Intelligent customer service dialogue generation optimization method and system based on knowledge graph

Abstract

The invention provides an intelligent customer service dialogue generation optimization method and system based on a knowledge graph, and relates to the technical field of knowledge graphs, and the method comprises the steps: employing a bidirectional long-short-term memory network to recognize the entity and intention of a user question; performing multi-hop query in the knowledge graph based on the question entity to obtain an initial knowledge sub-graph; constructing a user feature vector containing a dialogue state, problem solving and service preference; performing hierarchical reconstruction on the initial knowledge sub-graph based on the user feature vector; calculating the structural similarity of the domain knowledge graph, and extracting a general knowledge organization mode; decomposing the reconstructed knowledge sub-graph into a structure template and a content filling item, and determining a knowledge mapping relation between fields to obtain a migration optimization knowledge graph; and generating candidate replies based on the optimized knowledge graph and the initial intention, evaluating the accuracy, suitability and correlation of the candidate replies, and selecting the optimal reply output.

Description

Technical Field

[0001] The present invention relates to knowledge graph technology, and in particular to a method and system for optimizing the generation of intelligent customer service dialogues based on knowledge graphs. Background Art

[0002] Existing intelligent customer service systems primarily use retrieval-based or generative approaches for dialogue generation. Retrieval-based approaches rely on a pre-set Q&A database to match user questions and are unable to handle complex or novel questions. While generative approaches can produce fluent responses, they often suffer from inaccurate knowledge and fabricated information. Existing systems lack a mechanism for effectively integrating domain knowledge with dialogue strategies, have a limited understanding of user intent, and are unable to dynamically adjust the content and style of responses based on user characteristics. Furthermore, their ability to transfer knowledge across domains is insufficient, requiring extensive manual adjustments when the system is applied in new domains. Therefore, there is an urgent need for an intelligent customer service dialogue generation optimization method that can deeply integrate knowledge graphs with dialogue generation and possess cross-domain knowledge transfer capabilities. Summary of the Invention

[0003] In response to the shortcomings of the existing technology, the present invention provides an intelligent customer service dialogue generation optimization method and system based on knowledge graph, which can solve the problems in the existing technology.

[0004] In a first aspect, the present invention provides a method for optimizing intelligent customer service dialogue generation based on a knowledge graph, comprising: A bidirectional long short-term memory network is used to perform entity recognition and intent recognition on the question text input by the user to obtain the question entity and the initial intent; a multi-hop query is performed in the preset knowledge graph based on the question entity to obtain the initial knowledge subgraph; a user feature vector containing dialogue state features, problem-solving features and service preference features is constructed based on the question text input by the user; the initial knowledge subgraph is hierarchically reconstructed based on the user feature vector to obtain a reconstructed knowledge subgraph; structural similarity calculation and element recognition are performed on the existing domain knowledge graph to extract a general knowledge organization pattern; based on the general knowledge organization pattern, the reconstructed knowledge subgraph is decomposed into structural templates and content filling items, and the knowledge mapping relationship between domains is determined through semantic similarity matching to obtain a migration optimized knowledge graph; candidate replies are generated based on the migration optimized knowledge graph and the initial intent, the candidate replies are evaluated for knowledge accuracy, expression adaptability and domain relevance, and the candidate reply with the highest comprehensive score is selected as the output.

[0005] Optionally, performing a multi-hop query in a preset knowledge graph based on the problem entity, obtaining an initial knowledge subgraph includes: constructing an entity relationship importance scoring matrix based on the problem entity; generating an adaptive attenuation factor related to the query path hop count and path score according to the entity relationship importance scoring matrix; calculating the expected benefit of the query path starting from the problem entity based on the entity relationship importance scoring matrix and the adaptive attenuation factor, and generating a prioritized query path set; selecting multiple complementary paths with the highest scores from the prioritized query path set, allocating computing resources based on the path scores of the complementary paths for parallel query, and obtaining multiple candidate knowledge subgraphs and their feature matrices; calculating the complementarity and redundancy between the candidate knowledge subgraphs based on the feature matrix, weightedly fusing the complementary regions in the candidate knowledge subgraphs, and pruning redundant branches in the overlapping regions to obtain an initial knowledge subgraph.

[0006] Optionally, constructing a user feature vector containing dialogue state features, problem-solving features and service preference features based on the question text input by the user includes: performing multimodal feature extraction based on the user's question text, including: identifying professional terms in the question text to obtain a professionalism score, extracting the user message length sequence and response time sequence to obtain an interaction pattern feature vector, analyzing the frequency of emoticon use to obtain the emotional polarity strength; fusing the professionalism score, interaction pattern feature vector and emotional polarity strength through an attention mechanism to obtain a fused feature vector; constructing a user intention-topic transfer matrix based on the initial intention, and calculating adjacent topics based on the user intention-topic transfer matrix. Similarity, generate a conversation coherence index based on the similarity of the adjacent topics; calculate the complexity of the problem and the completeness of the answer to obtain a conversation depth index; construct a multi-round conversation state transition network based on the conversation coherence index and the conversation depth index; construct a business knowledge association map, analyze the user consultation scope based on the business knowledge association map, identify potential service needs, and calculate the problem-solving path; perform self-organizing map clustering on the user interaction vector sequence to obtain interaction behavior categories, and construct a service preference vector based on the interaction behavior category; combine the fusion feature vector, the multi-round conversation state transition network, the problem-solving path and the service preference vector to form a user feature vector.

[0007] Optionally, the initial knowledge subgraph is hierarchically reconstructed based on the user feature vector to obtain a reconstructed knowledge subgraph, including: hierarchically reconstructing the nodes of the initial knowledge subgraph according to the fused feature vector to obtain a preliminary hierarchical structure; constructing an inter-layer transition probability matrix using the multi-round dialogue state transition network, and dynamically adjusting the hierarchical structure based on the inter-layer transition probability matrix to obtain an optimized hierarchical structure; constructing a knowledge transfer graph based on the problem-solving path, mapping the optimized hierarchical structure to the knowledge transfer graph to obtain multiple possible transfer paths; reordering the transfer paths based on the service preference vector to obtain an optimal transfer path; constructing a multidimensional knowledge network based on the optimal transfer path, eliminating and supplementing knowledge redundancy, and obtaining a reconstructed knowledge subgraph.

[0008] Optionally, a multidimensional knowledge network is constructed based on the optimal transmission path, and knowledge redundancy is eliminated and supplemented to obtain a reconstructed knowledge subgraph, including: constructing a node semantic similarity matrix and a structural correlation matrix based on the optimal transmission path, and obtaining a hybrid correlation matrix through weighted fusion; using the hybrid correlation matrix to identify knowledge clusters, and calculating the importance weights of nodes within the cluster based on a hierarchical attention mechanism; constructing a multidimensional knowledge network in semantic, structural and functional dimensions according to the node importance weights and the hybrid correlation matrix; using the multidimensional knowledge network to perform cross-dimensional reasoning, identify and supplement knowledge gaps, and optimize the overall network structure through weighted graph embedding to obtain a reconstructed knowledge subgraph.

[0009] Optionally, structural similarity calculation and element identification are performed on existing domain knowledge graphs, and a general knowledge organization pattern is extracted, including: constructing a structural feature vector of the existing domain knowledge graph, including node degree distribution, path length distribution and local clustering coefficient; embedding nodes of the existing domain knowledge graph based on a graph neural network, and obtaining a graph-level representation through attention pooling; calculating the structural similarity matrix between knowledge graphs in different domains using the structural feature vector and graph-level representation; performing pattern matching on knowledge graph substructures whose similarity is higher than a preset similarity threshold, and identifying structural units and semantic elements whose co-occurrence frequency is higher than a preset frequency threshold; abstracting and summarizing the identified structural units and semantic elements to form a general knowledge organization pattern.

[0010] Optionally, the reconstructed knowledge subgraph is decomposed into structural templates and content filling items based on the general knowledge organization model, and the inter-domain knowledge mapping relationship is determined through semantic similarity matching to obtain a migration-optimized knowledge graph, including: constructing a structural matching matrix for the reconstructed knowledge subgraph based on the general knowledge organization model, the structural matching matrix fuses local structural similarity and semantic similarity, and decomposing the reconstructed knowledge subgraph into structural templates and corresponding content filling items through maximum bipartite matching; calculating the semantic similarity between the content filling items and the concepts of the target application domain using word vector similarity, knowledge base path similarity and context similarity, and generating an inter-domain knowledge mapping matrix based on the semantic similarity; constructing an initial knowledge structure in the target application domain based on the structural template and the inter-domain knowledge mapping matrix, identifying the concept types and relationship patterns of the target application domain, and fusing and supplementing them with the initial knowledge structure to form a domain-enhanced initial knowledge structure; jointly optimizing the structural preservation, semantic consistency and constraint satisfaction of the domain-enhanced initial knowledge structure to obtain a migration-optimized knowledge graph.

[0011] Optionally, candidate replies are generated based on the migration-optimized knowledge graph and the initial intent, the candidate replies are evaluated for knowledge accuracy, expression adaptability and domain relevance, and the candidate reply with the highest comprehensive score is selected as the output, including: extracting knowledge path representation, entity relationship representation and attribute feature representation from the migration-optimized knowledge graph, performing hierarchical attention fusion with the semantic vector of the initial intent, and constructing a multi-level context vector with knowledge perception capability; adjusting the knowledge-enhanced dialogue generation model based on the multi-level context vector to generate multiple candidate replies; performing multi-dimensional evaluation on the candidate replies: calculating the knowledge accuracy score using knowledge graph path verification and entity relationship constraint verification; calculating the language style matching and professional matching of the candidate reply based on the user feature vector to obtain the expression adaptability score; calculating the domain coverage and professional terminology accuracy of the candidate reply in combination with the structural characteristics of the migration-optimized knowledge graph and the domain ontology to obtain the domain relevance score; weighted fusion of the knowledge accuracy score, expression adaptability score and domain relevance score to obtain a comprehensive score, and selecting the candidate reply with the highest comprehensive score as the output.

[0012] In a second aspect, the present invention provides an intelligent customer service dialogue generation and optimization system based on a knowledge graph, comprising: The first unit is used to use a bidirectional long short-term memory network to perform entity recognition and intent recognition on the question text input by the user to obtain the question entity and initial intent; based on the question entity, a multi-hop query is performed in a preset knowledge graph to obtain an initial knowledge subgraph; The second unit is configured to construct a user feature vector including dialogue state features, problem-solving features, and service preference features based on the question text input by the user; hierarchically reconstruct the initial knowledge subgraph based on the user feature vector to obtain a reconstructed knowledge subgraph; The third unit is used to calculate the structural similarity and identify the elements of the existing domain knowledge graph to extract the general knowledge organization model; based on the general knowledge organization model, the reconstructed knowledge subgraph is decomposed into a structural template and content filling items, and the knowledge mapping relationship between domains is determined by semantic similarity matching to obtain a migration-optimized knowledge graph; The fourth unit is used to generate candidate responses based on the migration-optimized knowledge graph and the initial intent, evaluate the candidate responses in terms of knowledge accuracy, expression adaptability, and domain relevance, and select the candidate response with the highest comprehensive score as output.

[0013] The user feature vector construction mechanism provided by the present invention integrates dialogue status, problem solving and service preference features into the knowledge subgraph reconstruction process, enabling the system to dynamically adjust the reply strategy according to different user characteristics, significantly improving user satisfaction.

[0014] The general knowledge organization model extraction method proposed in this paper achieves the separation of knowledge structure templates from content filling items, significantly improving the efficiency of knowledge transfer and shortening the adaptation time to new domains. It also comprehensively evaluates candidate responses through a multi-dimensional evaluation mechanism to ensure the knowledge accuracy, expression adaptability, and domain relevance of the responses, effectively improving the accuracy rate of responses and the rate of user problem resolution. The overall solution achieves a deep integration of knowledge graphs, user characteristics, and dialogue generation, possessing cross-domain knowledge transfer capabilities, significantly reducing the labor and time costs of deploying intelligent customer service in new domains, while improving service quality and generating good technical and economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flow chart of a method for optimizing intelligent customer service dialogue generation based on a knowledge graph according to an embodiment of the present invention. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the present invention will be described below in conjunction with the drawings in the embodiments of the present invention. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0017] Figure 1 This is a flow chart of the optimization method for intelligent customer service dialogue generation based on knowledge graph of the present invention. Figure 1 As shown, the method includes: A bidirectional long short-term memory network is used to perform entity recognition and intent recognition on the question text input by the user to obtain the question entity and the initial intent; a multi-hop query is performed in the preset knowledge graph based on the question entity to obtain the initial knowledge subgraph; a user feature vector containing dialogue state features, problem-solving features and service preference features is constructed based on the question text input by the user; the initial knowledge subgraph is hierarchically reconstructed based on the user feature vector to obtain a reconstructed knowledge subgraph; structural similarity calculation and element recognition are performed on the existing domain knowledge graph to extract a general knowledge organization pattern; based on the general knowledge organization pattern, the reconstructed knowledge subgraph is decomposed into structural templates and content filling items, and the knowledge mapping relationship between domains is determined through semantic similarity matching to obtain a migration optimized knowledge graph; candidate replies are generated based on the migration optimized knowledge graph and the initial intent, the candidate replies are evaluated for knowledge accuracy, expression adaptability and domain relevance, and the candidate reply with the highest comprehensive score is selected as the output.

[0018] Optionally, performing a multi-hop query in a preset knowledge graph based on the problem entity, obtaining an initial knowledge subgraph includes: constructing an entity relationship importance scoring matrix based on the problem entity; generating an adaptive attenuation factor related to the query path hop count and path score according to the entity relationship importance scoring matrix; calculating the expected benefit of the query path starting from the problem entity based on the entity relationship importance scoring matrix and the adaptive attenuation factor, and generating a prioritized query path set; selecting multiple complementary paths with the highest scores from the prioritized query path set, allocating computing resources based on the path scores of the complementary paths for parallel query, and obtaining multiple candidate knowledge subgraphs and their feature matrices; calculating the complementarity and redundancy between the candidate knowledge subgraphs based on the feature matrix, weightedly fusing the complementary regions in the candidate knowledge subgraphs, and pruning redundant branches in the overlapping regions to obtain an initial knowledge subgraph.

[0019] For example, in the entity recognition process, the user's question text is used as the input sequence, and the text is converted into a low-dimensional dense vector representation through the word vector embedding layer, and then input into the bidirectional long short-term memory network. The network contains two long short-term memory units, forward and backward, which process information from the left and right sides of the sentence respectively to capture contextual dependencies. The network output layer uses conditional random fields to optimize the sequence labeling results and finally outputs the entity type label of each word. For intent recognition, a classifier based on the attention mechanism is used, and the hidden layer state of the bidirectional long short-term memory network is used as input. The key information in the question text is captured through the multi-head self-attention mechanism. The probability distribution of each intent category is then calculated through the fully connected layer and the softmax function, and the category with the highest probability is selected as the initial intent.

[0020] Based on the identified problem entity, a multi-hop query is performed within a pre-built knowledge graph to obtain an initial knowledge subgraph. The pre-built knowledge graph is a pre-built structured graph database containing domain entities, relationships, attributes, and other knowledge, used to support knowledge retrieval and reasoning for intelligent customer service. For each relationship type directly connected to the problem entity in the knowledge graph, three metrics are calculated: its frequency in historical queries, user feedback score, and a priori weight defined by domain experts. For the question "How to deal with bank card fraud?" as an example, the identified problem entities are "bank card" and "fraudulent use," and related relationships include "preventative measures," "handling process," and "legal liability." The "handling process" relationship has a high historical query frequency (0.8), an excellent user feedback score (0.9), and a high weight defined by domain experts (0.9). The calculated importance score for this relationship is 0.87. All relationships are then combined into an entity relationship importance score matrix based on entity pairs. Each element in the matrix represents the importance score of a specific relationship type between a specific entity pair.

[0021] The decay factor controls the degree to which the importance of relationships decreases as the number of hops increases during a multi-hop query. The decay rate is dynamically adjusted based on the cumulative score of the current query path, the current number of hops, and the complexity of the query domain. For simple domains (such as FAQs), a larger decay factor (such as 0.7) can be used, while for complex domains (such as professional technical consultations), a smaller decay factor (such as 0.4) can be used to retain more in-depth knowledge.

[0022] Based on the entity relationship importance scoring matrix and an adaptive decay factor, the expected return of query paths starting from the question entity is calculated, generating a prioritized set of query paths. The calculation begins with the question entity and expands to connected nodes in the knowledge graph, forming a path tree. For each path in the path tree, the weighted sum of the importance of each hop relationship is accumulated, with the weight being the decay factor corresponding to the number of hops. All paths are sorted according to expected return, and the top N paths are selected as the prioritized set of query paths.

[0023] In the complementary path selection process, in addition to considering path scores, the complementarity between paths is calculated to avoid selecting paths with overlapping information. The entity and relationship overlap ratio between paths is calculated, and paths with an overlap ratio below a threshold are considered complementary. For example, in a bank card fraud case, three highly complementary paths might be selected: "Bank card - Type of fraud - Emergency response measures," "Bank card - Lost and reported process - Customer service contact information," and "Fraud - Legal liability - Bank obligations." Computing resources are allocated to these selected complementary paths based on their score ratios, with higher-scoring paths receiving more resources for parallel querying. This results in multiple candidate knowledge subgraphs. For each subgraph, structural features (such as node degree distribution and path length distribution), content features (such as entity type distribution and relationship type distribution), and semantic features (graph embedding vectors) are extracted to form a feature matrix.

[0024] The complementarity between candidate knowledge subgraphs is determined by comparing the orthogonal components of the subgraph feature matrices; larger orthogonal components indicate stronger complementarity. Redundancy is determined by comparing the overlap of entity relationships between subgraphs; higher overlap indicates greater redundancy. For example, in the case of bank card fraud, the complementarity between the "Emergency Measures" subgraph and the "Loss Reporting Process" subgraph is 0.72 (high complementarity), and the redundancy with the "Customer Service Contact Information" subgraph is 0.08 (low redundancy). Complementary regions within the candidate knowledge subgraphs are weighted and fused, with the weighting coefficient proportional to the region's information richness and relevance to the query intent. For overlapping regions, redundant branches are pruned, retaining branches with high information density and connectivity while removing redundant branches. The resulting initial knowledge subgraph contains both core knowledge directly related to the problem and supplementary information from multiple perspectives, while avoiding information redundancy.

[0025] The present invention guides multi-hop queries through an entity relationship importance scoring matrix and an adaptive attenuation factor, making the knowledge acquisition process more accurate and efficient; the priority mechanism based on path expected benefits and the complementary path selection strategy significantly improve the knowledge coverage and diversity; through complementarity fusion and redundancy pruning processing, the structure of the knowledge subgraph is optimized, information redundancy is reduced, and the quality and efficiency of subsequent dialogue generation are improved.

[0026] Optionally, constructing a user feature vector containing dialogue state features, problem-solving features and service preference features based on the question text input by the user includes: performing multimodal feature extraction based on the user's question text, including: identifying professional terms in the question text to obtain a professionalism score, extracting the user message length sequence and response time sequence to obtain an interaction pattern feature vector, analyzing the frequency of emoticon use to obtain the emotional polarity strength; fusing the professionalism score, interaction pattern feature vector and emotional polarity strength through an attention mechanism to obtain a fused feature vector; constructing a user intention-topic transfer matrix based on the initial intention, and calculating adjacent topics based on the user intention-topic transfer matrix. Similarity, generate a conversation coherence index based on the similarity of the adjacent topics; calculate the complexity of the problem and the completeness of the answer to obtain a conversation depth index; construct a multi-round conversation state transition network based on the conversation coherence index and the conversation depth index; construct a business knowledge association map, analyze the user consultation scope based on the business knowledge association map, identify potential service needs, and calculate the problem-solving path; perform self-organizing map clustering on the user interaction vector sequence to obtain interaction behavior categories, and construct a service preference vector based on the interaction behavior category; combine the fusion feature vector, the multi-round conversation state transition network, the problem-solving path and the service preference vector to form a user feature vector.

[0027] For example, this embodiment describes in detail how to construct a user feature vector based on the user's input question text, including a fusion feature vector, conversation state features, problem-solving features, and service preference features. First, multimodal feature extraction is performed based on the user's question text, including technical terminology recognition, interaction mode feature extraction, and sentiment polarity analysis.

[0028] During the terminology recognition phase, a terminology database containing industry terminology, specialized vocabulary, and technical terms is maintained. This database is categorized by industry and includes specialized vocabulary from major fields such as finance, technology, and healthcare. When a user enters a question, a hybrid approach based on vocabulary matching and contextual semantic analysis is used to identify specialized terms within the text. Specifically, this approach first identifies explicit specialized terms through vocabulary matching, then uses a pre-trained terminology recognition model to capture contextually relevant specialized expressions. For example, in the question "My structured deposit with an annualized yield of 4.5% has matured. How do I automatically roll it over?", three financial terms are identified: "annualized yield," "structured deposit," and "automatic rollover." A professionalism score is calculated based on the number of identified terms, their proportion in the question, and the complexity of the terms. For this example, the calculated professionalism score is 0.78 (high professionalism).

[0029] To extract interaction pattern features, we analyze the message length and response time sequences in a user's historical conversation logs. The message length sequence is the length of each message sent across multiple consecutive conversations; the response time sequence is the time interval between a user sending their next message after receiving a reply. For new users, the average value of a similar user group is used as the initial value. For example, in a financial services consulting scenario, the message length sequence for a user's last five conversations is [45, 12, 78, 25, 36] characters, and the response time sequence is [15, 60, 25, 40] seconds. Statistical features (mean, variance, trend) and pattern features (volatility, persistence) are calculated for these sequences to form an interaction pattern feature vector. This vector reflects the user's expression habits and patience. For this example user, we judged them to be "detailed and moderately patient."

[0030] Sentiment polarity analysis detects emojis, sentiment words, and sentence structure features in user question text. Emoji frequency is calculated by calculating the ratio of emoji appearances to text length. A sentiment lexicon is also used to assess the sentiment of the text. For example, the question "This financial product is terrible (angry emoji), I want to complain!" contains the negative sentiment word "bad" and an angry emoji, resulting in a sentiment polarity strength of -0.85 (strongly negative). The professionalism score, interaction pattern feature vector, and sentiment polarity strength are fused using an attention mechanism to generate a fused feature vector. The attention mechanism dynamically adjusts the weight of each feature based on the current conversation context and service type. For example, in a complaint handling scenario, sentiment polarity is given a higher weight.

[0031] To construct a user intent-topic transition matrix based on initial intent, we maintain a predefined set of intent categories, such as "inquiry information," "problem resolution," and "express dissatisfaction," along with corresponding topic domains, such as "product information," "operational procedures," and "account security." We construct the user intent-topic transition matrix by analyzing the transition patterns between intent and topic in historical user conversations. Each element in the matrix represents the probability of transitioning from one intent-topic combination to another. For example, in a banking customer service situation, the transition probability from "inquiry information-product information" to "problem resolution-operational procedures" is 0.65, indicating that after inquiring about product information, users are highly likely to inquire about specific operational procedures.

[0032] The similarity of adjacent topics is calculated based on the user intent-topic transition matrix, employing a combination of a semantic similarity algorithm based on word vectors and a structural similarity algorithm based on the knowledge graph. Semantic similarity between adjacent topics is calculated by calculating the cosine similarity of the word vectors of the topic keywords; structural similarity is calculated by calculating the path overlap between topics in the knowledge graph. A weighted average of these two similarities is used to obtain a comprehensive similarity. For example, the semantic similarity between "credit card application" and "credit card activation" is 0.75, the structural similarity is 0.82, and the comprehensive similarity is 0.78 (high similarity). A conversation coherence index is generated based on adjacent topic similarity, reflecting the naturalness and logicality of topic transitions within a user conversation. The coherence index calculation considers the average value, volatility, and frequency of unusual topic transitions in adjacent topic similarity.

[0033] Question complexity is assessed based on three dimensions: the complexity of the question's syntactic structure, the number of knowledge points covered, and the length of the solution path. For example, the question "How do I choose a suitable portfolio of financial products based on my risk appetite and investment horizon?" has a complexity of 0.82 (high complexity). Answer completeness is calculated by assessing the proportion of knowledge points covered in the response and the user's confirmation of completion. A multi-turn conversation state transition network is constructed based on the conversation coherence metric and the conversation depth index. This network is a directed graph, with nodes representing conversation states (such as "initial consultation," "in-depth discussion," and "solution confirmed") and edges representing state transition probabilities. The network structure reflects the user's conversation progression pattern and preference for in-depth discussion.

[0034] Construct a business knowledge association graph, which organizes the core concepts, processes, and relationships within a business domain in a graph structure. For example, in the financial services sector, the graph includes key nodes such as "product type," "service process," and "risk level," as well as relationships between them such as "inclusion," "sequence," and "applicable conditions." Analyze user inquiry scopes based on the business knowledge association graph, mapping users' past inquiries to nodes and paths within the graph and identifying frequently accessed subgraph structures. For example, a user's past inquiries primarily focused on the "investment product - risk assessment - return calculation" subgraph area, indicating a focus on risk and return.

[0035] Based on the analysis of the scope of consultation, potential service needs are further identified. By comparing user consultation paths with typical service demand patterns, highly compatible service demand types are identified. For example, users who frequently inquire about "account security" and "abnormal transactions" are identified as potentially requiring "security services." Based on the identified potential service needs, the problem-solving path is calculated by combining the service process nodes and relationship edges in the business knowledge association graph. Taking the complexity of the problem, the user's expertise level, and historical solution patterns into consideration, the optimal solution path is planned within the business knowledge association graph. Taking "How to apply for a credit card installment" as an example, for highly specialized users, the path is "installment conditions → application channels → fee calculation"; for less specialized users, the path is "basic explanation → installment conditions → application steps → precautions."

[0036] The user interaction vector sequence contains multidimensional characteristics of each user interaction, such as the questioning method (direct / indirect), the completeness of information provided, and the frequency of confirmation requests. A self-organizing map algorithm is used to cluster historical user interaction vectors, forming interaction behavior categories such as "efficient and concise," "detailed confirmation," and "exploratory learning." New users' interaction vectors are then matched against these categories for similarity to determine their category. A service preference vector is constructed based on interaction behavior categories, encompassing four dimensions: response speed, information detail, jargon usage, and interaction frequency. For example, the service preference vector for an "efficient and concise" user is [0.9, 0.3, 0.7, 0.2], indicating a high emphasis on response speed, a preference for concise information, a tolerance for a certain level of jargon, and a low need for frequent interaction.

[0037] Finally, the fused feature vector, the multi-turn dialogue state transition network, the problem-solving path, and the service preference vector are combined to form a user feature vector. This combination utilizes a hierarchical encoding structure, effectively integrating the various components through feature embedding and a fusion network. First, the fused feature vector undergoes dimension normalization to ensure that all feature components are within the same numerical range. Key topological features of the multi-turn dialogue state transition network are extracted, including node connectivity distribution, centrality metrics, and the principal eigenvalues ​​of the transition probability matrix, to form a fixed-dimensional network representation vector. The problem-solving path is converted into a sequence representation using a path encoder, capturing information about node types, relationship types, and their order within the path. The service preference vector is then normalized with the aforementioned features and directly used as a subcomponent of the user feature vector. A multilayer perceptron then performs nonlinear mapping on these preprocessed features to generate an intermediate representation. During this mapping process, a feature selection gating mechanism is implemented to dynamically adjust the importance of each feature component based on the current dialogue stage and question type. Finally, the intermediate representation is fused with the original features using residual connections and layer normalization to form the final user feature vector.

[0038] This technical solution extracts and integrates multi-dimensional features to construct user feature vectors rich in semantic and behavioral information. This accurately captures users' expertise, interaction habits, emotional state, and service needs, enabling intelligent customer service to deeply understand user characteristics and conduct personalized knowledge reconstruction and dialogue generation. Compared to traditional methods, this solution significantly improves user experience and problem-solving efficiency, making intelligent customer service more intelligent and user-friendly.

[0039] Optionally, the initial knowledge subgraph is hierarchically reconstructed based on the user feature vector to obtain a reconstructed knowledge subgraph, including: hierarchically reconstructing the nodes of the initial knowledge subgraph according to the fused feature vector to obtain a preliminary hierarchical structure; constructing an inter-layer transition probability matrix using the multi-round dialogue state transition network, and dynamically adjusting the hierarchical structure based on the inter-layer transition probability matrix to obtain an optimized hierarchical structure; constructing a knowledge transfer graph based on the problem-solving path, mapping the optimized hierarchical structure to the knowledge transfer graph to obtain multiple possible transfer paths; reordering the transfer paths based on the service preference vector to obtain an optimal transfer path; constructing a multidimensional knowledge network based on the optimal transfer path, eliminating and supplementing knowledge redundancy, and obtaining a reconstructed knowledge subgraph.

[0040] For example, the user's fused feature vector is analyzed for its expertise, cognitive style, and affective state components. The expertise component is used to determine the user's ability to accept professional knowledge. For example, a user in the financial industry with an expertise component value of 0.85 is able to understand relatively complex financial terms and concepts. The cognitive style component is used to determine the user's preferred information organization method, such as linear or divergent thinking. The affective state component is used to adjust the way knowledge is presented, such as simplifying the knowledge structure when in a negative mood.

[0041] Based on the above analysis, the nodes of the initial knowledge subgraph are divided into different hierarchies. Taking the investment and financial consulting scenario as an example, the initial knowledge subgraph includes multiple knowledge nodes, such as "Fund Type," "Risk Rating," "Return Calculation," "Tax Policy," and "Subscription Process." For more specialized users, a three-tier structure is constructed: the top layer is the basic concepts layer, including "Fund Type" and "Risk Rating"; the middle layer is the deep knowledge layer, including "Return Calculation" and "Tax Policy"; and the bottom layer is the operational guidance layer, including "Subscription Process." For less specialized users, four or more layers may be constructed, adding an explanation layer between basic concepts and deep knowledge to refine the connections between knowledge points. Nodes in each layer are clustered based on semantic similarity and structural connectivity between nodes to form knowledge clusters. This process results in a preliminary hierarchical structure, which is a hierarchical directed acyclic graph.

[0042] Analyze the user's multi-turn dialogue state transition network to extract state transition patterns during the conversation. For example, a user's dialogue state transition network shows a 0.7 probability of transitioning from "basic consultation" to "detailed exploration" and a 0.8 probability of transitioning from "detailed exploration" to "practical verification," indicating that the user prefers to first understand the concept and then delve into the details. Based on these state transition patterns, a transition probability matrix between knowledge levels is constructed. Taking the aforementioned three-layer structure as an example, the transition probability from the basic concept layer to the deep knowledge layer is 0.75, the transition probability from the deep knowledge layer to the operational guidance layer is 0.65, and the transition probability from the basic concept layer directly to the operational guidance layer is 0.15.

[0043] Based on the inter-layer transition probability matrix, the hierarchical relationships of the preliminary hierarchical structure are dynamically adjusted. This adjustment process considers three key factors: transition probability, node correlation within a layer, and inter-layer connection density. When the transition probability between two layers is above a threshold (e.g., 0.7), the connection strength between the two layers is strengthened. When the correlation between nodes within a layer is below a threshold (e.g., 0.4), the layer is considered for splitting. When the inter-layer connection density is too high, causing the structural complexity to exceed the user's cognitive load, some layers are merged or simplified. Through these adjustments, an optimized hierarchical structure is achieved. In the investment and financial consulting example, the "Risk Rating" and "Return Calculation" nodes, which were originally located on different layers, might be relocated to the same layer because user conversation patterns indicate that these two types of information are often queried and discussed together.

[0044] A knowledge transfer graph is constructed based on the problem-solving path. Key nodes and transition steps along the path are extracted. These nodes typically represent the knowledge points required to solve the problem, while the transition steps reflect the causal or dependency relationships between these knowledge points. For example, the problem-solving path for "how to choose a suitable fund product" consists of five key steps: risk tolerance assessment → fund type screening → performance comparison → fee analysis → purchase decision. A knowledge transfer graph is constructed based on these steps. Nodes represent related knowledge points, and edges represent the transfer relationships between knowledge points. During the construction process, three types of transfer relationships are considered: premise relationships (understanding A is a prerequisite for understanding B), derivation relationships (from A, B can be derived), and complement relationships (A and B complement each other).

[0045] The optimized hierarchical structure is mapped to a knowledge transfer graph, generating multiple possible transfer paths. This mapping process is achieved by calculating semantic similarity and functional equivalence between nodes in the hierarchical structure and those in the transfer graph. In the investment and wealth management example, the "Risk Rating" node in the hierarchical structure might be mapped to the "Risk Tolerance Assessment" node in the transfer graph. This mapping yields multiple possible knowledge transfer paths, such as "Fund Type → Risk Rating → Return Calculation → Subscription Process" and "Risk Rating → Fund Type → Tax Policy → Subscription Process." These transfer paths are re-ranked based on the service preference vector to determine the optimal transfer path. The service preference vector includes components such as response speed preference and information detail preference, and the degree of fit between each path and these preferences is calculated. For example, users who prioritize efficiency (with a response speed preference component of 0.9) prioritize concise paths with fewer nodes; users who prioritize detailed information (with an information detail preference component of 0.8) prioritize paths that cover more comprehensive knowledge points. The optimal transfer path is selected based on a comprehensive evaluation.

[0046] This technical solution achieves personalized optimization of knowledge subgraphs through hierarchical reconstruction based on user characteristics, closely matching the knowledge organization structure with user cognitive characteristics and conversational preferences. Compared to traditional static knowledge organization methods, this solution dynamically adjusts knowledge presentation based on user expertise, conversation status, and service preferences, significantly improving the relevance and effectiveness of knowledge transfer.

[0047] Optionally, a multidimensional knowledge network is constructed based on the optimal transmission path, and knowledge redundancy is eliminated and supplemented to obtain a reconstructed knowledge subgraph, including: constructing a node semantic similarity matrix and a structural correlation matrix based on the optimal transmission path, and obtaining a hybrid correlation matrix through weighted fusion; using the hybrid correlation matrix to identify knowledge clusters, and calculating the importance weights of nodes within the cluster based on a hierarchical attention mechanism; constructing a multidimensional knowledge network in semantic, structural and functional dimensions according to the node importance weights and the hybrid correlation matrix; using the multidimensional knowledge network to perform cross-dimensional reasoning, identify and supplement knowledge gaps, and optimize the overall network structure through weighted graph embedding to obtain a reconstructed knowledge subgraph.

[0048] For example, all knowledge nodes in the optimal transfer path are analyzed, and the semantic similarity and structural relevance between pairs of nodes are calculated. Semantic similarity is calculated by comparing the semantic representation vectors of the node content. A deep learning pre-trained model is used to extract semantic features of the node text. For example, in the investment product consultation scenario, the cosine similarity of the semantic vectors of the "fixed deposit" node and the "structured deposit" node is 0.75, indicating a high degree of semantic relevance between the two.

[0049] Structural relevance is based on an analysis of node connectivity patterns within the knowledge graph, considering three dimensions: direct connections between nodes, the number of shared neighbors, and path distance. For example, while the semantic similarity between the "Fund Redemption" and "Fund Subscription" nodes is 0.62, their structural relevance is calculated to be 0.85 because they are frequently connected to the "Fund Trading" node in the knowledge graph and share similar connectivity patterns. By setting a semantic similarity weight α and a structural relevance weight β (α + β = 1), the two matrices are weighted and fused into a hybrid relevance matrix. For specialized domain knowledge, α = 0.4 and β = 0.6 can be set to emphasize structural relationships; for general domain knowledge, α = 0.7 and β = 0.3 can be set to emphasize semantic associations.

[0050] A knowledge cluster is a collection of highly semantically and structurally related nodes in a knowledge graph that collectively represent a complete knowledge topic or concept. By applying a community discovery algorithm to a mixed correlation matrix, nodes with correlations above a preset threshold (e.g., 0.65) are clustered into knowledge clusters. For example, in the field of investment and financial management, a "risk assessment" knowledge cluster might be identified, containing nodes related to "risk tolerance," "risk preference testing," and "risk grading," while a "return analysis" knowledge cluster might include nodes such as "expected rate of return," "historical performance comparison," and "return calculation method." For each identified knowledge cluster, a hierarchical attention mechanism is applied to calculate the importance weights of nodes within the cluster. This mechanism first calculates the semantic importance of each node at the node level and then, at the knowledge cluster level, calculates the node's contribution to the overall cluster topic. Node-level attention is based on the information entropy of the node content and the term frequency-inverse document frequency value; cluster-level attention is based on the node's position in the optimal transmission path and its connectivity centrality. For example, within the "Risk Assessment" knowledge cluster, the "Risk Tolerance" node, located at a critical position in the transmission path and connecting multiple other nodes, has an importance weight of 0.82; while the "Risk Preference Test," as an implementation method, has an importance weight of 0.65. The hierarchical attention mechanism not only identifies knowledge clusters—a collection of semantically and structurally highly related nodes—but also accurately calculates the importance weights of each node within the cluster, laying the foundation for subsequent knowledge network construction.

[0051] Multidimensional knowledge networks contain not only binary relationships between nodes but also high-order connections across dimensions. In the semantic dimension, connections are established based on semantic similarity between nodes. For example, "fixed deposit" and "structured deposit" are semantically similar, thus establishing a semantic connection. In the structural dimension, connections are established based on the original structural relationships within the knowledge graph. For example, "fund subscription" and "fund redemption" are sub-operations of "fund trading," thus establishing a structural connection. In the functional dimension, connections are established based on the functional roles of nodes in the problem-solving process. For example, "risk assessment" and "return expectation" belong to the evaluation and expectation functions in investment decision-making, thus establishing a functional connection. During the construction process, node centrality and connection priority are determined based on their importance weights. Nodes with high importance (e.g., weight ≥ 0.7) are given greater visibility and more cross-dimensional connections. The values ​​in the mixed correlation matrix are also used to determine connection strength. Node pairs with correlations above 0.8 are strongly connected, those with correlations between 0.5 and 0.8 are moderately connected, and those with correlations between 0.3 and 0.5 are weakly connected. Taking investment product consulting as an example, a multidimensional knowledge network consisting of 30 nodes across four main categories: "Product Type," "Risk Level," "Return Calculation," and "Operational Process." In this network, the "Risk Level" node, due to its high importance weight (0.85), establishes connections with nodes from multiple dimensions; the "Return Calculation Formula" node, on the other hand, primarily connects to nodes from the functional dimension.

[0052] Cross-dimensional reasoning involves analyzing node relationship patterns across different dimensions to infer potential implicit relationships or missing knowledge. Cross-dimensional reasoning is performed using three reasoning modes: transitive reasoning (if A is related to B, and B is related to C, then A is likely related to C), symmetric reasoning (if A and B have a specific relationship, then B and A are likely to have a symmetric relationship), and compositional reasoning (inferring new relationships based on multiple known relationships).

[0053] During the reasoning process, knowledge gaps are identified, including structural gaps (connections that should exist in the network but are missing) and content gaps (information that should exist at a node but is missing). For example, in the investment and wealth management network, it was discovered that the "Risk Assessment" and "Investment Advice" nodes were missing an intermediate connection node. Functionally, these two nodes should be connected through the concept of "risk-return matching." Relevant content is retrieved from the complete knowledge base to supplement the "risk-return matching" node and its connections. For content gaps, such as the "Structured Deposit" node lacking information on the "Impact of Early Redemption," relevant content is also supplemented through retrieval.

[0054] After knowledge enrichment is completed, the overall network structure is optimized using weighted graph embedding technology. Weighted graph embedding considers node and edge weights to ensure that key network structures and semantic relationships are preserved during dimensionality reduction. A fixed-dimensional (typically 128 or 256-dimensional) vector representation is generated for each node in the knowledge network. The relative positions of these vectors in space reflect the semantic and structural relationships between nodes. During weighted graph embedding, a feature representation is first constructed for each node, containing the semantic vector of the node's content and structural position information. The embedding process is then differentiated based on the previously calculated node importance weights, resulting in more accurate vector representations for important nodes. This is achieved by adjusting the optimization intensity for different nodes during the learning process. Edge weights are also set based on the degree of association between node pairs in the hybrid correlation matrix. Edges between highly correlated node pairs receive higher weights, while weakly correlated edges may be pruned. Within the embedding space, the node distribution is further optimized, and structural constraints are imposed to ensure that the hierarchical relationships and semantic associations of the knowledge network are preserved. For example, the subordination between higher-level and lower-level concepts is expressed through specific directional relationships between their embedding vectors.

[0055] Through the optimized network structure, redundant nodes are identified, including those with highly overlapping information and those containing technical details that are unnecessary for user understanding. Next, redundant connections are identified—connections that can be effectively reached via other paths and do not provide additional information. This optimization process significantly reduces network complexity while preserving knowledge integrity, ultimately resulting in a streamlined and efficient reconstructed knowledge subgraph.

[0056] This technical solution utilizes a weighted fusion of a node semantic similarity matrix and a structural relevance matrix to more accurately capture the complex relationships between knowledge elements. A hierarchical attention mechanism and knowledge cluster identification technology significantly enhance the rationality of the network structure and the efficiency of knowledge transfer. Weighted graph embedding optimization further improves the clarity of knowledge organization and access efficiency. The overall solution enables personalized reconstruction of knowledge subgraphs, making knowledge presentation more aligned with user cognitive characteristics and problem-solving needs, effectively improving the response quality and user satisfaction of intelligent customer service.

[0057] Optionally, structural similarity calculation and element identification are performed on existing domain knowledge graphs, and a general knowledge organization pattern is extracted, including: constructing a structural feature vector of the existing domain knowledge graph, including node degree distribution, path length distribution and local clustering coefficient; embedding nodes of the existing domain knowledge graph based on a graph neural network, and obtaining a graph-level representation through attention pooling; calculating the structural similarity matrix between knowledge graphs in different domains using the structural feature vector and graph-level representation; performing pattern matching on knowledge graph substructures whose similarity is higher than a preset similarity threshold, and identifying structural units and semantic elements whose co-occurrence frequency is higher than a preset frequency threshold; abstracting and summarizing the identified structural units and semantic elements to form a general knowledge organization pattern.

[0058] For example, a structural feature vector of an existing domain knowledge graph is constructed, including node degree distribution, path length distribution, and local clustering coefficient. For each node in the knowledge graph, its degree (i.e., the number of edges directly connected to it) is calculated, and then the node degree distribution characteristics of the entire graph are statistically analyzed. For example, a financial domain knowledge graph contains 500 entity nodes and 1,200 relationship edges. Its node degree distribution shows that 25% of the nodes have degrees 1-2 (lowly connected nodes), 60% have degrees 3-8 (medium-connected nodes), and 15% have degrees greater than 8 (highly connected nodes). This distribution feature is encoded as a fixed-dimensional vector, such as a 20-dimensional vector, where each element represents the percentage of nodes within a specific degree range. Path length distribution characteristics are obtained by calculating the shortest path length between any two nodes in the graph. 1,000 pairs of nodes are randomly sampled from the graph, the shortest path lengths between them are calculated, and the distribution of different lengths is statistically analyzed. In the financial domain graph, sampling results show that 15% of node pairs have a distance of 1, 35% have a distance of 2, 30% have a distance of 3, 15% have a distance of 4, and 5% have a distance greater than 4. This distribution is also encoded as a fixed-dimensional vector. The local clustering coefficient reflects the closeness of the node neighborhood in the graph. The local clustering coefficient is calculated for each node and its distribution characteristics are statistically analyzed. In the financial graph, the average local clustering coefficient is 0.42, indicating a moderate degree of local clustering. The three feature vectors of node degree distribution, path length distribution, and local clustering coefficient distribution are concatenated to form a complete structural feature vector.

[0059] The node embedding process uses a graph convolutional network (GCN), which consists of multiple convolutional layers, each of which aggregates and updates node features. Initial node features include a one-hot encoding of the node type and a node attribute vector. For example, in a medical knowledge graph, the initial features of a "disease" node include the node type encoding and disease attributes; "symptom" nodes include attributes such as symptom severity and frequency. The first layer of the GCN aggregates features of each node's immediate neighbors, the second layer aggregates features of second-order neighbors, and so on. After three layers of graph convolution, each node receives a 128-dimensional embedding vector that encodes the node's local structural information and semantic features.

[0060] Attention pooling is a key step in aggregating node-level features into a graph-level representation. First, an importance score is calculated for each node. Importance is calculated based on the node's centrality metrics and domain relevance. Centrality metrics consider a node's degree centrality, closeness centrality, and betweenness centrality; domain relevance is based on the node's association with core concepts in the domain. In the medical graph, the "heart disease" node receives an importance score of 0.85 (high importance) because it connects to multiple symptom and treatment nodes and is a core disease concept. Attention weights are assigned based on the node's importance scores, and the embedding vectors of all nodes are aggregated into a single graph-level representation vector through weighted summation. The graph-level representation of the medical graph is a 256-dimensional vector that encodes the overall structural characteristics of the graph and the distribution of domain knowledge.

[0061] A two-stage similarity calculation method is used: first, pure structural similarity is calculated based on the structural feature vectors; then, comprehensive similarity is calculated by combining graph-level representations. Pure structural similarity is obtained by calculating the cosine similarity of the structural feature vectors of two graphs, reflecting the degree of similarity in the graphs' topological structures. Comprehensive similarity, by considering both structural features and graph-level representations, captures deeper similarities. Similarity calculations were performed on knowledge graphs in four fields: finance, healthcare, education, and e-commerce, resulting in a 4×4 similarity matrix. The results show that the structural similarity between the finance and education graphs is 0.72, indicating a high degree of similarity in knowledge organization between the two fields. In contrast, the similarity between finance and healthcare is only 0.45, indicating significant structural differences.

[0062] Pattern matching was performed on knowledge graph substructures with similarities exceeding a preset similarity threshold to identify structural units and semantic elements with co-occurrence frequencies exceeding the preset frequency threshold. A similarity threshold of 0.7 was set to extract highly similar substructures from the knowledge graphs in the finance and education domains. Substructure extraction employed a frequent subgraph mining algorithm, which first decomposed the graph into multiple subgraphs and then identified frequently occurring patterns based on the subgraphs' hash features. In the finance and education graphs, two high-frequency structural units were identified: star structures (where a central concept connects multiple related concepts) and hierarchical structures (where concepts are organized by level of abstraction). Co-occurrence frequency analysis focused on the semantic roles and relationship types within these structural units. The frequency of occurrence of each semantic role (e.g., "core concept," "attribute description," and "instance description") and relationship type (e.g., "inclusion relationship," "precondition relationship," and "causality relationship") across different domains was calculated. A frequency threshold of 80% was set, meaning that semantic elements that appeared in more than 80% of the domains were considered common. The analysis results show that the semantic combination of "definition-example-application" appears in 95% of all analysis fields and is identified as a highly universal semantic element; the tree-like organizational structure of "core concepts and attribute relationships" appears in 92% of the fields and is identified as a universal structural unit.

[0063] The identified structural units and semantic elements are abstracted and summarized. This process involves three steps: pattern cleaning, generalization, and formalization. Pattern cleaning removes domain-specific terminology and expressions, retaining the essential characteristics of the structure and relationships. Generalization transforms similar structures and semantic expressions into a unified abstract representation. Formalization transforms the abstract representation into a formal description that can be recognized and applied.

[0064] For example, six common knowledge organization models were extracted from knowledge graphs in five fields: finance, healthcare, education, e-commerce, and technology: the hierarchical decomposition model (decomposing complex concepts into a hierarchical structure); the domain matrix model (organizing related concepts across dimensions); the sequential process model (organizing knowledge units in chronological or logical order); the comparison and contrast model (highlighting conceptual differences through contrast); the problem-solving model (organizing knowledge around problems and their solutions); and the case illustration model (illustrating abstract concepts through typical cases). Taking the hierarchical decomposition model as an example, its abstract representation is a four-layer structure of "core concept - classification dimension - subconcept - instance," which is suitable for knowledge domains with clear organizational classification systems. Examples in the financial field of "investment products - risk level - fixed income category - government bonds" and in the educational field of "discipline - difficulty - advanced mathematics - calculus" both follow this model.

[0065] For each general pattern, its applicable conditions, organizational rules, and evaluation metrics are defined. Applicable conditions indicate the knowledge types and application scenarios for which the pattern is suitable; organizational rules specify the associations and presentation order between knowledge units; and evaluation metrics include dimensions such as knowledge coverage, structural complexity, and learning efficiency. For example, the hierarchical decomposition pattern is suitable for conceptual systems with clear classifications. Its organizational rules require maintaining a unidirectional inclusion relationship between hierarchies, and the evaluation emphasizes the rationality and completeness of the hierarchical divisions. The domain matrix pattern is suitable for organizing knowledge with multi-dimensional attributes, such as comparing product features, prices, and applicable scenarios. Its organizational rules emphasize the orthogonality and consistency of dimensions. After the general knowledge organization pattern is formed, a pattern library is constructed. Each pattern includes a formal description, example templates, and application guidelines. The formal description uses a graph structure to define the pattern's node types, relationship types, and topology; the example templates provide specific application cases in multiple domains; and the application guidelines explain how to select and adapt the pattern to meet specific knowledge organization needs.

[0066] This technical solution achieves efficient identification and abstraction of cross-domain knowledge patterns; provides a basic framework and rule system for knowledge reconstruction for intelligent customer service, and improves the standardization and effectiveness of knowledge organization.

[0067] Optionally, the reconstructed knowledge subgraph is decomposed into structural templates and content filling items based on the general knowledge organization model, and the inter-domain knowledge mapping relationship is determined through semantic similarity matching to obtain a migration-optimized knowledge graph, including: constructing a structural matching matrix for the reconstructed knowledge subgraph based on the general knowledge organization model, the structural matching matrix fuses local structural similarity and semantic similarity, and decomposing the reconstructed knowledge subgraph into structural templates and corresponding content filling items through maximum bipartite matching; calculating the semantic similarity between the content filling items and the concepts of the target application domain using word vector similarity, knowledge base path similarity and context similarity, and generating an inter-domain knowledge mapping matrix based on the semantic similarity; constructing an initial knowledge structure in the target application domain based on the structural template and the inter-domain knowledge mapping matrix, identifying the concept types and relationship patterns of the target application domain, and fusing and supplementing them with the initial knowledge structure to form a domain-enhanced initial knowledge structure; jointly optimizing the structural preservation, semantic consistency and constraint satisfaction of the domain-enhanced initial knowledge structure to obtain a migration-optimized knowledge graph.

[0068] For example, when constructing a structural matching matrix, the reconstructed knowledge subgraph is compared with each pattern in the general knowledge organization pattern library. This comparison considers two key dimensions: local structural similarity and semantic similarity. Local structural similarity is calculated using the graph edit distance, which measures the number of operations required to transform the local structure in the knowledge subgraph into the pattern structure. For example, comparing the three-layer structure of "product type-risk level-return characteristics" in the financial product knowledge subgraph with the hierarchical decomposition pattern of "concept-attribute-feature" in the general pattern library requires two node replacements and one edge attribute modification, with a normalized edit distance of 0.15 (smaller distances indicate higher similarity). Semantic similarity is calculated based on semantic matching of node and relationship types. A pre-trained semantic representation model is used to extract semantic features of nodes and relationships, and their cosine similarity with the corresponding elements in the pattern is calculated. The local structural similarity and semantic similarity are combined into a comprehensive similarity through weighted averaging to construct an N×M structural matching matrix, where N is the number of local structures in the knowledge subgraph and M is the number of patterns in the pattern library.

[0069] Based on the previously constructed structural matching matrix, the knowledge subgraph decomposition problem is transformed into a bipartite graph maximum weight matching problem. Specifically, a bipartite graph is constructed, where the left-side node set represents all local structures in the knowledge subgraph, and the right-side node set represents all patterns in the general knowledge organization pattern library. The edge weights are the similarity values ​​in the corresponding structural matching matrix. An optimized implementation based on the Hungarian algorithm is used to compute the maximum weight matching. This algorithm first initializes feasible top labels, then iteratively constructs an alternating path tree, updating the matching and feasible top labels until the optimal match is found. The matching results form a mapping table indicating which pattern should be used to represent each local structure. For example, the "wealth management product - risk level - return performance" local structure is matched to the "concept - attribute - metric" hierarchical decomposition pattern with a similarity of 0.87; the "fixed deposit vs. wealth management product" local structure is matched to the "project A - comparison dimension - project B" comparison pattern with a similarity of 0.82. Based on the matching results, each local structure in the knowledge subgraph is structurally abstracted and content separated according to the corresponding pattern. Structural abstraction extracts the topological relationships and node types of the local structure to form a structural template. Content separation extracts specific conceptual entities, attribute values, and relationship descriptions as content-filling items. For example, for the example "Fund Investment - Risk Level - Level R3 (Medium Risk) - Expected Annualized Return 4%-6%," the resulting structural template is "Investment Product Type - Risk Attributes - Risk Level - Return Attributes - Return Range," with the corresponding content-filling items being "Fund Investment," "Level R3 (Medium Risk)," and "Expected Annualized Return 4%-6%." To handle edge cases, structural overlap detection and conflict resolution mechanisms are implemented. When a local structure may match multiple patterns, the pattern with the highest similarity is selected. When the similarity difference is below a threshold (e.g., 0.05), multiple patterns are retained for subsequent processing.

[0070] A concept set is extracted from the target application domain (e.g., insurance). The multi-dimensional semantic similarity between content-filled items in the source domain (e.g., banking and wealth management) and concepts in the target domain is then calculated. Word vector similarity is calculated by comparing the word vectors of concept names using a domain-adapted word embedding model trained on a comprehensive corpus and domain-specific documents to accurately capture the semantics of specialized terms. For example, the word vector similarity between "principal-protected products" in the banking and wealth management domain and "principal-guaranteed insurance" in the insurance domain is 0.82, indicating a high degree of semantic correlation. Knowledge base path similarity examines the structural similarity of concepts within their respective knowledge bases. Concepts' hierarchical and hyponymous relationships and associated concept paths are extracted, and their structural similarity is calculated. In the above example, "principal-protected products" and "principal-guaranteed insurance" both occupy similar positions in their respective domain classification systems and are subcategories of low-risk investment instruments, resulting in a path similarity of 0.75. Contextual similarity analyzes the usage and functional roles of concepts in real-world application scenarios. Contextual fragments of concepts are extracted from customer service conversation transcripts, and their usage contexts are compared using a textual semantic similarity algorithm. For example, "principal protection products" and "principal-guaranteed insurance" often appear in customer inquiries in the context of security inquiries and risk aversion, with a calculated contextual similarity of 0.79. The word vector similarity, path similarity, and contextual similarity are weighted averaged to form a comprehensive semantic similarity. A threshold screening (e.g., similarity ≥ 0.7) is then used to generate an inter-domain knowledge mapping matrix. This matrix is ​​P × Q dimensional, where P is the number of source domain content entries and Q is the number of target domain concepts. The matrix element values ​​represent the mapping strength of the corresponding concept pairs.

[0071] Based on the structural template and the inter-domain knowledge mapping matrix, an initial knowledge structure is constructed in the target application domain. Concept types and relationship patterns in the target application domain are identified and integrated with the initial knowledge structure to form a domain-enhanced initial knowledge structure. When constructing the initial knowledge structure, the structural template of the source domain serves as the skeleton. Based on the knowledge mapping matrix, the source domain's content items are replaced with corresponding concepts in the target domain. For example, the hierarchical decomposition template "Product Type - Risk Level - Return Characteristics - Applicable Population" from the banking wealth management domain can be applied to the insurance domain. Using the mapping matrix, "wealth management products" are replaced with "insurance products," and "fixed income" is replaced with "life insurance," thus constructing the initial structure for the insurance domain.

[0072] During implementation, different mapping scenarios were handled by adjusting confidence thresholds: high-confidence mappings (e.g., similarity ≥ 0.8) were directly replaced; medium-confidence mappings (similarity between 0.6 and 0.8) were double-labeled for verification; low-confidence or unmapped items were marked as pending. For example, in the migration of bank wealth management products to the insurance sector, the concept of "annualized rate of return" was mapped to "expected rate of return" in the insurance sector with a confidence of 0.85, resulting in a direct replacement; "Risk Level - R3" was mapped to "Risk Level - Medium" with a confidence of 0.72, resulting in a double label of "Risk Level - R3 / Medium"; and "Wealth Manager" had no high-similarity mapping in the insurance sector and was marked as pending.

[0073] Identify the unique concept types and relationship patterns of the target application domain. This step is achieved by analyzing the target domain knowledge base and domain documents to extract important concepts and relationships not covered by the initial structure. In the insurance domain, unique concepts such as "policy terms," ​​"exemptions," and "claims process" are identified, as well as unique relationship patterns such as "underwriting-claims." The importance of these unique elements is assessed, and a comprehensive score is calculated based on three dimensions: frequency of use, centrality, and customer attention. Elements with importance above a threshold are included in the integration and supplementation phase. This integration and supplementation process adheres to three principles: maintaining structural consistency (new elements should conform to the existing structural pattern), the principle of minimal modification (minimizing changes to the initial structure), and domain adaptability (adjustments should reflect the knowledge characteristics of the target domain). Through this series of processes, an initial domain-enhanced knowledge structure is formed.

[0074] Joint optimization employs an iterative optimization approach, evaluating and improving three key metrics in each iteration. Structural preservation assesses the consistency of the knowledge structure with the original template, ensuring that the knowledge organization logic is not disrupted. This metric is quantified by calculating graph isomorphism and path preservation. Semantic consistency evaluates the correctness and coherence of relationships between concepts, assessed through relationship triple verification and semantic conflict detection. Constraint satisfaction assesses whether the structure conforms to domain-specific knowledge constraints, such as mutually exclusive and dependent relationships. During the optimization process, appropriate correction strategies are implemented for different issues. Structural conflicts are resolved by adjusting node positions or adding transition nodes; semantic inconsistencies are resolved by modifying relationship types or remapping concepts; and constraint violations are resolved by adding missing nodes or adjusting relationship directions. Taking the insurance sector migration optimization as an example, it was discovered that in the initial structure, "Insurance Product" was directly connected to "Application Process," lacking an intermediate link, resulting in structural incoherence. Based on the structural preservation assessment, a "Requirements Analysis" node was added as an intermediate link, restoring the integrity of the process in the original template. Regarding the semantic consistency issue, it was found that the "directly generate" relationship was used between "insurance products" and "expected rate of return", which is inaccurate in the insurance context. The relationship was modified to "may provide" to improve semantic accuracy.

[0075] During the constraint satisfaction optimization phase, specific constraints in the insurance domain were identified, such as "life insurance is not applicable to corporate customers." However, in the initial structure, "corporate customers" was incorrectly linked to "life insurance products." This connection was removed, and constraint descriptions were added to ensure that the knowledge adhered to domain rules. The optimization process was iterative, with the three metrics recalculated after each round of adjustments until the preset threshold or the maximum number of iterations was reached. The resulting migrated and optimized knowledge graph retained the structural advantages of the general knowledge organization model while fully adapting to the specific requirements of the target domain.

[0076] This technical solution achieves efficient migration and optimization of knowledge graphs and effective reuse of knowledge structures through structural decomposition and cross-domain knowledge mapping based on a general knowledge organization model; multi-dimensional semantic similarity calculation and inter-domain knowledge mapping technology ensure the accuracy of cross-domain concept correspondence, while the joint optimization of structure preservation, semantic consistency and constraint satisfaction ensures the quality and applicability of the migration results.

[0077] Optionally, candidate replies are generated based on the migration-optimized knowledge graph and the initial intent, the candidate replies are evaluated for knowledge accuracy, expression adaptability and domain relevance, and the candidate reply with the highest comprehensive score is selected as the output, including: extracting knowledge path representation, entity relationship representation and attribute feature representation from the migration-optimized knowledge graph, performing hierarchical attention fusion with the semantic vector of the initial intent, and constructing a multi-level context vector with knowledge perception capability; adjusting the knowledge-enhanced dialogue generation model based on the multi-level context vector to generate multiple candidate replies; performing multi-dimensional evaluation on the candidate replies: calculating the knowledge accuracy score using knowledge graph path verification and entity relationship constraint verification; calculating the language style matching and professional matching of the candidate reply based on the user feature vector to obtain the expression adaptability score; calculating the domain coverage and professional terminology accuracy of the candidate reply in combination with the structural characteristics of the migration-optimized knowledge graph and the domain ontology to obtain the domain relevance score; weighted fusion of the knowledge accuracy score, expression adaptability score and domain relevance score to obtain a comprehensive score, and selecting the candidate reply with the highest comprehensive score as the output.

[0078] For example, the initial intent is already a classification result. An intent-semantics mapping table is maintained, mapping each intent type to a predefined semantic vector. Knowledge path representation extracts multiple possible paths from the user query entity to the target answer entity, each consisting of a series of entity nodes and relationship edges. First, the query entity, such as "critical illness insurance," is identified. Then, starting from this entity in the knowledge graph, a path search algorithm of a limited depth (typically 3-4 hops) is used to find all possible related paths. Representation learning is performed on each path using a path encoder, implemented by a unidirectional LSTM network. This encoder sequentially processes the nodes and edges along the path, ultimately outputting a vector representation of the path. For example, in an insurance consultation scenario, if a user queries "What are the exclusions for critical illness insurance?", paths such as "critical illness insurance - includes - insurance terms - provisions - exclusions" and "critical illness insurance - belongs to - insurance product - has - exclusions" are extracted and encoded as 128-dimensional vectors. Entity-relationship representation focuses on the core entities and their relationships relevant to the query. A local relationship graph is constructed, consisting of the entities identified in the query (e.g., "critical illness insurance" and "disclaimer") and their first- and second-order neighbors in the knowledge graph. For each entity node, its type and attributes are extracted; for the edges between entities, the relationship type and attributes are extracted. The local relationship graph is encoded using a graph convolutional network, which updates node and edge representations through a message passing mechanism. The network consists of two convolutional layers, each followed by a Reluctant Unit (ReLU) activation function, ultimately outputting a 64-dimensional entity representation and a 32-dimensional relationship representation. Attribute features focus on the specific attribute values ​​of the entity. All attribute triplets of query-related entities are extracted from the knowledge graph, such as "critical illness insurance - payout conditions - confirmed specific disease" and "critical illness insurance - waiting period - 90 days." Textual attribute values ​​are converted to vectors using a bag-of-words model or TF-IDF. Numerical attribute values ​​are normalized, and categorical attribute values ​​are one-hot encoded. All attribute vectors are combined to form the attribute feature representation of the entity.

[0079] Hierarchical attention fusion is the core step in combining the three knowledge representations with the initial intent semantic vector. This process uses a three-level stacked attention network: first, at the path level, the relevance of each knowledge path to the initial intent is calculated, and an attention weight is assigned to each path; then, at the entity relationship level, the importance weights of each entity and relationship in the local relationship graph are calculated; finally, at the attribute level, the attribute features most relevant to the current intent are identified. Each layer of the attention network is implemented by a two-layer feedforward neural network. The input is the concatenation of the current layer features and the intent vector, and the output is the normalized attention score. Through weighted summation, a multi-level context vector with a dimension of 512 is generated that combines the three knowledge representations and the initial intent.

[0080] The knowledge-enhanced dialogue generation model employs an encoder-decoder architecture, with the encoder processing multi-level context vectors and the decoder generating responses. The encoder, consisting of a three-layer bidirectional LSTM network with 128 hidden units per layer, converts the context vectors into encoded representations. The decoder, consisting of a three-layer unidirectional LSTM network, accesses the encoder's hidden state at each time step through an attention mechanism to generate a response word sequence. The attention mechanism, implemented by a feedforward neural network, calculates the degree of match between the decoder's current hidden state and each encoder state to generate an attention distribution. To increase response diversity, a beam search strategy with a diversity penalty is employed. The beam width is set to 5, multiple candidate word sequences are retained during the decoding process, and a diversity score is introduced to reduce duplication. By adjusting the sampling temperature (e.g., 0.7, 0.8, and 0.9), 5-10 candidate responses are generated for each query, encompassing diverse expression styles and content emphasis.

[0081] Knowledge accuracy assessment first utilizes rule-based information extraction techniques to extract entity and relationship triplets from candidate responses. A domain-adaptive named entity recognition module, based on a conditional random field model, identifies entity mentions within the responses. Dependency analysis and relation template matching are then used to identify relationships between entities. For example, from the response "Critical illness insurance will not pay claims if diagnosed within the waiting period," two relationship triplets were identified: "Critical illness insurance has a waiting period" and "Diagnosis within the waiting period leads to a non-payment of claims."

[0082] Knowledge graph path verification checks whether these triples can find supporting paths in the knowledge graph. A bidirectional search strategy is used, starting with the head and tail entities of the triple to find the path connecting them. For each extracted triple, if a fully matching supporting path is found, a score of 1.0 is assigned; if a partially matching path (including similar relationships or intermediate entities) is found, a score of 0.5-0.9 is assigned depending on the degree of match; if there is no matching path, the score is 0. Entity relationship constraint verification verifies whether the triples comply with predefined domain constraint rules. A rule base is maintained, including type constraints (such as "insurance products cannot be used as attributes of people"), mutually exclusive relationship constraints (such as "the same product cannot be both high-risk and low-risk"). Each triple in the response is checked for constraint violations, with the corresponding score deducted for each rule violated. The average score of the two tests is combined to calculate the knowledge accuracy score, with a maximum score of 10.

[0083] The expression adaptability assessment calculates the matching degree of candidate replies based on the user feature vector. The language style matching assessment uses a text style feature extractor to extract features such as sentence length distribution, vocabulary complexity, and frequency of modal particles in the reply, and calculates cosine similarity with the user's language preference vector. For example, replies to ordinary users should be concise and straightforward, with sentence length limited to 15 words or less. The professionalism matching degree is calculated by calculating the density of professional terms in the reply and comparing it with the user's professional level. A terminology difficulty grading table is maintained, which divides the terms into elementary, intermediate, and advanced levels, and a weighted professionalism score is calculated based on the frequency of use of terms at each level in the reply. If the professionalism of the reply matches the user's level, full marks are awarded; the greater the deviation, the lower the score.

[0084] Domain relevance assessment is calculated by combining the structural features of the transferred and optimized knowledge graph with the domain ontology. Domain coverage measures the extent to which a response covers knowledge relevant to the query. First, based on the initial intent and query entity, a relevant subgraph in the knowledge graph is identified, containing core nodes and edges directly related to the query. The response is then analyzed to determine how many nodes and edges of the relevant subgraph are covered. Coverage is calculated based on the degree of match between the identified entity and relationship triples and the relevant subgraph. For example, for the query "critical illness insurance exemption clause," the relevant subgraph includes the "critical illness insurance" and "exempt clause" nodes, along with their associated nodes such as "waiting period" and "compensation limit." If a response mentions 80% of these core nodes, domain coverage is high; if only 20% is covered, coverage is low. Terminology accuracy assesses whether the terminology used in the response is standardized and accurate. A standard dictionary of domain terms is maintained, containing standardized terms and their variants. Terms identified in the response are checked for standardized expressions. For example, in the insurance field, "surance cancellation" is a standard term, while "refund" is used irregularly; "insurance liability" is a standard expression, while "coverage" is understandable but unprofessional. The accuracy score is calculated by calculating the proportion of standardized terms. The domain relevance score is obtained by taking a weighted average of the domain coverage and the accuracy of professional terms, with a maximum score of 10.

[0085] A comprehensive score is calculated by weighting the knowledge accuracy score, expression adaptability score, and domain relevance score. This weighted fusion uses a linear combination approach, with weights assigned based on the user's query type. For knowledge consultation queries, the weights for knowledge accuracy are 0.5, for expression adaptability 0.3, and for domain relevance 0.2; for advice queries, the weights are 0.3, 0.5, and 0.2, respectively; and for terminology queries, the weights are 0.4, 0.2, and 0.4, respectively. The query type is determined based on the initial intent classification; for example, "inquiry about insurance terms" falls into the knowledge consultation category. The comprehensive score is calculated as the weighted sum of the scores of the three dimensions and their corresponding weights. The comprehensive scores of all candidate responses are compared, and the highest score is selected as the output. When the difference between multiple responses is less than 0.2 points, response length (preferring responses between 100-300 words) and the degree of difference from previous responses are considered as auxiliary judgment criteria. A feedback collection mechanism is also designed to dynamically adjust the scoring weights based on user satisfaction and follow-up questions, continuously optimizing response quality.

[0086] This technical solution achieves knowledge-aware dialogue generation and refined response evaluation by extracting multi-level knowledge representation from the migration-optimized knowledge graph and integrating it with the initial intent; the dynamic weight adjustment mechanism enables the system to flexibly adjust the evaluation criteria according to different scenarios and user needs, providing a personalized service experience.

[0087] Optionally, the initial knowledge structure is jointly optimized in terms of structure preservation, semantic consistency and constraint satisfaction to obtain a migration optimized knowledge graph, including: constructing a dynamic structure preservation evaluation module, the dynamic structure preservation evaluation module generates a dynamic structure deviation matrix based on a graph structure similarity measurement function and a temporal attention mechanism, calculating the structure change trend according to the dynamic structure deviation matrix and dynamically adjusting the optimization target weight coefficient to obtain a structure preservation score; fusing text semantic vectors, structural feature vectors and attribute feature vectors to construct a multimodal knowledge representation, learning the association between modalities through a cross-modal attention mechanism to obtain a unified knowledge representation, calculating the semantic consistency within and between modalities based on the unified knowledge representation, and obtaining a semantic consistency score using contrastive learning; mining implicit constraint rules from historical migration data and extracting local constraint features using a graph neural network, designing an adaptive threshold determination mechanism, and calculating a constraint satisfaction score based on a soft constraint mechanism; substituting the structure preservation score, semantic consistency score and constraint satisfaction score into a multi-objective optimization function for joint optimization, iteratively updating the knowledge structure by dynamically adjusting the optimization step size to obtain a migration optimized knowledge graph.

[0088] For example, the graph structure similarity metric function considers node correspondence, edge consistency, and path preservation, calculating a feature vector for each optimized structure. The temporal attention mechanism processes the structural feature sequences of multiple rounds using LSTM and self-attention layers to identify key change points and trends. A structural deviation matrix is ​​calculated based on the current features and temporal weighted representation. This matrix reflects the degree of structural deviation between nodes. In insurance knowledge transfer, it was found that the preservation rate of the "insurance product-inclusion-terms" path decreased, and temporal attention assigned a high weight, indicating that the structure deviated from the original template.

[0089] Based on the deviation matrix, the overall deviation, key node deviation, and change acceleration are calculated, and optimization weights are dynamically adjusted. When the structure changes rapidly or key node deviations increase, the structural preservation weight is increased; when the structure is stable, the weight is reduced to increase adaptability. For example, if a rapid change in the "claims process" is detected, the structural preservation weight is increased from 0.4 to 0.6 to ensure process integrity.

[0090] Three cross-attention modules are implemented to handle text-structure, text-attribute, and structure-attribute interactions. Each module computes bidirectional attention via a query-key-value mechanism to capture complementary modal information. For the "critical illness insurance" node, cross-modal attention is used to semantically align the text description with its structural position (connecting "insurance product" and multiple "disease types").

[0091] Three contrastive tasks were designed to assess semantic consistency: intra-modality comparison of positive and negative samples, inter-modality comparison of the same node, and inter-modality comparison of different nodes. A cross-modal attention network was trained by minimizing the contrastive loss to obtain a semantically consistent unified representation. Node semantic consistency scores were calculated based on this unified representation, and the overall consistency score was calculated as a weighted average of the node scores.

[0092] The constraint-aware graph neural network, comprised of three layers of graph convolution and a constraint checking module, evaluates whether node representations satisfy predefined constraints. An adaptive threshold mechanism adjusts the constraint judgment criteria based on the optimization phase: initially loose to allow for structural adjustments, and later tightened to ensure that key constraints are met. A soft constraint mechanism quantifies constraint satisfaction into a continuous score, which is then used to calculate an overall constraint satisfaction score.

[0093] The three scores are fed into a multi-objective optimization function for joint optimization. The knowledge structure is updated iteratively using a dynamic step size. The optimization process includes adjusting node relationships, modifying attributes, and introducing new nodes. Operation probabilities are dynamically adjusted based on the current scores. The step size adjustment mechanism adjusts the step size based on the optimization progress and gradient changes: initially, a large step size is used for accelerated adjustment, while a small step size is used for refined optimization near convergence. Multiple stopping conditions can be set, such as reaching the maximum number of iterations, objective improvement falling below a threshold, or meeting business requirements.

[0094] In a second aspect, the present invention provides an intelligent customer service dialogue generation and optimization system based on a knowledge graph, comprising: The first unit is used to use a bidirectional long short-term memory network to perform entity recognition and intent recognition on the question text input by the user to obtain the question entity and initial intent; based on the question entity, a multi-hop query is performed in a preset knowledge graph to obtain an initial knowledge subgraph; The second unit is configured to construct a user feature vector including dialogue state features, problem-solving features, and service preference features based on the question text input by the user; hierarchically reconstruct the initial knowledge subgraph based on the user feature vector to obtain a reconstructed knowledge subgraph; The third unit is used to calculate the structural similarity and identify the elements of the existing domain knowledge graph to extract the general knowledge organization model; based on the general knowledge organization model, the reconstructed knowledge subgraph is decomposed into a structural template and content filling items, and the knowledge mapping relationship between domains is determined by semantic similarity matching to obtain a migration-optimized knowledge graph; The fourth unit is used to generate candidate responses based on the migration-optimized knowledge graph and the initial intent, evaluate the candidate responses in terms of knowledge accuracy, expression adaptability, and domain relevance, and select the candidate response with the highest comprehensive score as output.

Claims

1. An optimization method for intelligent customer service dialogue generation based on knowledge graph, characterized by: include: Use a bidirectional long short-term memory network to perform entity recognition and intent recognition on the question text entered by the user to obtain the question entity and initial intent; Perform multi-hop query in a preset knowledge graph based on the problem entity to obtain an initial knowledge subgraph; Constructing a user feature vector including dialogue state features, problem-solving features, and service preference features based on the question text input by the user; hierarchically reconstructing the initial knowledge subgraph based on the user feature vector to obtain a reconstructed knowledge subgraph; Calculate structural similarity and identify elements of existing domain knowledge graphs to extract common knowledge organization patterns; Decomposing the reconstructed knowledge subgraph into a structural template and content filling items based on the general knowledge organization model, determining the knowledge mapping relationship between domains through semantic similarity matching, and obtaining a migration optimized knowledge graph; Candidate responses are generated based on the migration-optimized knowledge graph and the initial intent, the candidate responses are evaluated for knowledge accuracy, expression adaptability, and domain relevance, and the candidate response with the highest comprehensive score is selected as output.

2. The method according to claim 1, characterized in that Performing a multi-hop query in a preset knowledge graph based on the problem entity to obtain an initial knowledge subgraph includes: An entity relationship importance scoring matrix is ​​constructed based on the problem entity; an adaptive attenuation factor related to the query path hop count and the path score is generated based on the entity relationship importance scoring matrix; an expected return of the query path starting from the problem entity is calculated based on the entity relationship importance scoring matrix and the adaptive attenuation factor, and a prioritized query path set is generated; Selecting multiple complementary paths with the highest scores from the prioritized query path set, allocating computing resources based on the path scores of the complementary paths to perform parallel queries, and obtaining multiple candidate knowledge subgraphs and their feature matrices; The complementarity and redundancy between the candidate knowledge subgraphs are calculated based on the feature matrix, the complementary regions in the candidate knowledge subgraphs are weightedly fused, and the redundant branches of the overlapping regions are pruned to obtain the initial knowledge subgraph.

3. The method according to claim 1, characterized in that The user feature vector constructed based on the question text entered by the user includes the following: Performing multimodal feature extraction based on the user's question text, including: identifying professional terms in the question text to obtain a professionalism score, extracting user message length sequences and response time sequences to obtain an interaction pattern feature vector, analyzing the frequency of emoticon use to obtain sentiment polarity strength; and fusing the professionalism score, interaction pattern feature vector, and sentiment polarity strength through an attention mechanism to obtain a fused feature vector. Constructing a user intention-topic transfer matrix based on the initial intention, calculating adjacent topic similarities based on the user intention-topic transfer matrix, and generating a conversation coherence index based on the adjacent topic similarities; calculating question complexity and answer completeness to obtain a conversation depth index; and constructing a multi-round conversation state transition network based on the conversation coherence index and the conversation depth index; Construct a business knowledge association map, analyze the user's consultation scope based on the business knowledge association map, identify potential service needs, and calculate the problem-solving path; Performing self-organizing map clustering on the user interaction vector sequence to obtain interaction behavior categories, and constructing a service preference vector based on the interaction behavior categories; The fused feature vector, the multi-round dialogue state transition network, the problem-solving path, and the service preference vector are combined to form a user feature vector.

4. The method according to claim 3, characterized in that The initial knowledge subgraph is hierarchically reconstructed based on the user feature vector to obtain a reconstructed knowledge subgraph comprising: hierarchical the nodes of the initial knowledge subgraph according to the fused feature vector to obtain a preliminary hierarchical structure; constructing an inter-layer transition probability matrix using the multi-round dialogue state transition network, and dynamically adjusting the hierarchical structure based on the inter-layer transition probability matrix to obtain an optimized hierarchical structure; constructing a knowledge transfer graph based on the problem-solving path, mapping the optimized hierarchical structure to the knowledge transfer graph to obtain multiple possible transfer paths; reordering the transfer paths based on the service preference vector to obtain an optimal transfer path; A multidimensional knowledge network is constructed based on the optimal transfer path, knowledge redundancy is eliminated and supplemented, and a reconstructed knowledge subgraph is obtained.

5. The method according to claim 4, characterized in that Based on the optimal transfer path, a multi-dimensional knowledge network is constructed to eliminate and supplement knowledge redundancy, and the reconstructed knowledge subgraph includes: Constructing a node semantic similarity matrix and a structural correlation matrix based on the optimal transfer path, and obtaining a hybrid correlation matrix through weighted fusion; Identifying knowledge clusters using the hybrid relevance matrix and calculating importance weights of nodes within the cluster based on a hierarchical attention mechanism; Constructing a multidimensional knowledge network in semantic, structural and functional dimensions according to the node importance weights and the hybrid relevance matrix; The multidimensional knowledge network is used to perform cross-dimensional reasoning, identify and supplement knowledge gaps, and optimize the overall network structure through weighted graph embedding to obtain a reconstructed knowledge subgraph.

6. The method according to claim 1, characterized in that Calculate structural similarity and identify elements of existing domain knowledge graphs to extract common knowledge organization patterns, including: Construct the structural feature vector of the existing domain knowledge graph, including node degree distribution, path length distribution and local clustering coefficient; Embed nodes in the existing domain knowledge graph based on a graph neural network and obtain a graph-level representation through attention pooling; Calculating the structural similarity matrix between knowledge graphs in different fields using the structural feature vector and graph-level representation; Perform pattern matching on knowledge graph substructures with similarity higher than a preset similarity threshold, and identify structural units and semantic elements with co-occurrence frequencies higher than a preset frequency threshold; The identified structural units and semantic elements are abstracted and summarized to form a general knowledge organization model.

7. The method according to claim 1, characterized in that Based on the general knowledge organization model, the reconstructed knowledge subgraph is decomposed into a structural template and content filling items, and the knowledge mapping relationship between domains is determined by semantic similarity matching to obtain a migration optimized knowledge graph including: Based on the general knowledge organization model, a structural matching matrix is ​​constructed for the reconstructed knowledge subgraph, wherein the structural matching matrix integrates local structural similarity and semantic similarity, and decomposes the reconstructed knowledge subgraph into structural templates and corresponding content filling items through maximum bipartite matching; Calculating semantic similarity between the content filling item and the target application domain concept using word vector similarity, knowledge base path similarity, and context similarity, and generating an inter-domain knowledge mapping matrix based on the semantic similarity; Constructing an initial knowledge structure in the target application domain based on the structural template and the inter-domain knowledge mapping matrix, identifying concept types and relationship patterns in the target application domain, and fusing and supplementing them with the initial knowledge structure to form a domain-enhanced initial knowledge structure; The initial knowledge structure enhanced in the domain is jointly optimized in terms of structure preservation, semantic consistency and constraint satisfaction to obtain a migration optimized knowledge graph.

8. The method according to claim 1, characterized in that Generating candidate responses based on the migration-optimized knowledge graph and the initial intent, evaluating the candidate responses for knowledge accuracy, expression adaptability, and domain relevance, and selecting the candidate response with the highest comprehensive score as output includes: Extracting knowledge path representations, entity relationship representations, and attribute feature representations from the transfer-optimized knowledge graph, performing hierarchical attention fusion with the semantic vector of the initial intent, and constructing a multi-level context vector with knowledge perception capabilities; adjusting the knowledge-enhanced dialogue generation model based on the multi-level context vector to generate multiple candidate responses; The candidate responses are evaluated in multiple dimensions: knowledge accuracy scores are calculated using knowledge graph path verification and entity relationship constraint verification; the language style matching and professionalism matching of the candidate responses are calculated based on the user feature vector to obtain an expression adaptability score; and the domain coverage and professional terminology accuracy of the candidate responses are calculated by combining the structural features of the migration-optimized knowledge graph and the domain ontology to obtain a domain relevance score. The knowledge accuracy score, expression adaptability score and domain relevance score are weightedly fused to obtain a comprehensive score, and the candidate response with the highest comprehensive score is selected as the output.

9. An intelligent customer service dialogue generation and optimization system based on knowledge graph, used to implement the method of any one of claims 1 to 8, characterized in that: include: The first unit is used to use a bidirectional long short-term memory network to perform entity recognition and intent recognition on the question text input by the user to obtain the question entity and initial intent; Perform multi-hop query in a preset knowledge graph based on the problem entity to obtain an initial knowledge subgraph; The second unit is configured to construct a user feature vector including dialogue state features, problem-solving features, and service preference features based on the question text input by the user; hierarchically reconstruct the initial knowledge subgraph based on the user feature vector to obtain a reconstructed knowledge subgraph; The third unit is used to calculate the structural similarity and identify elements of the existing domain knowledge graph and extract the general knowledge organization model; Decomposing the reconstructed knowledge subgraph into a structural template and content filling items based on the general knowledge organization model, determining the knowledge mapping relationship between domains through semantic similarity matching, and obtaining a migration optimized knowledge graph; The fourth unit is used to generate candidate responses based on the migration-optimized knowledge graph and the initial intent, evaluate the candidate responses in terms of knowledge accuracy, expression adaptability, and domain relevance, and select the candidate response with the highest comprehensive score as output.

Images