and generate a connection pattern. Thereby, it is possible to automatically create the summary template 15 with high accuracy, or to assist the user such as an administrator to create the summary template 15.

  As mentioned above, the invention made by the present inventor has been specifically described based on the embodiment. However, the present invention is not limited to the embodiment, and various modifications can be made without departing from the scope of the invention. Needless to say.

  INDUSTRIAL APPLICABILITY The present invention is applicable to a discourse summary template creation system and a discourse summary template creation program for creating a template for generating a summary of a discourse from the result of analyzing the structure of discourse data obtained by speech recognition processing.

DESCRIPTION OF SYMBOLS 1 ... Discourse summary template creation system, 2 ... Discourse summary generation system, 3 ... Discourse structure analysis system, 4 ... Speech recognition engine,
DESCRIPTION OF SYMBOLS 11 ... Clustering part, 12 ... Candidate sentence selection part, 13 ... Template making part, 14 ... Cluster information, 15 ... Summary template, 16 ... Candidate sentence, 17 ... Word ranking,
21 ... Discussion summary section,
DESCRIPTION OF SYMBOLS 31 ... Pre-processing part, 32 ... Flow analysis part, 33 ... Cohesiveness analysis part, 34 ... Morphological analysis part, 35 ... Specific expression analysis part, 36 ... Discourse reduction part, 37 ... Flow analysis rule, 38 ... Specific expression analysis rule ,
51 ... Discourse data, 52 ... Simple information, 53 ... Flow information, 54 ... Cohesiveness information, 55 ... Morphological information, 56 ... Specific expression information, 57, 57 '... Discourse semantics, 58 ... Discourse data, 59 ... Summary,
61 ... Summary history.

Claims (6)

It is used when generating a summary about the discourse based on the discourse data in which the content of the discourse has been converted into text, and a portion to be included in the summary that is assumed to be a utterance of a predetermined content in the discourse data. When a list of correspondences between word concatenation patterns for identification and templates of summary sentences to be included in the summary is specified, and the concatenation pattern and the discourse data match, A discourse summary template creation system for creating a summary template that is used to generate the summary text from the corresponding summary text template and use it as the summary,
A clustering unit that obtains the summary records from the summary history in which summary records that are past summaries are accumulated, classifies them into similar ones, and clusters them;
For each cluster, a candidate sentence selection unit that selects a representative summary record from the summary records included in the cluster and sets it as a candidate sentence;
Wherein for each candidate sentence and template of the summary sentence by replacing the specific expression parameter, it possesses a template unit for creating the summary template,
The clustering unit includes:
For each word included in each summary record in the summary history, the importance level of each word in each summary record is scored, and a word vector having the score for each word as an element is generated for each summary record. Clustering each summary record based on the distance between the word vectors;
The templating unit
For each cluster, each word is ranked according to the score, and a combination of one or more high-order words is combined with the connected sentence corresponding to the template of the summary sentence created from the candidate sentences in the cluster. A discourse summary template creation system , wherein the summary template is set as a pattern . In the discourse summary template creation system according to claim 1 ,
The candidate sentence selection unit
The discourse summary template creation system characterized by summing up the scores calculated for each word for each summary record, and selecting the summary record that maximizes the sum as the candidate sentence. In the discourse summary template creation system according to claim 1 ,
The clustering unit
A discourse summary template creation system, wherein the word vector is generated by excluding a word corresponding to a specific expression from each word included in the summary record. It is used when generating a summary about the discourse based on the discourse data in which the content of the discourse has been converted into text, and a portion to be included in the summary that is assumed to be a utterance of a predetermined content in the discourse data. When a list of correspondences between word concatenation patterns for identification and templates of summary sentences to be included in the summary is specified, and the concatenation pattern and the discourse data match, A discourse summary template creation system for creating a summary template that is used to generate the summary text from the corresponding summary text template and use it as the summary,
A clustering unit that obtains the summary records from the summary history in which summary records that are past summaries are accumulated, classifies them into similar ones, and clusters them;
For each cluster, a candidate sentence selection unit that selects a representative summary record from the summary records included in the cluster and sets it as a candidate sentence;
A template for creating a summary template by replacing a specific expression with a parameter for each candidate sentence to form a template of the summary sentence;
The clustering unit includes:
For each word included in each summary record in the summary history, the importance level of each word in each summary record is scored, and a word vector having the score for each word as an element is generated for each summary record. Clustering each summary record based on the distance between the word vectors;
The candidate sentence selection unit
Summing the score calculated for each word in each of the summary records, with respect to the total, had a row a predetermined weight according to the number of named entities in the summary record, the summary records the sum is the maximum Is selected as the candidate sentence, and a discourse summary template creation system. It is used when generating a summary about the discourse based on the discourse data in which the content of the discourse has been converted into text, and a portion to be included in the summary that is assumed to be a utterance of a predetermined content in the discourse data. When a list of correspondences between word concatenation patterns for identification and templates of summary sentences to be included in the summary is specified, and the concatenation pattern and the discourse data match, A discourse summary template creation system for creating a summary template that is used to generate the summary text from the corresponding summary text template and use it as the summary,
A clustering unit that obtains the summary records from the summary history in which summary records that are past summaries are accumulated, classifies them into similar ones, and clusters them;
For each cluster, a candidate sentence selection unit that selects a representative summary record from the summary records included in the cluster and sets it as a candidate sentence;
A template for creating a summary template by replacing a specific expression with a parameter for each candidate sentence to form a template of the summary sentence;
The clustering unit includes:
For each word included in each summary record in the summary history, the importance level of each word in each summary record is scored, and a word vector having the score for each word as an element is generated for each summary record. Clustering each summary record based on the distance between the word vectors;
The candidate sentence selection unit
Summing the score calculated for each word in each of the summary records, with respect to the total, had a row a predetermined weight in accordance with the short text of the summary records, the summary records the sum is the maximum Is selected as the candidate sentence, and a discourse summary template creation system. It is used when generating a summary about the discourse based on the discourse data in which the content of the discourse is converted into text. When a list of correspondences between word concatenation patterns for identification and templates of summary sentences to be included in the summary is specified, and the concatenation pattern and the discourse data match, A discourse summary template creation program that causes a computer to operate as a discourse summary template creation system that creates a summary template used to generate the summary sentences from the corresponding summary sentence template,
A clustering process for acquiring the summary records from the summary history in which summary records that are past summaries are accumulated, classifying them into similar ones, and clustering them;
For each cluster, a candidate sentence selection process that selects a representative summary record from the summary records included in the cluster as a candidate sentence;
For each of the candidate sentences, replace the specific expression with a parameter to form a template of the summary sentence, and execute a template processing for creating the summary template ,
In the clustering process,
For each word included in each summary record in the summary history, the importance level of each word in each summary record is scored, and a word vector having the score for each word as an element is generated for each summary record. Clustering each summary record based on the distance between the word vectors;
In the template processing,
For each cluster, each word is ranked according to the score, and a combination of one or more high-order words is combined with the connected sentence corresponding to the template of the summary sentence created from the candidate sentences in the cluster. A discourse summary template creation program characterized in that the summary template is set as a pattern .

Images