Response proposal device, response proposal method, and response proposal program

Abstract

Provides context-appropriate search and response suggestions in chat. [Solution] The response suggestion device acquires chat text including the user's chat, extracts pre-specified context elements from the chat text, generates context element information including the extracted context elements, generates search query information based on the context element information, searches for chat history information that matches the search query information from a chat history storage unit that stores chat history information which is the history of past chats, generates context-related information based on the search results, inputs the chat text, context element information and context-related information into a predetermined LLM, and generates a response suggestion for the user using the output from the LLM.

Description

Technical Field

[0004] , , , , , ,

[0003] , , ,

[0001] The present disclosure relates to a response proposal device, a response proposal method, and a response proposal program.

Background Art

[0002] Patent Document 1 discloses obtaining user classification items related to a question sentence, converting the question sentence into a question vector, and searching for a specified number of document vectors having a high degree of relevance to the question vector from among a plurality of document vectors, including searching for a document vector to which a content classification item corresponding to the user classification item related to the question vector is attached in the search, generating a prompt for input to a large language model, including the question sentence and the content that is the source of the specified number of document vectors serving as context in the prompt, inputting the prompt into the large language model, and outputting a response sentence generated based on the sentence output by the large language model. Patent Document 1 also discloses classifying some or all of a plurality of employees (users) based on a user classification (for example, department, job type, assigned region, position) corresponding to content classification.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] Patent Document 1 performs searches based on pre-known user classifications (such as department affiliation), making it difficult to perform searches appropriate to the context of a chat. For example, Patent Document 1 cannot generate response text that takes into account information obtained from the context, such as attributes of users not included in the user classification, or the urgency of the content being discussed in the chat.

[0005] This disclosure aims to provide technology that enables context-appropriate search and response suggestions in chat. [Means for solving the problem]

[0006] One aspect of this disclosure provides a response suggestion device comprising: a chat history storage unit that stores chat history information which is a history of past chats; an acquisition unit that acquires chat text including user chats; a context element extraction unit that extracts pre-specified context elements from the chat text and generates context element information including the extracted context elements; a search query generation unit that generates search query information based on the context element information; a context-related information generation unit that searches the chat history storage unit for chat history information that matches the search query information and generates context-related information based on the search results; and a response suggestion generation unit that inputs the chat text, the context element information and the context-related information into a predetermined LLM (Large Language Model) and generates a response suggestion for the user using the output from the LLM.

[0007] One aspect of this disclosure provides a response suggestion method that acquires chat text including a user's chat, extracts pre-specified context elements from the chat text, generates context element information including the extracted context elements, generates search query information based on the context element information, searches a chat history storage unit that stores chat history information which is a history of past chats for chat history information that matches the search query information, generates context-related information based on the search results, inputs the chat text, the context element information and the context-related information into a predetermined LLM (Large Language Model), and generates a proposed response to the user using the output from the LLM.

[0008] One aspect of this disclosure provides a response suggestion program that causes a computer to perform the following actions: acquire chat text including user chats; extract pre-specified context elements from the chat text; generate context element information including the extracted context elements; generate search query information based on the context element information; search a chat history storage unit that stores chat history information, which is a history of past chats, for chat history information that matches the search query information; generate context-related information based on the search results; input the chat text, the context element information, and the context-related information into a predetermined LLM (Large Language Model); and generate a proposed response to the user using the output from the LLM.

[0009] These comprehensive or specific embodiments may be implemented as systems, devices, methods, integrated circuits, computer programs, or recording media, or as any combination of systems, devices, methods, integrated circuits, computer programs, and recording media. [Effects of the Invention]

[0010] According to this disclosure, it is possible to perform searches and suggest responses that are appropriate to the context of the chat. [Brief explanation of the drawing]

[0011] [Figure 1] Block diagram showing an example of the configuration of the functions of the response proposal device according to Embodiment 1. [Figure 2] A diagram showing an example of information included in the context element information according to Embodiment 1. [Figure 3] A diagram showing an example of information included in the search query information according to Embodiment 1. [Figure 4] A diagram showing an example of information included in context-related information according to Embodiment 1. [Figure 5] A diagram showing an example of the information included in the response style information according to Embodiment 1. [Figure 6] A diagram illustrating an example of a method for selecting context element information based on the element scores and element weights of context elements according to Embodiment 1. [Figure 7] A diagram showing an example of context element information related to the first use case. [Figure 8] A diagram showing an example of search query information related to the first use case. [Figure 9] A diagram showing an example of context-related information for the first use case. [Figure 10] A diagram showing an example of response style information related to the first use case. [Figure 11] A diagram showing an example of context element information related to the second use case. [Figure 12] A diagram showing an example of search query information related to the second use case. [Figure 13] A diagram showing an example of context-related information for the second use case. [Figure 14] A diagram showing an example of response style information related to the second use case. [Figure 15] Block diagram showing an example of the hardware configuration of the response proposal device related to this disclosure. [Modes for carrying out the invention]

[0012] Hereinafter, embodiments of the present disclosure will be described in detail with appropriate reference to the drawings. However, overly detailed descriptions may be omitted. For example, detailed descriptions of well-known matters and duplicate descriptions of substantially identical configurations may be omitted. This is to avoid making the following description unnecessarily redundant and to facilitate the understanding of those skilled in the art. Note that the accompanying drawings and the following description are provided to enable those skilled in the art to fully understand the present disclosure, and are not intended to limit the subject matter described in the claims. Even if the functions of one configuration shown in this embodiment are realized by two or more physical configurations, or the functions of two or more configurations are realized by, for example, one physical configuration, it is acceptable.

[0013] (Embodiment 1) FIG. 1 is a block diagram showing a functional configuration example of a response proposal device 10 according to Embodiment 1.

[0014] The response proposal device 10 is a device that can propose an appropriate response answer 80 to user A when, for example, user A and user B are chatting in chat area 90 and user A responds to an inquiry from user B. This reduces the response workload of user A in the chat. The number of users participating in the chat can be any number. Also, the content of the chat is not limited to inquiries and their responses, and can be anything.

[0015] The chat area 90 is an area where chats between users are displayed and is displayed on a predetermined screen. The chat area 90 may be displayed on a display device 1005 (see FIG. 15) included in the response proposal device 10, or may be displayed on another terminal (for example, a smartphone or a tablet terminal).

[0016] The response suggestion device 10 includes, as functions, an acquisition unit 11, a context element extraction unit 12, a search query generation unit 13, a context-related information generation unit 14, a response style generation unit 15, a response proposal generation unit 16, an output unit 17, and a chat history DB 20. These functions may be realized by the processor 1001 (see Figure 15) of the response suggestion device 10 working in cooperation with the memory 1002 to execute a program.

[0017] The chat history DB20 is a database that holds chat history information 21, which records past chats between users as a history.

[0018] The acquisition unit 11 acquires text data (hereinafter referred to as chat text 30) that includes the content of the chat currently taking place between users in the chat area 90.

[0019] Figure 2 shows an example of the information included in the context element information 40 according to Embodiment 1.

[0020] The context element extraction unit 12 analyzes the chat text 30 acquired by the acquisition unit 11, extracts context elements to be expanded based on the analysis results, and generates context element information 40 as illustrated in Figure 2. The context elements to be expanded may be pre-set as defaults. For example, at least one of the following (A1) and (A2) to (A6) may be pre-set as context elements to be expanded. Note that (A1) is essential because if the user cannot be identified, it will be unclear which user's response should be proposed. (A2) to (A6) may be selected depending on which information to extract from the chat text 30. (A1) Identification information of each user participating in chat text 30 (A2) Roles of each user participating in chat text 30 (A3) Topic and / or subtopic of the conversation in chat text 30 (A4) Priority (urgency) of conversations in chat text 30 (e.g., inquiries) (A5) The emotions of each user participating in chat text 30 (A6) Potential tasks or requests included in the conversation in chat text 30

[0021] Here, "enhancement" includes the process of extracting information directly stated in the chat text 30 and the process of supplementing information that is not directly stated in the chat text 30 but can be inferred from the chat text 30. Examples of information directly stated in the chat text 30 include the user identification information in (A1). On the other hand, information from (A2) to (A6) may not be directly stated in the chat text 30. The context element extraction unit 12 infers and supplements information not directly stated in the chat text 30 from the context of the chat text 30 so that all pre-specified context elements are acquired without omission. The context element extraction unit 12 may also record that a context element is "unknown" if it contains information that cannot be inferred from the context of the chat text 30, or information that can be inferred but is likely to be inaccurate. This is because it is more likely that an appropriate response will be generated when it is clear that a context element is unknown than when it is incorrect.

[0022] Furthermore, in addition to at least one of (A1) and (A2) to (A6) above, at least one of the following (A7) to (A9) may be added as contextual elements to be expanded. The items to be added may be selected by the user. Note that (A7) to (A9) are for recording the time and region in which the chat text 30 was created. Generally, this information does not significantly affect the chat response, so it does not need to be included in the contextual element information 40. However, when dealing with topics specific to a particular region or time, this information may also affect the accuracy of the response, so it may be included in the contextual elements. (A7) Timestamp of chat text 30 (A8) The region where the conversation in chat text 30 is taking place (A9) The time period during which the conversation in chat text 30 is taking place

[0023] Furthermore, the context element extraction unit 12 may also analyze non-verbal data contained in the chat text 30 and extract context elements to be expanded, taking the analysis results into consideration. Examples of non-verbal data include emojis or images.

[0024] For example, the context element extraction unit 12 analyzes the Unicode corresponding to the emojis in the chat text 30 and converts the meaning of those emojis (e.g., happy, troubled, angry) into text information. This analysis clarifies the user's emotions and intentions from the conversation in the chat text 30 and reflects this in the extraction of context elements.

[0025] For example, the context element extraction unit 12 performs OCR (Optical Character Reader) on the image in the chat text 30 to extract text information from the image. Alternatively, the context element extraction unit 12 analyzes the visual elements contained in the image in the chat text 30 using VLM technology and converts the meaning and intent of the image into text information. VLM stands for Vision Language Model.

[0026] The context element extraction unit 12 analyzes the chat text 30, including the character information obtained from the non-verbal data mentioned above, and can extract the context elements (A1) to (A9) that you want to expand, with greater accuracy.

[0027] The context element extraction unit 12 may also extract the above-mentioned context elements from the chat text 30 using a predetermined LLM. LLM stands for Large Language Models. For example, the context element extraction unit 12 inputs the chat text 30 and an instruction (prompt) to extract at least one of the above-mentioned (A1) and (A2) to (A6) from the chat text 30 into the LLM, obtains the context element corresponding to at least one of the above-mentioned (A1) and (A2) to (A6) from the LLM, and generates context element information 40.

[0028] Figure 3 shows an example of the information included in the search query information 50 according to Embodiment 1.

[0029] The search query generation unit 13 organizes and structures the contents of one or more context element information 40 and generates search query information 50 as illustrated in Figure 3.

[0030] For example, the search query generation unit 13 generates search query information 50 which includes at least one of the following (B1) to (B8) as a search query. The search query information 50 may be a JSON file. (B1) Target Topic (B2) Subtopic (B3) Related User Roles (B4) Priority (urgency) (B5) Current Active Users (B6) Required information

[0031] The search query generation unit 13 may generate the search query information 50 using a predetermined LLM. For example, the context element extraction unit 12 inputs context element information 40 and an instruction (prompt) to generate search query information 50 that includes at least one of the above-mentioned (B1) to (B6) as a search query based on the context element information 40 to the LLM, and obtains the search query information 50 that includes at least one of the above-mentioned (B1) to (B6) from the LLM.

[0032] Figure 4 shows an example of the information included in the context-related information 60 according to Embodiment 1.

[0033] The context-related information generation unit 14 uses the search query information 50 generated by the search query generation unit 13 to search the chat history DB 20 for chat history information 21 that matches the search query information 50. This search may be a hybrid search that uses both full-text search and vector search of the information contained in the search query information 50. The context-related information generation unit 14 aggregates the search results to generate one or more context-related information 60 as illustrated in Figure 4. In other words, the context-related information generation unit 14 uses the search query information 50 to search the past chat history information 21 for conversation content related to the context element to be expanded, and generates one or more context-related information 60. The context-related information 60 may be a JSON file. The context-related information 60 may include at least one of the following items (C1) and (C2) to (C5). Note that, since the context is usually different for each conversation, if conversations with different contexts are included indiscriminately in the search results, the context read from the search results will become unclear. Therefore, (C1) is essential to identify each conversation. (C2) through (C5) may be selected depending on which information you want to retrieve from the search. (C1) Conversation ID that identifies the searched conversation (C2) Topics and subtopics of the searched conversations (C3) Identification information of participants in the searched conversation (C4) Content of the searched conversation (C5) Key elements in the searched conversations

[0034] Furthermore, the context-related information 60 may include, in addition to at least one of the above (C1) and (C2) to (C5), the following (C6). (C6) Timestamp

[0035] This information is used to record the time the searched conversation was recorded. Generally, it is unlikely that the suggested response will change based on this information, so it is not essential as context-related information 60. However, when dealing with topics specific to a particular time, this information may also affect the accuracy of the suggested response, so it may be included as a context element.

[0036] The context-related information generation unit 14 may generate context-related information 60 using a predetermined LLM. For example, the context-related information generation unit 14 inputs chat history information 21 obtained by the search and an instruction (prompt) to generate context-related information 60 from the chat history information 21 that includes at least one of the above-mentioned (C1) and (C2) to (C5) into the LLM, and obtains the context-related information 60 output from the LLM that includes at least one of the above-mentioned (C1) and (C2) to (C5).

[0037] Furthermore, if the search yields search results corresponding to multiple conversation IDs, the context-related information 60 may be recorded as multiple files. For example, the context-related information 60 may be recorded as a different file for each conversation ID.

[0038] Figure 5 shows an example of the information included in the response style information 70 according to Embodiment 1.

[0039] The response style generation unit 15 generates response style information 70, which includes information about the response style of user A making a chat response, as illustrated in Figure 5, based on the context-related information 60 generated by the context-related information generation unit 14. The response style information 70 may be a JSON file. The response style generation unit 15 is not mandatory and may be executed only at the request of user A. For example, the response style information 70 may include at least one of the following pieces of information: (D1), (D2), and (D3) to (D8). (D1) and (D2) are mandatory to distinguish from which user the response was generated. (D3) to (D8) may be selected depending on which information should be reflected in the response. (D1) User ID of the user making the response (D2) Username of the user making the response (D3) The tone of voice preferred by the user making the response (D4) Phrases frequently used by users making a response (D5) Level of directness between the responding user and the other user (D6) The user's area of ​​expertise (D7) User interaction preferences for the responding user (D8) The emotional tendencies of the user making the response

[0040] The response style generation unit 15 may generate the response style information 70 using a predetermined LLM. For example, the response style generation unit 15 inputs context-related information 60 and an instruction (prompt) to generate response style information 70 from the context-related information 60 that includes at least one of (D1), (D2), and (D3) to (D8) described above, and obtains the response style information 70 from the LLM that includes at least one of (D1), (D2), and (D3) to (D8) described above.

[0041] The response proposal generation unit 16 integrates the context element information 40, the context-related information 60, and the response style information 70 to generate integrated data. Then, the response proposal generation unit 16 generates a response proposal 80 to propose to user A based on the integrated data. The response proposal generation unit 16 may generate the response proposal 80 using a predetermined LLM. For example, the response proposal generation unit 16 inputs the integrated data and an instruction (prompt) to generate a response proposal 80 to propose to user A based on the integrated data into the LLM, and obtains the response proposal 80 from the LLM. Alternatively, RAG (Search Enhanced Generation) may be used to make the LLM refer to the integrated data.

[0042] Here, the context-related information 60 is extracted by searching the chat history DB 20, and depending on the search results, it may consist of a huge number of files. Generally, the number of tokens that LLM can process is limited, so if all of the context-related information 60 obtained from the search is included in the integrated data, the number of tokens in the integrated data may exceed the upper limit of the number of tokens that LLM can process. Therefore, the response draft generation unit 16 selects the context-related information 60 files to be included in the integrated data based on the following procedure. In the following explanation, the context-related information 60 is described as being composed of different files for each conversation ID. However, this configuration is just an example, and a single file may contain context-related information 60 corresponding to multiple conversation IDs. In this case, the ranges of a single file corresponding to each conversation ID can be treated as separate files, and the following processing can be performed.

[0043] The response generation unit 16 selects context-related information 60 to be integrated into the integrated data based on a score that quantifies the relationship between the context element information 40 and each of the multiple context-related information 60. In other words, the response generation unit 16 selects search results (context-related information 60) that have a context closely related to the context (context element information 40) that can be read from the chat text 30 and includes them in the integrated data. The score is determined, for example, based on a value (element score) that indicates the relationship of a predetermined context element between the context element information 40 and the context-related information 60. For example, the similarity (topic similarity) between the topic included in the context element information 40 and the topic included in each context-related information 60 is calculated as an element score, and the score is set such that the higher the element score, the higher the probability of it being selected as context-related information 60 to be integrated into the integrated data. This score may use the element score itself, or it may use another value calculated based on the element score. For example, the element score may be classified into several stages (e.g., high, medium, low), and the numerical value of each stage may be used as the score. The response generation unit 16 selects context-related information 60 with a score above a predetermined threshold to be included in the integrated data, and excludes context-related information 60 with a score below a predetermined threshold. This reduces the amount of context-related information 60 included in the integrated data, thereby reducing the number of tokens in the integrated data.

[0044] In the example above, topic similarity was used as the element score to calculate the score, but the score may be calculated from other context elements as long as the relationship between the context element information 40 and the context-related information 60 can be evaluated.

[0045] Furthermore, the response generation unit 16 may calculate a score by comprehensively evaluating multiple context elements included in the context element information 40 and the context-related information 60. For example, the score could be the sum of the element scores calculated for each of the multiple context elements between the context element information 40 and each context-related information 60. Note that when considering multiple context elements, the units of the element scores may differ. For example, the degree of matching for "topics" is expressed as a percentage, while the relevance of "user identification information" is expressed as a binary value of matching or not matching. Therefore, the element score for "topics" and the element score for "user identification information" cannot be added together directly. Thus, when considering multiple context elements, the element scores for each context element may be classified into a unified standard such as high, medium, or low, and then the high, medium, and low values ​​may be quantified and added together. In this case, "medium" is assigned to a standard range that is likely to appear as an element score.

[0046] The response generation unit 16 may provide a customization function as an optional function for selecting context-related information 60 to be included in the integrated data, which allows for prioritizing the consideration of specific context elements. Specifically, it may be possible to assign weights to element scores when calculating the score. User A can adjust the score calculation result by assigning higher weights to context elements that they want to evaluate more carefully. Through this adjustment, User A can influence the selection of context-related information 60 to be integrated into the integrated data, and the response generation unit 16 can generate a response 80 that better suits User A's needs.

[0047] When calculating a score that takes multiple context elements into account between context element information 40 and each context-related information 60, the score may be calculated using, for example, the following formula.

[0048]

number

[0049] Here, the element score represents the score based on each context element (for example, the topic relevance mentioned above). Similar to the above, the element score is set as a numerical value representing high, medium, or low. The weight represents the weight value for each context element set by the user. For example, the weight is set as a numerical value representing high, medium, or low. Here, the numerical values ​​representing high, medium, and low weights are pre-set. n represents the total number of context elements.

[0050] Figure 6 is a diagram illustrating an example of a method for selecting context-related information 60 based on the element scores and element weights of context elements according to Embodiment 1.

[0051] For example, as shown in Figure 6, suppose user A sets the urgency match to "high" as a scoring weight adjustment. In this case, the response generation unit 16 increases the weight of the urgency match set to "high" and calculates the scores for each context-related information 60A and 60B. Suppose the result is that the score of context-related information 60A is 79 and the score of context-related information 60B is 71. If the threshold is 75, then the score of context-related information 60A, "79", is above the threshold of "75", and the score of context-related information 60B, "71", is below the threshold of "75". Therefore, the response generation unit 16 selects context-related information 60A, whose score is above the threshold, and does not select context-related information 60B, whose score is below the threshold. In this way, by adjusting the weight of context elements, context-related information 60 selected based on context elements that are high priority for user A is included in the integrated data, enabling efficient selection of highly relevant search results.

[0052] In Equation 1, the impact on the score is roughly the same whether the element score is high and the weight is low, or whether the element score is low and the weight is high. Therefore, depending on the values ​​of the element score and weight, multiple context-related information 60 may appear with the same score despite having different element scores for the context elements. In this case, the response generation unit 16 includes the context-related information 60 with high element scores for context elements that have been given a high weight in the integrated data. This is because elements with higher weights are considered to be more important to the user.

[0053] Returning to the explanation of Figure 1, the output unit 17 displays the proposed response 80 on the display device 1005. In other words, the output unit 17 presents the proposed response 80 to user A. User A modifies the presented proposed response 80 as needed via the input device 1004 and instructs it to send it to the chat area 90. Upon receiving the instruction to send, the output unit 17 sends the proposed response 80 to the chat area 90 as a chat response from user A.

[0054] This allows User A to easily provide chat responses that are appropriate to the context of the chat currently taking place in chat area 90, User A's characteristics in the current chat, and User A's past response style.

[0055] Furthermore, the output unit 17 adds the chat response sent to the chat area 90 to the chat history information 21 in the chat history DB 20. As a result, past chats are accumulated in the chat history information 21.

[0056] The context element extraction unit 12, the search query generation unit 13, and the context-related information generation unit 14 may be collectively referred to as the context enhancement module. Furthermore, the response style generation unit 15 and the response proposal generation unit 16 may be collectively referred to as the personalized response generation module.

[0057] As described above, according to this embodiment, by using search query information 50 generated based on the chat text 30 to search past chat history information 21, it is possible to extract information not directly written in the chat text 30 and expand the context. For example, the user's knowledge level and preferred response style can be extracted from past conversation logs and reflected in the context.

[0058] Furthermore, according to this embodiment, since the context elements to be parsed and expanded from the chat text 30 are specified in advance, it is possible to reliably obtain the context elements to be included in the search query information 50. It is also possible to have the chat text 30 parsed by the LLM without specifying context elements. However, in that case, the LLM will obtain context elements that it deems important based on its own criteria, so there is a possibility that the context information necessary for subsequent processing may not be obtained. For example, information that is not necessary for the search query information 50 but appears frequently in the chat text is likely to be judged as important by the LLM. Therefore, if context elements are not specified, there is a risk that unnecessary information will be included in the search query information 50.

[0059] <First Use Case> A first use case utilizing the response suggestion device 10 described above will now be explained. Figure 7 shows an example of context element information 40 related to the first use case. Figure 8 shows an example of search query information 50 related to the first use case. Figure 9 shows an example of context-related information 60 related to the first use case. Figure 10 shows an example of response style information 70 related to the first use case.

[0060] For example, suppose in chat area 90, multiple users are discussing a problem related to Wi-Fi settings for an extended period of time, and user A, who is participating in this chat, is asking the response suggestion device 10 for an appropriate solution to this problem.

[0061] In this case, the conventional method, which does not utilize the context expansion module, only searches for basic information regarding Wi-Fi security settings and presents user A with only general solutions as suggested responses, regardless of the user's role. In other words, the conventional method presents suggested responses that do not take into account user A's expertise and needs.

[0062] In contrast, the method of this embodiment, which utilizes the context enhancement module, can identify, for example, that User A's role is "Network Administrator" by analyzing the context of the real-time conversation in the chat area 90. In that case, the context element extraction unit 12 generates context element information 40, including that User A's role is "Network Administrator," as shown in Figure 7. Next, the search query generation unit 13 generates search query information 50 for searching for information appropriate for "Network Administrator" based on the context element information 40, as shown in Figure 8. Therefore, the context-related information generation unit 14 can use the search query information 50 to search for and generate context-related information 60 appropriate for the response of User A, who is a "Network Administrator," as shown in Figure 9. Furthermore, the response style generation unit 15 can generate response style information 70 appropriate for the tone and preferences of User A, who is a "Network Administrator," as shown in Figure 10. As a result, the response suggestion device 10 can propose a response suggestion 80 optimized for User A's role as "Network Administrator" (i.e., a response suggestion 80 according to User A's level of expertise) to User A.

[0063] Therefore, the response suggestion device 10 can suggest a response that user A truly desires in the context of the chat.

[0064] Furthermore, even with conventional methods, if words indicating the user's role frequently appear in the chat text, the LLM may coincidentally estimate the user's role and present response suggestions that take into account the user's expertise and needs. However, in this embodiment, by pre-specifying "role" as a contextual element to be expanded, the user's role can be reliably extracted, and truly desired response suggestions can be proposed with greater certainty.

[0065] <Second Use Case> A second use case utilizing the response suggestion device 10 described above will now be explained. Figure 11 shows an example of context element information 40 related to the second use case. Figure 12 shows an example of search query information 50 related to the second use case. Figure 13 shows an example of context-related information 60 related to the second use case. Figure 14 shows an example of response style information 70 related to the second use case.

[0066] For example, in chat area 90, several users are discussing the inability to connect to Wi-Fi during a meeting, and user A, who is participating in this chat, is concerned about this issue. In response, an urgent response is required.

[0067] In this case, the conventional method, which does not utilize the context extension module, would only search for basic information about the Wi-Fi connection and present user A with only general solutions as suggested responses, regardless of the urgency of the situation. In other words, the conventional method would present suggested responses that do not take into account the current situation or urgency.

[0068] In contrast, the method of this embodiment, which utilizes the context enhancement module, can identify, for example, that the user's situation is "in a meeting" by analyzing the context of the real-time conversation in the chat area 90. In that case, as shown in Figure 11, the context element extraction unit 12 generates context element information 40 that includes requests for countermeasures with high priority (high urgency) and long-term improvement measures. Next, the search query generation unit 13 generates search query information for searching for information regarding the high priority (high urgency) and long-term improvement measures based on the context element information 40, as shown in Figure 12. Therefore, the context-related information generation unit 14 can use the search query information 50 to search for and generate context-related information 60 appropriate for the situation in the meeting room and the situation of high urgency, as shown in Figure 13. As a result, the response suggestion device 10 can propose a response suggestion 80 optimized for the user's situation of "in a meeting" and high urgency to user A.

[0069] Therefore, the response suggestion device 10 can suggest a response 80 that is appropriate to the situation and urgency of the user who is chatting.

[0070] Furthermore, even with conventional methods, if explanations of the situation and urgency frequently appear in chat text, LLM may incidentally estimate the user's situation and urgency, and suggest appropriate responses. However, as in this embodiment, by pre-specifying "situation and urgency" as a contextual element to be expanded, it becomes possible to more reliably suggest appropriate responses.

[0071] (Hardware configuration) Figure 15 is a block diagram showing an example of the hardware configuration of the response suggestion device 10 according to this disclosure.

[0072] The response suggestion device 10 comprises, as hardware, a processor 1001, memory 1002, storage 1003, input device 1004, display device 1005, and communication device 1006. The response suggestion device 10 may be interpreted as a computer or server device.

[0073] The processor 1001 realizes the functions of the response suggestion device 10 described above by reading a predetermined program from the memory 1002 and executing it. For example, the processing of the acquisition unit 11, context element extraction unit 12, search query generation unit 13, context-related information generation unit 14, response style generation unit 15, response suggestion generation unit 16, and output unit 17 described above may be performed by the processor 1001. The processor 1001 may also be read as a Central Processing Unit (CPU), controller, control device, Large Scale Integration (LSI), etc. Furthermore, the processor 1001 may include a Graphics Processing Unit (GPU) and / or a Neural Network Processing Unit (NPU).

[0074] Memory 1002 stores programs and data for realizing the functions of the response suggestion device 10 described above. Memory 1002 may be composed of a volatile storage medium and / or a non-volatile storage medium.

[0075] The storage device 1003 stores programs and data for realizing the functions of the response suggestion device 10 described above. The storage device 1003 may be composed of a non-volatile storage medium. Examples of the storage device 1003 include a Hard Disk Drive (HDD), a Solid State Drive (SSD), and flash memory.

[0076] For example, the chat history DB20, chat history information21, chat text30, context element information40, search query information50, context-related information60, response style information70, and response draft80 mentioned above may be stored in memory1002 and / or storage1003.

[0077] The input device 1004 is a device that receives input from the user. Examples of the input device 1004 include a keyboard, mouse, touchpad, touch panel, and microphone.

[0078] The display device 1005 is a device that displays characters, images, etc. Examples of the display device 1005 include liquid crystal displays and organic EL displays.

[0079] The communication device 1006 is a device that connects the response suggestion device 10 to a predetermined communication network. Examples of communication networks include wired LAN (e.g., Ethernet®), wireless LAN (e.g., Wi-Fi®), the Internet, mobile communication networks (4G, 5G), and Bluetooth®.

[0080] For example, chat information in chat area 90 may be sent and received via communication device 1006 and the communication network. In addition, user input and display of information as described above may be performed from other terminals (e.g., smartphones, tablet devices, PCs, etc.) via communication device 1006 and the communication network.

[0081] (Other variations) In the embodiment described above, an example was explained in which "role" is used as a user attribute. However, the user attribute can be any attribute that may be related to the content of the response proposal or the user's preference for interaction. For example, the user's age, gender, nationality, religion, address, hobbies, etc., are possible.

[0082] In the embodiment described above, the response suggestion device 10 was configured to suggest a response to user B based on chat text, including the chat between user A, who is the respondent, and user B, who is the questioner. However, the response suggestion device 10 may also suggest a response based on chat text, including only the chat between user B and user B. For example, if user B was initially conversing with an automated chatbot, and a human user A takes over the conversation from the chatbot and responds, the response suggestion device 10 could suggest a response based on the chat text, including the chat between user B and the chatbot. Furthermore, by having the chatbot output the response suggested by the response suggestion device 10 as is, a chatbot that performs automated responses while considering the context can be realized.

[0083] In the embodiment described above, the response suggestion device 10 generated response style information 70 indicating the response style of user A, who is the respondent, and used it to generate the response suggestion 80. However, the response suggestion device 10 may also generate response style information 70 indicating the respondent's response style preferred by user B, who is the questioner, and use it to generate the response suggestion 80. The respondent's response style preferred by user B can be generated, for example, based on user B's chat history included in the chat history information 21. More specifically, the respondent's response style preferred by user B can be estimated from user B's feelings towards the respondent's response style in the chat history information 21.

[0084] In the embodiment described above, an example was explained in which the data format for search query information 50, context-related information 60, and response style information 70 was JSON format. However, the data format for this information may be any other format as long as it is a structured data format. For example, XML format could be used. It is also possible to use an unstructured format as the data structure for this information. However, with an unstructured format, it may not be possible to accurately convey to the LLM what each piece of information means. For example, if a user is chatting under the name "administrator," it is difficult to distinguish in an unstructured data format whether this "administrator" is user identification information or represents an attribute indicating that the user possesses advanced knowledge. As a result, the response suggestion device 10 may generate a response suggestion intended for a user with advanced knowledge, even if the user is merely a novice using the name "administrator." On the other hand, when a structured data format is used, it becomes clear that the user identification information is "administrator," and that this "administrator" does not represent an attribute. As a result, the response suggestion device 10 can generate a response suggestion intended for a user with general knowledge.

[0085] In the embodiment described above, the chat history information 21 was a history of past chats. However, the chat history information may also include the chat history of a virtual case created manually by the user or using LLM, etc.

[0086] In the above-described embodiment, the LLM may be located inside the response suggestion device 10, or it may be located in an external cloud or the like. If it is located externally, the response suggestion device 10 invokes the functions of the LLM using an API.

[0087] (Summary of Embodiment 1) Based on the description of Embodiment 1 above, the following technology is disclosed.

[0088] <Technology 1> The response suggestion device (10) includes a chat history storage unit (e.g., chat history DB20) that stores chat history information (21) which is a history of past chats; an acquisition unit (11) that acquires chat text (30) including the user's chats; a context element extraction unit (12) that extracts pre-specified context elements from the chat text and generates context element information (40) that includes the extracted context elements; a search query generation unit (13) that generates search query information (50) based on the context element information; a context related information generation unit (14) that searches the chat history storage unit for chat history information that matches the search query information and generates context related information (60) based on the search results; and a response suggestion generation unit (16) that inputs the chat text, the context element information and the context related information into a predetermined LLM (Large Language Model) and generates a response suggestion (80) for the user using the output from the LLM. This allows you to reliably obtain the contextual elements you want to include in your search query information, because the contextual elements you want to expand upon by analyzing the chat text are specified in advance.

[0089] <Technology 2> In the response suggestion device described in Technology 1, the context element extraction unit extracts at least one of the pre-specified context elements by expanding the context elements based on the analysis results of the chat text. This allows us to extract information not directly written in the chat text and expand the context.

[0090] <Technology 3> In the response proposal generation device described in Technology 2, the chat text includes at least a chat between a first user and a second user, and the response proposal generation unit generates a response proposal from the first user to the second user. This allows the first user to efficiently respond to the second user via chat.

[0091] <Technology 4> The response suggestion device described in Technical 3 further comprises a response style generation unit (15) that generates response style information indicating the style of the first user's response based on the context-related information, the response suggestion generation unit further inputs the response style information to the LLM and generates the response suggestion based on the style of the first user's response using the output from the LLM. This ensures that the suggested response is tailored to the first user's response style, allowing the first user to more efficiently engage in response chats with the second user.

[0092] <Technology 5> In the response suggestion device described in Technology 3 or 4, the chat history information includes the chat history of the first user. This allows for the enrichment of contextual elements based on the first user's past chat history, enabling the generation of appropriate response suggestions for the first user.

[0093] <Technology 6> In the response suggestion device described in any one of technologies 2 to 5, the context element includes user identification information and includes at least one of the roles of each user participating in the chat text, the topic and / or subtopic of the conversation in the chat text, the priority or urgency of the conversation in the chat text, the emotions of each user participating in the chat text, and potential tasks or requests included in the conversation in the chat text. This allows for the expansion of contextual elements from a desired perspective.

[0094] <Technology 7> In the response suggestion device described in any one of the technologies 1 to 6, the response suggestion generation unit calculates a score for each of the multiple context-related information based on predetermined criteria, selects at least one context-related information based on the calculated score, and inputs the selected at least one context-related information into the LLM. This allows for a reasonable reduction in the number of tokens that need to be input into the LLM.

[0095] <Technology 8> The response proposal method acquires chat text (30) including the user's chat, extracts pre-specified context elements from the chat text, generates context element information (40) including the extracted context elements, generates search query information (50) based on the context element information, searches for chat history information (21), which is the history of past chats, from a chat history storage unit (e.g., chat history DB20) that matches the search query information, generates context-related information (60) based on the search results, inputs the chat text, the context element information, and the context-related information into a predetermined LLM (Large Language Model), and generates a proposed response (80) to the user using the output from the LLM. This allows you to reliably obtain the contextual elements you want to include in your search query information, because the contextual elements you want to expand upon by analyzing the chat text are specified in advance.

[0096] <Technology 9> The response suggestion program causes the computer to perform the following actions: acquire chat text (30) including the user's chat; extract pre-specified context elements from the chat text; generate context element information (40) including the extracted context elements; generate search query information (50) based on the context element information; search the chat history storage unit (e.g., chat history DB20) which stores chat history information (21) which is a history of past chats, for chat history information that matches the search query information; generate context-related information (60) based on the search results; input the chat text, the context element information, and the context-related information into a predetermined LLM (Large Language Model); and generate a suggested response (80) for the user using the output from the LLM. This allows you to reliably obtain the contextual elements you want to include in your search query information, because the contextual elements you want to expand upon by analyzing the chat text are specified in advance.

[0097] While embodiments have been described above with reference to the attached drawings, this disclosure is not limited to such examples. It is clear to those skilled in the art that various modifications, alterations, substitutions, additions, deletions, and equivalents can be conceived within the scope of the claims, and these are also understood to fall within the technical scope of this disclosure. Furthermore, the components of the embodiments described above can be combined in any way without departing from the spirit of the invention. [Industrial applicability]

[0098] The technology disclosed herein is useful for improving the efficiency of user-to-user chats and reducing the workload. [Explanation of Symbols]

[0099] 10 Response Proposal Device 11 Acquisition Department 12. Context element extraction unit 13 Search Query Generation Unit 14. Context-related information generation unit 15 Response Style Generation Unit 16 Response Draft Generation Unit 17 Output section 20 Chat History Database 21 Chat history information 30 chat texts 40 Contextual Element Information 50 Search Query Information 60, 60A, 60B Context-related information 70 Response Style Information 80. Proposed Response 90 Chat Area 1001 Processor 1002 memory 1003 Storage 1004 Input device 1005 Display device 1006 Communication device

Claims

1. A chat history storage unit that stores chat history information, which is the history of past chats, A unit that retrieves chat text, including user chat messages, A context element extraction unit extracts pre-specified context elements from the chat text and generates context element information that includes the extracted context elements. A search query generation unit that generates search query information based on the aforementioned context element information, A context-related information generation unit searches the chat history storage unit for chat history information that matches the search query information and generates context-related information based on the search results, The system includes a response generation unit that inputs the chat text, context element information, and context-related information into a predetermined LLM (Large Language Model), and generates a response proposal for the user using the output from the LLM. Response suggestion device.

2. The response suggestion device according to claim 1, wherein the context element extraction unit extracts at least one of the pre-specified context elements by expanding the context elements based on the analysis results of the chat text.

3. The chat text includes at least the chat between the first user and the second user, The response proposal device according to claim 2, wherein the response proposal generation unit generates a response proposal from the first user to the second user.

4. The system further includes a response style generation unit that generates response style information indicating the style of the first user's response based on the context-related information, The response proposal generation unit further inputs the response style information to the LLM and generates the response proposal based on the response style of the first user using the output from the LLM. The response suggestion device according to claim 3.

5. The chat history information includes the past chat history of the first user. The response suggestion device according to claim 3.

6. The aforementioned context element includes the user's identification information, The roles of each user participating in the aforementioned chat text, The topic and / or subtopic of the conversation in the aforementioned chat text, Priority or urgency of the conversation in the aforementioned chat text, The emotions of each user participating in the aforementioned chat text, and, The conversation in the chat text includes at least one of the potential tasks or requests, The response suggestion device according to claim 2.

7. The response proposal generation unit, For each of the multiple context-related pieces of information, a score is calculated based on predetermined criteria. Based on the calculated score, select at least one context-related piece of information. The selected at least one context-related piece of information is entered into the LLM. The response suggestion device according to claim 1.

8. Retrieve chat text including user chat messages, From the chat text, the system extracts pre-specified context elements and generates context element information that includes the extracted context elements. Based on the aforementioned context element information, search query information is generated. From the chat history storage unit, which stores chat history information that is a history of past chats, the system searches for chat history information that matches the search query information, and generates context-related information based on the search results. The chat text, context element information, and context-related information are input into a predetermined LLM (Large Language Model), and a proposed response to the user is generated using the output from the LLM. Method of proposing a response.

9. Retrieve chat text including user chat messages, From the chat text, the system extracts pre-specified context elements and generates context element information that includes the extracted context elements. Based on the aforementioned context element information, search query information is generated. From the chat history storage unit, which stores chat history information that is a history of past chats, the system searches for chat history information that matches the search query information, and generates context-related information based on the search results. The chat text, context element information, and context-related information are input into a predetermined LLM (Large Language Model), and a proposed response to the user is generated using the output from the LLM. A response suggestion program that instructs a computer to perform a specific action.

Images