A multi-model collaborative customer service intelligent response method and system based on a hierarchical architecture

Abstract

The application discloses a multi-model collaborative customer service intelligent response method and system based on a layered architecture, relates to the technical field of natural language processing, and comprises the following steps: performing semantic analysis processing on user input, extracting intent categories and business constraint information, and forming a structured intent representation; triggering at least two logical heterogeneous executors to run in parallel according to the structured intent representation, each executor outputting a corresponding candidate reply set, while adding an initial confidence mark to each candidate reply to constitute a cross-model candidate pool; performing semantic consistency comparison on any candidate reply in the candidate pool, identifying semantic conflict relationships between candidate replies, and mapping the conflict relationships into a constraint structure; and under the constraint propagation effect of the constraint structure, adjusting the initial confidence of each candidate reply for consistency, providing a mathematical and logical basis for subsequent conflict resolution, and enabling the system to have the premise of fact checking and logical self-consistency.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a multi-model collaborative intelligent customer service response method and system based on a hierarchical architecture. Background Technology

[0002] Intelligent customer service systems have become a key infrastructure for enterprise digital transformation. Current technological development is shifting from traditional keyword matching and retrieval-based question answering to generative interaction based on deep learning and large language models. In order to improve the accuracy of semantic understanding while ensuring response speed, the industry is gradually exploring multi-model collaborative solutions, attempting to cope with complex and ever-changing user inquiries by integrating models with different parameter levels or different training paradigms. Existing technical solutions usually adopt serial processing or simple routing and distribution mechanisms, that is, assigning user requests to specific models for processing according to preset rules, or using large models as central controllers to schedule various specialized small models in vertical fields.

[0003] When faced with complex customer service scenarios requiring high concurrency and high reliability, existing multi-model collaboration technologies often rely on static routing strategies or simple chained calls. They lack the ability to deeply align semantics and resolve conflicts in the outputs of heterogeneous models. When different models generate responses with factual contradictions or logical conflicts for the same intent, traditional systems often lack effective arbitration mechanisms, which can easily lead to the transmission of "illusory" content and seriously affect the user experience. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a multi-model collaborative customer service intelligent response method based on a layered architecture. This addresses the problem that existing collaborative mechanisms often rely on static routing strategies or simple chained calls, lacking the ability to deeply align semantics and resolve conflicts in the output results of heterogeneous models. When different models generate responses with factual contradictions or logical conflicts for the same intent, traditional systems often lack an effective arbitration mechanism, which can easily lead to the transmission of "illusory" content and seriously affect the user experience.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a multi-model collaborative intelligent customer service response method based on a hierarchical architecture, which includes performing semantic parsing processing on user input, extracting intent category and business constraint information, and forming a structured intent representation; Based on the structured intent representation, at least two logically heterogeneous executors are triggered to run in parallel. Each executor outputs a corresponding set of candidate responses and attaches an initial confidence label to each candidate response to form a cross-model candidate pool. Perform semantic consistency comparison on any candidate response in the candidate pool, identify semantic conflict relationships between candidate responses, and map the conflict relationships into constraint structures; Under the constraint propagation effect of the constraint structure, the initial confidence of each candidate response is adjusted for consistency. A conflict suppression mechanism is introduced to weaken the weight of candidate responses with semantic contradictions, while enhancing the confidence transmission strength on the semantically consistent path. This forms a calibrated confidence distribution modulated by the structural constraints. Based on the calibrated confidence distribution, the structured screening and fusion decision of candidate responses are performed, and the fusion response result is output. The fused response is subjected to compliance rule verification. When a response that violates the constraints is detected, the violation constraint is located to the corresponding conflict relationship and the constraint structure is updated. At the same time, the confidence correction and conflict resolution fusion process is re-triggered, and the response result that meets the compliance constraints is taken as the final output.

[0007] As a preferred embodiment of the multi-model collaborative customer service intelligent response method based on a hierarchical architecture described in this invention, the structured intent representation includes: The system performs text standardization and noise removal on user input, and encodes the input text based on a pre-trained semantic understanding model, outputting the corresponding intent probability distribution vector. The intent label with the highest probability is selected on the probability distribution vector. The intent label is then compared with the intent identifier field of each rule in the business rule base. When the comparison is consistent, the business restriction field in the rule is extracted using the rule parser. The intent label is filled into the intent field and the business restriction field is filled into the constraint field using a predefined structured template to obtain a unified data structure.

[0008] As a preferred embodiment of the multi-model collaborative customer service intelligent response method based on a hierarchical architecture described in this invention, the cross-model candidate pool includes: By mapping intent tags to execution path identifiers in the policy routing table, the retrieval execution path is activated when the found entry contains a retrieval identifier, and the inference execution path is activated when it contains an inference identifier. The retrieval execution path generates candidate responses and outputs corresponding confidence scores based on knowledge base vector retrieval, while the inference execution path generates candidate responses and outputs corresponding confidence scores based on multi-step tool calls and logical deduction. Candidate responses and confidence scores from different execution paths are uniformly formatted, encapsulated, and aggregated for storage, ensuring that candidate responses from different sources retain their source identifiers and initial confidence attributes within the same candidate pool, thereby forming a cross-model candidate set for the model.

[0009] As a preferred embodiment of the multi-model collaborative intelligent customer service response method based on a hierarchical architecture described in this invention, the constraint structure includes: Semantic vector encoding is performed on each candidate response in the candidate pool, and a similarity matrix is ​​constructed based on the similarity between the vectors. Candidate pairs below a preset threshold are marked as conflicting, while highly similar candidate pairs retain consistent association relationships. By using the conflict marking relationship and the consistency association relationship, the candidate responses are represented as nodes in the graph structure, and the consistency relationship and the conflict relationship are mapped to different types of edges, thereby forming a structured graph model to describe the constraints and competition relationships between candidate responses.

[0010] As a preferred embodiment of the multi-model collaborative intelligent customer service response method based on a hierarchical architecture described in this invention, the fused response result includes: The candidate response set is vector-encoded to obtain the corresponding semantic vector, and the cosine similarity matrix between any two candidate response vectors is calculated. For candidate response pairs with similarity greater than a preset threshold, undirected consistent edges are established in the graph structure with the similarity value as the edge weight. Perform pairwise semantic contradiction judgment between candidate responses. When mutually exclusive conclusions or semantic negation are detected, establish conflict edges between corresponding nodes and use the conflict intensity score as the edge weight. Perform text alignment matching on candidate responses output by different models. Establish cross-model alignment edges between successfully matched candidate response nodes and use the matching confidence score as the edge weight. During the graph structure initialization phase, the original model output score of each candidate response node is normalized and probability-mapped. The mapped probability distribution is used as the initial confidence state of the node and stored in the node attributes. During the graph structure iterative update phase, the node confidence states are updated with distribution constraints according to edge types. Specifically: for adjacent nodes connected by consistent edges, a normalized consistency influence distribution is constructed based on edge weights, and the node confidence states are updated with probability redistribution based on this distribution; for adjacent nodes connected by conflicting edges, conflict edge weights are mapped to gating suppression factors, and the node confidence states are updated with suppression through a nonlinear gating function; for nodes connected by cross-model aligned edges, the confidence states from different sources are projected to the same representation space using a shared embedding mapping. The alignment fusion update of the confidence states is obtained by solving the maximum consistency mapping of the aligned node confidence distribution in the space and using the Lagrange multiplier method for constraint optimization. In each round of propagation, the propagation path between nodes is serialized and recorded, and the conflict edges in the path are labeled segment by segment. At the same time, a path position decay function is introduced to assign different influence weights to conflict edges at different positions, so that the path pollution coefficient is jointly determined by the conflict edge weights and the position decay function. When the path pollution coefficient exceeds the preset threshold, the path is not deleted as a whole. Instead, the decay node is determined based on the position index of the conflict edge in the path. An exponential decay function is used to reduce the weight of the conflict edge along the path direction, and the reduced weight is redistributed to the adjacent consistent edges on the path according to the consistency contribution ratio.

[0011] As a preferred embodiment of the multi-model collaborative intelligent customer service response method based on a layered architecture described in this invention, the compliance rule verification includes: In the graph structure, a depth-first search is used to enumerate the paths for each candidate response node, obtaining multiple propagation paths from the source node to the sink node. The confidence values ​​of the nodes on each propagation path are calculated using a segmented recursive fusion method. The complete path is divided into several semantically consistent sub-segments. Confidence aggregation is performed within the sub-segments, and a consistency constraint function is introduced between the sub-segments for recursive correction. The path pollution intensity is obtained by calculating the weighted sum of all conflicting edges on the path after position decay. When the path pollution intensity exceeds the preset threshold, the path is removed and the remaining paths participate in the subsequent fusion calculation. The path confidence values ​​of the retained paths are normalized based on the path length, and the paths are serialized. The path pollution intensity is calculated based on the location of conflict edges, and a location-related attenuation coefficient is generated through an exponential gating function. The attenuation coefficient is applied to the path confidence propagation process to perform differentiated modulation on different path segments, and finally normalized to generate the path weight distribution under global constraints. A path competition relationship model is introduced for the set of retained paths. The structural conflicts and consistency relationships between paths are constructed into a path competition graph, and consistency propagation calculation is performed on the path competition graph to generate path contribution coefficients. The variance between paths is calculated based on the path weight distribution, and the gating coefficient is calculated in combination with the path pollution intensity. The gating coefficient is then used as an adjustment factor to perform a secondary scaling update on the path weights. The path weight distribution is back-mapped to the graph structure edge set. Consistent edges are redistributed based on the path contribution ratio through multiplicative enhancement. Conflicting edges are weakened based on the path contamination contribution through an exponential decay function. Cross-model aligned edges are adjusted by mixing the path consistency contribution ratio with the original weight of the aligned edge through a nonlinear gating function to obtain updated edge weights. When the path weight fluctuates below a preset threshold in continuous iterations, a convergence judgment is triggered. Otherwise, the edge weights of the subgraph in the region with dense conflict edges are re-initialized and the propagation calculation process is restarted.

[0012] As a preferred embodiment of the multi-model collaborative intelligent customer service response method based on a hierarchical architecture described in this invention, the adaptive structural update includes: By using a pre-set compliance rule base, the fusion response results are matched with rules and checked for logical consistency. When violations are detected, the candidate nodes that trigger the violations and the associated conflict relationships are located. When a compliance violation is detected in the fusion response result, the candidate nodes and associated conflict subgraphs corresponding to the violation are located, and a structural reconstruction operation is performed on the conflict subgraph, including disconnecting the conflict edge that triggered the violation, reconstructing alternative connection edges based on the semantic similarity between candidate nodes, updating the path weight allocation associated with the conflict edge, correcting the affected propagation paths in the global graph, and re-executing the confidence propagation and fusion calculation process on the reconstructed graph structure.

[0013] Secondly, the present invention provides a multi-model collaborative intelligent customer service response system based on a layered architecture, including an intent parsing module, a heterogeneous execution module, a conflict mapping module, a confidence fusion module, and a compliance verification module. The intent parsing module is used to perform semantic parsing processing on user input, extract intent categories and business constraint information, and form a structured intent representation; The heterogeneous execution module is used to trigger at least two logical heterogeneous executors to run in parallel based on the structured intent representation. Each executor outputs a corresponding set of candidate responses and attaches an initial confidence label to each candidate response to form a cross-model candidate pool. The conflict mapping module is used to perform semantic consistency comparison on any candidate response in the candidate pool, identify semantic conflict relationships between candidate responses, and map the conflict relationships into a constraint structure. The confidence fusion module is used to adjust the initial confidence of each candidate response in a consistent manner under the constraint propagation effect of the constraint structure, introduce a conflict suppression mechanism to weaken the weight of candidate responses with semantic contradictions, enhance the confidence transmission strength on the semantically consistent path, form a calibrated confidence distribution modulated by the structural constraints, and perform structured screening and fusion decision of candidate responses based on the calibrated confidence distribution, and output the fusion response result. The compliance verification module is used to perform compliance rule verification on the fused response. When a response that violates the constraints is detected, the violation constraint is located to the corresponding conflict relationship and the constraint structure is updated. At the same time, the confidence correction and conflict resolution fusion process is re-triggered, and the response result that meets the compliance constraints is taken as the final output.

[0014] Thirdly, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the multi-model collaborative customer service intelligent response method based on a hierarchical architecture as described in the first aspect of the present invention.

[0015] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the multi-model collaborative customer service intelligent response method based on a hierarchical architecture as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: By triggering at least two logically heterogeneous executors to run in parallel based on the structured intent representation, outputting a set of candidate responses and attaching an initial confidence level label, a cross-model candidate pool is formed. This achieves precise allocation of computing power based on task attributes and parallel acquisition of multi-source information, allowing models of different scales and specialties to work independently within their respective "comfort zones." This breaks the bottleneck of a single model handling all requests, effectively avoids the infinite loop problem, and significantly improves the processing efficiency of complex mixed intents. By performing semantic consistency comparison on any candidate response in the candidate pool, semantic conflict relationships are identified and mapped to a constraint structure. This enables the explicit and topological expression of the "information source heterogeneity" and "factual conflict" problems that inevitably arise when multiple models are generated in parallel. This provides a mathematical and logical basis for subsequent conflict resolution, enabling the system to have the prerequisite of fact verification and logical self-consistency. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a multi-model collaborative intelligent customer service response method based on a layered architecture; Figure 2 This is a schematic diagram of a multi-model collaborative intelligent customer service response system based on a layered architecture. Detailed Implementation

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0021] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0022] Reference Figures 1-2 This is one embodiment of the present invention, which provides a multi-model collaborative customer service intelligent response method based on a hierarchical architecture, including the following steps: S1. Perform semantic parsing processing on user input, extract intent category and business constraint information, and form a structured intent representation.

[0023] S1.1 Perform text standardization and noise removal on user input, and encode the input text based on a pre-trained semantic understanding model to output the corresponding intent probability distribution vector.

[0024] Furthermore, the design of text standardization and noise removal is based on the fact that the input sources in customer service scenarios are complex, the expression forms are highly unstructured, and there are cross-language issues. If they are not standardized, the semantic coding space will be shifted, thus affecting the stability of subsequent intent recognition.

[0025] First, the input text undergoes character-level unified encoding processing, including Unicode standardization, full-width and half-width character unification, and special control character filtering, to eliminate structural inconsistencies caused by differences in input from different terminals. Then, based on a dictionary mapping mechanism, high-frequency colloquial expressions are semantically equivalently replaced, mapping different expression forms to a unified semantic representation to reduce semantic noise caused by expression drift. To address semantic dilution and local semantic loss issues in long text inputs, a sliding window segmentation mechanism based on semantic breakpoint detection is introduced, dividing the input text into multiple semantically continuous units, which are then input into a pre-trained semantic encoding model for vectorization. On this basis, positional encoding and sequence alignment mechanisms are used to fuse and concatenate the local semantic vectors, ensuring that local semantic information remains consistent with the global semantic structure, thereby improving the stability and robustness of intent representation.

[0026] S1.2 Select the intent label with the highest probability on the probability distribution vector. By comparing the intent label with the intent identifier field of each rule in the business rule base, when the comparison is consistent, use the rule parser to extract the business restriction field in the rule. Fill the intent label into the intent field and the business restriction into the constraint field through the predefined structured template to obtain a unified data structure.

[0027] Furthermore, using a single maximum probability intent in the intent determination process may lead to misjudgment under low confidence distribution conditions. Therefore, a confidence difference constraint judgment mechanism is designed to improve decision reliability.

[0028] After obtaining the intent probability distribution, the difference between the highest probability and the second highest probability is calculated. When the difference is lower than a preset threshold, the input is marked as an intent ambiguity state, and multiple candidate intents are retained for subsequent processing to avoid the propagation of single-path decision errors. In the business rule matching process, to solve the problem of complex rule base structure and unstable string parsing, an index-based hierarchical rule storage structure is designed. Rules are split into intent identifier fields and business constraint fields, and a hash index is established to achieve fast location matching. In the structured encapsulation process, a unified schema mapping mechanism is introduced to map intent tags and business constraint conditions to fixed structure fields, so that data from different rule sources can be expressed under a unified data model, thereby ensuring the consistency and scalability of subsequent executor call interfaces. In the state of multiple candidate intent retention, parallel scoring consistency detection is performed on each candidate intent. By calculating the degree of divergence between the output results of different intent paths, secondary semantic correction is triggered for candidate intents with divergence exceeding the threshold to reduce the risk of intent misassignment.

[0029] S2. Based on the structured intent representation, trigger at least two logically heterogeneous executors to run in parallel. Each executor outputs a corresponding set of candidate responses and attaches an initial confidence label to each candidate response to form a cross-model candidate pool.

[0030] S2.1. By mapping intent tags to execution path identifiers in the policy routing table, when the found entry contains a retrieval identifier, the retrieval execution path is activated; when it contains an inference identifier, the inference execution path is activated. The retrieval execution path generates candidate responses and outputs corresponding confidence scores based on knowledge base vector retrieval, while the inference execution path generates candidate responses and outputs corresponding confidence scores based on multi-step tool calls and logical deduction.

[0031] Furthermore, the design of the policy routing mechanism is based on the differences in computational complexity, tool dependency, and response structure among different task types. Therefore, it is necessary to achieve task decoupling and resource allocation optimization of the execution path through the routing mechanism. The intent label is used as the main index, and a multi-dimensional routing index table is constructed by combining the task complexity label and resource constraint label. The set of active executors is determined through joint retrieval. In the retrieval-type execution path, in order to solve the semantic drift problem of single vector retrieval, a three-stage structure of "vector recall + keyword precise recall + re-ranking model" is adopted. Vector recall is responsible for semantic coverage, keyword recall is responsible for precise matching, and the re-ranking model is used to correct the semantic consistency of candidate results. In the inference-type execution path, in order to solve the context loss problem during multi-step tool calls, an intermediate state caching mechanism is introduced. The input and output results of each step of the tool call are serialized and stored, and used as the input context for the next step of inference, thereby ensuring the continuity and traceability of the inference chain.

[0032] S2.2. The candidate responses and confidence scores output by different execution paths are uniformly formatted, encapsulated, and aggregated for storage, so that the candidate responses from each source maintain the source identifier and initial confidence attribute in the same candidate pool, thereby forming a cross-model candidate set of the model.

[0033] Furthermore, in the process of constructing the cross-model candidate pool, considering the differences in text structure, scoring system and semantic granularity of different executor outputs, direct fusion will lead to distortion in subsequent conflict modeling. Therefore, a unified expression layer processing is required.

[0034] All candidate responses are structurally encapsulated, uniformly including text content fields, source identifier fields, and confidence score fields. A normalization function maps the outputs of different executors to a unified probability space to eliminate scale inconsistencies between different scoring systems. A source isolation storage mechanism is introduced to ensure that the outputs of different execution paths remain independent at the structural level and comparable at the semantic level, thus providing a reliable input foundation for subsequent conflict identification and graph structure construction. To address the issue of inconsistent scoring scales across different executor outputs, a distribution normalization mechanism is introduced for the confidence scores of candidate responses. The original scores output by retrieval-type and inference-type executors are segmented and normalized, mapping scores from different sources to a unified probability space. The normalization process standardizes scores based on the moving mean and standard deviation statistics and performs nonlinear compression using an exponential normalization function. This ensures that high-confidence scores maintain a relative advantage in the unified space while avoiding scale-dominated effects, thereby guaranteeing score comparability across model candidate pools.

[0035] S3. Perform semantic consistency comparison on any candidate response in the candidate pool, identify semantic conflict relationships between candidate responses, and map the conflict relationships into constraint structures.

[0036] S3.1. Semantic vector encoding is performed on each candidate response in the candidate pool, and a similarity matrix is ​​constructed based on the similarity between the vectors. Candidate pairs below the preset threshold are marked as conflicting, while the consistency association is preserved for highly similar candidate pairs.

[0037] Furthermore, in the process of semantic similarity calculation, considering the distribution offset problem in the encoding space of different models, vector normalization and temperature scaling mechanisms are introduced before similarity calculation to make the semantic vector distribution more consistent, thereby improving the stability of similarity calculation. In terms of threshold design, a hierarchical semantic judgment strategy is adopted, which divides candidate pairs into three categories: strong consistency relationship, weak consistency relationship and irrelevant relationship. Among them, weak consistency relationship does not directly participate in conflict judgment, but is retained as potential semantic association for supplementing semantic path information in the subsequent graph propagation stage, thereby avoiding information loss caused by binarization judgment.

[0038] S3.2. By using conflict labeling relationships and consistency association relationships, candidate responses are represented as nodes in a graph structure, and consistency relationships and conflict relationships are mapped to different types of edges, thereby forming a structured graph model to describe the constraints and competition relationships between candidate responses.

[0039] Furthermore, during the graph structure construction process, the semantic relationships between candidate responses from multiple models are explicitly and structurally expressed, thus providing a computable graph model foundation for subsequent confidence propagation. Each candidate response is represented as a graph node, and typified edge structures are introduced for different semantic relationships, including consistent edges, conflicting edges, and cross-model aligned edges. Simultaneously, independent type encoding is introduced for each edge type, enabling the graph structure to distinguish different semantic mechanisms during propagation. During node initialization, candidate responses from different sources undergo unified embedding space alignment processing, allowing different model outputs to be mapped to the same semantic space, thereby ensuring the consistency and executability of graph structure propagation computation. During graph structure construction, candidate response nodes undergo unified index encoding processing, mapping candidate responses output by different executors to unique node identifiers. A node alignment mechanism is established based on the semantic vector space consistency principle, ensuring that semantically similar but different-origin candidate responses remain structurally distinguishable yet semantically associative. Simultaneously, during edge generation, multi-condition judgment rules are introduced, including semantic similarity threshold judgment, logical negation relation judgment, and cross-source consistency judgment. Edge types are determined through multi-condition joint judgment, thereby achieving explicit distinction of different semantic relationships in the graph structure.

[0040] Consistent edges refer to the structural connections established between two candidate responses with a high degree of similarity in the semantic space, used to represent the support and consistency of the two candidate responses in semantic expression; the edges are generated by the similarity judgment results after semantic encoding.

[0041] Conflict edges refer to the structural connections established between candidate responses that have obvious semantic contradictions, negations, or mutually exclusive relationships in the semantic space. They are used to represent the opposition between two candidate responses in terms of semantic conclusions and are generated through semantic comparison and conflict identification rules.

[0042] Cross-model alignment edges refer to the connections established between candidate responses that originate from different actuators or different models but express the same or highly similar semantic content. They are used to align the correspondences of different model outputs in a unified semantic space.

[0043] The determination of logical negation is achieved through semantic recognition algorithms, including the identification of negation words in candidate responses, the judgment of polarity classification models, and the comparison of business action directions. When two candidate responses output processing results with opposite polarities for the same business object, they are determined to be logically negated.

[0044] S4. Under the constraint propagation effect of the constraint structure, the initial confidence of each candidate response is adjusted for consistency. A conflict suppression mechanism is introduced to weaken the weight of candidate responses with semantic contradictions. At the same time, the confidence transmission strength on the semantically consistent path is enhanced, forming a calibrated confidence distribution modulated by the structural constraints. Based on the calibrated confidence distribution, the structured screening and fusion decision of candidate responses are performed, and the fusion response result is output.

[0045] S4.1. The candidate response set is vector-encoded to obtain the corresponding semantic vector, and the cosine similarity matrix between any two candidate response vectors is calculated. For candidate response pairs with similarity greater than a preset threshold, undirected consistent edges are established in the graph structure with the similarity value as the edge weight.

[0046] Furthermore, the design of constructing a graph structure that includes consistency relationships, conflict relationships, and cross-model alignment relationships stems from the fact that there are not only semantic similarity relationships between the outputs of different models in a multi-model customer service system, but also semantic conflict and expression alignment relationships. If only a single similarity is used to build the graph, different semantic functions will be mixed and expressed, thereby weakening the ability to distinguish subsequent propagation.

[0047] First, candidate responses are encoded using a unified semantic vector, and a judgment module is constructed for different semantic relationships. Consistent relationships are obtained through high similarity judgment, conflicting relationships are obtained through semantic negation detection and low similarity joint judgment, and aligned relationships are obtained through entity-level matching and source mapping consistency detection. Then, the three types of relationships are mapped to different types of edge structures, and independent weight parameters and propagation coefficients are set for each type of edge, so that different semantic relationships have distinguishable propagation behavior in the graph.

[0048] Considering the inconsistencies in magnitude and scale drift among candidate responses generated by different models in the semantic space, it is necessary to first map the candidate responses to a unified semantic embedding space, and take "directional consistency" in semantic differences as the core discriminant factor, rather than magnitude differences. Therefore, the semantic vector of the candidate response is used as the semantic feature carrier, the vector direction is used as the semantic consistency representation term, and semantic similarity is modeled as the relationship between vector angles. By normalizing the vector inner product and norm, the "semantic consistency strength" is transformed into a computable scalar index, which is ultimately used to characterize the degree of semantic association between candidate responses. The cosine similarity matrix between any two candidate response vectors is calculated as follows: ; in, Let represent the semantic cosine similarity between candidate responses i and j. , These are the semantic vectors of candidate responses i and j, respectively; , These are the L2 norms of the corresponding vectors.

[0049] Considering that semantic similarity relationships need to be transformed into propagable information enhancement channels in graph structures, high-confidence semantic consistency between candidate responses is used as a positive supporting relationship for graph propagation, and semantic similarity is used as the consistency contribution strength term. A threshold filtering mechanism is used to filter low-quality semantic associations, thus mapping the highly reliable part of the original semantic similarity matrix to consistency edge weights in the graph structure, which are then used as weight amplification terms in the subsequent information enhancement propagation process. Undirected consistency edges are established with similarity values ​​as edge weights. ; in, Consistent edge weights; This is an indicator function that takes the value 1 when the condition is true and 0 otherwise. Based on a preset consistency similarity threshold, the traditional "candidate set" is transformed into a "multi-relational semantic propagation network," providing a structural foundation for subsequent path-level evaluation.

[0050] S4.2. Perform pairwise semantic contradiction determination between candidate responses. When mutually exclusive conclusions or semantic negation are detected, establish conflict edges between corresponding nodes and use the conflict intensity score as the edge weight. Perform text alignment matching on candidate responses output by different models. Establish cross-model alignment edges between successfully matched candidate response nodes and use the matching confidence score as the edge weight.

[0051] Furthermore, the design of heterogeneous edge modulation propagation in the graph structure message passing mechanism stems from the fact that traditional graph propagation methods usually use a unified propagation function when dealing with multi-relation graphs, which can easily lead to conflicting relationships being covered by consistent propagation, thus causing semantic oversmoothing problems.

[0052] Propagation rules are defined for different edge types: consistent edges use an enhanced propagation function to amplify the information contribution of similar nodes; conflicting edges use an inhibited propagation function to reduce the influence of semantically contradictory nodes; and aligned edges use a spatial projection function to solve the problem of inconsistent semantic spaces between different models. In each round of propagation, the information received by nodes is split and calculated according to the edge type, and weighted fusion is performed during the node update stage to ensure that different semantic relationships do not interfere with each other during propagation. In the graph structure propagation process, differentiated information transmission functions are defined for different edge types. Consistent edges use a normalized message passing function to scale and transmit adjacent node information according to the edge weight ratio; conflicting edges use a gated inhibition function to map the conflict edge weights into nonlinear inhibition factors that act on the node state update process; and cross-model aligned edges use a spatial projection function to embed nodes from different sources into a shared semantic space for linear projection before fusion calculation, thereby realizing a heterogeneous information propagation mechanism in multi-relation graph structures.

[0053] Considering that low similarity in semantic space not only indicates semantic irrelevance but may also imply semantic opposition or logical conflict, it is necessary to reverse-map semantic similarity to semantic conflict intensity. "Reduced similarity" is used as a conflict enhancement signal, and the conflict signal is used as a negative modulation factor to suppress propagation. This transforms low-consistency regions in semantic space into conflict constraints in a graph structure, which are then used to establish conflict edges to suppress the spread of erroneous information in the propagation path, with the conflict intensity score serving as the edge weight. ; in, For conflict edge weights, The sigmoid function maps the input to the interval (0,1). Let be the cosine similarity.

[0054] Considering the distribution offset and representation inconsistency issues among different actuators or models in the semantic representation space, it is necessary to map the outputs of different models to a unified semantic alignment space. The semantic consistency between different sources is used as an indicator of cross-model alignment strength. By calculating the relative directional consistency between the semantic vectors of different models, the differences between models are modeled as learnable alignment terms, thereby achieving unified representation and alignment modeling across the cross-model semantic space. ; in, To align edge weights across models, Indicates candidate responses In the model The semantic vector below, Indicates candidate responses In the model The semantic vector below, , The L2 norm of the corresponding vector is used to improve the information representation capability and propagation stability in multi-relation graph structures.

[0055] The conflict intensity score is a quantitative indicator used to characterize the degree of semantic conflict between two candidate responses. It is derived from the comprehensive judgment results of the degree of contradiction of conclusions, semantic negation relationship and logical inconsistency during the semantic analysis process. It is used to indicate the severity of the conflict relationship. The higher the degree of conflict, the stronger the inhibition effect on the confidence state of nodes during the propagation of graph structure.

[0056] Cross-model alignment confidence refers to the degree of semantic consistency among candidate responses from different actuators or different models. It measures whether outputs from different sources express the same semantic intent. It is obtained through semantic similarity detection and source consistency verification and is used to represent the reliability of semantic alignment between cross models.

[0057] Conclusion mutual exclusion is achieved through text classification models or rule matching methods, classifying candidate responses into categories such as allow, reject, restrict, and transfer to human intervention. Different conclusion categories for the same business object are compared, and conclusions are determined to be mutually exclusive when a predefined mutual exclusion relationship exists between categories. Semantic negation detection is achieved through sentiment polarity recognition, negation word recognition, and syntactic dependency analysis techniques in natural language processing, used to identify negation expressions and their scope of action in candidate responses. Gating inhibition functions are implemented through nonlinear mapping methods, including the Sigmoid function, exponential decay function, or piecewise threshold function, used to convert conflict edge weights into inhibition coefficients, thereby reducing the influence of conflict nodes on the confidence state of the current node. Spatial projection functions are implemented through vector linear transformation or neural network mapping models, used to transform semantic vectors output by different models into a unified semantic space, thereby achieving cross-model semantic alignment.

[0058] S4.3 During the graph structure initialization phase, the original model output score of each candidate response node is normalized and mapped to a probability. The mapped probability distribution is used as the initial confidence state of the node and stored in the node attributes.

[0059] Furthermore, the design of the initial confidence state construction of the node stems from the problem of inconsistent output scoring systems of different actuators. Direct fusion would lead to the inability to compare scores from different sources on a unified scale. Therefore, the output scores of each actuator are input into the temperature scaling softmax function for probabilistic transformation, so that the scores are mapped to a unified probability space. At the same time, a model source confidence coefficient is introduced, which is preset or dynamically updated based on historical execution accuracy or task adaptability, and fused with the probability results to form the initial confidence state of the node, which is stored in the node attributes as the starting point of graph propagation.

[0060] Considering the inconsistencies in numerical scale and distribution range of output scores from different actuators, they cannot be directly used for unified confidence calculation. Therefore, the original scores need to be used as inputs for non-standardized confidence values, and the distribution smoothness is adjusted by a temperature scaling parameter to ensure that score differences are controllably expressed in the probability space. Subsequently, exponential normalization is used to map scores from different sources to a unified probability space, thereby converting the original scores into uniformly comparable confidence probabilities. Variable normalization probability mapping: ; in, For nodes The initial confidence probability, For nodes The original model output score corresponds to the candidate response. Temperature scaling parameters are used to adjust the smoothness of the distribution, establish a unified probability benchmark, and make candidates from different sources comparable and propagable.

[0061] S4.4 During the graph structure iterative update phase, the node confidence state is updated with distribution constraints according to the edge type. Specifically: for adjacent nodes connected by consistent edges, a normalized consistency influence distribution is constructed based on the edge weights, and the node confidence state is updated with probability redistribution based on the distribution; for adjacent nodes connected by conflicting edges, the conflict edge weights are mapped to gating suppression factors, and the node confidence state is updated with suppression through a nonlinear gating function; for nodes connected by cross-model aligned edges, the confidence states from different sources are projected to the same representation space using a shared embedding mapping, and the alignment fusion updated confidence state is obtained by solving the maximum consistency mapping of the aligned node confidence distribution in the space and using the Lagrange multiplier method for constraint optimization.

[0062] Furthermore, the design of introducing a path-level evaluation mechanism stems from the fact that node-level confidence propagation cannot reflect the overall structural quality of information propagation, especially when there is multi-path competition, a single node score cannot reflect the overall reliability of the propagation chain.

[0063] First, all reachable paths from the source node to the sink node are enumerated in the graph structure, and the path is represented as a sequence of nodes. Then, the confidence state of each node in the path is serialized and calculated, and a path influence model is constructed in combination with the edge type distribution in the path. Consistent edges are used to provide positive cumulative gain, conflict edges are used to introduce propagation attenuation, and alignment edges are used to correct the bias of different models. In order to solve the problem that nodes at different positions in the path have different influences, a position weight function is introduced so that the impact of conflicts in the early part of the path on the overall path quality is greater than that in the later part, thus conforming to the actual law that the propagation error gradually amplifies.

[0064] Considering that consistent edges represent semantic support relationships, it is necessary to regard neighboring nodes as a set of support information sources and model the node confidence state as the aggregation result of information from multiple support sources. By weighted aggregation of the confidence values ​​of adjacent nodes, local semantic support information is accumulated and integrated, thereby realizing the enhanced propagation of consistent information in the graph structure. This allows the node confidence to gradually increase as the support information is enhanced, and consistent edges undergo probability redistribution updates. ; in, For nodes In the The confidence state vector after rounds of iterations. For nodes The set of all consistent neighbor nodes, For consistent edge weights, For neighboring nodes In the The confidence state vector of the wheel.

[0065] Considering that conflict edges represent semantic mutual exclusion and logical contradiction, they cannot participate in positive information propagation. Instead, they should act as negative constraint signals on the node's confidence state. By mapping the conflict intensity to a gating suppression factor, the node's current confidence state is nonlinearly compressed, thereby reducing the misleading influence of the conflict source node on the current node. This achieves a conflict information propagation suppression mechanism, and the conflict edge suppression update is implemented as follows: ; in, , These represent the confidence states of node i before and after the update, respectively. For the sigmoid function, The weights are for conflict edges.

[0066] Considering that different model outputs have spatial differences in semantic representation but the target semantics are consistent, it is necessary to map the semantic representations from different sources to a unified fusion space and achieve semantic information fusion through weighted interpolation. The fusion coefficients are used to control the contribution ratio of different source models to the final semantic representation, thereby achieving smooth integration of cross-model semantic information and edge fusion update: ; in, To align the merged nodes The confidence state, , The models are respectively Next node ,Model Next node The semantic vector controls the contribution ratio of different sources.

[0067] Considering that the cross-model semantic alignment problem not only needs to maximize semantic consistency but also satisfy graph structure propagation constraints, the semantic alignment process needs to be transformed into a constraint optimization problem. The consistency term characterizes the degree of semantic matching, while the constraint term ensures the stability of graph structure propagation. By introducing Lagrange multipliers to embed the constraints into the objective function, a joint optimization of consistency maximization and structural constraints is achieved. ; in, Let Lagrange loss function be used. To align edge weights, , The confidence representation of the nodes to be optimized is given by the first term, which is a weighted sum of the squares of the Euclidean distances between aligned nodes, used to maximize consistency. The regularization coefficient is . To constrain the regularization term and ensure the stability of graph structure propagation, the confidence of local nodes is extended to path-level overall quality assessment, thus elevating the evaluation object from "points" to "propagation chains".

[0068] The node confidence state refers to the degree of credibility of each candidate response in the current propagation round. It is used to measure the relative reliability of the candidate response in the overall candidate set. It is derived from the unified and standardized processing result of the original model output score and is continuously updated during the graph structure propagation process. It is used to reflect the dynamic credibility change of the candidate response after multi-source information interaction.

[0069] Path enumeration is implemented using graph search algorithms, including depth-first search or breadth-first search methods, traversing all reachable paths within a preset path length range; the position weight function is implemented by mapping the indices of nodes or edges in the path, and assigning different weights to nodes based on their positions in the path using a decreasing function or a piecewise function.

[0070] S4.5 In each round of propagation, the propagation path between nodes is serialized and recorded, and the conflict edges in the path are labeled segment by segment. At the same time, a path position decay function is introduced to assign different influence weights to conflict edges at different positions, so that the path contamination coefficient is jointly determined by the conflict edge weights and the position decay function.

[0071] Furthermore, the design of applying the path evaluation results in reverse to the edge weights of the graph structure stems from the fact that traditional graph models only have unidirectional propagation capabilities and cannot adaptively optimize the structure based on the overall path quality, thus leading to the problem of persistent erroneous propagation paths.

[0072] First, the path quality score is calculated based on the path evaluation results, and the score is decomposed to each edge involved in the path. For consistent edges in high-quality paths, weight enhancement is performed to strengthen reliable propagation channels. For conflicting edges in low-quality paths, weight decay is performed to reduce the impact of error propagation. For cross-model aligned edges, dynamic adjustments are made based on the degree of path consistency.

[0073] Considering the cumulative amplification effect of conflict information in the propagation path, and the varying impacts of conflicts at different locations on the final output, it is necessary to treat conflict edges in the path as pollution sources, conflict intensity as a cumulative variable, and introduce a location attenuation function to simulate the diminishing effect of propagation distance on influence intensity. This transforms path-level conflict information into an overall pollution energy index, with the path pollution coefficient determined jointly by the conflict edge weights and the location attenuation function. ; in, For path The pollution coefficient, Representing a path The set of edges, For the conflict side The weight, The position decay constant, For the conflict side Position index in the path.

[0074] Considering the gradual attenuation of information during multi-hop propagation, conflicting information closer to the beginning of the path has a greater impact on the final result, while the impact of information further down the path gradually weakens. Therefore, an exponential decay function is used to model the propagation distance, making the information impact decrease non-linearly with the path length. This better reflects the actual propagation error diffusion pattern. The position decay function is as follows: ; in, For the edge Position attenuation factor, To unify the attenuation constant, For the edge The number of propagation steps from the source node is updated after each round of propagation, enabling the edge weights of the graph structure to change dynamically with the path quality distribution, thus forming an adaptive evolution mechanism for the graph structure and realizing the transformation of the graph structure from static modeling to dynamic optimization based on path feedback.

[0075] The path quality score is obtained by statistically calculating the number of consistent edges, the number of conflicting edges, and the confidence status of nodes in the path, and is comprehensively evaluated in combination with the path pollution coefficient.

[0076] S4.6 When the path pollution coefficient exceeds the preset threshold, the path is not deleted as a whole. Instead, the decay node is determined based on the position index of the conflict edge in the path. The exponential decay function is used to reduce the weight of the conflict edge along the path direction, and the reduced weight is redistributed to the adjacent consistent edges on the path according to the consistency contribution ratio.

[0077] Furthermore, to avoid the loss of effective information due to the direct deletion of highly polluted paths, and to address the problem of unrepairable local conflicts, a path structure-level local repair mechanism needs to be introduced. When the path pollution coefficient exceeds a threshold, path deletion is not performed. Instead, attenuation nodes are determined based on the position index of the conflicting edges in the path. Subsequently, an exponential decay function is used to gradually reduce the weight of the conflicting edges along the path propagation direction, and the weakened weights are redistributed to adjacent consistent edges according to the consistency contribution ratio. This reduces the impact of conflict propagation while preserving the path structure, achieving path-level local repair and stabilization.

[0078] Considering that path-level contamination has a holistic impact on the propagation quality of the graph structure, the path contamination intensity needs to be used as a global modulation factor and applied to the weight of each conflicting edge in the path. This is achieved through exponential nonlinear suppression, thereby weakening the conflicting edges in highly contaminated paths as a whole during propagation. The exponential decay function reduces the conflicting edge weights edge by edge: ; in, The weights of the conflict edges after reduction. The original conflict edge weights, To unify the attenuation constant, The pollution intensity of the path.

[0079] Considering that the weakened conflict edge weights are not information loss but rather suppressed semantic contributions, it is necessary to redistribute the reduced conflict energy to consistent edges according to the proportion of contribution to consistent paths, while maintaining the conservation of total information. This will achieve the redistribution of conflict information to effective semantic support paths and to consistent edges. ; in, The weights of the new consistent edges after redistribution. Original consistent edge weights To assign a proportional factor, it is typically based on the contribution percentage of consistent edges in the path. This represents the total weight increment reduced from the conflict edge.

[0080] The consistency contribution ratio is calculated by the weight ratio of each consistent edge in the statistical path and the confidence status of the connected nodes, and is used to reflect the relative contribution of each consistent edge in the path.

[0081] S5. Perform compliance rule verification on the fused response. When a response that violates the constraint is detected, locate the violation constraint to the corresponding conflict relationship and drive the constraint structure update. At the same time, re-trigger the confidence correction and conflict resolution fusion process, and take the response result that meets the compliance constraint as the final output.

[0082] S5.1 In the graph structure, depth-first search is used to enumerate the paths for each candidate response node, obtain multiple propagation paths from the source node to the sink node, calculate the confidence value of the nodes on each propagation path using a segmented recursive fusion method, divide the complete path into several semantically consistent sub-segments, perform confidence aggregation within the sub-segments, and introduce consistency constraint functions between the sub-segments for recursive correction.

[0083] Furthermore, the design of path-level parsing for the fused response stems from the fact that compliance violations usually originate from local propagation path anomalies rather than overall graph structure errors. Therefore, it is necessary to back-map the output results to the propagation path for localization, map the fused response text back to the candidate nodes in the graph structure, and extract the corresponding propagation path set. Subsequently, the path is segmented into semantically consistent sub-segments, and the confidence contribution value of each sub-segment is calculated to identify the propagation substructure that has the greatest impact on the final output.

[0084] Considering that the propagation path consists of multiple semantically continuous substructures, and different substructures contribute differently to the final semantics, it is necessary to divide the path into semantically consistent segments and perform local confidence aggregation on each segment. Simultaneously, segment weights are introduced to characterize the importance of different semantic units, thereby achieving hierarchical fusion of path-level semantic information and segmented recursive fusion. ; in, The fusion confidence representation for the path, The total number of segments into which the path is divided. For sub-field indexing, for The weight of each sub-segment For the first The representation of each segment after internal confidence aggregation establishes a traceable mapping relationship from the output result to the graph structure path.

[0085] Semantically consistent segment division is achieved by calculating the semantic similarity between adjacent nodes and judging the continuity of edge types. When the semantic similarity between adjacent nodes is higher than the threshold and the edge types are consistent, they are divided into the same segment.

[0086] S5.2. The path pollution intensity is obtained by calculating the weighted sum of all conflicting edges on the path after position decay. When the path pollution intensity exceeds the preset threshold, the path is removed and the remaining paths participate in the subsequent fusion calculation.

[0087] Furthermore, the design of identifying violation propagation chains through path contamination intensity stems from the cumulative amplification effect of conflict information in the propagation path, which can easily lead to local errors being amplified into global violations at the end of the path. The path contamination intensity value is obtained by combining the conflict edge weights in the path with the position decay function. When the value exceeds the preset threshold, the path is not directly deleted, but the path structure is retained as the analysis object and marked as a high-risk propagation chain for subsequent structural reconstruction input.

[0088] Considering that conflict information in the path gradually accumulates and amplifies during propagation, and that conflicts at different locations have different impacts on the final output, it is necessary to treat the conflict edge weights as pollution source terms and modulate the impact using a location attenuation function to obtain an overall path pollution intensity index. Path pollution intensity: ; in, For path pollution intensity, Let be the weights of the conflicting edges in the path. To unify the attenuation constant, This provides an index for the position of conflict edges in the path, enabling the preservation and location of violation propagation chains without losing information.

[0089] High-risk transmission chains are determined by a combination of path contamination intensity and conflict edge density. When the proportion of conflict edges in a path exceeds a preset ratio or the contamination intensity exceeds a threshold, it is identified as a high-risk transmission chain.

[0090] S5.3. The path confidence value of the retained path is normalized based on the path length, and the path is serialized. The path pollution intensity is calculated based on the location of the conflict edge distribution, and the location-related attenuation coefficient is generated through an exponential gating function. The attenuation coefficient is applied to the path confidence propagation process to perform differentiated modulation on different path segments, and finally normalized to generate the path weight distribution under global constraints.

[0091] Furthermore, the design of introducing a path competition graph and a compliance-driven structural reconstruction mechanism stems from the fact that relying solely on rule filtering cannot correct deep-propagation structural deviations. Therefore, it is necessary to achieve graph structure correction through path-level structural feedback. The retained paths are used as nodes to construct a path competition graph, and path-level competition edges are established based on the proportion of shared nodes and conflict relationships between paths. Subsequently, consistency propagation calculations are performed on the path competition graph to obtain path contribution coefficients, which are used to measure the degree of influence of each path on the final output. When the contribution coefficient of a non-compliant path is detected to exceed a threshold, structural reconstruction operations are performed on the original subgraph mapped by the path, including disconnecting conflicting edges, reconstructing alternative edges based on semantic similarity, and adjusting the weights of cross-model alignment edges. The reconstruction process adopts a local priority strategy to reduce disturbances to the stability of the global structure.

[0092] Considering the varying contributions of different propagation paths to the final output, it is necessary to map the degree of path contamination to a path reliability penalty term and convert the path quality into a probability distribution using exponential normalization. This achieves a unified quantitative expression of the contributions of different paths, generating a normalized path weight distribution. ; in, For path Normalized weights For path The intensity of pollution, The scaling factor controls the sensitivity of the impact of pollution intensity on the weights, realizing a closed-loop optimization mechanism from path-level anomaly detection to graph structure repair, enabling the system to maintain semantic consistency and propagation stability while meeting compliance constraints.

[0093] S5.4 Introduce path competition relationship modeling to the set of retained paths, construct a path competition graph by constructing the structural conflicts and consistency relationships between paths, and perform consistency propagation calculation on the path competition graph to generate path contribution coefficients.

[0094] Furthermore, since confidence calculation for a single path cannot reflect the mutual influence between paths, it can easily lead to different inference paths being independent of each other and lacking global constraints. Therefore, it is necessary to introduce a competition modeling mechanism between paths to characterize the cooperative and conflict relationships between paths. First, each retained path is abstracted into a path node, and the structural similarity and conflict degree between paths are calculated based on the composition of nodes in the path and the degree of semantic overlap. When there is high semantic overlap between paths, they are judged as a consistent relationship; when there is a conflict of conclusions or opposition of key nodes, they are judged as a conflict relationship. On this basis, a path competition graph is constructed, where nodes represent paths, edges represent the consistent or conflict relationship between paths, and positive weights are assigned to edges with consistent relationships, while suppression weights are assigned to edges with conflict relationships. Then, consistency propagation calculation is performed on the path competition graph, and the contribution of path nodes is diffused and updated iteratively to obtain the path contribution coefficient of each path in the global structure, which is used for subsequent path weight allocation.

[0095] Considering that different propagation paths are competitive and mutually influential, rather than independent, it is necessary to abstract the paths as nodes in a graph structure, and model the structural similarities and conflicts between paths as edges, thereby constructing a path-level competition network to characterize the global game relationship between paths. Path competition graph: ; in, This is a path competition graph. This is a set of path nodes, where each node represents a reserved path. Edge set.

[0096] Considering that path contribution is not a static value but is influenced by other path structures, iterative information propagation is needed on the path competition graph. Dynamic adjustment of path weights is achieved through the interaction of contributions from neighboring paths, resulting in a globally consistent optimized path contribution distribution. Path consistency propagation: ; in, Path node In the Contribution coefficient after round of iterations Path node and The boundary rights between them Path node In the The contribution coefficient of the wheel.

[0097] The contribution coefficient of the path competition graph refers to the importance of each propagation path in the overall path set in the path competition graph structure. It is obtained by comprehensively analyzing the mutual support and conflict relationships between paths and is used to measure the relative influence of a certain path among all candidate propagation paths. The path competition graph establishes edge relationships by calculating the node overlap rate, semantic similarity and conclusion consistency between paths, and generates a path-level relationship network using graph construction methods.

[0098] S5.5 Calculate the variance between paths based on the path weight distribution and calculate the gating coefficient in combination with the path pollution intensity, and use the gating coefficient as an adjustment factor to perform secondary scaling and updating of the path weights.

[0099] Furthermore, the stability between paths depends not only on the average weight but also on the dispersion of path distribution and the intensity of conflict. Therefore, linear weighting alone cannot accurately reflect the differences in path reliability. A joint adjustment mechanism of statistical variance and contamination intensity needs to be introduced. First, the weight distribution of all retained paths is statistically analyzed to calculate the mean of path weights and the variance between paths to characterize the dispersion of path distribution. At the same time, combined with the path contamination intensity obtained in S5.4, a weighted cumulative calculation is performed on areas with dense conflict edges. Then, the path variance and path contamination intensity are input into a gating function model, such as a sigmoid or normalized exponential function, to generate path-level gating coefficients. Finally, the gating coefficients are used as scaling factors on the original path weights to perform a secondary adjustment and update of the path weights. This suppresses the weights of high-conflict or highly discrete paths while enhancing the weights of stable paths, thereby improving the robustness of the overall fusion result.

[0100] Considering that path stability is affected by both the dispersion of weight distribution and the degree of conflict contamination, it is necessary to use statistical variance as an indicator of distribution instability and path contamination intensity as an indicator of structural conflict. These two factors are then fused using a nonlinear function to obtain the path reliability gating coefficient. The gating coefficient is as follows: ; Considering that path weights reflect the overall contribution of the path to the output, and that a path is essentially composed of a set of edges, it is necessary to construct a path-edge association matrix to decompose and map path-level weights to the edge-level structure. This enables backpropagation updates from path-level evaluation to graph-structure edge weights, and a second-order scaling update. ; in, The path weights are the results of the second scaling. The gating coefficient, This represents the original path weight.

[0101] The path structure complexity index is used to characterize the comprehensive features of a propagation path in terms of structural connection complexity and semantic change frequency. Structural connection complexity describes the density of connections between nodes in the path; semantic change frequency describes the frequency of semantic consistency changes between adjacent nodes in the path, and is used to reflect whether there are too many jumps or semantic instability in the path during propagation.

[0102] The path quality correction factor is an adjustment parameter generated based on the path structure complexity index. It is used to adjust the contribution of the path in the propagation process and reduce the influence of paths with high structural complexity in confidence propagation, thereby avoiding excessive influence of long paths or paths with frequent semantic jumps on the final fusion result. The gating coefficient is calculated through a nonlinear function, including the Sigmoid function, exponential function, or normalization function, and is used to map the path variance and path contamination intensity into adjustment coefficients.

[0103] S5.6. The path weight distribution is back-mapped to the graph structure edge set. Consistent edges are redistributed based on the path contribution ratio through multiplicative enhancement. Conflicting edges are weakened based on the path contamination contribution through an exponential decay function. Cross-model aligned edges are adjusted by mixing the path consistency contribution ratio with the original weight of the aligned edge through a nonlinear gating function to obtain updated edge weights. When the path weight fluctuates below a preset threshold in continuous iterations, a convergence judgment is triggered. Otherwise, the edge weights of the subgraph in the dense conflict edge region are re-initialized and the propagation calculation process is restarted.

[0104] Furthermore, updating weights at the path level alone cannot directly affect the graph structure propagation process. Therefore, it is necessary to establish a reverse feedback mechanism from path-level evaluation results to graph structure edge weights in order to achieve closed-loop optimization between paths and graphs. Considering that different types of edges play different roles in semantic propagation, uniform updates may lead to semantic shifts. Therefore, differentiated update strategies are needed for different edge types.

[0105] First, after path-level computation is completed, an association mapping relationship is established between the path and the edge set in the graph. This is achieved by recording the node sequence and edge index of each path, thus clarifying the contribution ratio of each path to each edge in the graph. Then, for consistent edges, a multiplicative enhancement update is performed based on the path contribution coefficient, that is, the path contribution ratio is multiplied by the original edge weight as an amplification factor to enhance the consistency propagation capability of high-contribution paths. For conflict edges, an attenuation coefficient is calculated based on the path pollution contribution intensity, and an exponential attenuation function is used to nonlinearly weaken the conflict edge weight to reduce the impact of conflict information in graph propagation. For cross-model alignment edges, the path consistency contribution ratio and the original alignment edge weight are considered comprehensively, and a nonlinear gating function is used for fusion adjustment to balance the semantic alignment intensity between different models.

[0106] To prevent graph structures from oscillating or failing to converge during iteration, path weight changes are continuously monitored. By calculating the difference in path weight distribution between adjacent iterations, a convergence judgment is triggered when the difference falls below a preset threshold. If the convergence condition is not met, a local edge weight re-initialization operation is performed on the subgraph in the region with dense conflict edges. That is, the local edge weights are recalculated based on the semantic similarity of the current nodes to avoid the accumulation of local extreme weights that could lead to propagation offset. After re-initialization, the graph propagation and path calculation process is re-entered, thus forming a path feedback-driven adaptive optimization closed-loop mechanism for graph structures.

[0107] Considering that the graph structure gradually tends to a stable state during iterative propagation, it is necessary to determine whether the system has converged by calculating the degree of change in global edge weights in adjacent iterations. When the change is below a threshold, it indicates that the structure has reached a stable equilibrium state, and the path weights are mapped back to the graph structure. ; in, This represents the updated weight vector of a given edge. This is the path-edge incidence matrix, representing the path Includes edges The proportion of contribution, Let k be the weight of path k.

[0108] Considering that the graph structure gradually tends to a stable state during iterative propagation, it is necessary to determine whether the system has converged by calculating the degree of change in global edge weights in adjacent iterations. When the change is below a threshold, it indicates that the structure has reached a stable equilibrium state, and convergence is determined as follows: ; in, , The first Wheel and First The path weight vector of the wheel, This is the convergence threshold.

[0109] The path contribution ratio refers to the relative proportion of the impact of a certain path on the final fusion result during the overall propagation process. It is determined by the degree of participation of each edge in the path during the propagation process and is used to represent the importance weight among all candidate paths.

[0110] The path-edge relationship is established by recording the edge indices traversed during the path traversal, which is used to establish the correspondence between the path and the edge of the graph structure.

[0111] S5.7. The fusion response results are matched with rules and checked for logical consistency through a preset compliance rule base. When illegal content is detected, the candidate nodes that trigger the illegality and the associated conflict relationships are located.

[0112] Furthermore, since the multi-model fusion process may introduce output results that do not conform to business rules or logical constraints, an explicit compliance verification mechanism is needed to ensure the security and consistency of the system output. First, the fused response results are input into a preset compliance rule base, which includes business constraint rules, logical consistency rules, and prohibited expression rules. The system uses a rule matching engine to check each rule of the candidate responses and combines the node attribute information in the graph structure to locate the source of the violation. When a violation is detected, not only is the violation text marked, but the corresponding candidate nodes and conflicting edges in the graph structure are also located, thereby enabling traceable analysis of the violation propagation path and providing a basis for subsequent structural repair.

[0113] Considering that compliance constraints are essentially legal semantic set constraints, it is necessary to map the output to the legal set defined by the rule, and determine whether the output meets the constraint conditions by judging the set members. Rule matching: ; in, For compliance determination results, 1 represents compliance and 0 represents non-compliance. To respond to fusion Key semantic elements extracted from the text A set of legal semantics defined for the compliance rule base.

[0114] Violation location strength is an indicator used to measure the role of a candidate node in the formation of a violation result. It is determined by analyzing the degree of participation of the node in the propagation path and the degree of influence on the final output result. It is used to locate the key nodes and key propagation paths that lead to compliance violations.

[0115] S5.8 When a compliance violation is detected in the fusion response result, the candidate nodes and associated conflict subgraphs corresponding to the violation are located, and a structural reconstruction operation is performed on the conflict subgraph, including disconnecting the conflict edge that triggered the violation, reconstructing alternative connection edges based on the semantic similarity between candidate nodes, updating the path weight allocation associated with the conflict edge, correcting the affected propagation paths in the global graph, and re-executing the confidence propagation and fusion calculation process on the reconstructed graph structure.

[0116] Furthermore, when violations or inconsistencies are detected, simply deleting or filtering them would result in information loss. Therefore, it is necessary to achieve partial repair through structural reconstruction to retain as much valid information as possible while ensuring compliance.

[0117] Based on the candidate nodes for violations and the associated conflict subgraphs, a structural reconstruction operation is performed on the subgraph. This includes disconnecting the conflicting edges that trigger the violation to eliminate the violation propagation path; simultaneously, based on the semantic vector similarity calculation results of the nodes, alternative connection edges are regenerated under the condition of meeting the threshold to restore reasonable semantic relationships; then, the weight distribution relationship of the affected paths is recalculated, and the path propagation structure is locally updated; the propagation path state related to the subgraph in the global graph is corrected, and the confidence propagation and fusion calculation process is re-executed on the reconstructed graph structure, so that the system reconverges to a stable output state that meets compliance constraints after structural repair.

[0118] Considering that breaking conflicting edges will disrupt the connectivity of the graph structure, it is necessary to reconstruct alternative connections based on the semantic similarity of nodes to restore the continuity of the semantic propagation path. Alternative edges are generated as follows: ; in, The weights of the newly created replacement edges after reconstruction. , They are nodes , The semantic vector; the numerator is the inner product of vectors, and the denominator is the L2 norm product.

[0119] Considering that the graph structure changes after reconstruction, it is necessary to re-execute the propagation computation to update the node confidence state. Therefore, a graph propagation operator F is defined to re-iterate the computation on the reconstructed graph structure, thereby restoring the stable propagation state, and then propagating again after reconstruction: ; in, To reconstruct the diagram structure The node confidence state matrix of the round, in the reconstructed graph Perform a round of confidence propagation update. Given the confidence state of the previous moment, after the structural reconstruction is completed, the updated graph structure is re-input into the confidence propagation calculation process, and an iterative convergence judgment mechanism is introduced. The system is judged to have reached a stable state by calculating the change in the confidence state of nodes in adjacent iteration rounds. When the change is lower than a preset threshold, the system is judged to have converged; otherwise, the propagation and fusion update process continues. At the same time, the structural change log is recorded during the reconstruction process for backtracking analysis of conflict propagation paths, thereby forming a compliance constraint-driven closed-loop optimization mechanism for graph structures.

[0120] Alternate connection edge weights refer to the strength of new connection relationships generated based on the semantic similarity of nodes during graph structure reconstruction. They are used to represent the proximity of two nodes in the semantic space and to restore the semantic connectivity lost due to the disconnection of conflict edges, thereby ensuring the continuous and stable propagation capability of the graph structure after repair. The rule matching engine is implemented through keyword matching, entity matching, and logical rule matching methods to detect whether candidate responses meet preset business rules.

[0121] This embodiment also provides a multi-model collaborative customer service intelligent response system based on a layered architecture, including: an intent parsing module, a heterogeneous execution module, a conflict mapping module, a confidence fusion module, and a compliance verification module; The intent parsing module is used to perform semantic parsing processing on user input, extract intent categories and business constraint information, and form a structured intent representation; The heterogeneous execution module is used to trigger at least two logical heterogeneous executors to run in parallel based on the structured intent representation. Each executor outputs a corresponding set of candidate responses and attaches an initial confidence label to each candidate response to form a cross-model candidate pool. The conflict mapping module is used to perform semantic consistency comparison on any candidate response in the candidate pool, identify semantic conflict relationships between candidate responses, and map the conflict relationships into a constraint structure. The confidence fusion module is used to make consistent adjustments to the initial confidence of each candidate response under the constraint propagation effect of the constraint structure, introduce a conflict suppression mechanism to weaken the weight of candidate responses with semantic contradictions, and at the same time enhance the confidence transmission strength on the semantically consistent path, forming a calibrated confidence distribution modulated by the structural constraints, and performing structured screening and fusion decision of candidate responses based on the calibrated confidence distribution, and outputting the fusion response result. The compliance verification module is used to perform compliance rule verification on the fused response. When a response that violates the constraints is detected, the violation constraint is located to the corresponding conflict relationship and the constraint structure is updated. At the same time, the confidence correction and conflict resolution fusion process is re-triggered, and the response result that meets the compliance constraints is taken as the final output.

[0122] This embodiment also provides a computer device applicable to the multi-model collaborative customer service intelligent response method based on a hierarchical architecture, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the multi-model collaborative customer service intelligent response method based on a hierarchical architecture as proposed in the above embodiment.

[0123] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0124] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the multi-model collaborative customer service intelligent response method based on a hierarchical architecture as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0125] In summary, this invention achieves precise computational power allocation based on task attributes and parallel acquisition of multi-source information by triggering at least two logically heterogeneous executors to run in parallel based on structured intent representation, outputting a set of candidate responses and attaching an initial confidence level label to form a cross-model candidate pool. This allows models of different scales and specializations to work independently within their respective "comfort zones," breaking the bottleneck of a single model handling all requests, effectively avoiding infinite loops and significantly improving the processing efficiency of complex mixed intents. By performing semantic consistency comparison on any candidate response in the candidate pool, semantic conflict relationships are identified and mapped to constraint structures. This enables an explicit and topological expression of the "information source heterogeneity" and "factual conflict" problems that inevitably arise when multiple models are generated in parallel, providing a mathematical and logical basis for subsequent conflict resolution and enabling the system to have the prerequisites for fact-checking and logical self-consistency.

[0126] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multi-model collaborative intelligent customer service response method based on a hierarchical architecture, characterized in that: include: Perform semantic parsing on user input, extract intent categories and business constraint information, and form a structured intent representation; Based on the structured intent representation, at least two logically heterogeneous executors are triggered to run in parallel. Each executor outputs a corresponding set of candidate responses and attaches an initial confidence label to each candidate response to form a cross-model candidate pool. Perform semantic consistency comparison on any candidate response in the candidate pool, identify semantic conflict relationships between candidate responses, and map the conflict relationships into constraint structures; Under the constraint propagation effect of the constraint structure, the initial confidence of each candidate response is adjusted for consistency. A conflict suppression mechanism is introduced to weaken the weight of candidate responses with semantic contradictions, while enhancing the confidence transmission strength on the semantically consistent path. This forms a calibrated confidence distribution modulated by the structural constraints. Based on the calibrated confidence distribution, the structured screening and fusion decision of candidate responses are performed, and the fusion response result is output. The fused response is subjected to compliance rule verification. When a response that violates the constraints is detected, the violation constraint is located to the corresponding conflict relationship and the constraint structure is updated. At the same time, the confidence correction and conflict resolution fusion process is re-triggered, and the response result that meets the compliance constraints is taken as the final output.

2. The multi-model collaborative intelligent customer service response method based on a hierarchical architecture as described in claim 1, characterized in that: The structured intent representation includes: The system performs text standardization and noise removal on user input, and encodes the input text based on a pre-trained semantic understanding model, outputting the corresponding intent probability distribution vector. The intent label with the highest probability is selected on the probability distribution vector. The intent label is then compared with the intent identifier field of each rule in the business rule base. When the comparison is consistent, the business restriction field in the rule is extracted using the rule parser. The intent label is filled into the intent field and the business restriction field is filled into the constraint field using a predefined structured template to obtain a unified data structure.

3. The multi-model collaborative customer service intelligent response method based on a hierarchical architecture as described in claim 2, characterized in that: The cross-model candidate pool includes: By mapping intent tags to execution path identifiers in the policy routing table, the retrieval execution path is activated when the found entry contains a retrieval identifier, and the inference execution path is activated when it contains an inference identifier. The retrieval execution path generates candidate responses and outputs corresponding confidence scores based on knowledge base vector retrieval, while the inference execution path generates candidate responses and outputs corresponding confidence scores based on multi-step tool calls and logical deduction. Candidate responses and confidence scores from different execution paths are uniformly formatted, encapsulated, and aggregated for storage, ensuring that candidate responses from different sources retain their source identifiers and initial confidence attributes within the same candidate pool, thereby forming a cross-model candidate set for the model.

4. The multi-model collaborative intelligent customer service response method based on a hierarchical architecture as described in claim 3, characterized in that: The constraint structure includes: Semantic vector encoding is performed on each candidate response in the candidate pool, and a similarity matrix is ​​constructed based on the similarity between the vectors. Candidate pairs below a preset threshold are marked as conflicting, while highly similar candidate pairs retain consistent association relationships. By using the conflict marking relationship and the consistency association relationship, the candidate responses are represented as nodes in the graph structure, and the consistency relationship and the conflict relationship are mapped to different types of edges, thereby forming a structured graph model to describe the constraints and competition relationships between candidate responses.

5. The multi-model collaborative customer service intelligent response method based on a hierarchical architecture as described in claim 4, characterized in that: The fusion response results include: The candidate response set is vector-encoded to obtain the corresponding semantic vector, and the cosine similarity matrix between any two candidate response vectors is calculated. For candidate response pairs with similarity greater than a preset threshold, undirected consistent edges are established in the graph structure with the similarity value as the edge weight. Perform pairwise semantic contradiction judgment between candidate responses. When mutually exclusive conclusions or semantic negation are detected, establish conflict edges between corresponding nodes and use the conflict intensity score as the edge weight. Perform text alignment matching on candidate responses output by different models. Establish cross-model alignment edges between successfully matched candidate response nodes and use the matching confidence score as the edge weight. During the graph structure initialization phase, the original model output score of each candidate response node is normalized and probability-mapped. The mapped probability distribution is used as the initial confidence state of the node and stored in the node attributes. During the graph structure iterative update phase, the node confidence states are updated with distribution constraints according to edge types. Specifically: for adjacent nodes connected by consistent edges, a normalized consistency influence distribution is constructed based on edge weights, and the node confidence states are updated with probability redistribution based on this distribution; for adjacent nodes connected by conflicting edges, conflict edge weights are mapped to gating suppression factors, and the node confidence states are updated with suppression through a nonlinear gating function; for nodes connected by cross-model aligned edges, the confidence states from different sources are projected to the same representation space using a shared embedding mapping. The alignment fusion update of the confidence states is obtained by solving the maximum consistency mapping of the aligned node confidence distribution in the space and using the Lagrange multiplier method for constraint optimization. In each round of propagation, the propagation path between nodes is serialized and recorded, and the conflict edges in the path are labeled segment by segment. At the same time, a path position decay function is introduced to assign different influence weights to conflict edges at different positions, so that the path pollution coefficient is jointly determined by the conflict edge weights and the position decay function. When the path pollution coefficient exceeds the preset threshold, the path is not deleted as a whole. Instead, the decay node is determined based on the position index of the conflict edge in the path. An exponential decay function is used to reduce the weight of the conflict edge along the path direction, and the reduced weight is redistributed to the adjacent consistent edges on the path according to the consistency contribution ratio.

6. The multi-model collaborative customer service intelligent response method based on a hierarchical architecture as described in claim 5, characterized in that: The compliance rule verification includes: In the graph structure, a depth-first search is used to enumerate the paths for each candidate response node, obtaining multiple propagation paths from the source node to the sink node. The confidence values ​​of the nodes on each propagation path are calculated using a segmented recursive fusion method. The complete path is divided into several semantically consistent sub-segments. Confidence aggregation is performed within the sub-segments, and a consistency constraint function is introduced between the sub-segments for recursive correction. The path pollution intensity is obtained by calculating the weighted sum of all conflicting edges on the path after position decay. When the path pollution intensity exceeds the preset threshold, the path is removed and the remaining paths participate in the subsequent fusion calculation. The path confidence values ​​of the retained paths are normalized based on the path length, and the paths are serialized. The path pollution intensity is calculated based on the location of conflict edges, and a location-related attenuation coefficient is generated through an exponential gating function. The attenuation coefficient is applied to the path confidence propagation process to perform differentiated modulation on different path segments, and finally normalized to generate the path weight distribution under global constraints. A path competition relationship model is introduced for the set of retained paths. The structural conflicts and consistency relationships between paths are constructed into a path competition graph, and consistency propagation calculation is performed on the path competition graph to generate path contribution coefficients. The variance between paths is calculated based on the path weight distribution, and the gating coefficient is calculated in combination with the path pollution intensity. The gating coefficient is then used as an adjustment factor to perform a secondary scaling update on the path weights. The path weight distribution is back-mapped to the graph structure edge set. Consistent edges are redistributed based on the path contribution ratio through multiplicative enhancement. Conflicting edges are weakened based on the path contamination contribution through an exponential decay function. Cross-model aligned edges are adjusted by mixing the path consistency contribution ratio with the original weight of the aligned edge through a nonlinear gating function to obtain updated edge weights. When the path weight fluctuates below a preset threshold in continuous iterations, a convergence judgment is triggered. Otherwise, the edge weights of the subgraph in the region with dense conflict edges are re-initialized and the propagation calculation process is restarted.

7. The multi-model collaborative customer service intelligent response method based on a hierarchical architecture as described in claim 6, characterized in that: The adaptive structural update includes: By using a pre-set compliance rule base, the fusion response results are matched with rules and checked for logical consistency. When violations are detected, the candidate nodes that trigger the violations and the associated conflict relationships are located. When a compliance violation is detected in the fusion response result, the candidate nodes and associated conflict subgraphs corresponding to the violation are located, and a structural reconstruction operation is performed on the conflict subgraph, including disconnecting the conflict edge that triggered the violation, reconstructing alternative connection edges based on the semantic similarity between candidate nodes, updating the path weight allocation associated with the conflict edge, correcting the affected propagation paths in the global graph, and re-executing the confidence propagation and fusion calculation process on the reconstructed graph structure.

8. A multi-model collaborative intelligent customer service response system based on a hierarchical architecture, based on the multi-model collaborative intelligent customer service response method based on a hierarchical architecture as described in any one of claims 1 to 7, characterized in that: This includes an intent parsing module, a heterogeneous execution module, a conflict mapping module, a confidence fusion module, and a compliance verification module; The intent parsing module is used to perform semantic parsing processing on user input, extract intent categories and business constraint information, and form a structured intent representation; The heterogeneous execution module is used to trigger at least two logical heterogeneous executors to run in parallel based on the structured intent representation. Each executor outputs a corresponding set of candidate responses and attaches an initial confidence label to each candidate response to form a cross-model candidate pool. The conflict mapping module is used to perform semantic consistency comparison on any candidate response in the candidate pool, identify semantic conflict relationships between candidate responses, and map the conflict relationships into a constraint structure. The confidence fusion module is used to adjust the initial confidence of each candidate response in a consistent manner under the constraint propagation effect of the constraint structure, introduce a conflict suppression mechanism to weaken the weight of candidate responses with semantic contradictions, enhance the confidence transmission strength on the semantically consistent path, form a calibrated confidence distribution modulated by the structural constraints, and perform structured screening and fusion decision of candidate responses based on the calibrated confidence distribution, and output the fusion response result. The compliance verification module is used to perform compliance rule verification on the fused response. When a response that violates the constraints is detected, the violation constraint is located to the corresponding conflict relationship and the constraint structure is updated. At the same time, the confidence correction and conflict resolution fusion process is re-triggered, and the response result that meets the compliance constraints is taken as the final output.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the multi-model collaborative customer service intelligent response method based on a hierarchical architecture as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the multi-model collaborative customer service intelligent response method based on a hierarchical architecture as described in any one of claims 1 to 7.

Images