JP6208794B2 - Conversation analyzer, method and computer program - Google Patents

Description

  The present invention relates to a technique for analyzing interactive text.

  Various information received from customers and potential customers (hereinafter referred to as customers) is useful for companies. For example, there is a technique for detecting a sign of a problem relating to a product or service that may occur in the future by converting an opinion received from an existing customer into text and analyzing the text (for example, Patent Document 1).

Japanese Patent No. 5078960

  In recent years, due to the development of Internet technology and the spread of SNS (Social Network Service), means for collecting customers' voices have been diversified. However, the telephone handling operation in which the customer and the operator talk directly with each other to deal with troubles of the product and answer the questions of the customer is still important as a means of collecting the voice of the customer. This is because the operator may be able to obtain information that cannot be obtained by other means by appropriately responding. To that end, it is important to train operators with high conversation skills.

  By the way, conversations between customers and operators are usually used in spoken language (spoken language), and the grammar is inaccurate and ambiguous compared to written sentences. It is characterized by a large amount of words that are not related to the context, such as lyrics. For this reason, automatic analysis technology for conversational text, which is a text conversation between a customer and an operator, has not been put to practical use.

  Therefore, an object of the present invention is to provide a technique for automatically analyzing a conversation text.

  Another object of the present invention is to automatically analyze the conversation text and extract the features of the speaker's conversation.

  According to an embodiment of the present invention, there is provided a conversation analysis apparatus for storing conversation record data including text describing at least the contents of a conversation between a first speaker and a second speaker; Segmenting means for dividing the text into a plurality of units and dividing the plurality of units into a plurality of segments based on the similarity between the units; and the plurality of segments formed by the segmenting means in a plurality of categories Categorizing means for classifying into: and a means for storing conversation pattern data indicating a plurality of categories included in the conversation and a transition of categories in the conversation based on a result of classification by the categorizing means; Prepare.

  In a preferred embodiment, a conversation pattern for extracting a category transition characteristic for each first speaker with reference to conversation pattern data of a plurality of first speakers stored in the means for storing the conversation pattern data. An analysis means may be further provided.

  In a preferred embodiment, the segmenting means performs a clustering process for separating the plurality of units into a predetermined number of clusters, and forms one segment with a plurality of units belonging to the same cluster and continuous. May be.

  In a preferred embodiment, there may be further provided means for calculating a time required for each category included in the conversation based on a result of classification by the categorizing means.

  In a preferred embodiment, the apparatus further comprises unnecessary word deleting means for deleting unnecessary words from the text of the conversation record data, and the segmenting means performs segmentation based on the text from which unnecessary words have been deleted by the unnecessary word deleting means. Dividing may be performed.

  In a preferred embodiment, the unnecessary word deleting means divides the plurality of utterances into a plurality of processing units, obtains a degree of association between the processing unit and the deletion target word using a concept search function, and determines the unnecessary words. You may specify.

It is a functional lineblock diagram of conversation analysis device 1 concerning one embodiment of the present invention. It is a figure which shows an example of the conversation attribute data. It is a figure which shows an example of the conversation text 351 and the edit text 353. FIG. It is a figure which shows an example of the conversation pattern data 370. It is a figure which shows an example of the category transition data 390. It is a figure which shows an example of the unnecessary word dictionary. It is a figure which shows the relationship between the unit of the clustered conversation text, and a segment. It is a figure which shows an example of the category dictionary 530. FIG.

  Hereinafter, a conversation analysis device according to an embodiment of the present invention will be described with reference to the drawings.

  FIG. 1 is a functional configuration diagram of a conversation analysis apparatus 1 according to the present embodiment. The conversation analysis apparatus 1 may be realized by causing a general-purpose information processing apparatus including a processor, a memory, an input / output device, and the like to execute a predetermined computer program.

  As shown in the figure, the conversation analysis apparatus 1 has a conversation voice data storage unit 31, a conversation attribute data storage unit 33, a conversation record data storage unit 35, and a conversation pattern data storage unit 37 as means for storing data. And a category transition storage unit 39.

  The conversation voice data storage unit 31 stores a voice data file of conversations uttered by a plurality of speakers. The file name of the audio data file may include a conversation ID described later.

  The conversation attribute data storage unit 33 stores data indicating the attributes of the conversation voice data stored in the conversation voice data storage unit 31. The conversation attribute data may include, for example, information such as a speaker having a conversation, the date and time of the conversation, and the situation.

  FIG. 2 is a diagram illustrating an example of the conversation attribute data 330.

  As shown in the figure, the conversation attribute data 330 has a conversation ID 331 and a speaker ID 333 as data items.

  The conversation ID 331 is identification information for identifying a plurality of conversations to be processed by the conversation analysis apparatus 1.

  The speaker ID 333 is identification information of a speaker participating in the conversation with the conversation ID 331. Registration for the number of participants is possible. If the speaker cannot be identified, it may be unknown.

  Returning to FIG. 1, the conversation record data storage unit 35 stores text data of conversations performed by a plurality of speakers. For example, the conversation record data storage unit 35 stores a conversation data file including text describing the content of a conversation including at least a first speaker and an utterance of a second speaker. The conversation ID may be included in the file name of the conversation text data. The conversation record data storage unit 35 stores, for example, conversation text 351 and edited text 353.

  FIG. 3 is a diagram illustrating an example of the conversation text 351 and the edit text 353.

  As shown in FIG. A, the conversation text 351 has a text part 351A and a time code part 351B.

  The text part 351A is a text describing the content of a conversation including the utterances of a plurality (at least two) speakers. In conversation, the utterances of a plurality of speakers may be alternated. During the conversation, there are many unnecessary expressions to specify the context such as greetings and messages. As will be described later, the unnecessary word deletion processing unit 13 may remove these unnecessary expressions.

  The time code unit 351B has, for each utterance, a time code at which each utterance starts and an end time code. The time code unit 351B may be generated when the voice recognition processing unit 11 performs voice recognition based on conversation voice data.

  The conversation text 351 may be text other than the text extracted from the voice data by the voice recognition processing unit 11. For example, text entered by a person may be used, or text extracted from an image by OCR (Optical Character Reader) or the like may be used.

  FIG. 5B shows the edit text 353. The edited text 353 is text data in which an unnecessary word is deleted from the conversation text 351 by an unnecessary word deletion process described later. The edit text 353 has a text part 353A and a time code part 353B.

  Returning to FIG. 1, the conversation pattern data storage unit 37 stores data indicating a conversation pattern that is clarified by analysis of the conversation text 351. The conversation pattern data may indicate, for example, a plurality of categories included in the conversation and category transitions in the conversation.

  FIG. 4 is a diagram illustrating an example of the conversation pattern data 370.

  As shown in the figure, the conversation pattern data 370 has a conversation ID 371, a segment 373, a time 374, and a category 375 as a data structure.

  The segment 373 is segment identification information formed by dividing the conversation text 351 into a plurality of pieces. Here, each segment is assigned a unique serial number for one conversation ID 371 in the order of appearance in the conversation.

  Time 374 is the time that each segment lasts. The duration of the segment will be described later.

  A category 375 is a category assigned to each segment. There may be a plurality of categories 375 assigned to one segment.

  In the conversation pattern data 370, the number of the segment 373 corresponds to the order in which the categories appear in the conversation. For example, in the conversation with the conversation ID “1001” in FIG. 4, the category transitions in the order of category A → category B or C → category C → category D or A in the order of the segment 373.

  Returning to FIG. 1, the category transition storage unit 39 stores data indicating the characteristics of the category transition in the conversation. For example, the category transition storage unit 39 may store statistical data indicating characteristics of category transitions by speaker.

  FIG. 5 is a diagram illustrating an example of the category transition data 390.

  As shown in the figure, the category transition data 390 includes, as data items, a category 391, a speaker ID 392, the number of cases 393, a time 394, and a transition category 395. The number of cases 393 may be the number of cases for each category 391 and speaker ID 392. The time 394 is a total sum of durations of segments assigned to the respective categories 391 by the speaker ID 392. The transition category 395 is a transition destination category transitioned from the category 391. The transition category 395 further includes a category name, the number of cases, and a probability as data items. The transition category 395 not only counts the category that has directly transitioned from the category 391, that is, the category that appears next to the category 391, but also the category that has transitioned from the category 391 via another category, that is, the category 391. The next category that appears next may be counted.

  Referring back to FIG. 1, the conversation analysis device 1 includes a speech recognition processing unit 11, an unnecessary word deletion processing unit 13, a segment processing unit 15, a categorization processing unit 17, and a conversation pattern analysis processing unit 19. .

  The speech recognition processing unit 11 performs speech recognition processing that recognizes speech data and extracts text. For example, the voice recognition processing unit 11 performs voice recognition of a conversation voice file stored in the conversation voice data storage unit 31 and generates a conversation text 351 including text describing the contents of the conversation. The data of the conversation text 351 is stored in the conversation record data storage unit 35.

  For example, the voice recognition processing unit 11 extracts text from conversation voice data by voice recognition, and generates a text part 315A including text obtained by dividing the conversation in units of utterances. Furthermore, the voice recognition processing unit 11 specifies the time code of the start and end of the utterance in the conversation voice data, and generates a time code unit 351B corresponding to the text of the text unit 351A.

  The voice recognition processing unit 11 may identify which speaker is making each utterance by voice recognition. Furthermore, the speech recognition processing unit 11 may register the identified speaker ID in the conversation attribute data 330 if the identified speaker can be identified.

  The unnecessary word deletion processing unit 13 deletes unnecessary words from the conversation text 351. For example, the unnecessary word deletion processing unit 13 may identify an unnecessary word among words and sentences included in the text portion 351A of the conversation text 351 stored in the conversation record data storage unit 35 using the concept search function. good. The unnecessary word deletion processing unit 13 may have an unnecessary word dictionary, for example, and may specify an unnecessary word to be deleted with reference to this dictionary.

  FIG. 6 is a diagram illustrating an example of the unnecessary word dictionary 510.

  As shown in the figure, the unnecessary word dictionary 510 may have an expression example to be deleted and an expression example not to be deleted for each type of fixed phrase such as greetings.

  The unnecessary word deletion processing unit 13 may divide the text of the conversation text 351 into, for example, morphemes, and calculate the degree of association with the expression registered in the unnecessary word dictionary 510 by concept search. The unnecessary word deletion processing unit 13 may delete from the conversation text 351, as a unnecessary word, a text in a processing unit having a degree of relevance of concept search of a predetermined level or more. The unnecessary word deletion processing unit 13 deletes unnecessary words from the conversation text 351 and generates an edit text 353. The concept search may be performed using any concept search algorithm.

  Even when a part of one utterance is deleted as an unnecessary word in the text of the text portion 351A, the time code corresponding to the utterance need not be changed. For example, even when an unnecessary word is deleted from an utterance and the text corresponding to the utterance becomes shorter, the time code part 353B is generated without changing the time code of the time code part 351B corresponding to the utterance. Also good.

  Note that it is not essential to apply the unnecessary word deletion process to the conversation text 351. For example, the unnecessary word deletion process may not be applied to the conversation text in which the chat is recorded.

  The segment processing unit 15 divides the edited text 353 into a plurality of segments. For example, the segment processing unit 15 may perform processing by dividing a text into units that are predetermined processing units. The segment processing unit 15 may measure the time of each unit. Further, the segment processing unit 15 may classify a plurality of units in conversation into a plurality of segments based on the similarity between the units. The segment processing unit 15 may set the text portion of the utterance that has continued from when one speaker starts speaking until another speaker starts speaking as one unit, or the range delimited by punctuation marks as one unit. Or a sentence as a unit.

  The segment processing unit 15 may further calculate the time that the segment has continued. For example, the segment processing unit 15 may specify the utterance time assigned to the same segment with reference to the time code unit 353B, and add these times to make the duration of the segment.

  The segment processing unit 15 may divide the conversation text 351 into a plurality of segments.

  The segment processing unit 15 may classify a large number of units included in the conversation into a predetermined number of clusters by clustering processing. The number of clusters classified by the clustering process can be arbitrarily determined. Units that are not classified into any cluster may be classified as “others”. The segment processing unit 15 can use any clustering algorithm.

  The segment processing unit 15 may perform clustering by paying attention to the similarity between units without performing text syntax analysis or semantic analysis.

  The segment processing unit 15 divides a plurality of units into a plurality of segments based on the result of clustering.

  FIG. 7 is a diagram illustrating the relationship between clustered conversation text units and segments.

  In the figure, an item number 410 is assigned to a plurality of units 420 in which the edit text 353 is divided. Further, the cluster 430 includes four clusters (C1 to C4) and others, and indicates a cluster to which each unit 420 is allocated.

  Here, the segment processing unit 15 may set a plurality of continuous units belonging to the same cluster as one segment. One or more units are assigned to one segment. In the example of the figure, item number 1-2, item number 3-6, item number 7,..., Item number 25-27, item number 28-29 and item number 30 surrounded by one are respectively It becomes one segment.

  Referring to FIG. 1, the categorization processing unit 17 may classify a plurality of segments formed by the segment processing unit 15 into a plurality of categories. For example, the categorization processing unit 17 may have a category dictionary and classify each segment into one of a plurality of predetermined categories while referring to the category dictionary. The category may be, for example, a classification indicating a topic that can become a topic in conversation.

  FIG. 8 is a diagram illustrating an example of the category dictionary 530. The category dictionary 530 includes a category 531, an exemplary sentence 533, and an important word 535 as items.

  The categorization processing unit 17 compares the unit 420 classified into each segment with the model sentence 533 or the important word 535 of the category dictionary 530, and assigns any category to each segment. Two or more categories may be assigned to one segment. The categorization processing unit 17 generates conversation pattern data 370 (see FIG. 4) in which a category is assigned to each segment, and stores it in the conversation pattern data storage unit 37.

  The conversation pattern analysis processing unit 19 analyzes conversation pattern data and extracts conversation features.

  The conversation pattern analysis processing unit 19 may extract, for example, the characteristics of category transition for each first speaker by referring to the conversation pattern data of a plurality of first speakers. For example, when the conversation pattern data storage unit 37 stores the conversation pattern data 370 obtained by processing the conversation text 351 when a large number of operators correspond to various customers or the like, the conversation pattern analysis processing unit 19 May count the number of transitions from one category to another in the conversation for each operator, or may calculate the probability of transition. The calculated probability may be stored in the transition category 395 of the category transition data 390.

  The conversation pattern analysis processing unit 19 may calculate the time required for each category for each first speaker based on the conversation pattern data. For example, the conversation pattern analysis processing unit 19 may calculate the time required for each category for each operator by accumulating the time 374 for each operator category. The calculated time may be stored in the time 394 of the category transition data 390.

  The conversation pattern analysis processing unit 19 may generate category transition data 390 (see FIG. 5) as a result of calculating the category transition. The category transition data 390 is stored in the category transition storage unit 39.

  This makes it possible to extract categories that are discussed in conversation. Furthermore, focusing on the duration of the category and the category transition, the conversation characteristics for each speaker can be extracted. For example, by determining the duration of a category, the frequency of appearance of a category, the probability of transition from a specific category to another specific category, etc., it is possible to extract a characteristic tendency for each speaker.

  In particular, when conversation text data or category transition data when an operator corresponds to a customer or the like is extracted, the customer reception skill for each operator can be visualized. For example, if the tendency of category transition of highly evaluated operators can be grasped, there is a possibility that it can be used for new operator education.

  The above-described embodiments of the present invention are examples for explaining the present invention, and are not intended to limit the scope of the present invention only to those embodiments. Those skilled in the art can implement the present invention in various other modes without departing from the gist of the present invention.

DESCRIPTION OF SYMBOLS 1 Conversation analysis apparatus 11 Speech recognition process part 13 Unnecessary word deletion process part 15 Segment process part 17 Categorize process part 19 Conversation pattern analysis process part 31 Conversation voice data storage part 33 Conversation attribute data storage part 35 Conversation record data storage part 37 Conversation pattern Data storage unit 39 Category transition storage unit

Claims (7)

Means for storing conversation record data including text describing the content of a conversation between a first speaker who is a customer-facing operator and a second speaker including the customer ;
A clustering process is performed for dividing the text of the conversation recording data into a plurality of units, measuring the time of the units, and separating the text of the conversation recording data into a predetermined number of clusters based on the similarity between the units. The text of the conversation recording data is divided into a plurality of segments so as to form one segment with a plurality of continuous units belonging to the same cluster, and the time of the units of the same segment is added to form a segment Segment processing means for calculating the duration of
A category Rise processing means for classifying the plurality of segments formed by the segment processing means, into a plurality of categories respectively indicating topics related to the requirements or conditions of the second speaker,
And the duration of the segment, a storage means for storing the speech pattern data indicating a result of classification by the category rise processing means,
Referring to the conversation pattern data of the first speaker stored in the storage means, the conversation pattern analysis for extracting the category included in the conversation of the first speaker, the time required for the category, and the category transition Processing means;
Conversation analyzer with   The conversation analysis apparatus according to claim 1, wherein the category transition is indicated by a category name, a number of cases, and a probability of a transition destination category that has transitioned from the category.   The conversation analysis apparatus according to claim 1, wherein the time required for the category is a total sum of durations of segments classified into the category in the conversation pattern data. An unnecessary word dictionary in which a deletion target word is registered is held in advance, unnecessary words are identified from the text of the conversation recording data with reference to the unnecessary word dictionary , and further includes unnecessary word deletion processing means for deleting,
The segment processing means, said performing segmentation based on the text of unnecessary words are removed by the unnecessary word deletion section, conversation analysis apparatus according to any one of claims 1-3.   5. The unit according to claim 1, wherein the unit is a text portion of an utterance that lasts from the time when one speaker starts speaking to the time when another speaker starts speaking, a range separated by punctuation marks, or a sentence. The conversation analysis device described in 1. A conversation analysis method performed by an information processing apparatus including means for storing conversation record data including text describing the contents of a conversation between a first speaker who is a customer-facing operator and a second speaker including the customer There,
Segment processing means divides the text of the conversation recording data into a plurality of units, measures the time of the units, and separates the text of the conversation recording data into a predetermined number of clusters based on the similarity between the units. The conversation recording data text is divided into a plurality of segments so that a plurality of continuous units belonging to the same cluster and a plurality of continuous units form one segment. To calculate the duration of the segment by adding together ,
Category Rise processing means, a step of classifying a plurality of segments formed by the segment processing means, into a plurality of categories respectively indicating topics related to the requirements or conditions of the second speaker,
And recording the duration of the segment, the speech pattern data indicating a result of classification by the category rise processing means,
The conversation pattern analysis processing means refers to the recorded conversation pattern data of the plurality of first speakers, the category included in the conversation of the first speaker, the time required for the category, and the category transition Extracting the
Having a method. In order to perform conversation analysis in an information processing apparatus comprising means for storing conversation record data including text describing the contents of a conversation between a first speaker who is a customer-facing operator and a second speaker including the customer Computer program,
When executed on the information processing apparatus,
Segment processing means divides the text of the conversation recording data into a plurality of units, measures the time of the units, and separates the text of the conversation recording data into a predetermined number of clusters based on the similarity between the units. The conversation recording data text is divided into a plurality of segments so that a plurality of continuous units belonging to the same cluster and a plurality of continuous units form one segment. To calculate the duration of the segment by adding together ,
Category Rise processing means, a step of classifying a plurality of segments formed by the segment processing means, into a plurality of categories respectively indicating topics related to the requirements or conditions of the second speaker,
And recording the duration of the segment, the speech pattern data indicating a result of classification by the category rise processing means,
The conversation pattern analysis processing means refers to the recorded conversation pattern data of the plurality of first speakers, the category included in the conversation of the first speaker, the time required for the category, and the category transition And a step of extracting a computer program.

Images