Account dialogue response methods, devices, computer equipment, readable storage media, and program products

By using large language models and dynamic tagging technology to process multi-turn dialogue data, the problem of automated customer service systems being unable to understand complex user dialogues has been solved, achieving highly accurate and personalized dialogue responses.

CN123115296APending Publication Date: 2026-11-17BEIJING & AIWEIWU TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610385515.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-03-26
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing automated customer service systems cannot understand users' complex, multi-turn natural language expressions and struggle to extract high-dimensional user features from unstructured dialogue data, resulting in low semantic matching accuracy.

Method used

By acquiring multi-turn dialogue data and behavioral trajectory data of the target account, natural language understanding is performed using a large language model to output dialogue semantic vectors. Dynamic tags are extracted based on preset prompt word templates to construct dynamic tag vectors. The dialogue semantic vectors, dynamic tag vectors, and behavioral feature vectors are fused to generate an intention assessment score. Dialogue response data is then generated based on the intention assessment score.

Benefits of technology

It achieves deep semantic understanding of multi-turn dialogue data, captures the true intent and emotions of the target account, forms a dynamic account profile, and improves the semantic matching degree of dialogue response data, making it more in line with the current intention and needs of the target account.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115296A_ABST
    Figure CN123115296A_ABST
Patent Text Reader

Abstract

This application relates to an account dialogue response method, apparatus, computer device, readable storage medium, and program product. The method includes: acquiring multi-turn dialogue data and behavioral trajectory data of a target account; performing natural language understanding processing on the multi-turn dialogue data using a large language model to output corresponding dialogue semantic vectors; extracting dynamic tags from the multi-turn dialogue data using the large language model based on preset prompt word templates, and constructing dynamic tag vectors based on the dynamic tags; constructing behavioral feature vectors based on the behavioral trajectory data; fusing the dialogue semantic vectors, dynamic tag vectors, and behavioral feature vectors to obtain a multimodal fusion vector, and generating an intention assessment score for the target account based on the multimodal fusion vector; and generating dialogue response data corresponding to the target account based on the intention assessment score. This method can improve the semantic matching degree between dialogue response data and the target account.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the fields of natural language processing and human-computer interaction technology, and in particular to an account dialogue response method, apparatus, computer equipment, computer-readable storage medium, and computer program product. Background Technology

[0002] With the widespread adoption of mobile internet and social media, businesses can interact with users through various private channels such as instant messaging groups, public information platforms, and applications. For example, businesses can use rule-based and keyword-matching automated customer service systems to categorize users using static data such as registration information and source channels. When a user's question matches specific keywords, the system automatically replies with a pre-defined, standardized answer corresponding to the tag. However, because existing automated customer service systems cannot understand users' complex, multi-turn natural language expressions, they struggle to extract high-dimensional user features from unstructured dialogue data. This results in a low semantic match between the templated answers provided by the automated customer service system and the user. Summary of the Invention

[0003] Therefore, it is necessary to provide an account dialogue response method, apparatus, computer device, computer-readable storage medium, and computer program product that can improve semantic matching accuracy in response to the above-mentioned technical problems.

[0004] Firstly, this application provides a method for responding to account conversations, including:

[0005] Obtain multi-turn conversation data and behavioral trajectory data of the target account;

[0006] The multi-turn dialogue data is processed using a large language model to perform natural language understanding and output the corresponding dialogue semantic vector.

[0007] Dynamic tags are extracted from the multi-turn dialogue data using the large language model based on a preset prompt word template, and a dynamic tag vector is constructed based on the dynamic tags;

[0008] Construct a behavioral feature vector based on the behavioral trajectory data;

[0009] The dialogue semantic vector, the dynamic tag vector, and the behavioral feature vector are fused to obtain a multimodal fusion vector, and the intention evaluation score of the target account is generated based on the multimodal fusion vector;

[0010] Based on the intention assessment score, generate dialogue response data corresponding to the target account.

[0011] In one embodiment, generating dialogue response data corresponding to the target account based on the intention assessment score includes:

[0012] Based on the dialogue semantic vector, determine the intent recognition result and emotion recognition result corresponding to the multi-turn dialogue data;

[0013] Using at least one of the intent recognition result, the emotion recognition result, and the dynamic tag, response material data is retrieved from a preset knowledge base;

[0014] The static tags and dynamic tags of the target account are combined to generate account profile data for the target account;

[0015] The dialogue response data is generated based on the comparison result between the intention assessment score and the preset threshold, the response material data, and the account profile data.

[0016] In one embodiment, the process of fusing the dialogue semantic vector, the dynamic tag vector, and the behavioral feature vector to obtain a multimodal fusion vector, and generating an intention assessment score for the target account based on the multimodal fusion vector, includes:

[0017] The multimodal fusion vector is obtained by linearly weighting and fusing the dialogue semantic vector, the dynamic label vector, and the behavioral feature vector using a machine learning model based on preset weight parameters.

[0018] The intention evaluation score is obtained by performing a nonlinear transformation on the multimodal fusion vector using a preset activation function.

[0019] In one embodiment, the prompt word template includes a triple extraction template, which is used to describe the triple extraction mechanism of tag-intent-entity.

[0020] The step of extracting dynamic tags from the multi-turn dialogue data using the large language model based on a preset prompt word template, and constructing a dynamic tag vector based on the dynamic tags, includes:

[0021] The triple extraction template is combined with the multi-turn dialogue data and then input into the large language model. The large language model extracts triples consisting of label attributes, intent attributes and entity values ​​based on the triple extraction template, and uses the triples as the dynamic labels.

[0022] The dynamic tags are vectorized to obtain the dynamic tag vector.

[0023] In one embodiment, obtaining multi-turn dialogue data and behavioral trajectory data of the target account includes:

[0024] The account information of the target account is received through multiple preset channels;

[0025] The account information under each preset channel is standardized to obtain the multi-round dialogue data and the behavior trajectory data.

[0026] In one embodiment, the method further includes:

[0027] Obtain real-time conversation data of the target account;

[0028] Real-time tags are extracted from the real-time dialogue data using the prompt word template;

[0029] The dynamic tags are updated using the real-time tags.

[0030] Secondly, this application also provides an account dialogue response device, including:

[0031] The data acquisition module is used to acquire multi-turn dialogue data and behavioral trajectory data of the target account;

[0032] The semantic understanding module is used to perform natural language understanding processing on the multi-turn dialogue data using a large language model, output the corresponding dialogue semantic vector, extract dynamic tags from the multi-turn dialogue data based on a preset prompt word template using the large language model, and construct a dynamic tag vector based on the dynamic tags.

[0033] The feature extraction module is used to construct a behavior feature vector based on the behavior trajectory data;

[0034] The intention assessment module is used to fuse the dialogue semantic vector, the dynamic tag vector, and the behavioral feature vector to obtain a multimodal fusion vector, and to generate an intention assessment score for the target account based on the multimodal fusion vector;

[0035] The response generation module is used to generate dialogue response data corresponding to the target account based on the intention assessment score.

[0036] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the account dialogue response method described in any of the embodiments of the first aspect.

[0037] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the account dialogue response method described in any of the embodiments of the first aspect.

[0038] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the account dialogue response method described in any of the embodiments of the first aspect.

[0039] The aforementioned account dialogue response method, device, computer equipment, computer-readable storage medium, and computer program product, by acquiring multi-turn dialogue data of the target account, utilize a large language model to perform natural language understanding processing on the multi-turn dialogue data, and output corresponding dialogue semantic vectors, can capture the target account's true intent, emotion, and contextual information, achieving deep semantic understanding of multi-turn dialogue data. Furthermore, by using a large language model to extract dynamic tags from the multi-turn dialogue data based on preset prompt word templates, and constructing dynamic tag vectors based on these dynamic tags, it is possible to extract high-value, high-dimensional personalized tags from the multi-turn dialogue data in real time, which is helpful for... A dynamic account profile is created by acquiring the target account's behavioral trajectory data and constructing a behavioral trajectory vector based on this data. This vector is then fused with dialogue semantic vectors, dynamic tag vectors, and behavioral feature vectors to obtain a multimodal fusion vector. Based on this multimodal fusion vector, an intention assessment score for the target account is generated. By complementing and cross-validating multimodal information, the limitations of a single data source are overcome, significantly improving the accuracy of the intention assessment score. Furthermore, by generating dialogue response data corresponding to the target account based on the intention assessment score, the semantic matching degree of the dialogue response data is improved, making the generated dialogue response data more closely aligned with the target account's current dialogue intention needs. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a diagram illustrating the application environment of an account dialogue response method in one embodiment;

[0042] Figure 2 This is a flowchart illustrating an account dialogue response method in one embodiment;

[0043] Figure 3 This is a flowchart illustrating the dialogue response data generation steps in one embodiment;

[0044] Figure 4 This is a flowchart illustrating the real-time update steps of dynamic tags in one embodiment;

[0045] Figure 5This is a schematic diagram of the system architecture in one embodiment;

[0046] Figure 6 This is a structural block diagram of an account dialogue response device 600 in one embodiment;

[0047] Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0049] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the solutions, or any combination of multiple solutions. The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with relevant regulations.

[0050] The account dialogue response method provided in this application embodiment can be applied to, for example, Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or placed on a cloud or other network server. Specifically, server 104 can obtain multi-turn dialogue data and behavioral trajectory data of the target account from terminal 102; perform natural language understanding processing on the multi-turn dialogue data using a large language model, outputting corresponding dialogue semantic vectors; extract dynamic tags from the multi-turn dialogue data based on preset prompt word templates using the large language model, and construct dynamic tag vectors based on the dynamic tags; construct behavioral feature vectors based on behavioral trajectory data; fuse the dialogue semantic vectors, dynamic tag vectors, and behavioral feature vectors to obtain a multimodal fusion vector; generate an intention evaluation score for the target account based on the multimodal fusion vector; and generate dialogue response data corresponding to the target account based on the intention evaluation score. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, drones, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart in-vehicle devices, projection devices, etc. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted displays. Head-mounted displays can be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. Server 104 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server providing cloud computing services.

[0051] In one exemplary embodiment, such as Figure 2 As shown, an account dialogue response method is provided, which can be applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps S202 to S212. Wherein:

[0052] Step S202: Obtain multi-round dialogue data and behavioral trajectory data of the target account.

[0053] Among them, multi-turn dialogue data can be used to characterize a series of continuous context-related dialogue interactions between a target account and business systems through various private channels such as instant messaging groups, public information publishing platforms, and applications.

[0054] Behavioral trajectory data can be used to describe the historical behavior of a target account in the business system, such as clicking and browsing of business activity information publicly available or sent by the business system, joining instant messaging groups related to business activities of the business system, following the public information release platform of the business system, and downloading and registering applications of the business system.

[0055] Optionally, in some implementations, the server may respond to a dialogue response request sent by the terminal for the target account by receiving the current multi-turn dialogue data and behavioral trajectory data between the terminal and the target account. Alternatively, in other implementations, the server may also receive the multi-turn dialogue data and the target account's account identifier sent by the terminal, and retrieve the target account's behavioral trajectory data from the database based on the target account's account identifier.

[0056] Step S204: Use a large language model to perform natural language understanding processing on the multi-turn dialogue data and output the corresponding dialogue semantic vector.

[0057] For example, the server can deploy large language models with Natural Language Understanding (NLU) and Natural Language Generation (NLG) capabilities, such as Large Language Models (LLMs), Generative Pre-trained Transformers (GPTs), or various general and vertical domain large language models. These large language models can be used to process multi-turn dialogue data for natural language understanding: for example, the multi-turn dialogue data can be context-encoded according to the dialogue sequence, and then the context-encoded vectors can be processed using semantic feature extraction techniques based on intent and emotion to output the corresponding dialogue semantic vectors.

[0058] Step S206: Using a large language model, dynamic tags are extracted from multi-turn dialogue data based on preset prompt word templates, and dynamic tag vectors are constructed based on the dynamic tags.

[0059] For example, the server can pre-store preset prompt word templates. These prompt word templates can be generated based on historical dialogue data and dynamically labeled tags within that data. The large language model is fine-tuned using the preset prompt word templates, and then the prompt word templates are combined with multi-turn dialogue data and input into the fine-tuned large language model to output dynamic tags for the current multi-turn dialogue data. Optionally, in some implementations, the dialogue semantic vector output in step S204 can also be combined with the prompt word templates and input into the fine-tuned large language model to output corresponding dynamic tags.

[0060] Step S208: Construct a behavioral feature vector based on the behavioral trajectory data.

[0061] For example, the server can extract key features from behavioral trajectory data, such as the number of clicks within a preset period as a statistical feature; the click-through rate as a proportional feature; and the active periods of the target account as time-series features. The extracted key features are then vectorized to obtain the behavioral feature vectors corresponding to the behavioral trajectory data.

[0062] Step S210: The dialogue semantic vector, dynamic tag vector and behavioral feature vector are fused to obtain a multimodal fusion vector, and the intention evaluation score of the target account is generated based on the multimodal fusion vector.

[0063] Optionally, in some implementations, the server can use the dialogue semantic vector as the semantic dimension vector, the dynamic tag vector as the tag dimension vector, and the behavior feature vector as the behavior dimension vector. The dialogue semantic vector, dynamic tag vector, and behavior feature vector are then concatenated to obtain a multimodal fusion vector containing semantic, tag, and behavior dimensions. Alternatively, in other implementations, the server can store semantic weights, tag weights, and behavior weights. The semantic weights are weighted separately with the dialogue semantic vector, the tag weights are weighted with the dynamic tag vector, and the behavior weights are weighted with the behavior feature vector. Finally, the weighted vectors are summed to obtain the multimodal fusion vector.

[0064] Optionally, in some implementations, the server can match the multimodal fusion vector with pre-stored vectors corresponding to different intention assessment scores, and based on the vector matching results, select the intention assessment score corresponding to the most similar vector as the intention assessment score of the target account. Alternatively, in other implementations, the server can store a machine learning model. This machine learning model can be trained using multimodal fusion vectors corresponding to historical dialogue data and intention assessment scores obtained by labeling the historical dialogue data. The multimodal fusion vector is input into the machine learning model, and the machine learning model outputs the corresponding intention assessment score.

[0065] Step S212: Generate dialogue response data corresponding to the target account based on the intention assessment score.

[0066] The intention assessment score is used to characterize the target account's degree of intention to participate in business activities provided by the business system. The higher the intention assessment score, the higher the target account's intention and the higher the probability of participating in business activities. Conversely, the lower the intention, the lower the probability of participating in business activities.

[0067] For example, the server can store response strategies corresponding to different scores. A higher intention assessment score corresponds to a response strategy of transferring to a human agent, thus generating dialogue response data for the target account as "waiting to connect to a human agent." Medium and lower intention assessment scores correspond to a machine response strategy, which can retrieve corresponding business activity details data from a pre-built knowledge base as the current dialogue response data.

[0068] In the aforementioned account dialogue response method, by acquiring multi-turn dialogue data of the target account and using a large language model to perform natural language understanding processing on the multi-turn dialogue data, corresponding dialogue semantic vectors are output. This can capture the target account's true intent, emotions, and contextual information, achieving a deep semantic understanding of the multi-turn dialogue data. By using a large language model to extract dynamic tags from the multi-turn dialogue data based on preset prompt word templates, and constructing dynamic tag vectors based on these dynamic tags, high-value, high-dimensional personalized tags can be extracted from the multi-turn dialogue data in real time, which helps to form a dynamic account profile. By acquiring the target account's behavioral trajectory data and constructing behavioral trajectory vectors based on this data, a multimodal fusion vector is obtained by fusing dialogue semantic vectors, dynamic tag vectors, and behavioral feature vectors. Based on the multimodal fusion vector, an intention assessment score for the target account is generated. This overcomes the one-sidedness of a single data source through the complementarity and cross-validation of multimodal information, significantly improving the accuracy of the intention assessment score. By generating dialogue response data corresponding to the target account based on the intention assessment score, the semantic matching degree of the dialogue response data can be improved, making the generated dialogue response data more closely match the target account's current dialogue intention needs.

[0069] In one exemplary embodiment, such as Figure 3 As shown, step S212 includes steps S302 to S308. Wherein:

[0070] Step S302: Determine the intent recognition result and emotion recognition result corresponding to the multi-turn dialogue data based on the dialogue semantic vector.

[0071] For example, the server can connect to the intent classification head and the emotion classification head separately through a large language model. The dialogue semantic vectors extracted by the large language model are input into the intent classification head and the emotion classification head respectively. The intent classification head is used to determine the intent category corresponding to the dialogue semantic vector, and the intent category is output as the intent recognition result. The intent category can include, but is not limited to, categories such as price inquiry intent and time inquiry intent for business activities. The emotion classification head is used to determine the emotion category corresponding to the dialogue semantic vector, and the emotion category is output as the emotion recognition result. The emotion category can include, but is not limited to, positive, negative, anxious, and hesitant categories.

[0072] Step S304: Using at least one of the intent recognition results, emotion recognition results, and dynamic tags, retrieve response material data from the preset knowledge base.

[0073] The knowledge base can be used to store relevant information about business activities, response text templates, and other material information. Optionally, in some implementations, the material information in the knowledge base can be received through a client. Alternatively, in other implementations, the material information in the knowledge base can be extracted from historical multi-turn dialogue data.

[0074] For example, the server can store the correspondence between various material information in the knowledge base and intent categories, emotion categories, or dynamic tags. Using at least one of the intent recognition results, emotion recognition results, and dynamic tags as query keys, the server detects material information with corresponding relationships in the preset knowledge base and uses the detected material information as the current response material data.

[0075] Step S306: Combine the static and dynamic tags of the target account to generate account profile data for the target account.

[0076] Static tags for the target account can be used to represent attribute information based on the target account. For example, static tags can include, but are not limited to, tags generated from attribute information such as the target account's age and region.

[0077] For example, the server can pre-read the target account's attribute information from the client and use this information to generate static tags for the target account. The static tags of the target account are then combined with dynamically extracted tags in real time, and the combined tags are subjected to vectorization or natural language processing to form account profile data for the target account.

[0078] Step S308: Generate dialogue response data based on the comparison results of the intention assessment score and the preset threshold, the response material data, and the account profile data.

[0079] For example, the server can compare the current target account's intent assessment score with a preset threshold. Based on the comparison result, a corresponding response strategy is determined. The response strategy, response material data, and account profile data are combined to construct dialogue response prompts. These prompts are then input into a large language model, which performs a text generation task based on the prompts to generate corresponding dialogue response data. Subsequently, the server can send the dialogue response data to the client, which then replies to the target account using the dialogue response data.

[0080] Optionally, in other embodiments, the preset thresholds may include multiple thresholds, such as a first threshold, a second threshold, and a third threshold ordered from high to low. If the intention assessment score is higher than or equal to the first threshold, the corresponding response strategy can be determined as a strategy to transfer to a live agent. If the intention assessment score is lower than the first threshold but higher than or equal to the second threshold, the corresponding response strategy can be determined as a business activity promotion strategy. If the intention assessment score is lower than the second threshold but higher than or equal to the third threshold, the corresponding response strategy can be determined as a business scenario guidance strategy. If the intention assessment score is lower than the third threshold, the corresponding response strategy can be determined as a strategy to continue the conversation.

[0081] In this embodiment, the intent recognition result and emotion recognition result corresponding to the multi-turn dialogue data are determined based on the dialogue semantic vector; the response material data is retrieved from the preset knowledge base using at least one of the intent recognition result, emotion recognition result, and dynamic tags; the account profile data of the target account is generated by combining the static tags and dynamic tags of the target account; and the dialogue response data is generated based on the comparison result of the intent evaluation score and the preset threshold, the response material data, and the account profile data. This can understand the intent needs and emotional state of the target account, improve the semantic naturalness of the dialogue response data, and make it more personalized to the target account.

[0082] In an exemplary embodiment, step S210 may include: using a machine learning model to linearly weight and fuse the dialogue semantic vector, dynamic label vector, and behavioral feature vector based on preset weight parameters to obtain a multimodal fusion vector, and using a preset activation function to perform nonlinear transformation processing on the multimodal fusion vector to obtain an intention evaluation score.

[0083] For example, the server can store machine learning models for intention assessment, such as machine learning models based on Transformer and Multilayer Perceptron (MLP), Gradient Boosting Decision Tree (GBDT), Support Vector Machine (SVM), etc. Using the machine learning model, the dialogue semantic vector, dynamic label vector, and behavioral feature vector are linearly weighted and fused according to a preset intention assessment formula and preset weight parameters to obtain a multimodal fused vector. A preset activation function is then used to perform a nonlinear transformation on the multimodal fused vector to obtain an intention assessment score.

[0084] The formula for assessing intent can be referenced as follows:

[0085] ,

[0086] in, This is the score for the intention assessment. This is the Sigmoid activation function. These are the weight parameters corresponding to the dialogue semantic vector. This is the dialogue semantic vector. These are the weight parameters corresponding to the behavioral feature vector. This is the behavioral feature vector. These are the weight parameters corresponding to the dynamic label vector. This is a dynamic label vector. This is a bias term.

[0087] In this embodiment, a machine learning model is used to linearly weight and fuse the dialogue semantic vector, dynamic label vector, and behavioral feature vector based on preset weight parameters. A preset activation function is then used to perform nonlinear transformation on the result of the linear weighted fusion to obtain an intention assessment score, which can improve the accuracy of the intention assessment score.

[0088] In one exemplary embodiment, the prompt word template includes a triple extraction template, which describes a tag-intent-entity triple extraction mechanism.

[0089] Step S206 may include: combining the triple extraction template with multi-turn dialogue data and inputting it into a large language model; using the large language model to extract triples composed of label attributes, intent attributes, and entity values ​​based on the triple extraction template, and using the triples as dynamic labels; and vectorizing the dynamic labels to obtain dynamic label vectors.

[0090] For example, the server can combine a triplet extraction template with multi-turn dialogue data and input it into a large language model. The large language model then performs natural language understanding processing on the multi-turn dialogue data based on the triplet extraction template, identifying tag attributes, intent attributes, and entity values ​​in the multi-turn dialogue data. Based on semantics or context, it constructs triplet relationships between tag attributes, intent attributes, and entity values, forming triples composed of tag attributes, intent attributes, and entity values. These triples serve as the current dynamic tags for the target account. The dynamic tags are then vectorized to form dynamic tag vectors.

[0091] Optionally, in some implementations, the tags in the triple extraction template may include, but are not limited to, time constraint tags, obstacle tags, price focus tags, and business need tags. For example, when multi-turn dialogue data contains phrases such as "free in the evening," "no prior knowledge," "inquiring about prices," and "wanting to get certified," the following triples can be extracted: {tag attribute (time constraint tag); intent attribute (description); entity value (free in the evening)}, {tag attribute (obstacle tag); intent attribute (description); entity value (no prior knowledge)}, {tag attribute (price focus tag); intent attribute (inquiring about prices)}, and {tag attribute (business need tag); intent attribute (inquiring about prices); entity value (wanting to get certified)}.

[0092] In this embodiment, dynamic tags are obtained by using a tag-intent-entity triple extraction mechanism. Compared with the traditional static tag system, this can continuously reflect the changes in the status, interests and needs of the target account, which helps to achieve real-time and high-dimensional dynamic updates of the target account profile.

[0093] In an exemplary embodiment, step S202 includes: receiving account information of the target account through multiple preset channels. The account information from each preset channel is then standardized to obtain multi-round dialogue data and behavioral trajectory data.

[0094] The preset channels may include, but are not limited to, private domain channels such as instant messaging groups, public information publishing platforms, and applications.

[0095] Account information may include, but is not limited to, conversation information and behavioral information of the target account interacting with business systems or business activities through preset channels.

[0096] For example, the server can receive account information of the target account through multiple preset channels. Based on a preset standard data format, the account information under each preset channel is standardized to obtain multi-round dialogue data and behavioral trajectory data of the target account from different preset channels that conform to the standard data format.

[0097] Optionally, in some implementations, the server may record the preset channels through which the current business system communicates with the target account, and send the communication response data to the target account according to the preset channels.

[0098] In this embodiment, by standardizing the account information collected from different preset channels, multi-turn dialogue data and behavioral trajectory data are obtained, which can eliminate the data differences between different preset channels and facilitate providing a unified input for large language models and behavioral feature vectors.

[0099] In one exemplary embodiment, such as Figure 4As shown, the account dialogue response method provided in this application may further include steps S402 to S406. Wherein:

[0100] Step S402: Obtain real-time conversation data of the target account.

[0101] Step S404: Extract real-time tags from real-time dialogue data using prompt word templates.

[0102] Step S406: Update dynamic tags using real-time tags.

[0103] For example, after sending dialogue response data to the target account, the server can continuously monitor subsequent real-time dialogue data of the target account. The server combines the prompt word template with the real-time dialogue data and inputs the combined data into a large language model. Using the large language model and referring to the dynamic tag extraction method described above, real-time tags are extracted from the real-time dialogue data. The extracted real-time tags are then used to update the original dynamic tags. Subsequently, the server can use the updated dynamic tags to update the target account's account profile data.

[0104] In this embodiment, by continuously acquiring real-time dialogue data of the target account, using a large language model to extract real-time tags from the real-time dialogue data based on a preset prompt word template, and using real-time tags to update dynamic tags, the real-time performance and accuracy of dynamic tags can be improved.

[0105] In one exemplary embodiment, such as Figure 5 As shown, a schematic diagram of a server system architecture is provided, including a data and storage layer 510, a core processing layer 520, and an interface and access layer 530.

[0106] The data and storage layer 510 may include a knowledge base and a user profile library. The knowledge base can be used to store business knowledge content. The user profile library can be used to store static tags, dynamic tags, and account profile data generated by combining static and dynamic tags for target accounts.

[0107] The core processing layer 520 may include a multi-channel data access module, a large-model dialogue processing module, a dynamic tagging and intent assessment module, and an intelligent outreach execution module. The multi-channel data access module receives account information from multiple preset channels, standardizes the different formats of account information from these channels, and transmits it downwards in a unified manner. The large-model dialogue processing module utilizes a large language model to perform core tasks such as natural language understanding, intent recognition, sentiment analysis, response strategy decision-making, and dialogue response data generation on multi-turn dialogue data. The dynamic tagging and intent assessment module fuses dialogue semantic vectors, behavioral feature vectors, and dynamic tag vectors to calculate intent assessment scores in real time, and extracts real-time tags from real-time dialogue data and updates dynamic tags using these real-time tags. The intelligent outreach execution module executes response strategies, sending dialogue response data to the target account through the corresponding preset channels.

[0108] The interface and reach layer 530 may include a channel interface encapsulation layer. The channel interface encapsulation layer can be used to encapsulate the information transmission interfaces of multiple preset channels, ensuring that dialogue response data is accurately, stably, and in a standardized manner pushed to the target account.

[0109] In this embodiment, the dialogue response method provided in any of the above embodiments is implemented through the system architecture described above. This effectively overcomes the technical problems of inaccurate intention assessment caused by insufficient depth of dialogue understanding and the single dimension of static tags in the prior art. It accurately captures the emotions and true intentions of the target account, provides targeted and personalized dialogue response replies, and significantly improves the semantic matching degree between dialogue response data and the target account. This embodiment extracts high-dimensional dynamic tags from unstructured multi-turn dialogue data in real time, enriching the account profile data and providing a high-quality data foundation for subsequent dialogue responses.

[0110] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0111] Based on the same inventive concept, this application also provides an account dialogue response device for implementing the account dialogue response method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations of one or more account dialogue response device embodiments provided below can be found in the limitations of the account dialogue response method described above, and will not be repeated here.

[0112] In one exemplary embodiment, such as Figure 6 As shown, an account dialogue response device 600 is provided, including: a data acquisition module 602, a semantic understanding module 604, a feature extraction module 606, an intention assessment module 608, and a response generation module 610, wherein:

[0113] The data acquisition module 602 is used to acquire multi-round dialogue data and behavioral trajectory data of the target account.

[0114] The semantic understanding module 604 is used to perform natural language understanding processing on multi-turn dialogue data using a large language model, output the corresponding dialogue semantic vector, extract dynamic tags from the multi-turn dialogue data based on a preset prompt word template using the large language model, and construct dynamic tag vectors based on the dynamic tags.

[0115] The feature extraction module 606 is used to construct behavioral feature vectors based on behavioral trajectory data.

[0116] The intention assessment module 608 is used to fuse dialogue semantic vectors, dynamic tag vectors and behavioral feature vectors to obtain a multimodal fusion vector, and generate an intention assessment score for the target account based on the multimodal fusion vector.

[0117] The response generation module 610 is used to generate dialogue response data corresponding to the target account based on the intention assessment score.

[0118] In an exemplary embodiment, the response generation module 610 is further configured to determine the intent recognition result and emotion recognition result corresponding to the multi-turn dialogue data based on the dialogue semantic vector; retrieve response material data from a preset knowledge base using at least one of the intent recognition result, emotion recognition result, and dynamic tags; generate account profile data of the target account by combining the static tags and dynamic tags of the target account; and generate dialogue response data based on the comparison result of the intent evaluation score and a preset threshold, the response material data, and the account profile data.

[0119] In an exemplary embodiment, the intention assessment module 608 is further configured to use a machine learning model to linearly weight and fuse the dialogue semantic vector, dynamic label vector, and behavioral feature vector based on preset weight parameters to obtain a multimodal fusion vector; and to use a preset activation function to perform nonlinear transformation processing on the multimodal fusion vector to obtain an intention assessment score.

[0120] In an exemplary embodiment, the prompt word template includes a triple extraction template, which describes the triple extraction mechanism of tag-intent-entity. The semantic understanding module 604 is further configured to combine the triple extraction template with multi-turn dialogue data and input it into a large language model. Based on the triple extraction template, the large language model extracts triples consisting of tag attributes, intent attributes, and entity values, and uses these triples as dynamic tags. The dynamic tags are then vectorized to obtain dynamic tag vectors.

[0121] In an exemplary embodiment, the data acquisition module 602 is further configured to receive account information of the target account through multiple preset channels; and to standardize the account information under each preset channel to obtain multi-round dialogue data and behavior trajectory data.

[0122] In an exemplary embodiment, the semantic understanding module 604 is further configured to acquire real-time dialogue data of the target account; extract real-time tags from the real-time dialogue data using prompt word templates; and update dynamic tags using real-time tags.

[0123] Each module in the aforementioned account dialogue response device 600 can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0124] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 7As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores multi-turn dialogue data, behavioral trajectory data, dialogue semantic vectors, dynamic tags, dynamic tag vectors, behavioral feature vectors, multimodal fusion vectors, intention evaluation scores, dialogue response data, and other data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements an account-based dialogue response method.

[0125] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0126] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0127] In one exemplary embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0128] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0129] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0130] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0131] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An account dialogue response method, characterized by, The method includes: Obtain multi-turn conversation data and behavioral trajectory data of the target account; The multi-turn dialogue data is processed using a large language model to perform natural language understanding and output the corresponding dialogue semantic vector. Dynamic tags are extracted from the multi-turn dialogue data using the large language model based on a preset prompt word template, and a dynamic tag vector is constructed based on the dynamic tags; Construct a behavioral feature vector based on the behavioral trajectory data; The dialogue semantic vector, the dynamic tag vector, and the behavioral feature vector are fused to obtain a multimodal fusion vector, and the intention evaluation score of the target account is generated based on the multimodal fusion vector; Based on the intention assessment score, generate dialogue response data corresponding to the target account.

2. The method of claim 1, wherein, The step of generating dialogue response data corresponding to the target account based on the intention assessment score includes: Based on the dialogue semantic vector, determine the intent recognition result and emotion recognition result corresponding to the multi-turn dialogue data; Using at least one of the intent recognition result, the emotion recognition result, and the dynamic tag, response material data is retrieved from a preset knowledge base; The static tags and dynamic tags of the target account are combined to generate account profile data for the target account; The dialogue response data is generated based on the comparison result between the intention assessment score and the preset threshold, the response material data, and the account profile data.

3. The method of claim 1, wherein, The process of fusing the dialogue semantic vector, the dynamic tag vector, and the behavioral feature vector to obtain a multimodal fusion vector, and generating an intention assessment score for the target account based on the multimodal fusion vector, includes: The multimodal fusion vector is obtained by linearly weighting and fusing the dialogue semantic vector, the dynamic label vector, and the behavioral feature vector using a machine learning model based on preset weight parameters. The intention evaluation score is obtained by performing a nonlinear transformation on the multimodal fusion vector using a preset activation function.

4. The method of claim 1, wherein, The prompt word template includes a triple extraction template, which is used to describe the triple extraction mechanism of tag-intent-entity. The step of extracting dynamic tags from the multi-turn dialogue data using the large language model based on a preset prompt word template, and constructing a dynamic tag vector based on the dynamic tags, includes: The triple extraction template is combined with the multi-turn dialogue data and then input into the large language model. The large language model extracts triples consisting of label attributes, intent attributes and entity values ​​based on the triple extraction template, and uses the triples as the dynamic labels. The dynamic tags are vectorized to obtain the dynamic tag vector.

5. The method according to claim 1, characterized in that, The acquisition of multi-turn dialogue data and behavioral trajectory data of the target account includes: The account information of the target account is received through multiple preset channels; The account information under each preset channel is standardized to obtain the multi-round dialogue data and the behavior trajectory data.

6. The method according to claim 1, characterized in that, The method further includes: Obtain real-time conversation data of the target account; Real-time tags are extracted from the real-time dialogue data using the prompt word template; The dynamic tags are updated using the real-time tags.

7. An account dialogue response device, characterized in that, The device includes: The data acquisition module is used to acquire multi-turn dialogue data and behavioral trajectory data of the target account; The semantic understanding module is used to perform natural language understanding processing on the multi-turn dialogue data using a large language model, output the corresponding dialogue semantic vector, extract dynamic tags from the multi-turn dialogue data based on a preset prompt word template using the large language model, and construct a dynamic tag vector based on the dynamic tags. The feature extraction module is used to construct a behavior feature vector based on the behavior trajectory data; The intention assessment module is used to fuse the dialogue semantic vector, the dynamic tag vector, and the behavioral feature vector to obtain a multimodal fusion vector, and to generate an intention assessment score for the target account based on the multimodal fusion vector; The response generation module is used to generate dialogue response data corresponding to the target account based on the intention assessment score.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.