Conversation support system and conversation support method

Abstract

To support smooth conversation using a conversational model.SOLUTION: A conversation support system 40 is configured to: store conversation logs of a conversational model; estimate emotion indicated by each input text in the conversation logs; extract, based on the estimated emotion, from the conversation logs, conversation know-how material candidates, which indicate the scope of texts including the scope of input texts in which a predetermined emotion is indicated continuously and output texts corresponding to the input texts; receive an input of a text from a user; cause the conversational model to input a prompt formed by adding conversation know-how information to the input text; obtain a text output by the conversational model based on the prompt; and cause a predetermined output device to output the obtained text.SELECTED DRAWING: Figure 2

Description

[Technical Field]

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

[0002] Communication between humans and AI using generative AI (Artificial Intelligence) such as Large Language Models (LLM) is becoming widespread. However, it is not easy to identify and teach AI the communication know-how that is cultivated by human personality and sensibility, and this is one of the reasons why smooth communication between humans and AI is often difficult.

[0003] As a technology for providing AI with supplementary information useful for communication, for example, Patent Document 1 discloses a technique for generating a separate question from an input question, searching a text database using a feature vector calculated from the input question and the generated separate question, obtaining text associated with a feature vector that satisfies a predetermined condition for similarity with the feature vector as a candidate text to be used for generating additional text to be added as reference information to the input question, and then adding the additional text generated based on the obtained candidate text to the question as reference information. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent No. 7325152 Summary of the Invention [Problem to be solved by the invention]

[0005] Patent Document 1 provides additional information to AI by adding reference information extracted from a text database to a question. However, because the added reference information is not communication know-how itself, it does not necessarily achieve the expected effect, i.e., smooth communication.

[0006] The present invention has been made in consideration of the above circumstances, and its purpose is to provide a dialogue support system and a dialogue support method that can support the realization of smooth dialogue using a dialogue model. [Means for solving the problem]

[0007] One of the present inventions for solving the above-mentioned problems is a dialogue support system including: a storage device that stores a conversation log, which is data of repetitions of input texts previously input to a dialogue model that outputs text corresponding to text input by a user and output texts output by the dialogue model in response to the input text; and a computing device that executes the following steps: a sentiment analysis process that estimates the emotion indicated by each input text in the conversation log; a conversation know-how material candidate extraction process that extracts, from the conversation log based on the estimated emotion, conversation know-how material candidates, which are ranges of text including a range of input texts in which a predetermined emotion is continuously indicated and each output text corresponding to each input text in the range; a conversation know-how creation process that creates, by a predetermined algorithm, conversation know-how, which is data representing characteristics of the flow of conversation indicated by the extracted conversation know-how material candidates; and an information addition process that accepts text input from a user, inputs a prompt into the dialogue model by adding information of the created conversation know-how to the input text, obtains text output by the dialogue model based on the prompt, and outputs the obtained text to a predetermined output device. [Effects of the Invention]

[0008] According to the present invention, it is possible to support the realization of smooth dialogue using a dialogue model. Configurations and effects other than those described above will become apparent from the following description of the embodiments. [Brief explanation of the drawings]

[0009] [Figure 1] FIG. 1 is a diagram illustrating an example of a configuration of a dialogue system according to an embodiment of the present invention. [Figure 2] FIG. 2 is a diagram illustrating an example of functions provided in the dialogue support system. [Figure 3] FIG. 2 is a diagram illustrating an example of a hardware configuration of the dialogue support system. [Figure 4] FIG. 1 is a block diagram illustrating an overview of dialogue processing performed in a dialogue system. [Figure 5] FIG. 2 is a diagram illustrating an example of the data configuration of an emotion-attached conversation log DB. [Figure 6] FIG. 10 is a diagram showing an example of a prompt used to create conversation know-how candidates. [Figure 7] FIG. 10 is a diagram showing an example of a conversation know-how view screen. [Figure 8] FIG. 10 is a flowchart illustrating details of a conversation know-how material candidate extraction process. [Figure 9] FIG. 10 is a diagram showing an example of conversation know-how material candidates identified in the conversation know-how material candidate extraction process. [Figure 10] FIG. 10 is a flowchart illustrating details of an information addition process. [Figure 11] FIG. 2 is a diagram illustrating an example of a conversation know-how DB. DETAILED DESCRIPTION OF THE INVENTION

[0010] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described with reference to the drawings.

[0011] 1 is a diagram showing an example of the configuration of a dialogue system 1 according to this embodiment. The dialogue system 1 includes information processing devices: a dialogue model device 20 that stores a dialogue model 10 (described later) for conducting a dialogue using text; a user terminal 30 used by a user who conducts a dialogue using the dialogue model 10; and a dialogue support system 40 that supports input of input text (prompt) to the dialogue model 10.

[0012] The dialogue model 10 in the dialogue model device 20 accepts input of input data (prompt) including text (e.g., a question) as a utterance input from a user terminal 30 (user), and outputs text (e.g., an answer to the question) as a utterance corresponding to the input text. The dialogue model 10 can realize a series of dialogues between the user and the model by repeatedly executing input and output of such text between the user terminal 30. The dialogue model 10 is, for example, a large language model (LLM) such as BERT or GPT-3, but the type is not particularly limited as long as it is a model (machine-learned model) capable of processing natural language.

[0013] The dialogue support system 40 acquires communication know-how (such as how to proceed with a conversation and the rules for doing so) based on text previously transmitted and received between the user terminal 30 and the dialogue model device 20. After acquiring the know-how, the dialogue support system 40 creates a prompt by adding the know-how to the text received from the user terminal 30, and inputs the created prompt into the dialogue model 10 of the dialogue model device 20. This allows the user to communicate more naturally and smoothly with the dialogue model device 20.

[0014] The dialogue model device 20, the user terminal 30, and the dialogue support system 40 are communicatively connected to each other via a wired or wireless communication network 5 such as the Internet, a LAN (Local Area Network), a WAN (Wide Area Network), or a dedicated line.

[0015] 2 is a diagram showing an example of functions provided in the dialogue support system 40. First, the dialogue support system 40 stores data of a conversation log DB100, an emotion-attached conversation log DB200, a conversation know-how material candidate DB300, a conversation know-how candidate DB400, and a conversation know-how DB500.

[0016] The conversation log DB 100 is a database that accumulates text (conversation log) that has been previously transmitted and received between the user terminal 30 and the dialogue model device 20. In other words, the conversation log is a database that stores repetitive data of input text input by a user (user terminal 30) and output text output by the dialogue model 10 in response to the input text.

[0017] The emotion-attached conversation log DB200 is a database of conversation logs (hereinafter referred to as emotion-attached conversation logs) in which information on the emotions indicated by each text in the conversation log DB100 is added to the text.

[0018] The conversation know-how material candidate DB300 is a database that stores data on a range of text (hereinafter referred to as a conversation know-how material candidate) that includes a range of input text in which a predetermined emotion (e.g., negative emotion) is continuously shown among the texts in the conversation log DB100 and each output text corresponding to each input text in that range.

[0019] The conversation know-how candidate DB 400 is data (hereinafter referred to as conversation know-how candidate) indicating features (know-how) of the conversation flow indicated by the text in the conversation know-how material candidate DB 300.

[0020] The conversation know-how DB 500 is data (hereinafter referred to as conversation know-how) extracted from the text in the conversation know-how candidate DB 400.

[0021] Next, the conversation support system 40 includes functional units including a sentiment analysis unit 41, a conversation know-how material candidate extraction unit 42, a conversation know-how candidate creation unit 43, a conversation know-how candidate display unit 44, an information addition unit 45, and a conversation know-how generation unit 46.

[0022] The emotion analysis unit 41 estimates the emotion indicated by each input text in the conversation log DB100, and registers conversation log data (emotion-attached conversation log) including the estimated results in the emotion-attached conversation log DB200.

[0023] The conversation know-how material candidate extraction unit 42 extracts conversation know-how material candidates from the conversation log based on the emotions estimated by the emotion analysis unit 41, and registers the extracted conversation know-how material candidates in the conversation know-how material candidate DB300.

[0024] The conversation know-how candidate creation unit 43 uses a predetermined algorithm to create text (conversation know-how candidate) that indicates the characteristics of the conversation flow indicated by the conversation know-how material candidate extracted by the conversation know-how material candidate extraction unit 42. The conversation know-how candidate creation unit 43 registers the created conversation know-how candidate in the conversation know-how candidate DB 400.

[0025] The conversation know-how candidate display unit 44 determines conversation know-how to be finally registered in the conversation know-how DB 500 while displaying conversation know-how candidates on the screen.

[0026] The conversation know-how generating unit 46 registers the conversation know-how in the conversation know-how DB 500 based on the processing result of the conversation know-how candidate display unit 44 .

[0027] The information adding unit 45 performs processing related to RAG (Retrieval-Augmented Generation). That is, the information adding unit 45 accepts input of text (question) from the user via the user terminal 30, and inputs a prompt, which is text obtained by adding conversation know-how created by the conversation know-how generating unit 46 to the input text, into the dialogue model 10. Then, the information adding unit 45 acquires text (answer) output by the dialogue model 10 based on the prompt.

[0028] 3 is a diagram showing an example of the hardware configuration of the dialogue support system 40. The dialogue support system 40 includes an arithmetic device 51 such as a CPU (Central Processing Unit), a main memory device 52 such as a RAM (Random Access Memory) or a ROM (Read Only Memory), an external memory device 53 such as a HDD (Hard Disk Drive) or an SSD (Solid State Drive), an input device 54 such as a keyboard, a mouse, or a touch panel, an output device 55 such as a display or a touch panel, and a communication device 56 configured with a NIC (Network Interface Card), a wireless communication module, a USB (Universal Serial Interface) module, a serial communication module, or the like. The dialogue model device 20 and the user terminal 30 may also have a similar hardware configuration.

[0029] The functions of the functional units of the dialogue support system 40 described above are realized by the arithmetic unit 51 of the dialogue support system 40 reading out programs from the main memory device 52 or the external memory device 53. Each program can be recorded on a portable or fixed recording medium and distributed, for example. All or part of each program in the dialogue model 10 and the dialogue support system 40 may be realized using virtual information processing resources provided using virtualization technology, process space separation technology, or the like, such as a virtual server provided by a cloud system. All or part of these programs may be realized by a service provided by a cloud system via an API (Application Programming Interface), for example. Next, the processing performed by the dialogue system 1 will be described.

[0030] 4 is a block diagram illustrating an outline of the dialogue processing performed by the dialogue system 1. The dialogue processing is started, for example, at a predetermined timing (for example, at a predetermined time interval or a predetermined time), or when the dialogue support system 40 receives a predetermined input from the administrator.

[0031] First, the emotion analysis unit 41 of the dialogue support system 40 acquires all input texts from the conversation log DB 100. Then, the emotion analysis unit 41 estimates the emotion indicated by each input text (s101). The emotion analysis unit 41 associates the estimation results with the content of each text in the conversation log DB 100 and registers them in the emotion-attached conversation log DB 200.

[0032] In this embodiment, the emotion analysis unit 41 estimates whether the emotion indicated by each input text is "positive" (positive emotion such as joy, pleasure, etc.), "negative" (negative emotion such as sadness, anger, etc.), or "neutral" (emotion that is neither positive nor negative).

[0033] The emotion estimation algorithm is not particularly limited. For example, the emotion analysis unit 41 may match each text with a pre-stored text database related to emotions, or may input a prompt to a predetermined natural language processing model asking what emotion each text indicates. There are many known emotion estimation methods, such as the "Sentiment Analysis Service" (Hitachi, Ltd.) (https: / / www.hitachi.co.jp / products / it / appsvdiv / service / sentiment-analysis / index.html).

[0034] (Emotion-based conversation log database) 5 is a diagram showing an example of the data configuration of the emotion-attached conversation log DB 200. The emotion-attached conversation log DB 200 includes each text 201 (text input by the user and text output by the dialogue model 10) and its speaker (user or dialogue model device 20 (AI)), which are the contents of the conversation log DB 100, and an emotion 202 corresponding to each text. The emotion 202 is set to one of "neutral," "negative," or "positive."

[0035] 4, the conversation know-how material candidate extraction unit 42 executes conversation know-how material candidate extraction processing s102 to extract conversation know-how material candidates from the text and emotion data in the emotion-attached conversation log DB 200. The conversation know-how material candidate extraction unit 42 registers the extracted conversation know-how material candidates in the conversation know-how material candidate DB 300. Details of the conversation know-how material candidate extraction processing s102 will be described later.

[0036] Next, the conversation know-how candidate creating unit 43 creates conversation know-how candidates by inputting a prompt including a conversation know-how material candidate in the conversation know-how material candidate DB 300 into the dialogue model 10 of the dialogue model device 20 (s103). A plurality of conversation know-how candidates may be created. Furthermore, the conversation know-how candidate creating unit 43 estimates a task (conversation category) to which the conversation know-how candidate belongs by processing similar to that in s1072, which will be described later.

[0037] FIG. 6 is a diagram showing an example of a prompt 600 used to create a conversation know-how candidate. The conversation know-how candidate creation unit 43 creates the prompt 600, which includes each text 601 of the conversation know-how material candidate created in s102, an emotion 602 corresponding to each text, data 603 indicating the type of each text (whether it is input data (Human) from the user terminal 30 or output data (AI) output by the dialogue model 10), an estimation request 604 for rules or knowledge (i.e., conversation know-how) to be estimated from each text of the conversation know-how material candidate, and an instruction 605 to pay particular attention to text with a predetermined emotion (negative emotion) when estimating the know-how. The conversation know-how candidate creation unit 43 transmits the created prompt 600 to the dialogue model device 20, which inputs the prompt 600 into the dialogue model 10 to acquire the corresponding text (i.e., a response text representing the conversation know-how). The dialogue model device 20 transmits the acquired text to the conversation know-how candidate creation unit 43.

[0038] The conversation know-how candidate creation algorithm described here is an example. For example, the conversation know-how candidate creation unit 43 may create conversation know-how candidates based on a natural language processing model other than the dialogue model 10. Also, for example, the conversation know-how candidate creation unit 43 may display a predetermined input screen on the screen of the user terminal 30 and accept input of conversation know-how candidates corresponding to each conversation know-how material candidate from the user. Also, for example, the conversation know-how candidate creation unit 43 may create conversation know-how candidates by matching each word in the conversation know-how material candidate with a pre-stored text database (for example, a database that associates predetermined keywords representing know-how elements with text representing the content of know-how).

[0039] Furthermore, the conversation know-how candidate creating unit 43 may receive an input of selection (approval or denial) of conversation know-how candidates to be registered in the conversation know-how candidate DB 400 from among the created conversation know-how candidates.

[0040] 4, the conversation know-how candidate display unit 44 displays the contents of each conversation know-how candidate in the conversation know-how candidate DB 400 created in s103 on a conversation know-how view screen 700 (described later) of the user terminal 30 (s104). Then, the conversation know-how candidate display unit 44 receives, from the user, an input of selection (acceptance or rejection) of the conversation know-how candidate to be registered in the conversation know-how candidate DB 400 via the conversation know-how view screen 700.

[0041] Then, the conversation know-how generating unit 46 registers the conversation know-how candidates selected by the user in the conversation know-how DB 500 (s105).

[0042] The conversation know-how candidate display unit 44 and the conversation know-how generation unit 46 may select conversation know-how candidates using other algorithms. For example, the conversation know-how generation unit 46 may select only conversation know-how candidates that have or do not have a specific keyword.

[0043] (Conversation know-how view screen) 7 is a diagram showing an example of a conversation know-how view screen 700. The conversation know-how view screen 700 has a list display field 710 in which information on conversation know-how material candidates corresponding to each conversation know-how candidate is displayed, and a details display field 720 in which details of the conversation know-how material candidate corresponding to a conversation know-how candidate selected from the list display field 710 are displayed.

[0044] The list display field 710 displays a file name 711 of the conversation log DB 100 in which a conversation know-how material candidate corresponding to each conversation know-how candidate is registered, a position 712 of the first text of the conversation know-how material candidate in the conversation log DB 100, a position 713 of the last text of the conversation know-how material candidate in the conversation log DB 100, and a display field 714 displaying the selection (acceptance / denial) of each conversation know-how candidate. Of the conversation know-how candidates displayed in the list display field 710, the user can select a conversation know-how candidate whose detailed information is to be displayed in a detail display field 720.

[0045] The detailed display field 720 displays text 721 (i.e., part of the conversation) of each conversation know-how material candidate corresponding to the conversation know-how candidate selected in the list display field 710, together with its speaker 722 (user or dialogue model 10 (AI)). The detailed display field 720 also displays a selection field 723 for receiving a selection (acceptance or rejection) of the conversation know-how candidate from the user.

[0046] Note that conversation know-how candidates may also be displayed on the conversation know-how view screen 700.

[0047] 4, the user terminal 30 accepts input of text (input text) such as a question to be input to the dialogue model 10 from a user who wishes to have a conversation using the dialogue model device 20. The user terminal 30 transmits the input text to the dialogue support system 40 (s106).

[0048] Then, the information adding unit 45 of the dialogue support system 40 executes information adding processing s107 (RAG) to transmit a prompt to which conversation know-how corresponding to the input text has been added to the dialogue modeling device 20, and to transmit the text output from the dialogue modeling device 20 (output text) to the user terminal 30. Details of the information adding processing s107 will be described later.

[0049] The information adding unit 45 may add the input text from the user terminal 30 and the output text (conversation log) output from the dialogue modeling device 20 to the conversation log DB 100 in real time or periodically (s108).

[0050] Thereafter, the user terminal 30 displays the output text received from the dialogue support system 40 on the screen (s109).

[0051] Thereafter, the processes of s106 to s109 are repeatedly executed, thereby allowing the user to have a dialogue with the dialogue model 10. Meanwhile, the dialogue support system 40 repeatedly executes the processes of s101 to s105 for the conversation log DB 100 to which the conversation log has been added.

[0052] Next, the conversation know-how material candidate extraction process s102 and the information addition process s107 will be described in detail.

[0053] <Conversation know-how material candidate extraction process> 8 is a flow diagram illustrating details of the conversation know-how material candidate extraction process s102. The conversation know-how material candidate extraction unit 42 reads one emotion-annotated conversation log from the emotion-annotated conversation log DB 200 (s1021). The conversation know-how material candidate extraction unit 42 sets the search position for the conversation log to the beginning of the conversation log.

[0054] The conversation know-how material candidate extraction unit 42 sequentially searches through each input text (user's text) in the conversation log, identifies the first input text found that is associated with a negative emotion, and stores the identified input text as the target utterance (s1023).

[0055] The conversation know-how material candidate extraction unit 42 acquires the text immediately after the target utterance identified in s1023, that is, the output text, and determines whether the acquired output text includes the content of an apology (s1024).

[0056] For example, the conversation know-how material candidate extraction unit 42 may determine whether the acquired output text includes an apology by comparing the acquired output text with a predetermined database storing keywords related to an apology. Alternatively, for example, the conversation know-how material candidate extraction unit 42 may create a prompt asking whether the acquired output text includes an apology, input the created prompt to the dialogue model device 20, and acquire the output text, thereby determining whether the output text includes an apology. Alternatively, for example, the conversation know-how material candidate creation unit 43 may determine whether the output text includes an apology based on a natural language processing model other than the dialogue model 10. Alternatively, for example, the conversation know-how material candidate extraction unit 42 may receive a selection input from the user as to whether the output text includes an apology.

[0057] If the acquired output text includes the content of an apology (s1024: YES), the conversation know-how material candidate extraction unit 42 executes the process of s1025, and if the acquired output text does not include the content of an apology (s1024: NO), the conversation know-how material candidate extraction unit 42 executes the process of s1027.

[0058] In s1025, the conversation know-how material candidate extraction unit 42 identifies the input text immediately after the acquired output text and the output text immediately after that, by performing the same process as in s1023 and s1024, and determines whether the input text is associated with a negative emotion and whether the output text includes an apology. By repeatedly performing this type of determination, the conversation know-how material candidate extraction unit 42 identifies a range, starting from the target utterance, in which a pattern of an input text being associated with a negative emotion and an output text immediately thereafter including an apology continues.

[0059] Thereafter, the conversation know-how material candidate extraction unit 42 sets a series of texts in a range starting from the input text immediately before the target utterance and ending with the input text immediately after the output text, which is the end point of the range identified in s1025, as one of the conversation know-how material candidates (s1026).

[0060] The conversation know-how material candidate extraction unit 42 sets the search position for the conversation log to the input text immediately following the conversation know-how material candidate set in s1026, and repeats the processes from s1023 onwards.

[0061] The conversation know-how material candidate extraction unit 42 repeats the above process until the search position in the conversation log reaches the end point of the conversation log (s1022, s1027).

[0062] The conversation know-how material candidate extraction unit 42 registers each conversation know-how material candidate set in the above process in the conversation know-how material candidate DB 300. This completes the conversation know-how material candidate extraction process s102.

[0063] 9 is a diagram showing an example of a conversation know-how material candidate 900 identified in the conversation know-how material candidate extraction process s102. As shown in the figure, the output text 902 ("Sorry, I fixed it") immediately following the first negative emotion input text 901 ("That's not it") in the conversation log contains an apology. Therefore, the input text 903 ("Please do your job") immediately preceding the input text 901 ("That's not it") becomes the starting text of the conversation know-how material candidate 900.

[0064] On the other hand, immediately after pattern 904 of the input text 901 ("That's not it") and the output text 902 ("I'm sorry, I fixed it") that follows immediately thereafter are patterns 905 and 906 of negative emotions and apologies.

[0065] Then, the input text 907 ("Like") immediately following the last pattern of negative emotion and apology becomes the ending text of the conversation know-how material candidate 900.

[0066] <Information addition process> 10 is a flow diagram illustrating the details of the information addition process s 107. The information addition unit 45 receives input text from the user terminal 30 (s1071).

[0067] The information adding unit 45 estimates a task (hereinafter referred to as a target task) that is a category to which the input text received in s1071 belongs (s1072). Note that a task is, for example, a type of work or a type of conversation (such as a word chain game) that a user performs based on sending and receiving text using the dialogue model device 20.

[0068] For example, the information adding unit 45 creates a prompt asking which task the input text belongs to, inputs the created prompt to the dialogue modeling device 20, and obtains the output text, thereby estimating the category to which the input text belongs. Note that, for example, the information adding unit 45 may estimate the category to which the input text belongs based on a natural language processing model other than the dialogue model 10. Also, for example, the information adding unit 45 may estimate the task to which the input text belongs by comparing the input text with a predetermined database that stores keywords for each category. Also, for example, the information adding unit 45 may receive input of the task to which the input text belongs from the user.

[0069] The information adding unit 45 identifies a task corresponding to the target task identified in s1072 (a task that is the same as the target task or a task similar to the target task), and acquires conversation know-how related to the identified task from the conversation know-how DB 500 (s1073).

[0070] For example, the information adding unit 45 creates a prompt asking about tasks similar to the target task, inputs the created prompt to the dialogue modeling device 20, and acquires the output text, thereby inferring tasks similar to the target task. Note that, for example, the information adding unit 45 may infer tasks similar to the target task by referring to a predetermined database that stores keywords for each task.

[0071] The information adding unit 45 creates text (prompt) by adding the conversation know-how identified in s1073 to the input text received in s1072. Then, the information adding unit 45 transmits the created prompt to the dialogue modeling device 20. The dialogue modeling device 20 inputs the prompt into the dialogue model 10, thereby outputting corresponding output text. The dialogue modeling device 20 transmits the output text to the dialogue modeling device 20. The information adding unit 45 receives the output text (s1074).

[0072] The information adding unit 45 adds and registers the received output text in the conversation log DB 100 (s1075). Furthermore, the conversation know-how material candidate extracting unit 42 transmits the received output text to the user terminal 30 (s1076). The user terminal 30 displays the received output text on the screen.

[0073] (Conversation Know-How DB) 11 is a diagram showing an example of the conversation know-how DB 500. The conversation know-how DB 500 is a database to which the contents of the conversation log DB 100 and the conversation know-how material candidate DB 300 are added.

[0074] The conversation know-how DB 500 has a conversation log ID 501, an ID 502 of each conversation know-how material candidate in the conversation log, a start position 503 (position in the conversation log), an end position 504 (position in the conversation log), and a yes / no 505 (user selection of whether to adopt as a conversation know-how material candidate), an ID 506, a task 507, and a content 508 of the conversation know-how corresponding to each conversation know-how material candidate, a creation type 509 (whether the conversation know-how is output by the dialogue model device 20 (AI) or input by the user (user)), and a yes / no 510 (user specification of whether to adopt as conversation know-how).

[0075] As described above, the dialogue support system 40 of this embodiment estimates the emotion indicated by each input text in the conversation log, and based on the estimated emotion, extracts from the conversation log a range of text (conversation know-how material candidate) including a range of input text in which a predetermined emotion is continuously indicated and each output text corresponding to each input text in that range, and creates conversation know-how corresponding to the extracted conversation know-how material candidate. Thereafter, the dialogue support system 40 accepts text input from the user, inputs a prompt in which information about the conversation know-how is added to the input text, into the dialogue model 10, and outputs the text output by the dialogue model 10 based on the prompt to the user terminal 30.

[0076] That is, the dialogue support system 40 of this embodiment creates dialogue know-how based on dialogue portions (candidate dialogue know-how material) in which a predetermined emotion is continuously expressed by the user, and adds the created dialogue know-how to the input text from the user and inputs it into the dialogue model 10, so that the dialogue model 10 can output appropriate text (answer) taking into consideration not only the input text from the user but also the dialogue know-how.

[0077] In this way, the dialogue support system 40 of this embodiment can support the realization of smooth dialogue using a dialogue model.

[0078] Furthermore, the dialogue support system 40 of this embodiment estimates whether the emotion indicated by each input text in the conversation log is a negative emotion, and extracts conversation know-how material candidates including a range of input texts in which negative emotions are continuously indicated and each output text corresponding to each input text in that range.

[0079] In this way, by focusing on negative emotions and extracting examples of conversational failures (gaps between statements), it is possible to create more appropriate conversation know-how and help achieve smooth dialogue.

[0080] Furthermore, the dialogue support system 40 of this embodiment adds the text input by the user and the text output by the dialogue model 10 corresponding to that text to the conversation log.

[0081] This allows the dialogue support system 40 to create more appropriate conversation know-how based on the history of the dialogue between the user and the dialogue model 10, which is based on the conversation know-how.

[0082] Furthermore, the dialogue support system 40 of this embodiment inputs data including conversation know-how material candidates into the dialogue model 10, and sets data including output text corresponding to the data as conversation know-how.

[0083] In this way, by using the dialogue model 10, appropriate conversation know-how can be created.

[0084] At this time, the dialogue support system 40 of this embodiment inputs to the dialogue model 10 a prompt including a conversation know-how material candidate, an emotion corresponding to the text of the conversation know-how material candidate, information (user or AI) indicating whether the text of the conversation know-how material candidate is an input text or an output text, and information on the predetermined emotion.

[0085] In this way, by including in the prompt the emotion corresponding to the input text, the speaker of each text, and the emotion data characterizing the conversation know-how material candidate, appropriate conversation know-how can be obtained.

[0086] Furthermore, the dialogue support system 40 of this embodiment extracts, as candidate conversation know-how materials, a range that includes a range of text in which the combination of input text showing the above-mentioned specified emotion and output text having specified content (for example, content of an apology) output in response to the input text is continuous.

[0087] In this way, by selecting as candidates for conversation know-how material those output texts that correspond to the emotions related to the input text and have predetermined content (for example, the content of an apology), it is possible to extract conversation know-how patterns with high accuracy.

[0088] In addition, the dialogue support system 40 of this embodiment identifies the category (task) to which the conversation know-how belongs, and also identifies the category to which the text input by the user belongs, and inputs a prompt to the dialogue model 10, to which information on the conversation know-how of the category corresponding to the identified category has been added.

[0089] In this way, by using conversation know-how for a task that corresponds to (for example, is the same as or similar to) the task of the user's input text, it is possible to carry out an appropriate dialogue according to the category of the dialogue.

[0090] The present invention is not limited to the above-described embodiments, and can be implemented using any components within the scope of the present invention. The above-described embodiments and modifications are merely examples, and the present invention is not limited to these contents as long as the features of the invention are not impaired. Furthermore, although various embodiments and modifications have been described above, the present invention is not limited to these contents. Other aspects conceivable within the scope of the technical idea of ​​the present invention are also included within the scope of the present invention.

[0091] For example, part of the hardware provided in each device of this embodiment may be provided in another device.

[0092] Furthermore, each program of each device may be provided in another device, a program may consist of multiple programs, or multiple programs may be integrated into one program.

[0093] In addition, in this embodiment, the case of "Shiritori" is used as an example of a task in a conversation log, but the content of the task is not particularly limited, and for example, the task may be the type of business, the category to which the user belongs, or the type of dialogue model 10.

[0094] Furthermore, in the present embodiment, the data format of the conversation know-how and the like output by the dialogue support system 40 is text data, but other data formats such as vector data may also be used.

[0095] In addition, in this embodiment, input text of negative emotions is set as the target utterance, but input text of positive emotions or more specific and subdivided emotions (joy, anger, sadness, happiness, etc.) may also be set as the target utterance.

[0096] Furthermore, in this embodiment, conversation know-how material candidates are created based on combination patterns of input text with negative emotions and output text with the content of an apology immediately thereafter, but output text with content other than an apology (for example, text including content of a correction or amendment corresponding to the content of the input text) may also be used.

[0097] Furthermore, in this embodiment, the input text that the user terminal 30 transmits to the dialogue support system 40 may be a single utterance or a group of multiple utterances. [Explanation of symbols]

[0098] 1 Dialogue system, 10 Dialogue model, 40 Dialogue support system, 41 Sentiment analysis unit, 42 Conversation know-how material candidate extraction unit, 43 Conversation know-how candidate creation unit, 44 Conversation know-how candidate display unit, 45 Information addition unit, 46 Conversation know-how generation unit

Claims

1. a storage device that stores a conversation log, which is data of repetitions of input text previously input to a dialogue model that outputs text corresponding to text input by a user and output text output by the dialogue model in response to the input text; and a sentiment analysis process for estimating the sentiment indicated by each input text in the conversation log; a conversation know-how material candidate extraction process for extracting conversation know-how material candidates, which are text ranges including a range of input texts in which predetermined emotions are continuously shown and output texts corresponding to each input text in the range, from the conversation log based on the estimated emotions; a conversation know-how creation process for creating conversation know-how, which is data representing characteristics of the conversation flow indicated by the extracted conversation know-how material candidates, using a predetermined algorithm; an arithmetic device that executes an information addition process of accepting a text input from a user, inputting a prompt in which the information of the created conversation know-how is added to the input text into the dialogue model, acquiring text output by the dialogue model based on the prompt, and outputting the acquired text to a predetermined output device A dialogue support system comprising:

2. The computing device In the emotion analysis process, it is estimated whether the emotion indicated by each input text in the conversation log is a negative emotion; In the conversation know-how material candidate extraction process, conversation know-how material candidates are extracted that include a range of input texts in which negative emotions are continuously expressed and each output text corresponding to each input text in the range. The dialogue support system according to claim 1 .

3. The computing device In the information addition process, a text input from the user and a text output by the dialogue model corresponding to the text are added to the conversation log. The dialogue support system according to claim 1 .

4. The computing device In the conversation know-how creation process, data including the extracted conversation know-how material candidates is input to the dialogue model, thereby obtaining an output text corresponding to the data, and setting the data including the obtained text as the conversation know-how. The dialogue support system according to claim 1 .

5. The computing device In the conversation know-how creation process, a prompt including the extracted conversation know-how material candidate, an emotion corresponding to the text of the conversation know-how material candidate, information indicating whether the text of the conversation know-how material candidate is an input text or an output text, and information on the predetermined emotion is input to the dialogue model. The dialogue support system according to claim 4.

6. The computing device In the conversation know-how material candidate extraction process, a range including a range of text in which a combination of an input text showing the predetermined emotion and an output text having a predetermined content output in response to the input text is continuous is extracted as the conversation know-how material candidate. The dialogue support system according to claim 1 .

7. the storage device stores the conversation log in association with a category to which the conversation log belongs; The computing device In the conversation know-how creation process, the conversation know-how is created and a category to which the conversation know-how belongs is identified; In the information addition process, a category to which the input text belongs is identified, and a prompt to which information on conversation know-how of a category corresponding to the identified category has been added is input to the dialogue model. The dialogue support system according to claim 1 .

8. A dialogue support method using an information processing device including a storage device that stores a conversation log, which is repetitive data between an input text previously input to a dialogue model that outputs text corresponding to a text input by a user and an output text that the dialogue model outputs in response to the input text, and a computing device, The computing device a sentiment analysis process for estimating the sentiment indicated by each input text in the conversation log; a conversation know-how material candidate extraction process for extracting conversation know-how material candidates, which are text ranges including a range of input texts in which predetermined emotions are continuously shown and output texts corresponding to each input text in the range, from the conversation log based on the estimated emotions; a conversation know-how creation process for creating conversation know-how, which is data representing characteristics of the conversation flow indicated by the extracted conversation know-how material candidates, using a predetermined algorithm; receiving a text input from a user, inputting a prompt in which the information on the created conversation know-how has been added to the input text into the dialogue model, acquiring text output by the dialogue model based on the prompt, and outputting the output text to a predetermined output device; Dialogue support methods.

Images