Active dialogue robot system based on large language model

Abstract

The invention belongs to the technical field of intelligent dialogues, and particularly relates to an active dialogue robot system based on a large language model. The system comprises an active questioning module, a continuous dialogue guiding module and an active questioning module. The active questioning module is used for actively initiating a question to a user to start a dialogue with the user; the continuous dialogue guiding module is used for generating answers according to user input and historical dialogue records by using a large language model; the active query module is used for actively initiating a query to the user during the dialogue; according to the method, the powerful natural language processing capability of the large language model is used, the user is actively guided to perform dialogue and ask the user so that the user can more actively participate in the dialogue, intelligent suggestions and solutions are provided at a proper time, and the application effect of the talking robot in the fields of sales, customer service, presale and the like is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent dialogue, and in particular relates to an active dialogue robot system based on a large language model. Background Art

[0002] Currently, most common conversational robots on the market are passive, that is, only after the user initiates a conversation request, the system responds through rule matching or simple machine learning models. Traditional conversational robots have limitations when dealing with complex semantics and long conversation chains, and it is difficult to flexibly respond to user needs. At the same time, users often need to provide a lot of information on their own to get effective feedback, and the overall user experience is relatively average.

[0003] With the development of large language models (such as GPT, etc.), they have shown stronger understanding and generation capabilities in the field of natural language processing. However, most of the applications on the market currently only use these models in passive question and answer, and the active guidance and dynamic interaction capabilities have not been fully utilized. In view of the shortcomings of the prior art, the present invention proposes an active dialogue robot system based on a large language model, which aims to improve the user interaction experience and promote efficient communication in the fields of sales and customer service. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention proposes an active dialogue robot system based on a large language model, which includes: an active questioning module, a continuous dialogue guiding module and an active follow-up questioning module;

[0005] The active questioning module is used to actively ask questions to the user to start a dialogue with the user;

[0006] The continuous dialogue guiding module is used to generate answers based on user input and historical dialogue records using a large language model;

[0007] The active questioning module is used to actively ask questions to the user during the ongoing conversation.

[0008] Preferably, the active questioning module includes a user behavior monitoring module and a question output module;

[0009] The user behavior monitoring module determines whether the user triggers the conversation mode;

[0010] The question output module outputs guiding questions to the user when the user triggers the dialogue mode.

[0011] Furthermore, the guiding questions output by the question output module are obtained according to the content of the web page that the user is browsing when the dialogue mode is triggered.

[0012] Preferably, the continuous dialogue guiding module includes a user input receiving module, a dialogue context understanding module and a response generating module;

[0013] The user input receiving module is used to receive user input content;

[0014] The conversation context understanding module analyzes the historical conversation records and generates the user's current needs based on the user's input content;

[0015] The response generation module generates answers using a large model based on historical conversation records and the user's current needs.

[0016] Preferably, the active questioning module includes a time monitoring module, a dialogue delay determination module and a question generation module;

[0017] The time monitoring module monitors the user's input time and determines whether the user has timed out and has not responded;

[0018] The conversation delay determination module determines the user's behavior pattern when the user does not respond within a time limit, and determines whether to ask further questions and the time for asking further questions according to the user's behavior pattern;

[0019] The follow-up question generation module generates a follow-up question according to the historical dialogue record and the question time when the dialogue delay determination module determines that a follow-up question is needed.

[0020] Furthermore, the process of judging the user behavior pattern includes: judging the user behavior pattern according to the positivity of the user's response in the historical conversation records, including positive responses, occasional silence, and negative responses.

[0021] Furthermore, the process in which the conversation delay determination module determines whether follow-up questions are needed based on the user behavior pattern includes: if the user behavior pattern is a positive response, no follow-up question is needed; if the user behavior pattern is occasional silence, it is determined that a follow-up question is needed during the occasional silence; if the user behavior pattern is a negative response, it is determined that multiple follow-up questions are needed during the silence.

[0022] Furthermore, the conversation delay determination module determines the questioning time formula according to the user behavior pattern:

[0023]

[0024] Among them, T ask represents the inquiry time, α represents the maximum impact range coefficient, β represents the sensitivity of delay, Δt curren Indicates the user's current input delay, Δt k represents the user input delay cluster center, γ represents the weight of the content length on the question time, δ represents the sensitivity of the content length difference, and L curren Indicates the length of the user's current input content, L k The cluster center representing the length of user input content.

[0025] Furthermore, the active questioning module also includes a dialogue end judgment module; a threshold for the number of questions is set; when the number of questions asked by the speech delay judgment module exceeds the threshold for the number of questions, it is judged that the current dialogue is ended and the system no longer issues questions.

[0026] The beneficial effects of the present invention are as follows: the present invention effectively solves the problems of passive waiting, dialogue interruption and insufficient information in the existing dialogue system. By actively asking questions, continuous dialogue guidance and active follow-up questions, the system can guide users to actively provide more details when users do not provide enough information, thereby significantly improving the fluency of the dialogue and the user experience. The present invention further ensures the intelligence and flexibility of the dialogue through context, so that the system can cope with complex user scenarios such as product recommendations, and has high practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a flowchart of the active dialogue robot system in the present invention;

[0028] Figure 2 This is a flowchart of the inquiry response process of the active inquiry module in the present invention. DETAILED DESCRIPTION

[0029] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0030] The present invention proposes an active dialogue robot system based on a large language model, such as Figure 1 As shown, the system includes the following contents:

[0031] The active dialogue robot system designed by the present invention is implemented based on a large prediction model. The system includes an active questioning module, a continuous dialogue guiding module and an active follow-up questioning module.

[0032] Active question module:

[0033] The active questioning module is used to actively ask questions to users to start a dialogue with them. This module can automatically ask questions to users at the appropriate time to help activate the dialogue and encourage users to actively respond. Specifically, the active questioning module includes a user behavior monitoring module and a question output module. Among them:

[0034] The user behavior monitoring module determines whether the user triggers the conversation mode, and the user behavior monitoring module will continue to monitor the user behavior; for example, in the browser, when the user clicks to browse the web page, the conversation mode will be triggered, and the robot system will actively chat with the user; wherein, the method of triggering the conversation mode includes but is not limited to the web page click event.

[0035] The question output module outputs guiding questions to the user when the user triggers the dialogue mode. Preferably, the guiding questions output by the question output module can be obtained according to the content of the web page browsed when the user triggers the dialogue mode, for example: "Hello, do you want to know about xx product?"; or a list of fixed questions preset according to the scene can be output.

[0036] Continuous dialogue guidance module:

[0037] The continuous dialogue guidance module is used to use the large language model to generate answers based on user input and historical dialogue records. This module is mainly responsible for conducting a question-and-answer dialogue through interaction with the large language model after the user actively speaks. This module enables the system to generate responses based on user input and continuously adjust answers based on historical dialogue records to ensure the accuracy and natural fluency of the dialogue.

[0038] The continuous dialogue guiding module includes a user input receiving module, a dialogue context understanding module and a response generating module. Among them:

[0039] The user input receiving module receives the information input by the user to the system. The input content may be natural language text or inquiries in a specific field (such as related information on products and services). The system parses the user input and extracts key information from the input information (such as product name, inquiry intention, etc.). Natural language processing (NLP) technology is used to identify the input content.

[0040] The response generation module uses a large model to generate answers based on historical conversation records and the user's current needs.

[0041] The guiding question generation module generates guiding questions related to the current dialogue context based on context analysis and using a large language model. The guiding question generation module includes a model prompt construction module and a question diversity control module.

[0042] The continuous dialogue guidance module can prevent the system from ignoring the connection between previous and subsequent dialogues when answering, thereby providing coherent and targeted answers.

[0043] Active Questioning Module:

[0044] The active question module is used to proactively ask questions to users during a conversation. The main purpose of the active question module is to proactively ask questions at the appropriate time if the user does not respond for a long time after the user opens his mouth, so as to ensure that the conversation continues and avoid interruptions or user loss. This module relies on the system's time monitoring and understanding of the conversation context, and can intelligently determine when to ask questions and design appropriate question content based on the user's needs and the context of the conversation.

[0045] The active questioning module includes a time monitoring module, a dialogue delay determination module and a question generation module.

[0046] like Figure 2 As shown, the time monitoring module monitors the user's input time and determines whether the time from the last user input exceeds a preset time threshold. Once exceeded, it is determined that the user has timed out and has not responded.

[0047] The conversation delay determination module determines the user's behavior pattern when the user fails to respond within the time limit. Specifically, the user's behavior pattern is determined based on the degree of user response positivity in past conversation records, that is, based on the user's response time and response content, including positive responses, occasional silence, and negative responses.

[0048] First, the system needs to identify the user's behavior pattern and make corresponding decisions based on the user's input delay and content length. When the user is in a conversation, the system monitors their input time and input content in real time. The relationship between input time delay and content length provides the system with more information to help the system determine the user's current status and needs.

[0049] Input delay (Δt): The time delay of user input, that is, the time from the last question or prompt of the system to the user's response. The system will monitor this time delay. If the delay exceeds the set threshold, the system will determine that the user may have entered the "silent" state.

[0050] Input content length (L): The number of characters in each input by the user. Long input means that the user may be more intentional or making a detailed statement, while short input may mean that the user only makes a brief response.

[0051] The system combines the input delay and the length of the input content to determine the user's current reaction mode and decide whether to initiate the active questioning mechanism.

[0052] To better understand user behavior and reactions, the system uses a K-means clustering algorithm to analyze historical reaction patterns. Historical reaction patterns are classified based on the user's past conversation behavior (including reaction time and input length) to provide decision support for future conversations.

[0053] Specifically, the system analyzes the delay time and input content length of each input in the user's historical conversation, and then uses the K-means algorithm to quantify the user's behavior: Clustering algorithm formula:

[0054] Given the user's conversation history data {Δt1, Δt2, ..., Δtn} and the corresponding input content length {L1, L2, ..., Ln}, the system uses these two dimensions of data as input and clusters them using the K-means algorithm to obtain the user's response pattern. For each cluster k, calculate its center point (mean):

[0055]

[0056] Among them, μ k is the center point of cluster k, indicating the typical response pattern of the cluster. Through these clustering results, the system can identify the user's response pattern according to the previous range setting and make a more personalized response in subsequent conversations.

[0057] The dialogue delay determination module determines whether to ask questions and the time to ask questions based on the user's behavior pattern. When the system determines that active questions are needed, it generates corresponding question content and feeds it back to the user through the dialogue interface. The form of the question content is usually concise and polite, aiming to guide the user to continue to respond; specifically: if the user's behavior pattern is positive response, no question is needed; if the user's behavior pattern is occasional silence, it is determined that a question is needed during the occasional silence; if the user's behavior pattern is negative response, it is determined that multiple questions are needed during the silence.

[0058] Determine the questioning time based on the nonlinear questioning time calculation model:

[0059] When designing a nonlinear function for the question time task, consider the following factors:

[0060] User's current input delay Δt curren : If the user's current delay is long, the system will ask questions in advance; if the delay is short, the system will postpone asking questions.

[0061] User input content length L curren : Shorter content may indicate that the user has a negative attitude and needs to be followed up quickly; longer content indicates that the user is actively participating and follow-up can be postponed.

[0062] Cluster center (historical mode) μ k =(Δt k , L k ): Compare the current behavior with the cluster center and adjust the questioning time:

[0063]

[0064] The follow-up question time is a preset time delay from when the system's current question is not answered.

[0065] Parameter explanation: α represents the maximum influence range coefficient of the delay time, which determines the amplitude of nonlinear adjustment and can be 5-10 in practice; β represents the sensitivity of the delay and controls Δt curren -Δt k The speed of the impact on the Task; γ represents the weight of the impact of the content length on the questioning time; δ represents the sensitivity of the content length difference, controlling the speed of the impact of the length deviation on the Task; e represents a natural constant.

[0066] Function mechanism analysis:

[0067] Nonlinear adjustment of delay time: Use the Sigmoid function, whose output value changes smoothly between 0 and 1. When the user's current delay time is significantly greater than the historical average delay time Δt curren >Δt k When Δt curren <Δt k When , the value approaches 0, indicating that the questioning can be postponed.

[0068] Non-linear adjustment of content length: using an exponential decay function It means that the greater the difference between the user's current input length and the historical pattern, the shorter the questioning time.

[0069] When L curren With L k When it is close, the value is close to 1, indicating that the questioning can be postponed.

[0070] When it is determined that further questioning is necessary, the follow-up questioning mechanism is enabled. Specifically:

[0071] When the conversation delay determination module determines that a follow-up question is needed, the follow-up question generation module generates a follow-up question based on the historical conversation record and the question time. Using the generation capabilities of large language models such as GPT, the model dynamically generates follow-up questions that meet the emotional and content requirements based on the user input delay and content length. Each time the model generates, it automatically adjusts the tone and emotion of the output content based on the previous generated content and the delay results.

[0072] Dynamic emotion generation and adjustment algorithm: Assume that at a certain moment, the questioning time is Task, and the context and historical records are Context. Based on this, the generation of emotions and content can be achieved through the following process: Content = GPT (Context, Task).

[0073] The active questioning module also includes a dialogue end judgment module; a threshold for the number of questions is set; when the number of questions asked by the speech delay judgment module exceeds the threshold for the number of questions, it is judged that the current dialogue is over and the system will no longer issue questions.

[0074] The system generates appropriate follow-up questions through the follow-up question generation module, which can avoid conversation interruptions: by actively asking questions, the system can keep the conversation active when the user does not respond for a long time, thus avoiding conversation interruptions. Enhance user participation: Follow-up questions can help users reactivate their interests, making them feel that the system is paying attention to their needs, so that they can participate in the conversation more actively. Personalization and flexibility: The system can dynamically generate follow-up questions based on the user's interests, conversation context and historical behavior, making the questions more targeted and personalized.

[0075] To sum up, the present invention uses the powerful natural language processing capabilities of the large language model to actively guide users to have conversations, ask users questions so that they can participate in the conversation more actively, and provide intelligent suggestions and solutions at the right time, thereby improving the application effect of the talk robot in sales, customer service, pre-sales and other fields.

[0076] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An active dialogue robot system based on a large language model, characterized in that: include: Active questioning module, continuous dialogue guidance module and active follow-up questioning module; The active questioning module is used to actively ask questions to the user to start a dialogue with the user; The continuous dialogue guiding module is used to generate answers based on user input and historical dialogue records using a large language model; The active questioning module is used to actively ask questions to the user during the ongoing conversation.

2. The active dialogue robot system based on a large language model according to claim 1, characterized in that: The active questioning module includes a user behavior monitoring module and a question output module; The user behavior monitoring module determines whether the user triggers the conversation mode; The question output module outputs guiding questions to the user when the user triggers the dialogue mode.

3. An active dialogue robot system based on a large language model according to claim 2, characterized in that: The guiding questions output by the question output module are obtained according to the content of the web page that the user is browsing when the dialogue mode is triggered.

4. The active dialogue robot system based on a large language model according to claim 1, characterized in that: The continuous dialogue guiding module includes a user input receiving module, a dialogue context understanding module and a response generating module; The user input receiving module is used to receive user input content; The conversation context understanding module analyzes the historical conversation records and generates the user's current needs based on the user's input content; The response generation module generates answers using a large model based on historical conversation records and the user's current needs.

5. The active dialogue robot system based on a large language model according to claim 1, characterized in that: The active questioning module includes a time monitoring module, a dialogue delay determination module and a question generation module; The time monitoring module monitors the user's input time and determines whether the user has timed out and has not responded; The conversation delay determination module determines the user's behavior pattern when the user does not respond within a time limit, and determines whether to ask further questions and the time for asking further questions according to the user's behavior pattern; The follow-up question generation module generates a follow-up question according to the historical dialogue record and the question time when the dialogue delay determination module determines that a follow-up question is needed.

6. An active dialogue robot system based on a large language model according to claim 5, characterized in that: The process of judging the user behavior pattern includes: judging the user behavior pattern according to the positivity of the user's response in the historical conversation records, including positive responses, occasional silence and negative responses.

7. An active dialogue robot system based on a large language model according to claim 5, characterized in that: The process of the conversation delay determination module determining whether follow-up questions are needed based on the user behavior pattern includes: if the user behavior pattern is a positive response, no follow-up question is needed; if the user behavior pattern is occasional silence, it is determined that a follow-up question is needed during the occasional silence; if the user behavior pattern is a negative response, it is determined that multiple follow-up questions are needed during the silence.

8. An active dialogue robot system based on a large language model according to claim 5, characterized in that: The conversation delay determination module determines the question time formula based on the user behavior pattern: Among them, T ask represents the inquiry time, α represents the maximum impact range coefficient, β represents the sensitivity of delay, Δt curren Indicates the user's current input delay, Δt k represents the user input delay cluster center, γ represents the weight of the content length on the question time, δ represents the sensitivity of the content length difference, and L curren Indicates the length of the user's current input content, L k The cluster center representing the length of user input content.

9. An active dialogue robot system based on a large language model according to claim 5, characterized in that: The active questioning module also includes a dialogue end judgment module; a questioning number threshold is set; when the number of questionings performed by the speech delay judgment module exceeds the questioning number threshold, it is judged that the current dialogue is ended and the system no longer issues questioning.

Images