Method for optimizing multi-turn dialogue logic in intelligent question-answering system based on knowledge graph

Abstract

The application discloses a multi-round dialogue logic optimization method in an intelligent question-answering system based on a knowledge graph, and belongs to the technical field of artificial intelligence and natural language processing, and comprises the following steps: S1, constructing and dynamically updating a triple knowledge graph; S2, based on a graph neural network and a graph attention mechanism, modeling time sequence and emotion perception of cross-round semantic association in the triple knowledge graph, and outputting a dialogue context vector; S3, inputting the dialogue context vector and current user input features into a generation model, performing semantic mapping through an independent intent, entity and behavior embedding encoder, and combining logical constraints and emotion-driven weight adjustment from the triple knowledge graph to generate a robot behavior; S4, optimizing the generation model based on a task-related adversarial training mechanism; and S5, performing multi-modal emotion fusion and consistency verification.The application can realize dynamic, coherent and emotion-intelligent multi-round dialogue logic optimization.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing technology, specifically a method for optimizing multi-turn dialogue logic in a knowledge graph-based intelligent question-answering system. Background Technology

[0002] Optimizing the multi-turn dialogue logic of intelligent question-answering systems is a core technical direction for improving the quality of human-computer interaction. Its core objective is to solve problems such as weak cross-turn contextual association and inconsistent behavior prediction logic through structured semantic modeling and dynamic constraint control.

[0003] Current technical approaches primarily rely on the structured storage capabilities of knowledge graphs and the context-aware capabilities of generative models (such as GPT). However, a significant contradiction exists when these two approaches are combined: the static structure of knowledge graphs struggles to dynamically capture the evolution of dialogue and the continuous changes in user emotional states, while the open output of generative models can lead to behavioral predictions deviating from graph constraints and failing to adapt to immediate shifts in user emotions, thus hindering the formation of closed-loop logical control and emotionally intelligent responses. In particular, existing solutions generally neglect modeling the evolution path of user emotions in multi-turn dialogues, resulting in a lack of emotional adaptability in system responses, which is especially prominent in highly emotional interaction scenarios such as consultations and complaints. Therefore, there is an urgent need for a technical solution that can dynamically model dialogue context and emotional states and inject graph rules and emotion perception into generative models.

[0004] To address these issues, those skilled in the art have provided a method for optimizing multi-turn dialogue logic in knowledge graph-based intelligent question-answering systems. Summary of the Invention

[0005] The purpose of this invention is to provide a method for optimizing multi-turn dialogue logic in a knowledge graph-based intelligent question-answering system, which can achieve dynamic, coherent, and emotionally intelligent multi-turn dialogue logic optimization to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for optimizing multi-turn dialogue logic in a knowledge graph-based intelligent question-answering system includes the following steps:

[0008] S1. Construct and dynamically update the triplet knowledge graph, which is used to store the dialogue state and semantic associations containing user intent, user entity, robot behavior and user emotion.

[0009] S2. Based on graph neural networks and graph attention mechanisms, temporal and emotion perception modeling is performed on cross-round semantic associations in triplet knowledge graphs to output dialogue context vectors.

[0010] S3. Input the dialogue context vector and the current user input features into the generative model, perform semantic mapping through independent intent, entity and behavior embedding encoders, and combine logical constraints from the triple knowledge graph with emotion-driven weight adjustment to generate robot behavior;

[0011] S4. The adversarial training mechanism based on task association is used to optimize the generative model and dynamically update the triple knowledge graph.

[0012] S5. Perform multimodal emotion fusion and consistency verification.

[0013] As a further aspect of the present invention: step S1 specifically includes:

[0014] S101. Define the nodes and edges of the triplet knowledge graph; where user intent nodes have intent type attributes and emotion tag attributes, user entity nodes have entity category and entity value attributes, and robot behavior nodes have behavior type attributes; relation edges are used to connect nodes of different types;

[0015] S102. Parse user input, extract user intent, user entity and emotion tag for the current round, insert them as corresponding nodes into the triple knowledge graph, and establish corresponding relationship edges.

[0016] As a further aspect of the present invention: step S2 includes: using a graph attention mechanism to calculate the attention weights between nodes in the triplet knowledge graph, so as to aggregate neighbor node information and generate the dialogue context vector;

[0017] The process of calculating attention weights includes: for any two nodes i and j in the triplet knowledge graph, firstly, a time decay factor that decreases as the time interval increases is calculated based on the dialogue turn interval between nodes i and j; secondly, an emotion perception gain factor is calculated based on the difference in emotion intensity values ​​associated with nodes i and j; finally, the attention weight of node i on node j is calculated through an attention mechanism by combining node features, the time decay factor, and the emotion perception gain factor.

[0018] The attention weight The calculation formula is:

[0019] ;

[0020] in, This represents the attention weight of node i to node j in the t-th round of dialogue; The time decay factor is... This is the difference between the dialogue turn numbers corresponding to nodes i and j. The attenuation coefficient is configurable. This refers to the emotion perception gain factor. and These are the scalar values ​​of emotion intensity assigned to nodes i and j, respectively. A configurable emotion sensitivity coefficient; and Let be the feature vectors of nodes i and j in round t, respectively; W is the learnable weight matrix; and a is the learnable parameter vector of the attention mechanism. This indicates vector concatenation; This represents the set of neighboring nodes that are directly connected to node i through relation edges in the triplet knowledge graph.

[0021] As a further aspect of the present invention: the dialogue context vector generated in step S2 includes historical intent evolution information, entity association information, and emotional state transition information.

[0022] As a further aspect of the present invention: in step S3, the logical constraint is a rule mask constructed based on the mapping relationship between historical intent and behavior in the triplet knowledge graph, which is used to mask behavior options that do not conform to the mapping relationship when the generation model predicts behavior.

[0023] As a further aspect of the present invention: In step S3, the emotion-driven weight adjustment is achieved in the following way: In the output layer of the generative model, the emotion gain coefficient is calculated based on the current user's emotional state, and the initial probability distribution of the candidate behavior is adjusted using this coefficient.

[0024] As a further embodiment of the present invention, step S5 specifically includes:

[0025] S501. Obtain the text modal emotion feature vector and the speech modal emotion feature vector input by the user;

[0026] S502. Based on the attention mechanism, the emotion feature vectors of the text modality and the speech modality are fused to obtain a joint emotion representation;

[0027] S503. Calculate the consistency score between the joint emotion representation and the current emotion node representation in the triplet knowledge graph;

[0028] S504. If the consistency score is lower than a preset threshold, a clarification mechanism is triggered and the weights of relevant emotion nodes in the triplet knowledge graph are adjusted.

[0029] As a further aspect of the present invention: in step S503, the calculation process of the consistency score includes: calculating the cosine similarity between the joint emotion representation and the current emotion node representation in the triplet knowledge graph; calculating the difference between the emotion feature vectors of the text modality and the speech modality; weighting and combining the cosine similarity and the difference, and mapping them through the Sigmoid function to obtain the consistency score;

[0030] The formula for calculating the consistency score (C) is as follows:

[0031] ;

[0032] in, For the joint emotion representation, Let be the representation vector of the current sentiment node in the triplet knowledge graph. The cosine similarity function; and These are the text modal emotion feature vectors and the speech modal emotion feature vectors extracted by a pre-trained emotion encoder, respectively. γ represents the L2 norm; w is the logical consistency weight parameter, and γ is the conflict sensitivity factor parameter; This is the Sigmoid function.

[0033] As a further aspect of the present invention: the adversarial training mechanism in step S4 includes a generator and a discriminator;

[0034] The generator is used to generate robot responses based on the dialogue context vector;

[0035] The discriminator is used to determine the consistency between the robot's response and the dialogue task objective based on user feedback data;

[0036] The generator and discriminator are trained by optimizing the adversarial loss function. for:

[0037] ;

[0038] Wherein, G represents the generator, z is the dialogue context vector input to the generator; D represents the discriminator, and x is real user feedback data. For the true data distribution, This refers to the distribution of the dialogue context vectors.

[0039] As a further aspect of the present invention: the dynamic updating of the triplet knowledge graph includes: when a new user intent or a new user entity is detected, inserting a corresponding new node into the triplet knowledge graph and establishing relational edges;

[0040] The method also includes a dialogue termination step: when the generated behavior is a dialogue end marker, or when the user has no new input for more than a preset time, the graph nodes related to the current dialogue are cleared.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] This invention achieves accurate retention of multi-turn dialogue context and semantic association modeling through the deep fusion of dynamic knowledge graphs and emotion-aware graph neural networks. Independent encoders and a dual-constraint mechanism ensure the logical rationality and emotional adaptability of generated behaviors. Task-related adversarial training and multimodal verification further improve the system's response quality and robustness. This method significantly enhances the performance of intelligent question-answering systems in complex, multi-turn, and highly emotional interaction scenarios, enabling dynamic, coherent, and emotionally intelligent multi-turn dialogue logic optimization. Attached Figure Description

[0043] Figure 1 This is a flowchart of a method for optimizing multi-turn dialogue logic in a knowledge graph-based intelligent question-answering system. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] As mentioned in the background section of this application, research has found that the static structure of existing knowledge graphs struggles to dynamically capture the evolution of dialogue and the continuous changes in users' emotional states. Furthermore, the open output of generative models can easily lead to behavioral predictions deviating from graph constraints and failing to adapt to immediate shifts in user emotions, thus hindering the formation of closed-loop logical control and emotionally intelligent responses. In particular, existing solutions generally neglect modeling the evolution path of user emotions in multi-turn dialogues, resulting in a lack of emotional adaptability in system responses. This is especially prominent in highly emotional interaction scenarios such as consultations and complaints, presenting a significant deficiency.

[0046] To address the aforementioned shortcomings, this application discloses a method for optimizing multi-turn dialogue logic in a knowledge graph-based intelligent question-answering system. By constructing a dynamically updated triplet knowledge graph to structurally store dialogue states, and utilizing graph neural networks (GNNs) and graph attention mechanisms (GATs) to perform cross-turn semantic modeling of temporal and emotion perception, the method drives the generative model to make accurate decisions under the constraints of the knowledge graph and the guidance of emotion. This enables dynamic, coherent, and emotionally intelligent optimization of multi-turn dialogue logic.

[0047] The following will describe in detail, with reference to the accompanying drawings, how the solution of this application solves the above-mentioned technical problems.

[0048] Please see Figure 1 In this embodiment of the invention, a method for optimizing multi-turn dialogue logic in an intelligent question-answering system based on a knowledge graph includes the following steps: S1, constructing and dynamically updating a triplet knowledge graph, which stores dialogue states and semantic associations containing user intent, user entities, robot behavior, and user emotions; S2, based on graph neural networks and graph attention mechanisms, performing temporal and emotion perception modeling on cross-turn semantic associations in the triplet knowledge graph, and outputting a dialogue context vector; S3, inputting the dialogue context vector and current user input features into a generative model, performing semantic mapping through independent intent, entity, and behavior embedding encoders, and generating robot behavior by combining logical constraints and emotion-driven weight adjustments from the triplet knowledge graph; S4, optimizing the generative model based on a task-related adversarial training mechanism, and dynamically updating the triplet knowledge graph; S5, performing multimodal emotion fusion and consistency verification. This application systematically solves the problems of contextual coherence, logical consistency, and emotion adaptation in multi-turn dialogues by constructing a dynamic knowledge graph and combining graph neural networks and generative models.

[0049] In this embodiment, step S1 specifically includes: S101, defining the nodes and edges of the triplet knowledge graph; wherein, user intent nodes have intent type attributes and emotion tag attributes, user entity nodes have entity category and entity value attributes, and robot behavior nodes have behavior type attributes; relation edges are used to connect nodes of different types; S102, parsing user input, extracting the user intent, user entity, and emotion tag of the current round, inserting them as corresponding nodes into the triplet knowledge graph, and establishing corresponding relation edges. This setting concretizes the steps of knowledge graph construction and updating, providing a structured and dynamically expandable dialogue state representation foundation for subsequent semantic association modeling and decision optimization, ensuring that the intent, entity, and emotion information of each round of dialogue are accurately recorded and associated.

[0050] In this embodiment, step S2 includes: calculating the attention weights between nodes in the triplet knowledge graph using a graph attention mechanism to aggregate neighbor node information and generate a dialogue context vector; wherein, the process of calculating the attention weights includes: for any two nodes i and j in the triplet knowledge graph, firstly, calculating a time decay factor that decays as the time interval increases based on the dialogue round interval between nodes i and j; secondly, calculating an emotion perception gain factor based on the difference in emotion intensity values ​​associated with nodes i and j; finally, combining node features, the time decay factor, and the emotion perception gain factor, calculating the attention weight of node i on node j through the attention mechanism;

[0051] Attention weight The calculation formula is:

[0052] ;

[0053] in, This represents the attention weight of node i to node j in the t-th round of dialogue; The time decay factor, This is the difference between the dialogue turn numbers corresponding to nodes i and j. The attenuation coefficient is configurable. As an emotion perception gain factor, and These are the scalar values ​​of emotion intensity assigned to nodes i and j, respectively. A configurable emotion sensitivity coefficient; and Let be the feature vectors of nodes i and j in round t, respectively; W is the learnable weight matrix; and a is the learnable parameter vector of the attention mechanism. This indicates vector concatenation; This represents the set of neighboring nodes directly connected to node i via relation edges in the triplet knowledge graph. This setting clarifies the core computational mechanism for cross-round semantic modeling. By fusing dynamic attention weights based on time decay and emotion perception, the model can prioritize recent historical dialogue information with significant emotional changes, thereby accurately capturing the evolution path of user intent and emotion, and providing key contextual features for generating coherent and emotionally appropriate responses.

[0054] In this embodiment, the dialogue context vector generated in step S2 includes historical intent evolution information, entity association information, and emotional state transition information. This setting defines the specific content composition of the dialogue context vector, integrating and encoding key semantic elements (intent, entity) and emotional elements (emotion) scattered across multiple rounds of dialogue, providing a unified representation containing multi-dimensional and structured information for subsequent generation decisions.

[0055] In this embodiment, in step S3, the logical constraint is a rule mask constructed based on the mapping relationship between historical intents and behaviors in the triplet knowledge graph. This mask is used to filter out behavioral options that do not conform to the mapping relationship when the generation model predicts behavior. This setting specifically illustrates the implementation method of the logical constraint. The dialogue logic rules (such as the mapping from intent to behavior) accumulated in the knowledge graph are injected into the generation process as hard constraints, directly filtering out logically unreasonable behavioral options, reducing the logical error rate of responses from the source, and ensuring that the main dialogue line does not deviate.

[0056] In this embodiment, in step S3, the emotion-driven weight adjustment is implemented as follows: At the output layer of the generative model, an emotion gain coefficient is calculated based on the current user's emotional state, and this coefficient is used to weight and adjust the initial probability distribution of candidate behaviors. This setting specifically illustrates the implementation method of emotion-driven weight adjustment. Based on the real-time perceived user emotional state, the generation probability of different candidate behaviors is dynamically adjusted, so that when making decisions, the system tends to select response behaviors that are more likely to adapt to the current user's emotions (such as soothing, encouraging, or positive interaction), thereby improving the adaptability of emotional interaction.

[0057] In this embodiment, step S5 specifically includes: S501, obtaining the text modality emotion feature vector and the speech modality emotion feature vector input by the user; S502, fusing the emotion feature vectors of the text modality and the speech modality based on an attention mechanism to obtain a joint emotion representation; S503, calculating the consistency score between the joint emotion representation and the current emotion node representation in the triplet knowledge graph; S504, if the consistency score is lower than a preset threshold, triggering a clarification mechanism and adjusting the weights of relevant emotion nodes in the triplet knowledge graph. This setting concretizes the steps of multimodal emotion fusion and verification. By integrating text and speech information for cross-validation, it improves the robustness and accuracy of emotion state discrimination. When multimodal information conflicts or inconsistencies with historical context are detected, it actively initiates interactive clarification, thereby dynamically correcting the emotion model in the knowledge graph and enhancing the system's understanding and response reliability to complex emotional interactions.

[0058] In this embodiment, the consistency score calculation process in step S503 includes: calculating the cosine similarity between the joint emotion representation and the current emotion node representation in the triplet knowledge graph; calculating the difference between the emotion feature vectors of the text modality and the speech modality; weighting the cosine similarity and the difference, and mapping them through the Sigmoid function to obtain the consistency score.

[0059] The formula for calculating the consistency score (C) is:

[0060] ;

[0061] in, To express joint emotions, Let be the representation vector of the current sentiment node in the triplet knowledge graph. The cosine similarity function; and These are the text modal emotion feature vectors and the speech modal emotion feature vectors extracted by a pre-trained emotion encoder, respectively. γ represents the L2 norm; w is the logical consistency weight parameter, and γ is the conflict sensitivity factor parameter; This is the Sigmoid function. This setting provides a specific calculation method for quantifying sentiment consistency, offering a measurable metric that simultaneously considers the logical coherence between the current multimodal sentiment inference and the historical sentiment context, as well as the degree of inherent conflict between different modal sentiment signals. This provides a precise numerical basis for the system to decide whether clarification or adjustment is needed.

[0062] In this embodiment, the adversarial training mechanism in step S4 includes a generator and a discriminator; the generator is used to generate robot responses based on dialogue context vectors; the discriminator is used to determine the consistency between the robot responses and the dialogue task objectives based on user feedback data; the generator and discriminator are trained by optimizing the adversarial loss function. for:

[0063] ;

[0064] Where G represents the generator, z is the dialogue context vector input to the generator; D represents the discriminator, and x is the real user feedback data. For the true data distribution, This defines the distribution of dialogue context vectors. This setting clarifies the specific architecture and objectives of adversarial training. By introducing a discriminator based on real user feedback, the generator's responses not only conform to the logic and emotional constraints of the knowledge graph in terms of content, but also more closely align with real and effective dialogue goals in terms of task completion and user experience, thereby improving the overall relevance and usability of the system's responses.

[0065] In this embodiment, dynamically updating the triplet knowledge graph includes: when a new user intent or a new user entity is detected, inserting a corresponding new node into the triplet knowledge graph and establishing relational edges; the method also includes a dialogue termination step: when the generated behavior is a dialogue end marker, or when the user has no new input for more than a preset time, clearing the graph nodes related to the current dialogue. This setting specifies the knowledge graph update and dialogue termination mechanism, ensuring the system's adaptability and resource management efficiency. Incremental updates enable the knowledge graph to continuously expand to adapt to new dialogue scenarios, while the termination mechanism promptly clears the state of ended dialogues, avoiding interference from invalid historical information in subsequent dialogues and ensuring the long-term stable operation of the system in multi-turn dialogues.

[0066] This invention achieves accurate retention of multi-turn dialogue context and semantic association modeling through the deep fusion of dynamic knowledge graphs and emotion-aware graph neural networks. Independent encoders and a dual-constraint mechanism ensure the logical rationality and emotional adaptability of generated behaviors. Task-related adversarial training and multimodal verification further improve the system's response quality and robustness. This method significantly enhances the performance of intelligent question-answering systems in complex, multi-turn, and highly emotional interaction scenarios, enabling dynamic, coherent, and emotionally intelligent multi-turn dialogue logic optimization.

[0067] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0068] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for optimizing multi-turn dialogue logic in an intelligent question-answering system based on a knowledge graph, characterized in that, The method comprises the following steps: S1, constructing and dynamically updating a triple knowledge graph, the triple knowledge graph being used to store semantic association of dialogue state containing user intention, user entity, robot behavior and user emotion; specifically comprising: parsing user input, extracting user intention, user entity and emotion label of the current turn, inserting the corresponding nodes into the triple knowledge graph, and establishing the corresponding relationship edges; S2, calculating attention weight between nodes in the triple knowledge graph by using graph attention mechanism, to aggregate neighbor node information and generate dialogue context vector; wherein, the process of calculating the attention weight comprises: for any two nodes i and j in the triple knowledge graph, first, a time decay factor is calculated according to the dialogue turn interval between the nodes i and j, which increases with the increase of the time interval; second, an emotion perception gain factor is calculated according to the difference between the emotion intensity values associated with the nodes i and j; finally, the attention weight of the node i to the node j is calculated by the attention mechanism combined with the node features, the time decay factor and the emotion perception gain factor; S3, inputting the dialogue context vector and the current user input features into the generation model, the current user input features being the current turn user intention, user entity and emotion label extracted by parsing the user input, performing semantic mapping by independent intention, entity and behavior embedding encoder, and generating robot behavior combined with logical constraints and emotion driven weight adjustment from the triple knowledge graph; the emotion driven weight adjustment is realized by the following way: in the output layer of the generation model, an emotion gain coefficient is calculated according to the current user emotion state, and the coefficient is used to weight and adjust the initial probability distribution of the candidate behavior; S4, optimizing the generation model based on the task associated adversarial training mechanism, and dynamically updating the triple knowledge graph; S5, multi-modal emotion fusion and consistency verification, wherein the multi-modal emotion includes text modal emotion and speech modal emotion. 2.The method of claim 1, wherein, The step S5 specifically comprises: S501, obtaining text modal emotion feature vector and speech modal emotion feature vector of user input; S502, fusing the text modal and speech modal emotion feature vectors based on attention mechanism to obtain joint emotion representation; S503, calculating consistency score of the joint emotion representation and the current emotion node representation in the triple knowledge graph; S504, if the consistency score is lower than a preset threshold, triggering a clarification mechanism and adjusting the weight of the related emotion node in the triple knowledge graph. 3.The method of claim 2, wherein, In the step S503, the calculation process of the consistency score comprises: calculating the cosine similarity of the joint emotion representation and the current emotion node representation in the triple knowledge graph; calculating the difference degree between the text modal and speech modal emotion feature vectors; weighting and combining the cosine similarity and the difference degree, and mapping through the Sigmoid function to obtain the consistency score.

4. The method of claim 3, wherein the method further comprises: The adversarial training mechanism in the step S4 comprises a generator and a discriminator; the generator is used to generate a robot reply based on the dialogue context vector; The discriminator is used to judge consistency of the robot reply with the dialogue task target according to user feedback data; The generator and the discriminator are trained by optimizing an adversarial loss function. 5.The method of claim 4, wherein, The dynamic updating triple knowledge graph comprises: when a new user intention or a new user entity is detected, a corresponding new node is inserted into the triple knowledge graph and a relationship edge is established. The method further comprises a dialogue termination step: when the generated behavior is a dialogue end flag, or the time without new input of the user exceeds a preset length of time, the graph nodes related to the current dialogue are emptied.

Images