Interaction support system and interaction support method

Abstract

To automatically generate a question and answer pair in consideration of a case where an answer sentence depends on a question sentence and a case where an answer sentence indicates questioner's emotion.SOLUTION: An interaction support system retains question and answer pair candidates including question sentence vectors, emotion vectors indicating the emotion of a question, and answer sentence vectors; when determining that the variance of the emotion vectors included in a first group based on the similarity between the answer sentence vectors and the similarity between the question sentence vectors is large, outputs question and answer pairs based on the question and answer pair candidates included in the first group as sentence-dependent question and answer pairs; and when determining that the variance of the question sentence vectors included in a second group is large based on the similarity between the answer sentence vectors and the similarity between the emotion vectors is large, outputs the question and answer pairs based on the question and answer pair candidates included in the second group as emotion-dependent question and answer pairs.SELECTED DRAWING: Figure 5

Description

[Technical field]

[0001] The present invention relates to a dialogue support system and a dialogue support method. [Background technology]

[0002] In call centers, in order to assist new operators who lack the skills required, operator support systems are used that use voice recognition to convert spoken content into text, infer answers from the text using FAQs (Frequently Asked Questions), and present the inferred answers.

[0003] Background of the present invention is JP 2020-115244 A (Patent Document 1). This publication states that "the operator response support system is a system that understands the emotions of the customer and the operator and provides the operator with an appropriate response method and proposals (answers), and has a contact center 1 connected to an end user terminal 2 via a network. The operator response support system stores the contents of the utterances of the operator and the customer and the results of the emotion analysis in a dialogue data storage device, and enables the operator management device to check the data. In the operator management device, when a specific keyword is uttered, or when a certain threshold is set for the emotional level of each of the customer and the operator and the threshold is exceeded, the appropriate response is proposed to the operator or a warning is sent to the supervisor." (See abstract). [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2020-115244 A Summary of the Invention [Problem to be solved by the invention]

[0005] The technology described in Patent Document 1 searches for and presents FAQs based on the content of an utterance and the results of sentiment analysis, but a method for automatically generating such FAQs is not disclosed in Patent Document 1. Thus, one aspect of the present invention is to automatically generate question-answer pairs that take into consideration cases where the answer depends on the question, and cases where the answer evokes the feelings of the questioner. [Means for solving the problem]

[0006] In order to solve the above-mentioned problems, one embodiment of the present invention employs the following configuration: A dialogue support system includes a processor and a memory, and the memory holds a plurality of question-answer pair candidates, each of which includes a question sentence vector indicating a sentence of a question by a questioner, a question emotion vector indicating the emotion of the questioner when the question was asked, and an answer sentence vector indicating a sentence of an answer to the question by an answerer, and question-answer pair information storing question-answer pairs, and the processor generates, from the plurality of question-answer pair candidates, a first group based on the similarity between the answer sentence vectors and the similarity between the question sentence vectors, and a second group based on the similarity between the answer sentence vectors and the similarity between the emotion vectors, and When it is determined that the variance of the emotion vectors of the question-answer pair candidates included in the first group is large based on predetermined conditions, question-answer pairs based on the question-answer pair candidates included in the first group are stored in the question-answer pair information as sentence-dependent question-answer pairs in which an answer sentence to a question depends on the sentence of the question, and when it is determined that the variance of the question sentence vectors of the question-answer pair candidates included in the second group is large based on predetermined conditions, question-answer pairs based on the question-answer pair candidates included in the second group are stored in the question-answer pair information as emotion-dependent question-answer pairs in which an answer sentence to a question depends on the question emotion corresponding to the question. Effect of the Invention

[0007] According to one aspect of the present invention, it is possible to automatically generate question-and-answer pairs that take into consideration cases in which the answer sentence depends on the question sentence and cases in which the answer sentence affects the questioner's feelings.

[0008] Problems, configurations and effects other than those described above will become apparent from the following description of the embodiments. [Brief description of the drawings]

[0009] [Figure 1] 1 is a block diagram showing an example of the configuration of a dialogue support system according to a first embodiment. [Diagram 2] 2 is a block diagram showing an example of a hardware configuration of a computer constituting each system and each terminal included in the dialogue support system in the first embodiment. FIG. [Diagram 3] 4 is a diagram illustrating an example of a data structure of a response history DB according to the first embodiment; [Figure 4] FIG. 2 is a diagram illustrating an example of a data structure of a FAQ DB according to the first embodiment. [Diagram 5] 1 is a diagram illustrating an example of a functional configuration of an FAQ generation system according to a first embodiment, and an example of an outline of processing performed by the FAQ generation system. [Figure 6] 11 is a flowchart showing an example of a question-and-answer pair extraction process in the first embodiment. [Figure 7] 13 is a flowchart illustrating an example of an FAQ integration process according to the first embodiment. [Figure 8] 1 is a diagram illustrating an example of a functional configuration of an operator support system according to a first embodiment and an example of an outline of a process performed by the operator support system; [Figure 9] 11 is a flowchart showing an example of an FAQ search process in the first embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings. In this embodiment, the same components are generally designated by the same reference numerals, and repeated explanations are omitted. Note that this embodiment is merely an example for realizing the present invention, and does not limit the technical scope of the present invention. EXAMPLES

[0011] An embodiment of the present invention will be described below with reference to Fig. 1 to Fig. 9. In this embodiment, an FAQ is generated from a conversation voice between a customer and an operator at a call center, which indicates a response corresponding to at least one of a question by a customer and the customer's emotion (questioning emotion) when the customer asks the question. The process of collecting voice for generating the FAQ, the process of generating an FAQ database from the collected voice, and the process of searching for FAQs during a conversation between a customer and an operator will be described below.

[0012] In this embodiment, FAQ does not simply mean "frequently asked questions," but is a concept including questions (not necessarily frequently asked) and answers corresponding to the questions. In particular, in this embodiment, FAQ refers to at least one of a question sentence and a question sentiment, and an answer corresponding to at least one of the questions.

[0013] 1 is a block diagram showing an example of the configuration of a dialogue support system. The dialogue support system includes, for example, a customer terminal 101, an operator terminal 102, a response history accumulation system 103, a response history DB (DataBase) 104, an FAQ generation system 105, an FAQ DB 106 (an example of question and answer pair information), and an operator support system 107.

[0014] A conversation between a customer (an example of a questioner) and an operator (an example of a respondent) takes place between the customer terminal 101 and the operator terminal 102. That is, the customer terminal 101 inputs the customer's speech from a microphone or the like and transmits it to the operator terminal 102, and the operator terminal 102 outputs the speech from a speaker or the like. The operator terminal 102 inputs the operator's speech from a microphone or the like and transmits it to the customer terminal 101, and the customer terminal 101 outputs the speech from a speaker or the like.

[0015] Furthermore, the operator terminal 102 transmits the dialogue voice between the customer and the operator to the response history accumulation system 103. The response history accumulation system 103 records the dialogue voice and stores the recorded dialogue voice (for example, in a state where the voice spoken by the customer and the voice spoken by the operator are separated) in the response history DB 104.

[0016] The FAQ generation system 105 automatically generates FAQs based on the dialogue voice stored in the response history DB 104, and stores them in the FAQ DB 106. The operator support system 107 receives the dialogue voice between the operator and the customer from the operator terminal 102, searches the FAQ DB 106 for FAQs that correspond to the question sentence by the customer and the customer's question sentiment indicated by the dialogue voice, and presents them on the operator terminal 102, thereby utilizing the FAQ DB 106 to support the operator.

[0017] In addition, response history DB 104 and FAQ DB 106 may each be a computer independent of other systems as a database server, or, for example, response history DB 104 may be stored in an auxiliary storage device included in response history accumulation system 103, and FAQ DB 106 may be stored in an auxiliary storage device included in FAQ generation system 105 or an auxiliary storage device included in operator support system 107.

[0018] In addition, at least some of the terminals and systems included in the dialogue support system may be integrated; specifically, for example, at least some of the operator terminal 102, the response history accumulation system 103, the FAQ generation system 105, and the operator support system 107 may be integrated.

[0019] 2 is a block diagram showing an example of the hardware configuration of a computer constituting each system and each terminal included in the dialogue support system. The computer 100 includes, for example, a CPU (Central Processing Unit) 110, a memory 120, an auxiliary storage device 130, an input device 140, an output device 150, and a communication device 160.

[0020] CPU 110 includes a processor and executes programs stored in memory 120. Memory 120 includes ROM (Read Only Memory), which is a non-volatile storage element, and RAM (Random Access Memory), which is a volatile storage element. ROM stores immutable programs (e.g., BIOS (Basic Input / Output System)) and the like. RAM is a high-speed, volatile storage element such as DRAM (Dynamic Random Access Memory), and temporarily stores programs executed by CPU 110 and data used when the programs are executed.

[0021] The auxiliary storage device 130 is a large-capacity non-volatile storage device such as a magnetic storage device (HDD (Hard Disk Drive)) or a flash memory (SSD (Solid State Drive)), and stores programs executed by the CPU 110 and data used when the programs are executed. That is, the programs are read from the auxiliary storage device 130, loaded into the memory 120, and executed by the CPU 110.

[0022] In addition, some or all of the functions of the functional units included in each system included in the dialogue support system may be realized by hardware such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field-Programmable Gate Array).

[0023] Furthermore, the information used by the dialogue support system is not data structure dependent and may be represented in any data structure. For example, the information may be stored in a data structure appropriately selected from a table, a list, a database, or a queue.

[0024] The input device 140 is a device such as a keyboard or a mouse that receives input from an operator. The input device 140 of the computer 100 constituting the customer terminal 101 and the computer 100 constituting the operator terminal 102 may include a microphone for inputting voice. The output device 150 is a device such as a display device or a printer that outputs the results of program execution in a format that can be viewed by the operator. The input device 140 of the computer 100 constituting the customer terminal 101 and the computer 100 constituting the operator terminal 102 may include a speaker for outputting voice.

[0025] The communication device 160 is a network interface device that controls communication with other devices in accordance with a predetermined protocol, and includes a serial interface such as a Universal Serial Bus (USB).

[0026] A part or all of the programs executed by the CPU 110 may be provided to the computer 100 from a removable medium (such as a CD-ROM or a flash memory) which is a non-transitory storage medium, or from an external computer equipped with a non-transitory storage device via a network, and may be stored in the non-volatile auxiliary storage device 130 which is a non-transitory storage medium. Therefore, the computer 100 may be provided with an interface device which reads data from the removable medium.

[0027] Each system and each terminal included in the dialogue support system is a computer system configured on a single physical computer 100, or on multiple logically or physically configured computers 100, and may operate in separate threads on the same computer 100, or may operate on a virtual computer constructed on multiple physical computer resources.

[0028] 3 is a diagram showing an example of the data structure of the response history DB 104. The response history DB 104 includes, for example, an ID column 21, a customer voice column 22, an operator voice column 23, an operator number column 24, and a response evaluation column 25. The ID column 21 holds an ID for identifying the response history. The customer voice column 22 holds an audio file of the voice (voice uttered by the customer) input from the customer terminal 101 in the response history. The operator voice column 23 holds an audio file of the voice (voice uttered by the operator) input from the operator terminal 102 in the response history.

[0029] The operator number column 24 holds an operator number that identifies an operator who handled a customer in the handling history. The handling evaluation column 25 holds information indicating an evaluation of the customer handling by the operator in the handling history. In the example of Fig. 3, for example, one of two types of values, "good" and "bad", is stored in each handling evaluation column 25. Note that the handling evaluation column 25 may store an evaluation value indicating an evaluation of the customer handling using a predetermined number of stages.

[0030] 4 is a diagram showing an example of the data structure of the FAQ DB 106. The FAQ DB 106 includes, for example, an ID column 31, a title column 32, a question column 33, a question column 34, a column 35 dependent on a sentence, an emotion column 36, an emotion vector column 37, an emotion dependent flag column 38, an answer column 39, and an answer column 40. The ID column 31 holds an ID for identifying an FAQ. The title column 32 holds a character string indicating the title of the FAQ.

[0031] The question text field 33 holds a character string indicating the question text of the FAQ. The question text vector field 34 holds a question text vector obtained by vectorizing the question text. The text dependency flag field 35 holds a text dependency flag indicating whether the answer text depends on the question text (i.e., whether the answer text is determined based on the question text).

[0032] If the sentence dependent flag is "1", it indicates that the answer sentence depends on the question sentence, and if the sentence dependent flag is "0", it indicates that the answer sentence does not depend on the question sentence. Note that in a record in which the value of the sentence dependent flag column 35 of the FAQ DB 106 is "0", the question sentence may not be registered (that is, no value or a null value may be stored in the question sentence column 33 and the question sentence vector column 34).

[0033] The emotion column 36 holds a character string indicating the question emotion (i.e., the emotion of the customer when asking the question with the question sentence). The emotion vector column 37 holds an emotion vector obtained by vectorizing the question emotion. The emotion dependent flag column 38 holds an emotion dependent flag indicating whether the answer sentence depends on the question emotion (i.e., whether the answer sentence is determined based on the question emotion).

[0034] If the emotion-dependent flag is "1", it indicates that the answer sentence depends on the question emotion, and if the emotion-dependent flag is "0", it indicates that the answer sentence does not depend on the question emotion. Note that in a record in which the emotion-dependent flag column 38 in the FAQ DB 106 has a value of "0", the question emotion may not be registered (that is, no value or a null value may be stored in the emotion column 36 and the emotion vector column 37).

[0035] The answer sentence column 39 holds character strings indicating answer sentences corresponding to the question emotions indicated in the emotion column 36 and the question sentences indicated in the question sentence column 33. The answer sentence vector column 40 holds answer sentence vectors obtained by vectorizing the answer sentences.

[0036] 5 is a diagram showing an example of the functional configuration of the FAQ generation system 105 and an example of an outline of the processing by the FAQ generation system 105. The FAQ generation system 105 has, for example, a response selection unit 55, a voice recognition unit 42, an emotion recognition unit 46, a sentence vector conversion unit 44, a question-and-answer pair extraction unit 48, and an FAQ combination unit 54, which are all functional units. In addition, the FAQ generation system 105 is connected to the response history DB 104 and FAQ DB 106 described above.

[0037] 5 are included in a CPU 110 of a computer 100 constituting the FAQ generation system 105. For example, the CPU 110 of the computer 100 constituting the FAQ generation system 105 functions as a response selection unit 55 by operating according to a response selection program loaded into a memory 120 of the computer 100, and functions as a voice recognition unit 42 by operating according to a voice recognition program loaded into the memory 120.

[0038] For the other systems included in the dialogue support system, the relationship between the CPU 110 of the computer 100 constituting the system, the programs loaded into the memory 120 of the computer 100, and the functional units included in the system is similar.

[0039] An overview of the processing by the FAQ generation system 105 will be described. The response selection unit 55 extracts combinations of operator voices and customer voices from the response history DB 104. At this time, the response selection unit 55 can create an exemplary FAQ by extracting only combinations of operator voices and customer voices corresponding to experienced operators (for example, the FAQ generation system 105 accepts input of the experienced operator number from the user via the input device 140, or the FAQ generation system 105 holds in advance data in which the experienced operator numbers are registered) based on the operator number, or by extracting only combinations of operator voices and customer voices corresponding to good customer ratings (customer rating value is "good").

[0040] Furthermore, the response selection unit 55 may create FAQs for examples of poor customer responses by extracting only combinations of operator voice and customer voice that correspond to unskilled operators based on the operator number, or by extracting only combinations of operator voice and customer voice that correspond to poor customer ratings (customer rating values ​​that are "bad").

[0041] The response selection unit 55 inputs the extracted operator voice and customer voice to the voice recognition unit 42, and the voice recognition unit 42 generates an operator-customer sentence 43 including text transcribed from the operator voice and the customer voice.

[0042] The speech recognition unit 42 inputs the operator-customer sentence 43 to the sentence vector conversion unit 44, and the sentence vector conversion unit 44 converts it into an operator-customer sentence vector 45 (an example of a dialogue sentence vector) that includes sentence vectors corresponding to the text of the operator voice and the text of the customer voice contained in the operator-customer sentence 43.

[0043] The response selection unit 55 inputs the extracted operator voice and customer voice to the emotion recognition unit 46, and the emotion recognition unit 46 generates an operator / customer emotion 47 including emotion vectors corresponding to the operator voice and the customer voice, respectively.

[0044] The sentence vector conversion unit 44 inputs the operator-customer sentence vector 45 to the question-answer pair extraction unit 48, and the emotion recognition unit 46 inputs the operator-customer emotion 47 to the question-answer pair extraction unit 48, which generates an FAQ candidate 49 (one example of a question-answer pair candidate) including a question sentence vector 50, an answer sentence vector 51, and an emotion vector 52. Note that, although details will be described later, there may be an FAQ candidate 49 that does not include the question sentence vector 50. Also, in the example of FIG. 5, the FAQ generation system 105 generates the FAQ candidate 49, but the FAQ candidate 49 may be provided by manual input.

[0045] The FAQ candidates 49 generated by the question and answer pair extraction unit 48 include FAQs with overlapping contents. To eliminate the overlaps, the question and answer pair extraction unit 48 inputs the FAQ candidates 49 to the FAQ integrating unit 53, which then removes the FAQs with the overlapping contents to generate a formal FAQ. Details of the question and answer extraction process performed by the question and answer pair extraction unit 48 will be described later with reference to FIG. 6.

[0046] The FAQ integration section 53 stores the generated formal FAQ in the FAQ DB 106. Details of the FAQ integration process by the FAQ integration section 53 will be described later with reference to Fig. 7. The above is an overview of the series of processes by the FAQ generation system 105.

[0047] The voice recognition unit 42 receives voice as input and outputs text, and generates operator / customer sentences 43 from the operator / customer voices 41 by automatic voice recognition using deep learning. The voice recognition unit 42 may also generate operator / customer sentences 43 from the operator / customer voices 41 by other machine learning methods, or may obtain operator / customer sentences 43 transcribed manually by input via the input device 140.

[0048] In addition, the emotion recognition unit 46 inputs, for example, a voice and outputs a vector that numerically represents the degree of emotions such as joy, anger, sadness, and happiness, and generates operator / customer emotions 47 that indicate the transition of the operator's emotion vector and the transition of the customer's emotion vector (an example of the questioner's emotion) from the operator / customer voice 41 through automatic voice emotion recognition using deep learning.

[0049] In addition, the emotion recognition unit 46 may generate operator / customer emotions 47 from the operator / customer voices 41 by other machine learning, or may obtain operator / customer emotions 47 by inputting manually assigned emotion labels via the input device 140.

[0050] In addition, the sentence vector conversion unit 44 generates an operator-customer sentence vector 45 by, for example, performing morphological analysis on the text of the operator voice and the text of the customer voice contained in the operator-customer sentence 43 and generating a onehot vector reflecting the results of the morphological analysis.

[0051] 6 is a flowchart showing an example of the question-and-answer pair extraction process. The question-and-answer pair extraction unit 48 divides the dialogue data indicated by the operator-customer sentence vector 45 into one or more speech elements using a speech segment detection technique or the like (S601). Each of the one or more speech elements is represented by, for example, a onehot vector reflecting the result of the above-mentioned morphological analysis.

[0052] The question-and-answer pair extraction unit 48 extracts an utterance element corresponding to a question or an utterance element corresponding to an expression of emotion from the utterance elements included in the dialogue data (S602). Specifically, for example, the question-and-answer pair extraction unit 48 searches for the one or more utterance elements into which the dialogue is divided in order from the beginning (in order of the oldest uttered time) until it extracts an utterance element corresponding to a question or an utterance element corresponding to an expression of emotion (if it is not possible to extract these utterance elements, it searches until the last utterance element of the dialogue).

[0053] Furthermore, in step S602, the question and answer pair extraction unit 48 extracts questions from the customer's speech elements based on the onehot vector, and extracts speech elements at times when the change in the customer's emotion vector indicated by the operator / customer emotion 47 is large (for example, the change in the norm of the emotion vector is equal to or greater than a predetermined value) as an expression of emotion.

[0054] The question-and-answer pair extraction unit 48 determines whether a question or an emotional expression has been extracted in the most recent step S602 (S603). If the question-and-answer pair extraction unit 48 determines that a question or an emotional expression has not been extracted (S603: NO), the question-and-answer pair extraction unit 48 ends the question-and-answer pair extraction process.

[0055] When it is determined that a question or an expression of emotion has been extracted (S603: YES), the question-and-answer pair extraction unit 48 extracts an answer to the question or the expression of emotion (S604). Specifically, for example, the question-and-answer pair extraction unit 48 extracts, as the answer, an utterance element corresponding to an utterance by the operator immediately after the utterance element corresponding to the question or the expression of emotion.

[0056] Furthermore, the answer to the question or the expression of emotion may include an action other than the speech by the operator. For example, when a customer and an operator have a conversation over the phone, the question-and-answer pair extraction unit 48 may extract the action of the operator putting the phone on hold immediately after the speech corresponding to the question or the expression of emotion (the FAQ generation system 105 is assumed to hold information on the action taken by the operator in advance).

[0057] The question-and-answer pair extraction unit 48 determines whether an answer to the question or emotional expression has been extracted in the most recent step S605 (S606). If the question-and-answer pair extraction unit 48 determines that the answer has been extracted (S606: YES), it outputs the FAQ candidate 49 (S607).

[0058] Specifically, for example, the question-and-answer pair extraction unit 48 outputs, as an FAQ candidate 49, a combination of a question sentence vector 50 corresponding to the question extracted in the most recent step S602, an answer sentence vector 51 corresponding to the answer extracted in the most recent step S605, and an emotion vector 52 of the customer when the question extracted in step S602 was asked or when emotion was expressed.

[0059] When only the expression of emotion is extracted in the most recent step S602, the question-and-answer pair extraction unit 48 may not include the question sentence vector 50 in the FAQ candidate 49 in step S606, or may include in the FAQ candidate 49, as the question sentence vector 50, a sentence vector indicated by an utterance element of the customer at the time when the emotion is expressed.

[0060] The question-and-answer pair extraction unit 48 deletes from the dialogue data the customer's speech elements from the first utterance element to the utterance element corresponding to the question or expression of emotion extracted in step S602, and deletes the operator's speech elements from the first utterance element to the utterance element corresponding to the answer extracted in step S604 (S607), and returns to the process of step S602. Note that if the question-and-answer pair extraction unit 48 is unable to extract an answer in step S604, it may not delete any of the operator's utterance elements in step S607, or may delete all of the operator's utterance elements.

[0061] 7 is a flowchart showing an example of the FAQ integration process. The FAQ integration unit 53 receives an input of FAQ candidates 49 from the question and answer pair extraction unit 48 (S701). The FAQ integration unit 53 extracts FAQs indicated by each record (i.e., a combination of a question sentence vector, an answer sentence vector, and an emotion vector) from the FAQ DB 106 (S702). Hereinafter, the FAQs extracted in step S702 are also referred to as existing FAQs, and the FAQ candidates 49 and the existing FAQs are also collectively referred to as integrated FAQ candidates. After extracting the existing FAQs from the FAQ DB 106 in step S702, the FAQ integration unit 53 may delete all records in the FAQ DB 106.

[0062] The FAQ integration unit 53 groups the integrated FAQ candidates having similar answer sentence vectors (S703). Specifically, for example, the FAQ integration unit 53 calculates the cosine similarity between the answer sentence vectors for each combination of integrated FAQ candidates, and determines that a combination having a cosine similarity equal to or greater than a predetermined threshold is a combination having similar answer sentence vectors. Hereinafter, each group of integrated FAQ candidates generated in step S703 is also referred to as a large group.

[0063] For each of the large groups generated in step S703, the FAQ integration unit 53 divides the integrated FAQ candidates included in the large group into small groups whose question sentence vectors are similar (S704). Specifically, for example, the FAQ integration unit 53 calculates the cosine similarity between the question sentence vectors for each combination of integrated FAQ candidates included in the large group, and determines that combinations whose cosine similarity is equal to or greater than a predetermined threshold value are combinations whose question sentence vectors are similar.

[0064] The FAQ integration unit 53 calculates the variance of the emotion vectors of the integrated FAQ candidates included in each small group generated in step S704 (each small group is an example of a first group), and determines whether there is any small group for which the calculated variance is sufficiently large (e.g., larger than a predetermined threshold value) (S705).

[0065] Note that a large variance in the emotion vectors in the small group generated in step S704 indicates that although the similarity between the answer sentences and the question sentences is high in the integrated FAQ candidates included in the small group, the question sentences are made with diverse question emotions; in other words, the answer sentences included in the small group depend on the question sentence but do not depend on the question emotion.

[0066] When the FAQ integration unit 53 determines that there is a small group with a sufficiently large variance (S705: YES), it assigns an ID to each small group with a sufficiently large variance to identify the FAQ, integrates the integrated FAQ candidates included in the small group into one FAQ, stores the one FAQ in the FAQ DB 106 as an FAQ with a sentence-dependent flag of "1" and an emotion-dependent flag of "0" (an example of a sentence-dependent question-answer pair) (S706), and transitions to step S707. Note that the FAQ integration unit 53 deletes the integrated FAQ candidates included in the small groups corresponding to each of the FAQs integrated in step S706 from the large group.

[0067] In step S706, for example, the FAQ integration unit 53 determines the average of the question sentence vectors, the average of the emotion vectors, and the average of the answer sentence vectors of the integrated FAQ candidates included in the small group with the sufficiently large variance as the question sentence vector, the emotion vector, and the answer sentence vector of one integrated FAQ (i.e., to be stored in the FAQ DB 106), respectively.

[0068] Also, in step S706, for example, the FAQ integration unit 53 may determine the question sentence vector closest to the average among the question sentence vectors, the emotion vector closest to the average among the emotion vectors, and the answer sentence vector closest to the average among the answer sentence vectors of the integrated FAQ candidates included in the small group with the sufficiently large variance, as the question sentence vector, emotion vector, and answer sentence vector of one integrated FAQ (i.e., to be stored in the FAQ DB 106), respectively.

[0069] Also, in step S706, for example, the FAQ integration unit 53 may generate a question vector indicating a union or intersection (with words as elements) of utterances corresponding to the questions by synthesizing question vectors of integrated FAQ candidates included in a small group with a sufficiently large variance, and may determine the generated question vector as the question vector of one integrated FAQ (i.e., to be stored in the FAQ DB 106). The same applies to the emotion vector and the answer vector.

[0070] Furthermore, for example, for each small group with a sufficiently large variance, the FAQ integration unit 53 determines, as the question message, emotion, and answer message of one integrated FAQ (i.e., to be stored in the FAQ DB 106), the question message corresponding to the question message vector closest to the average among the question message vectors of the integrated FAQ candidates included in the small group with a sufficiently large variance, the emotion corresponding to the emotion vector closest to the average among the emotion vectors, and the answer message corresponding to the answer message vector closest to the average among the answer message vectors.

[0071] Also, for example, the FAQ integration unit 53 accepts, via the input device 140, an input of the title of the FAQ integrated for each small group with a sufficiently large variance, and thereby determines the title to be stored in the FAQ DB 106.

[0072] Note that since the emotion-dependent flag of the FAQ stored in the FAQ DB 106 in step S706 is “0” (that is, the answer in the FAQ does not depend on the emotion of the question in the FAQ), the FAQ integration unit 53 may store the emotion and emotion vector of the FAQ integrated in step S706 in the FAQ DB 106 as null values ​​or blanks.

[0073] If the FAQ integration unit 53 determines that there is no small group with a sufficiently large variance (S705: NO), the FAQ integration unit 53 transitions to step S707. The FAQ integration unit 53 disbands each small group that is not determined to have a sufficiently large variance in step S705 (S707).

[0074] The FAQ integration unit 53 divides, for each large group, integrated FAQ candidates contained in that large group that have similar emotion vectors into small groups based on the cosine similarity in the same manner as in step S704 (S708).

[0075] The FAQ integration unit 53 calculates the variance of the question sentence vectors of the integrated FAQ candidates included in each small group generated in step S708 (each small group is an example of a second group), and determines whether there is a small group for which the calculated variance is sufficiently large (e.g., larger than a predetermined threshold value) (S709).

[0076] Note that a large variance in question sentence vectors in the small group generated in step S708 indicates that, although the similarity between answer sentences and question sentiments is high in the integrated FAQ candidates included in the small group, a variety of question sentences have been uttered for the question sentiment; in other words, the answer sentences included in the small group depend on the question sentiment but do not depend on the question sentence.

[0077] When the FAQ integration unit 53 determines that there is a small group with a sufficiently large variance (S709: YES), it assigns an ID to each small group with a sufficiently large variance to identify the FAQ, integrates the integrated FAQ candidates included in the small group into one FAQ, stores the one FAQ in the FAQ DB 106 as an FAQ with a sentence-dependent flag of "0" and an emotion-dependent flag of "1" (an example of an emotion-dependent question-answer pair) (S710), and transitions to step S711. Note that the FAQ integration unit 53 deletes the integrated FAQ candidates included in the small groups corresponding to each of the FAQs integrated in step S710 from the large group.

[0078] The method by which the FAQ integrating unit 53 integrates the integrated FAQ candidates in step S710 is similar to the method by which the integrated FAQ candidates are integrated in step S706, and therefore a description thereof will be omitted.

[0079] In addition, since the text existence flag of the FAQ stored in the FAQ DB 106 in step S710 is "0" (that is, the answer in the FAQ does not depend on the question in the FAQ), the FAQ integration unit 53 may store the question and question vector of the FAQ integrated in step S710 in the FAQ DB 106 as null values ​​or blanks.

[0080] If the FAQ integration unit 53 determines that there is no small group with a sufficiently large variance (S709: NO), the FAQ integration unit 53 transitions to step S711. The FAQ integration unit 53 disbands each small group that is not determined to have a sufficiently large variance in step S709 (S711).

[0081] The FAQ integration unit 53 divides, for each large group, integrated FAQ candidates contained in the large group that are similar in both question sentence vectors and emotion vectors into small groups based on the cosine similarity in the same manner as in step S704 (S712).

[0082] The FAQ integration unit 53 assigns an ID to each of the small groups generated in step S712 (each small group is an example of a third group) to identify the FAQ, integrates the integrated FAQ candidates included in the small group into one FAQ, and stores the one FAQ in the FAQ DB 106 as an FAQ with a sentence-dependent flag of "1" and an emotion-dependent flag of "1" (an example of a sentence-emotion-dependent question-answer pair) (S713), and terminates the FAQ integration process.

[0083] In step S712, the FAQ integration unit 53 may generate a small group consisting of a single integrated FAQ candidate that does not have a similar question sentence vector and emotion vector. Also, the method of integrating the integrated FAQ candidates in step S713 by the FAQ integration unit 53 is similar to the method of integrating the integrated FAQ candidates in step S706, and therefore the description thereof will be omitted.

[0084] The FAQ integration unit 53 integrates FAQs having similar at least one of the question text, answer text, and question sentiment through the above-mentioned FAQ integration process, and automatically determines whether the integrated FAQ depends on the question sentiment or the question text, or on both the question sentiment and the question text, thereby making it possible to generate an FAQ DB 106 that can appropriately handle partial matches of the question sentiment and / or the question text.

[0085] In the above example, the FAQ integration unit 53 groups integrated FAQ candidates having similar answer sentence vectors into a large group in step S703, but may group integrated FAQ candidates having similar question sentences into a large group in step S703. In this case, the FAQ integration unit 53 groups integrated FAQ candidates having similar answer sentence vectors into small groups in step S704, determines whether the variance of the answer sentence vectors is sufficiently large in step S709, disbands small groups having small variance of the answer sentence vectors in step S711, and groups integrated FAQ candidates having similar answer sentence vectors and emotion vectors into small groups in step S712.

[0086] Furthermore, in the above example, the FAQ integration unit 53 first executes a series of processes from steps S704 to S706 and then executes a series of processes from steps S707 to S710, but it may also execute a series of processes from steps S707 to S710 first and then execute a series of processes from steps S704 to S706.

[0087] Furthermore, the FAQ integration unit 53 may calculate a numerical value indicating the degree of sentence dependency and a numerical value indicating the degree of emotion dependency and store them in the FAQ DB 106. For example, the FAQ integration unit 53 calculates the reciprocal of the variance of the vector not used for small grouping as these numerical values. That is, for example, the FAQ integration unit 53 calculates the reciprocal of the variance of the question sentence vector of the integrated FAQ candidate to be integrated in step S706 as a numerical value indicating the degree of sentence dependency, and calculates the reciprocal of the variance of the emotion vector of the integrated FAQ candidate to be integrated in step S710 as a numerical value indicating the degree of emotion dependency.

[0088] In addition, for example, in step S713, the FAQ integration unit 53 may calculate the inverse of the variance of the question sentence vectors of the integrated FAQ candidates to be integrated as a numerical value indicating the degree of sentence dependency, and may calculate the inverse of the variance of the emotion vectors of the integrated FAQ candidates as a numerical value indicating the degree of emotion dependency.

[0089] 8 is a diagram showing an example of a functional configuration of the operator support system 107 and an example of an outline of processing by the operator support system 107. The operator support system 107 includes, for example, a voice recognition unit 62, a sentence vector conversion unit 64, an emotion recognition unit 66, and an FAQ search unit 68, all of which are functional units.

[0090] The voice recognition unit 62 selects and acquires the question voice 61 to be searched that has been input to the operator support system 107 by the operator terminal 102, and generates a question sentence 63 from the question voice 61 using, for example, a method similar to the above-mentioned voice recognition method by the voice recognition unit 42 of the FAQ generation system 105.

[0091] The sentence vector conversion unit 44 converts the question sentence 63 into a question sentence vector 65 using an algorithm similar to that used by the sentence vector conversion unit 44 of the FAQ generation system 105, and inputs it to the FAQ search unit 68. The emotion recognition unit 66 generates a question emotion 67 (emotion vector) from the question voice 61 using an algorithm similar to that used by the emotion recognition unit 46 of the FAQ generation system 105, and inputs it to the FAQ search unit 68.

[0092] The FAQ search unit 68 searches the FAQ DB 106 based on the question sentence vector 65 and the question sentiment 67, and outputs the FAQ to the operator terminal 102. Details of the FAQ search process by the FAQ search unit 68 will be described later with reference to Fig. 9. The operator terminal 102 displays the FAQ output by the FAQ search unit 68 on the output device 150 of the operator terminal 102.

[0093] 9 is a flowchart showing an example of an FAQ search process. The FAQ search unit 68 receives an input of a question sentence vector 65 and a question emotion 67 (S901). The FAQ search unit 68 extracts records whose sentence-dependent flag is "1" and whose emotion-dependent flag is "0" from the FAQ DB 106 as a sentence-dependent FAQ list (S902).

[0094] The FAQ search unit 68 calculates the similarity (for example, cosine similarity, an example of the first similarity) between each of the question sentence vectors included in the sentence-dependent FAQ list and the question sentence vector input in step S901 (S903).

[0095] The FAQ search unit 68, for example, calculates a value by multiplying the highest similarity value calculated in step S903 by a sentence-dependent coefficient (a predetermined weight corresponding to sentence dependency), and temporarily stores the calculated value and the FAQ corresponding to the highest similarity included in the sentence-dependent FAQ list in the memory 120 of the operator support system 107 (S904).

[0096] The FAQ search unit 68 extracts records whose sentence-dependent flag is "0" and whose emotion-dependent flag is "1" from the FAQ DB 106 as an emotion-dependent FAQ list (S905). The FAQ search unit 68 calculates the similarity (for example, cosine similarity, an example of the second similarity) between each emotion vector included in the emotion-dependent FAQ list and the question emotion (emotion vector) input in step S901 (S906).

[0097] The FAQ search unit 68, for example, calculates a value by multiplying the highest similarity value calculated in step S906 by an emotion-dependent coefficient (a predetermined weight corresponding to emotion dependence), and temporarily stores the calculated value and the FAQ corresponding to the highest similarity included in the emotion-dependent FAQ list in the memory 120 of the operator support system 107 (S907).

[0098] The FAQ search unit 68 extracts records whose sentence-dependent flag is "1" and whose emotion-dependent flag is "1" from the FAQ DB 106 as a sentence-and-emotion-dependent FAQ list (S908). The FAQ search unit 68 calculates the similarity (for example, cosine similarity, an example of the third similarity) between each question sentence vector included in the sentence-and-emotion-dependent FAQ list and the question sentence input in step S901 (S909).

[0099] In addition, in step S909, the FAQ search unit 68 may calculate the similarity between each emotion vector included in the sentence-emotion dependent FAQ list and the question emotion (emotion vector) input in step S901, or may calculate a combined similarity between the question sentences and the emotion vectors.

[0100] The FAQ search unit 68, for example, calculates a value by multiplying the highest similarity value calculated in step S909 by a sentence / emotion dependency coefficient (a predetermined weight corresponding to sentence / emotion dependency), and temporarily stores the calculated value and the FAQ corresponding to the highest similarity included in the sentence / emotion dependent FAQ list in the memory 120 of the operator support system 107 (S910).

[0101] The FAQ search unit 68 outputs the FAQ corresponding to the highest one of the coefficient-multiplied similarities temporarily stored in steps S904, S907, and S910 to the operator terminal 102 (S911), and ends the FAQ search process.

[0102] In addition, in steps S904, S907, and S910, the FAQ search unit 68 may temporarily store similarities (or values ​​multiplied by a coefficient) that are equal to or greater than a predetermined value and FAQs corresponding to the similarities (or values ​​multiplied by a coefficient) that are equal to or greater than the predetermined value, or may temporarily store a predetermined number of similarities (or values ​​multiplied by a coefficient) in descending order of value and FAQs corresponding to the predetermined number of similarities (or values ​​multiplied by a coefficient).

[0103] Furthermore, the FAQ search unit 68 may compare and temporarily store the similarities without multiplying the similarities by a coefficient in steps S904, S907, and S910.

[0104] Furthermore, in step S911, the FAQ search unit 68 may output to the operator terminal 102 all of the FAQs temporarily stored in steps S904, S907, and S910.

[0105] Also, in step S911, the FAQ search unit 68 may, for example, transmit all column values ​​of a record corresponding to the FAQ in the FAQ DB 106 as the FAQ to be transmitted to the operator terminal 102, or may transmit at least one of the title, question sentence, emotion, and answer sentence indicated by the record (however, it is desirable that the answer sentence is always included). Also, in step S911, the FAQ search unit 68 may transmit to the operator terminal 102 the values ​​of the sentence-dependent flag and emotion-dependent flag indicated by the record.

[0106] By executing the above-mentioned FAQ search process, the FAQ search unit 68 can search for FAQs that match the question text and question sentiment from the customer and present them on the operator terminal 102. In particular, the FAQ search unit 68 can search for an appropriate FAQ from among FAQs that depend only on the question text, FAQs that depend only on the question sentiment, and FAQs that depend on the question text and the question sentiment.

[0107] The present invention is not limited to the above-mentioned embodiment, and various modifications are included. For example, the above-mentioned embodiment has been described in detail to clearly explain the present invention, and is not necessarily limited to those having all the configurations described. It is also possible to replace a part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add the configuration of another embodiment to the configuration of one embodiment. It is also possible to add, delete, or replace a part of the configuration of each embodiment with another configuration.

[0108] In addition, the above-mentioned configurations, functions, processing units, processing means, etc. may be realized in part or in whole by hardware, for example, by designing them as integrated circuits. In addition, the above-mentioned configurations, functions, etc. may be realized in software by a processor interpreting and executing a program that realizes each function. Information such as the program, table, file, etc. that realizes each function can be stored in a memory, a recording device such as a hard disk or SSD (Solid State Drive), or a recording medium such as an IC card, SD card, or DVD.

[0109] In addition, the control lines and information lines shown are those that are considered necessary for the explanation, and not all control lines and information lines in the product are necessarily shown. In reality, it can be considered that almost all components are connected to each other. [Explanation of symbols]

[0110] 48 question-answer pair extraction unit, 53 FAQ integration unit, 68 FAQ search unit, 100 computer, 104 response history DB, 105 FAQ generation system, 106 FAQ DB, 107 operator support system, 110 CPU, 120 memory, 130 auxiliary storage device, 140 input device, 150 output device, 160 communication device

Claims

1. A dialogue support system, A processor and a memory, The memory includes: a plurality of question and answer pair candidates each including a question sentence vector indicating a sentence of a question by a questioner, a question emotion vector indicating the emotion of the questioner when the question was asked, and an answer sentence vector indicating a sentence of an answer by an answerer to the question; A question and answer pair information storing the question and answer pairs is stored; The processor, generating a first group based on a similarity between the answer sentence vectors and a similarity between the question sentence vectors from the plurality of question-answer pair candidates, and a second group based on a similarity between the answer sentence vectors and a similarity between the emotion vectors; when it is determined that the variance of the emotion vectors of the question-answer pair candidates included in the first group is large based on a predetermined condition, storing the question-answer pairs based on the question-answer pair candidates included in the first group in the question-answer pair information as sentence-dependent question-answer pairs in which an answer sentence to a question depends on a sentence of the question; a dialogue support system that, when it is determined that the variance of the question sentence vectors of the question-answer pair candidates included in the second group is large based on predetermined conditions, stores the question-answer pairs based on the question-answer pair candidates included in the second group in the question-answer pair information as emotion-dependent question-answer pairs in which the answer sentence to a question depends on the question emotion corresponding to the question.

2. The dialogue support system according to claim 1, The processor, generating a third group from the plurality of candidate question-answer pairs based on a similarity between the answer sentence vectors, a similarity between the question sentence vectors, and a similarity between the emotion vectors; A dialogue support system that stores question and answer pairs based on the question and answer pair candidates included in the third group in the question and answer pair information as sentence / emotion-dependent question and answer pairs in which an answer sentence to a question depends on the sentence of the question and the question emotion corresponding to the question.

3. The dialogue support system according to claim 2, The processor, A search instruction including a question vector and a sentiment vector is received; a first similarity between a question sentence vector of the sentence-dependent question-answer pair stored in the question-answer pair information and a question sentence vector included in the search instruction; a second similarity between the emotion vector of the emotion-dependent question-answer pair stored in the question-answer pair information and the emotion vector included in the search instruction; selecting a question and answer pair as a search result for the search instruction from the question and answer pair information based on a third similarity between a question sentence vector of the sentence / emotion-dependent question and answer pair stored in the question and answer pair information and a question sentence vector included in the search instruction; A dialogue support system that outputs an answer sentence indicated by an answer sentence vector included in the selected question-answer pair.

4. The dialogue support system according to claim 1, the memory holds a dialogue sentence vector indicating a sentence of a dialogue between the questioner and the answerer, and a questioner emotion indicating an emotion of the questioner when an utterance is being made by the questioner in the dialogue; The processor, extracting an utterance indicating a question by the questioner in the dialogue or an expression of the questioner's emotion in the dialogue based on the dialogue text vector and the questioner's emotion; extracting an utterance indicating a response by the respondent corresponding to the extracted utterance or the expression of the emotion based on the dialogue vector; determining a vector indicating a sentence of an utterance by the questioner indicated by the dialogue sentence vector when the extracted utterance or the extracted emotion is expressed as the question sentence vector in the candidate question and answer pair; determining a vector indicating the questioner's emotion at the time when the extracted utterance or the extracted emotion was expressed as the emotion vector in the candidate question and answer pair; A dialogue support system that determines a vector indicating a sentence of an utterance indicating an answer by the answerer, which has been extracted, as the answer sentence vector in the candidate question and answer pair.

5. The dialogue support system according to claim 1, The processor, when storing the sentence-dependent question-answer pair in the question-answer pair information, calculating a sentence dependency degree indicating a degree to which an answer sentence to a question in the sentence-dependent question-answer pair depends on the question sentence based on the reciprocal of the variance of the question sentence vector included in the first group, and storing the sentence dependency degree in the question-answer pair information; When storing the emotion-dependent question and answer pair in the question and answer pair information, a degree of emotion dependency indicating the degree to which an answer sentence to a question in the emotion-dependent question and answer pair depends on the question emotion corresponding to the question is calculated based on the inverse of the variance of the emotion vectors included in the second group, and the degree of emotion dependency is stored in the question and answer pair information.

6. A dialogue support method using a dialogue support system, comprising: The dialogue support system includes a processor and a memory, The memory includes: a plurality of question and answer pair candidates each including a question sentence vector indicating a sentence of a question by a questioner, a question emotion vector indicating the emotion of the questioner when the question was asked, and an answer sentence vector indicating a sentence of an answer by an answerer to the question; A question and answer pair information storing the question and answer pairs is stored; The dialogue assistance method includes: the processor generates, from the plurality of question-answer pair candidates, a first group based on a similarity between the answer sentence vectors and a similarity between the question sentence vectors, and a second group based on a similarity between the answer sentence vectors and a similarity between the emotion vectors; when the processor determines that the variance of the emotion vectors of the question-answer pair candidates included in the first group is large based on a predetermined condition, the processor stores the question-answer pairs based on the question-answer pair candidates included in the first group in the question-answer pair information as sentence-dependent question-answer pairs in which an answer sentence to a question depends on a sentence of the question; a dialogue support method, wherein, when the processor determines that the variance of the question sentence vectors of the question-answer pair candidates included in the second group is large based on a predetermined condition, the processor stores the question-answer pairs based on the question-answer pair candidates included in the second group in the question-answer pair information as emotion-dependent question-answer pairs in which an answer sentence to a question depends on the question emotion corresponding to the question.

7. The dialogue support method according to claim 6, the processor generates a third group from the plurality of candidate question-answer pairs based on a similarity between the answer sentence vectors, a similarity between the question sentence vectors, and a similarity between the emotion vectors; the processor stores, in the question-answer pair information, question-answer pairs based on the question-answer pair candidates included in the third group as sentence-emotion-dependent question-answer pairs in which an answer sentence to a question depends on the sentence of the question and the question emotion corresponding to the question.

8. The dialogue support method according to claim 7, the processor accepts a search instruction including a question sentence vector and a sentiment vector; a first similarity between a question sentence vector of the sentence dependent question-answer pair stored in the question-answer pair information and a question sentence vector included in the search instruction; a second similarity between the emotion vector of the emotion-dependent question-answer pair stored in the question-answer pair information and the emotion vector included in the search instruction; the processor selects a question and answer pair as a search result for the search instruction from the question and answer pair information based on a third similarity between a question sentence vector of the sentence / emotion-dependent question and answer pair stored in the question and answer pair information and a question sentence vector included in the search instruction; The dialogue support method, wherein the processor outputs an answer sentence indicated by an answer sentence vector included in the selected question-answer pair.

9. The dialogue support method according to claim 6, the memory holds a dialogue sentence vector indicating a sentence of a dialogue between the questioner and the answerer, and a questioner emotion indicating an emotion of the questioner when an utterance is being made by the questioner in the dialogue; The dialogue assistance method includes: The processor extracts an utterance indicating a question by the questioner in the dialogue or an expression of the questioner's emotion in the dialogue based on the dialogue text vector and the questioner emotion, The processor extracts an utterance indicating a response by the respondent corresponding to the extracted utterance or the expression of the emotion based on the dialogue vector; the processor determines a vector indicating a sentence of an utterance by the questioner indicated by the dialogue sentence vector at the time when the extracted utterance or the extracted emotion is expressed, as the question sentence vector in the candidate question and answer pair; the processor determines a vector indicating the questioner's emotion at the time when the extracted utterance or the extracted emotion was expressed as the emotion vector in the candidate question and answer pair; The dialogue support method, wherein the processor determines a vector indicating a sentence of an utterance indicating an answer by the extracted answerer as the answer sentence vector in the question-answer pair candidate.

10. The dialogue support method according to claim 6, when the processor stores the sentence-dependent question-answer pair in the question-answer pair information, the processor calculates a sentence dependency degree indicating a degree to which an answer sentence to a question in the sentence-dependent question-answer pair depends on the question sentence based on the reciprocal of the variance of the question sentence vector included in the first group, and stores the sentence dependency degree in the question-answer pair information; a degree of emotion dependence indicating a degree to which an answer sentence to a question in the emotion-dependent question-answer pair depends on the question emotion corresponding to the question, based on the inverse of the variance of the emotion vectors included in the second group, and stores the degree of emotion dependence in the question-answer pair information.

Images