JP2019139280A - Text analyzer, text analysis method and text analysis program - Google Patents

Abstract

To improve accuracy of automatic summarization of text information converted from voice information for telephone answer.SOLUTION: A text analyzer comprises: a text acquisition unit which acquires answer text information obtained by converting voice information for telephone answer between a user and an operator into text; an identification unit 433 which identifies a character string indicating the start or the end of answering in which the operator checks identification information of the user in the answer text information; and a specifying unit 434 which specifies a front side text portion before the character string identified by the identification unit 433 in the answer text information or a rear side text portion after the character string identified by the identification unit 433.SELECTED DRAWING: Figure 2

Description

  The present invention relates to a text analysis apparatus, a text analysis method, and a text analysis program for causing a computer to function as a text analysis apparatus related to voice information for telephone reception.

  In the call center, the operator records the telephone answering history after the telephone answering such as an inquiry from the customer. There is a demand for reducing the burden on the operator due to this history recording operation. Recording information obtained by recording a telephone response is converted into text information by a voice recognition technique (see, for example, Patent Document 1).

JP 2001-2111245 A

  In order to make it easy for a third party to check the contents of the telephone reception or to confirm the contents, a telephone summary of the telephoneized text is automatically summarized. If the text part corresponding to the telephone response that the operator confirms whether or not the customer is the contractor of the service in the text of the telephone response content is the identity confirmation part, the identity confirmation information such as the user's name is included in the identity confirmation part. The frequency of words related to is increasing. Many automatic summaries employ an algorithm that preferentially picks up words that frequently appear in the text, so if the identity verification part is also subject to summarization, it will be the customer like the words and phrases related to the identity verification information. Content that is not the content of the inquiry itself was picked up preferentially, and there was a problem that the accuracy of the summary fell.

  The present invention has been made in view of the above circumstances, and provides a text analysis device, a text analysis method, and a text analysis program capable of improving the accuracy of automatic summarization of text information obtained by converting voice information for telephone reception. For the purpose.

  A text analysis apparatus according to a first aspect of the present invention includes a text acquisition unit that acquires response text information obtained by converting voice information of a telephone response between a user and an operator into text, and the operator in the response text information An identification unit for identifying a character string indicating the start or end of the response for confirming the personal identification information, and a front text portion preceding the character string identified by the identification unit in the response text information or the identification unit is identified A specifying unit for specifying a rear text portion after the character string.

  The text analysis apparatus may further include a summarizing unit that inputs the front text portion or the rear text portion to a machine learning model and outputs summary data output from the machine learning model.

  The identification unit identifies a start character string indicating the start of a response for confirming the identity confirmation information and an end character string indicating an end of the response for confirming the identity confirmation information, and the identification unit identifies the identified The end character string may be identified from a text portion after the start character string.

  The specifying unit specifies a front text portion before the start character string identified by the identifying unit and a rear text portion after the end character string identified by the identifying unit, and the summarizing unit May input the rear text portion and the front text portion into the machine learning model, and output the summary data of the rear text portion and the summary data of the front text portion in association with each other. The identification unit may identify a text part from the beginning of the response text information to a predetermined ratio as a front text part when the identification unit cannot identify the start character string.

  The specific unit is a ratio of the number of times that a speaker who has made a voice switch from the start of a telephone response to the number of times a speaker who makes a voice is switched in the whole telephone response when the identification unit cannot identify the start character string The text part until the value reaches a predetermined value may be specified as the front text part. The specifying unit may specify a text part corresponding to a predetermined number of sentences from the start of telephone reception as the front text part when the identification unit cannot identify the start character string. The specifying unit may specify a text part corresponding to a predetermined time from the start of telephone reception as the front text part when the identification unit cannot identify the start character string.

  The text analysis device further includes an extraction unit that extracts user information from a text portion sandwiched between the start character string and the end character string, and the summarization unit associates with the user information extracted by the extraction unit. The summary data may be output.

  The summarizing unit may output the summary data in association with an identifier indicating that the end character string cannot be identified when the identifying unit cannot identify the end character string. A conversion unit that converts voice information of a telephone reception between a user and an operator into the reception text information, and the identification unit is a reception text obtained by converting the voice uttered by the operator of the reception text information by the conversion unit. The end character string may be identified based on the information.

  The information processing apparatus further includes a storage unit storing feature information of an operator's voice, and the conversion unit converts the voice information into the response text information based on the feature information of the operator's voice stored in the storage unit. May be.

  According to a second aspect of the present invention, there is provided a text analysis method comprising: obtaining response text information obtained by converting voice information of a telephone response between a user and an operator into text; A step of identifying a character string indicating the start or end of reception for confirming the confirmation information, and a front text part before the character string identified in the reception text information or a rear text after the identified character string Identifying a portion.

  According to a third aspect of the present invention, there is provided a text analysis program comprising a computer, a text acquisition unit that acquires response text information obtained by converting voice information of a telephone response between a user and an operator into text, and the operator in the response text information. An identification unit for identifying a character string indicating the start or end of reception for confirming the user's identity confirmation information, and a front text part or the identification unit preceding the character string identified by the identification unit in the reception text information Is made to function as a specifying unit that specifies a rear text portion after the character string identified by.

  ADVANTAGE OF THE INVENTION According to this invention, the precision of the automatic summarization of the text information which converted the voice information of telephone reception can be improved.

It is a figure for demonstrating the outline | summary of the text analysis system S which concerns on embodiment of this invention. It is a figure which shows the structure of a text analyzer. It is a figure which shows reception text information. It is a figure which shows an example of the summary data which a summary part outputs. It is a flowchart which shows operation | movement of a text analyzer.

[Outline of Text Analysis System S]
FIG. 1 is a diagram for explaining an outline of a text analysis system S according to an embodiment of the present invention. The text analysis system S automatically summarizes the contents of user inquiries in order to support the work of creating user inquiry records in a call center.

  The text analysis system S includes a communication terminal 100, a telephone 200, a recording device 300, and a text analysis device 400. The communication terminal 100 is a user's mobile phone, for example, and performs voice communication with the call center phone 200 via the network N.

  The telephone 200 is a call operator terminal installed in a call center. The recording device 300 is connected to the telephone line of the telephone 200, and generates voice information in which the voices of the user and the operator are recorded while voice communication is performed between the communication terminal 100 and the telephone 200. At this time, the recording apparatus 300 stores the user's voice information acquired from the telephone output of the telephone 200 in association with the identification information of the communication terminal 100, and assigns the operator's voice information acquired from the microphone input of the telephone 200 to the telephone 200. And stored in association with the identified identification information. The identification information of the communication terminal 100 is, for example, a mobile phone number. The identification information of the telephone 200 is, for example, an extension number of the telephone 200 in the call center.

  The text analysis device 400 acquires the voice information of the user and the operator together with the identification information assigned to the voice information from the recording device 300, and converts the acquired voice information into response text information. The text analysis device 400 generates summary data of the converted response text information. The text analysis device 400 transmits the generated summary data to a management device (not shown).

  The text analysis apparatus 400 determines that the phrase repeatedly referred to in the response text information is important information, and includes this phrase in the summary data. In conventional text analyzers, words related to identity verification information such as the user's name are repeatedly referred to in the response text information. Therefore, it is determined that the words related to the identity verification information are important information and are summarized data. Was included. As a result, in the conventional text analysis device, since the words related to the identity verification information are determined as important information, it is difficult to determine information other than the identity verification information as important information, and the accuracy of the summary is low. .

  On the other hand, the text analysis apparatus 400 specifies an identity verification portion that is response text information for the operator to confirm the identity of the user, and excludes the identity verification portion from the summary data. For this reason, the text analysis apparatus 400 can improve the accuracy of the summary by making it difficult for the summary data to include words related to the personal identification information.

[Configuration of Text Analyzer 400]
FIG. 2 is a diagram illustrating a configuration of the text analysis device 400. The text analysis device 400 includes a communication unit 41, a storage unit 42, and a control unit 43. The communication unit 41 is a communication interface for communicating with the recording device 300 and a management device (not shown). The storage unit 42 is a storage medium such as a ROM (Read Only Memory) and a RAM (Random Access Memory). The storage unit 42 stores a program executed by the control unit 43. The control unit 43 is, for example, a CPU (Central Processing Unit). The control unit 43 functions as an acquisition unit 431, a conversion unit 432, an identification unit 433, an identification unit 434, an extraction unit 435, a summary unit 436, and a machine learning model 437 by executing a program stored in the storage unit 42. .

  The acquisition unit 431 acquires voice information for telephone reception between the user and the operator from the recording device 300 via the communication unit 41. The acquisition unit 431 outputs the acquired audio information to the conversion unit 432. The acquisition unit 431 functions as a text acquisition unit that acquires response text information obtained by converting the output voice information into text from the conversion unit 432. The acquisition unit 431 outputs the acquired reception text information to the identification unit 433.

  The conversion unit 432 converts the voice information acquired by the acquisition unit 431 into reception text information using a voice recognition technique. At this time, the conversion unit 432 individually converts the user's voice information and the operator's voice information into reception text information. The conversion unit 432 generates response text information in which the response text information corresponding to the user's voice information and the response text information corresponding to the operator's voice information are collected in time series.

  For example, the conversion unit 432 converts the voice information into reception text information based on the feature information of the operator's voice stored in the storage unit 42. The conversion unit 432 stores the characteristics of the operator's voice in the storage unit 42 by performing training for converting the voice information on the voice information issued by the operator in advance. When converting the voice information of the operator into the response text information, the conversion unit 432 reads the feature information of the operator's voice from the storage unit 42 and reflects the feature information in the conversion of the voice information. With such a configuration, the conversion unit 432 trains in advance on pronunciations and the like that are likely to cause errors in conversion of voice information, so that the accuracy of conversion of voice information can be improved.

  The identification unit 433 identifies, from the response text information, an end character string indicating the end of the response in which the operator confirms the user identification information. The identification unit 433 identifies the end character string based on the response text information obtained by converting the voice uttered by the operator from the response text information by the conversion unit 432. The identity confirmation information is information for confirming that the user is the subscriber of the service or the like, and is, for example, the user's name, address, date of birth, mobile phone number, user ID, or password. The operator confirms the identity confirmation information with the user to confirm that the user is the contractor.

  The operator notifies the user that the personal identification has been completed after the completion of the personal identification of the user, and the identification unit 433 identifies, for example, a part of the message that the operator notifies the user of the completion of the personal identification as the end character string. To do. Since phrases such as operator response procedures and questions have a certain regularity, the identification unit 433 identifies the end character string by identifying the end character string based on the voice response text information issued by the operator. Accuracy can be improved.

  The end character string may be a combination of a plurality of character strings for identifying a message from which the operator notifies the user that the personal identification has been completed. As an example, the first group includes “password” and “contract” as elements, the second group includes “number”, “content”, and “information” as elements, and the third group includes “confirmation” and “examination”. ”,“ Take ”and“ Inquiry ”as elements. The identification unit 433 identified the end character string on the condition that any element of the first group, any element of the second group, and any element of the third group were identified. Is determined.

  For example, when the conversion unit 432 converts the message “contract details have been inquired” into response text information, the identification unit 433 identifies the element “contract” as the first group and the element “contract” as the second group. "Content" is identified, and the element "query" or "taken" is identified as the third group. In this case, the identification unit 433 determines that the end character string has been identified because all of the elements of the first group, the second group, and the third group have been identified.

  These groups of elements may be included in the same phrase. For example, the identification unit 433 may identify the first group element “password” and the second group element “number” from the phrase “password” of the response text information.

  Further, the identification unit 433 may determine that the end character string has been identified on the condition that the elements of the two groups among the first to third groups have been identified. For example, the identification unit 433 determines that the end character string has been identified when the first group of elements “password” and the second group of elements “content” are identified among the first to third groups. May be.

  Further, the identification unit 433 identifies a start character string indicating the start of reception for confirming the identity confirmation information. Since the operator notifies the user that the identity confirmation is to be performed before starting the identity verification, the identification unit 433, for example, sends a part of a message notifying the user that the operator confirms the identity as a start character string. Identify as.

  The start character string may be a combination of a plurality of character strings for identifying a message that notifies the user that the operator confirms the identity. For example, the first group includes “name”, “contract contents”, and “ID” as elements, the second group includes “full name” as elements, and the third group includes “tell”, “confirm”, and “check”. ”As an element, the identification unit 433 recognizes that any element in the first group, any element in the second group, and any element in the third group have been identified. As a result, it is determined that the start character string is identified.

  The identification unit 433 may identify the end character string from a text portion after the identified start character string in order to suppress erroneously identifying the end character string. The identifying unit 433 notifies the identifying unit 434 of the identified end character string and start character string. In addition, when the start character string cannot be identified, the identification unit 433 notifies the specifying unit 434 that the start character string cannot be identified. Similarly, when the end character string cannot be identified, the identifying unit 433 notifies the specifying unit 434 that the end character string cannot be identified.

  The identification unit 433 may identify the start character string or the end character string based on the response text information obtained by converting the voice uttered by the user from the response text information. For example, when there is more response text information obtained by converting the voice uttered by the user than the response text information obtained by converting the voice uttered by the operator, the identification unit 433 converts the voice uttered by the user into the response text information obtained by converting the voice uttered by the user. Based on this, the start character string or the end character string may be identified.

  The identifying unit 434 identifies the front text portion that is the response text information before the start character string identified by the identifying unit 433. When the start character string is a combination of a plurality of character strings, the identifying unit 434 identifies the response text information before the forefront position among the plurality of character strings identified by the identifying unit 433 as the front text information. To do.

  Further, the specifying unit 434 specifies the rear text portion that is the response text information after the end character string identified by the identifying unit 433. When the ending character string is a combination of a plurality of character strings, the specifying unit 434 selects the response text information after the rearmost character string among the plurality of character strings identified by the identification unit 433 as the rear text information. As specified. When the identifying unit 433 cannot identify the end character string (that is, when receiving a notification from the identifying unit 433 that the end character string cannot be identified), the identifying unit 434 does not identify the rear text portion.

  FIG. 3 is a diagram showing the response text information. The storage unit 42 stores the response text information in association with the speaker. For example, the first line shows the message “Oh, is the operator” issued by the operator, and the second line is the message “I want to hear about the case of △△” issued by the user. Show.

  In the example of FIG. 3, the identifying unit 434 identifies, from the first message “I am XXX”, a message issued by the operator “Received requirements” as the front text portion. In the front text part, for example, the operator asks the user for the user's inquiry requirements.

  The identity confirmation portion is response text information sandwiched between the start character string and the end character string, and the operator confirms the identity confirmation information of the user. The identification unit 434 sends a message from the operator “Thank you for your patience, but we will confirm your identity.” To the message “The identity confirmation has been completed. Thank you.” Identified as the identity verification part.

  The identification unit 434 identifies, from the message “□□, how do you like the price plan?” From the operator to the last message “Thank you for using.” As the rear text part. In the rear text portion, for example, the operator makes a specific proposal or answer that meets the user's requirements. As described above, in the example of FIG. 3, the operator confirms the user's identity confirmation information after hearing the user's inquiry requirements, and makes a proposal or answer that meets the user's requirements after the identity confirmation is completed. For this reason, the contents of the front text part, the identity verification part, and the rear text part are different from each other.

  When the identifying unit 433 cannot identify the start character string of the response text information, the identifying unit 434 cannot identify the front text portion using the start character string. Therefore, when the identification unit 433 cannot identify the start character string of the response text information (that is, when the notification that the start character string cannot be identified is received from the identification unit 433), the specifying unit 434 starts from the beginning of the response text information. A text part up to a predetermined ratio is specified as a front text part. The ratio is, for example, the ratio of the number of characters from the beginning to the total number of characters of the response text information, the ratio of the time from the beginning to the total time, or the ratio of the number of sentences from the start to the total number of sentences. The number of sentences is the number delimited by the punctuation points in the response text information.

  As an example, the specifying unit 434 calculates the ratio of the number of characters from the beginning to the total number of characters of the response text information based on the ratio of the length of the front text portion to the total number of characters of the response text information acquired in the past. The text part up to the proportion (for example, 30%) is specified as the front text part. In this way, for example, the specifying unit 434 can recognize the front text portion even if the identifying unit 433 cannot identify the start character string due to an error in conversion of the speech information by the converting unit 432 or the like. Can be specified.

  Note that when the identifying unit 433 cannot identify the start character string, the identifying unit 434 may identify the corresponding text part from the start of the telephone reception until the predetermined time is reached as the front text part. The predetermined time can be obtained as a statistic of the time of the front text portion in a plurality of past inquiries. With such a configuration, the specifying unit 434 can specify the text portion corresponding to the predetermined time as the front text portion without being affected by the length of the response text portion other than the front text portion.

  Further, the specifying unit 434 determines that the ratio of the number of sentences from the beginning to the total number of sentences in the response text information is based on the ratio of the length of the front text portion to the total number of sentences in the response text information acquired in the past. The text part up to the calculated ratio (for example, 30%) may be specified as the front text part. Further, when the identifying unit 433 cannot identify the start character string, the identifying unit 434 may identify the corresponding text part from the start of the telephone reception until the predetermined number of sentences is reached as the front text part. The predetermined number of sentences is, for example, a statistic of the number of sentences in the past front text part.

  In the case where the identifying unit 433 cannot identify the start character string, the specifying unit 434 has a ratio of the number of times that the speaker who has spoken is switched from the start of the telephone response to the number of times that the speaker who has spoken is switched in the entire telephone response. The text part until the predetermined value is reached may be specified as the front text part. As shown in FIG. 3, the storage unit 42 stores reception text information in association with the speaker. The identification unit 434 counts the number of times the speaker is switched in the response text information. For example, the specifying unit 434 counts that the speaker is switched once when the user utters the voice after the operator utters the voice, and also speaks when the operator utters the voice after the user utters the voice. Count the person switched once.

  The specifying unit 434 obtains the number of times the speaker has been switched in the entire telephone reception from the start to the end of the telephone reception. The identification unit 434 identifies the response text part until the ratio of the number of times the speaker is switched from the start of the telephone response to the total number of telephone receptions reaches a predetermined value as the front text part. As an example, the specifying unit 434 specifies the text part until the ratio of the number of times the speaker is switched to the total number of telephone receptions reaches 30% as the front text part. If the number of times the speaker is switched in the entire telephone reception is 100 times, the specifying unit 434 specifies the response text information from the start of the telephone reception until the speaker is switched 30 times as the front text portion.

  It can be said that every time the speaker switches twice, the operator utters a voice once. Since the operator's telephone response procedure is predetermined, the ratio of the number of times the speaker in the front text part has switched to the total telephone response is compared with the percentage of the number of characters in the front text part or the percentage of time in the front text part. For example, it is difficult to be influenced by individual differences among users such as the length of a story. For this reason, the specifying unit 434 specifies the front text portion based on the ratio of the number of times the speaker is switched, thereby suppressing a reduction in accuracy in specifying the front text portion when the identifying unit 433 cannot identify the start character string. be able to.

  The extraction unit 435 extracts user information from an identification part that is a text part sandwiched between a start character string and an end character string. The extraction unit 435 extracts the user's name, address, date of birth, and the like as user information. For example, the extraction unit 435 recognizes a question issued by the operator in order to find out the user's name and the like, and extracts a word / phrase issued by the user after this question as the user's name and the like. The extraction unit 435 notifies the summary unit 436 of the extracted user information.

  The summary unit 436 generates summary data of the response text. In the example of the present specification, the summary unit 436 inputs the response text information to the learned machine learning model 437 and acquires the summary data of the response text information output by the machine learning model 437. For example, the control unit 43 functions as the machine learning model 437 by reading a library such as Summpy, which is an open source summary creation API, from the storage unit 42 and executing it. Note that an external server (not shown) may have a machine learning model 437.

  The machine learning model 437 is a learned model generated in advance by learning big data of the front text portion or the rear text portion as an input. The machine learning model 437 evaluates a phrase with a high repetition frequency as having high importance, and evaluates a phrase with a low repetition frequency as having low importance. The machine learning model 437 also includes repetition of related words in the repetition frequency. For example, the machine learning model 437 determines that the phrase “lost” and the phrase “lost” are related phrases. On the other hand, the machine learning model 437 evaluates that words included in any inquiry have low importance. For example, since the phrase “thank you” is often included in any inquiry, the machine learning model 437 evaluates the phrase “thank you” as low in importance.

  The summarization unit 436 inputs the front text portion or the rear text portion to the machine learning model 437, and outputs the summary data output from the machine learning model 437. The summary unit 436 outputs the summary data to a computer (not shown) used by an administrator who analyzes summary data, for example, via the communication unit 41. The summarizing unit 436 does not output the summary data of the person confirmation part.

  By the way, because the contents of the front text part and the back text part are different, if the summary data of the front text part and the back text part is created together, it is assumed that one of the main points is not sufficiently reflected in the summary data Is done. Therefore, the summarization unit 436 individually inputs the front text portion and the rear text portion to the machine learning model 437, and acquires the summary data of the front text portion and the rear text portion output from the machine learning model. Good. The summary unit 436 associates and outputs the summary data of the front text portion and the summary data of the back text portion.

  FIG. 4 is a diagram illustrating an example of summary data output from the summary unit 436. FIG. 4 shows summary data for the front text portion and summary data for the rear text portion. In the summary data in the front text part, “I want to know the details of the rate plan” is displayed as a summary of the user's inquiry requirements, and in the summary data in the rear text part, a summary of the results of the operator's response to the inquiry “The discount service has reported that △△ is available.” Is displayed. Since the machine learning model deletes interjections such as “Hello” in the response text information, the summary data does not include interjections.

  The summary unit 436 outputs summary data in association with the user information extracted by the extraction unit 435. For example, the summary unit 436 outputs the user's name, the summary data of the front text portion, and the summary data of the rear text portion in association with each other. With such a configuration, the administrator can manage the summary data for each user. Therefore, when there are a plurality of inquiries from the same user, a plurality of summary data associated with the user information of the user is displayed. Each can be confirmed.

  The summarizing unit 436 may associate and output type information indicating the type of the front text part and the summary data of the corresponding rear text part. The type is a classification for grouping the front text portions of a plurality of inquiries. For example, the summary unit 436 may output the inquiry type “charge plan confirmation” for confirming the details of the charge plan in association with the summary data of the rear text portion corresponding to the inquiry. The summarization unit 436 may output the service cancellation inquiry type “service cancellation” and the summary data of the rear text portion corresponding to the inquiry in association with each other.

  If the identifying unit 433 cannot identify the end character string, the summarizing unit 436 outputs summary data in association with an identifier indicating that the end character string cannot be identified. The identifying unit 434 does not identify the rear text portion when the identifying unit 433 cannot identify the end character string. In this case, the summarizing unit 436 inputs the front text portion specified by the specifying unit 434 and the response text information other than the front text portion to the machine learning model 437, and outputs the front text portion and the front side output by the machine learning model 437. Get summary data other than text. The summary unit 436 associates each summary data output from the machine learning model 437 with an identifier indicating that the end character string could not be identified so that the administrator can grasp that the rear text portion has not been specified. To the computer used by the administrator.

[Operation of Text Analyzer 400]
FIG. 5 is a flowchart showing the operation of the text analysis apparatus 400. This processing procedure starts when the recording apparatus 300 generates voice information in which the voices of the user and the operator are recorded.

  First, the acquisition unit 431 acquires voice information for telephone reception between the user and the operator from the recording device 300 via the communication unit 41 (step S101). Next, the conversion unit 432 converts the voice information acquired by the acquisition unit 431 into response text information (step S102). The identification unit 433 determines whether or not the start character string has been identified in the response text information (step S103). When the identification unit 433 can identify the start character string in the response text information (YES in S103), the identification unit 434 identifies the response text information before the identified start character string as the front text portion (step S104). ).

  The identification unit 433 determines whether or not the end character string has been identified in the response text information (step S105). When the identification unit 433 can identify the end character string in the response text information (YES in S105), the specification unit 434 specifies the response text information after the identified end character string as the rear text portion (step S104). S106). The summarizing unit 436 inputs the front text portion and the rear text portion to the learned machine learning model 437, and acquires the summary data of the front text portion and the rear text portion output from the machine learning model 437, respectively. The summarizing unit 436 associates and outputs the acquired summary data of the front text portion and the summary data of the rear text portion via the communication unit 41 (S107), and ends the processing.

  If the identification unit 433 cannot identify the start character string of the response text information in the determination in step S103 (NO in S103), the specification unit 434 determines the text portion from the beginning of the response text information to a predetermined ratio as the front text portion. (S108), and the process proceeds to step S105. When the identifying unit 433 cannot identify the end character string of the response text information in the determination in step S105 (NO in S105), the summarizing unit 436 receives the front text portion specified by the specifying unit 434 and the response other than the front text portion. The text information is input to the machine learning model 437, and the summary data of the front text portion and the summary data other than the front text portion output by the machine learning model 437 are acquired. The summarizing unit 436 associates each summary data output from the machine learning model 437 with an identifier indicating that the end character string could not be identified, and outputs it to the computer used by the administrator, and ends the process (step S109). .

[Preprocessing for machine learning]
In the present embodiment, an example has been described in which the summarizing unit 436 inputs the front text portion or the rear text portion to the learned machine learning model 437 and acquires the summary data output by the machine learning model 437. However, the present invention is not limited to an example in which reception text information is input to a learned machine learning model. For example, the text analysis device 400 may include a learning unit for generating a machine learning model.

  The learning unit learns a plurality of rear text portions specified by the specifying unit 434, thereby generating a machine learning model having the rear text portion as an input and the summary data of the rear text portion as an output. With this configuration, the learning unit improves the accuracy of summarization of the rear text part compared to the case where a machine learning model is generated by learning the entire response text information for a plurality of response text information. A model can be generated.

[Effects of text analysis device 400]
When automatic summarization is performed using a machine learning model, if response text information including an identification part is input to the machine learning model, identification information that is repeatedly referred to in the identification part is likely to be included in the summary. On the other hand, text analysis device 400 according to the present embodiment specifies a rear text portion after the end character string identified by identification unit 433 in the response text information. For this reason, the text analysis apparatus 400 can improve the accuracy of the summary content when automatically summarizing using the machine learning model.

[Modification]
In the present embodiment, an example has been described in which the text analysis apparatus 400 includes the conversion unit 432 that converts voice information into response text information. However, the present invention is not limited to this. For example, the conversion unit 432 may be provided separately from the text analysis device 400.

  As mentioned above, although this invention was demonstrated using embodiment, the technical scope of this invention is not limited to the range as described in the said embodiment, A various deformation | transformation and change are possible within the range of the summary. is there. For example, the specific embodiments of device distribution / integration are not limited to the above-described embodiments, and all or a part of them may be configured to be functionally or physically distributed / integrated in arbitrary units. Can do. In addition, new embodiments generated by any combination of a plurality of embodiments are also included in the embodiments of the present invention. The effect of the new embodiment produced by the combination has the effect of the original embodiment.

41 Communication Unit 42 Storage Unit 43 Control Unit 100 Communication Terminal 200 Telephone 300 Recording Device 400 Text Analysis Device 431 Acquisition Unit 432 Conversion Unit 433 Identification Unit 434 Identification Unit 435 Extraction Unit 436 Summarization Unit 437 Machine Learning Model

Claims (14)

A text acquisition unit that acquires response text information in which voice information of a telephone response between a user and an operator is converted into text;
An identification unit for identifying a character string indicating the start or end of reception in which the operator confirms the user identification information in the reception text information;
A specifying unit that specifies a front text part before the character string identified by the identifying unit in the response text information or a rear text part after the character string identified by the identifying unit;
Text analysis device. A summary unit that inputs the front text portion or the back text portion to a machine learning model and outputs summary data output from the machine learning model;
The text analysis apparatus according to claim 1. The identification unit identifies a start character string indicating the start of the response for confirming the identity confirmation information and an end character string indicating the end of the response for confirming the identity confirmation information,
The identification unit identifies the end character string from a text portion after the identified start character string;
The text analysis apparatus according to claim 2. The specifying unit specifies a front text portion before the start character string identified by the identification unit, and a rear text portion after the end character string identified by the identification unit,
The summary unit inputs the rear text portion and the front text portion to the machine learning model, and outputs the summary data of the rear text portion and the summary data of the front text portion in association with each other.
The text analysis apparatus according to claim 3. The specifying unit specifies a text part from the beginning of the response text information to a predetermined ratio as a front text part when the identification unit cannot identify the start character string.
The text analysis apparatus according to claim 3 or 4. The specific unit is a ratio of the number of times that a speaker who has made a voice switch from the start of a telephone response to the number of times a speaker who makes a voice is switched in the whole telephone response when the identification unit cannot identify the start character string Identifies the text part until is reached a predetermined value as the front text part,
The text analysis apparatus according to claim 3 or 4. The specifying unit specifies a text part corresponding to a predetermined text number from the start of a telephone response as a front text part when the identification unit cannot identify the start character string;
The text analysis apparatus according to claim 3 or 4. The specifying unit specifies a text part corresponding to a predetermined time from the start of telephone reception as a front text part when the identification unit cannot identify the start character string;
The text analysis apparatus according to claim 3 or 4. An extraction unit that extracts user information from a text portion sandwiched between the start character string and the end character string;
The summary unit outputs the summary data in association with the user information extracted by the extraction unit.
The text analysis apparatus according to any one of claims 3 to 8. The summary unit outputs the summary data in association with an identifier indicating that the end character string cannot be identified when the identification unit cannot identify the end character string;
The text analysis apparatus according to any one of claims 3 to 9. A conversion unit that converts voice information of a telephone reception between a user and an operator into the reception text information;
The identification unit identifies the end character string based on the response text information obtained by converting the voice uttered by an operator from the response text information.
The text analysis apparatus according to any one of claims 3 to 10. It further includes a storage unit that stores feature information of the operator's voice,
The conversion unit converts the voice information into the response text information based on the feature information of the operator's voice stored in the storage unit.
The text analysis apparatus according to claim 11. Obtaining response text information in which voice information of a telephone response between a user and an operator is converted into text;
Identifying a character string indicating the start or end of reception in which the operator confirms the user identification information in the reception text information;
Identifying a front text portion before the character string identified in the response text information or a rear text portion after the identified character string;
A text analysis method comprising: Computer
A text acquisition unit for acquiring response text information obtained by converting voice information of a telephone response between a user and an operator into text;
An identification unit for identifying a character string indicating the start or end of reception in which the operator confirms the user identification information in the reception text information, and a front side before the character string identified by the identification unit in the reception text information A specifying part for specifying a text part or a rear text part after the character string identified by the identifying part;
Text analysis program to make it function as

Images