Multi-agent cooperative workflow construction method for English customs consultation

Abstract

The invention relates to the field of English customs consultation, and discloses an English customs consultation-oriented multi-agent collaborative workflow construction method, which comprises the following steps of constructing a four-dimensional dynamic hypergraph; converting the four-dimensional dynamic hypergraph into tensor representation, performing tensor decomposition, and identifying key agent role nodes under a high weight dimension; constructing an alliance value function and carrying out optimization solution; respectively representing Chinese and English input by adopting an encoder, generating a semantic representation vector, and obtaining a structured semantic fusion result; and carrying out incremental updating on system weight parameters to realize model synchronization of dynamic policy contents. According to the invention, a multi-agent cooperation mechanism and a dynamic hypergraph construction technology are adopted, and real-time information integration in English customs affair consultation is realized. According to the technology, the accuracy of policy interpretation is remarkably improved, and compared with an existing information access information processing technology based on a static database, the problem of information isolated island is solved, and customer requirements are responded in time.

Description

Technical Field

[0001] The present invention relates to the technical field of English customs consultation, and in particular to a multi-agent collaborative workflow construction method for English customs consultation. Background Art

[0002] With the rapid development of global trade, the demand for English-language customs consulting is growing. Faced with complex and ever-changing customs policies, businesses urgently need timely and accurate information to ensure compliance and mitigate customs clearance risks. In this context, the effective integration and application of diverse information resources has become a crucial research topic.

[0003] Against this backdrop, existing technologies typically rely on static databases and simple information retrieval systems. These systems primarily rely on manual data entry and updates to provide customs consulting services. When users query policies or cases, they rely on manually curated information, resulting in lengthy processing cycles and the inability to achieve real-time responses. Furthermore, the accuracy and completeness of information processed are often limited, making it difficult for businesses to quickly adapt to policy changes.

[0004] Existing intelligent customs advisory systems also have significant shortcomings. Most lack cross-language processing capabilities and are unable to effectively address the semantic differences between Chinese and English information. Furthermore, traditional role allocation mechanisms are often based on fixed rules, which can easily lead to inaccurate role identification and affect resource utilization efficiency. Furthermore, in a dynamic policy environment, many systems are unable to flexibly respond to policy changes. These shortcomings have led to delayed responses and information distortion when businesses are navigating increasingly stringent and complex customs policies, seriously impacting the timeliness and effectiveness of decision-making. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a multi-agent collaborative workflow construction method for English customs consultation, which solves the problems of the existing English customs consultation system such as untimely information integration, inaccurate role identification and inability to quickly respond to policy changes.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a multi-agent collaborative workflow construction method for English customs consultation, comprising the following steps:

[0007] S1 collects and integrates English and Chinese customs policy texts, historical customs declaration case data, and real-time announcement API data to construct a four-dimensional dynamic hypergraph consisting of entity nodes, hyperedge relationships, relationship weights, and dynamic policy dimensions;

[0008] S2, converting the four-dimensional dynamic hypergraph into a tensor representation and performing tensor decomposition to identify key agent role nodes in high-weight dimensions;

[0009] S3, builds the alliance value function based on the identified key agent role nodes and performs optimization solution;

[0010] S4 uses an encoder to represent the Chinese and English inputs respectively, generates a semantic representation vector, and combines it with the output of the alliance value function to fuse the semantic information through a dynamic gating mechanism to obtain a structured semantic fusion result;

[0011] S5, based on the policy dimension semantic channel in the structured semantic fusion result, monitoring the policy dimension changes at consecutive time points, and using the projection operator to incrementally update the system weight parameters to achieve model synchronization of dynamic policy content;

[0012] S6, based on the output of the alliance value function, the structured semantic fusion results and the updated system parameters, generates an executable process including agents, tasks and confidence levels, and performs compliance verification based on predefined rules.

[0013] Preferably, the four-dimensional dynamic hypergraph includes:

[0014] Entity node collection Represents different customs semantic entities;

[0015] Hyperedge set Used to express multiple semantic or logical relationships between entities;

[0016] Weight Matrix Used to characterize the strength of the association between hyperedges;

[0017] Dynamic policy dimension set It is dynamically updated at different time points based on the TF-IDF anomaly detection rules.

[0018] Preferably, the update of the dynamic policy dimension set D is performed according to the following rules:

[0019]

[0020] in, represents the TF-IDF change of the kth dimension; τ is the threshold; Δd k Incremental for updates.

[0021] Preferably, in said S2, the four-dimensional hypergraph is tensor-constructed and decomposed, and the four-dimensional hypergraph tensor is The factor vectors representing different roles are extracted through the non-negative tensor decomposition method for subsequent role identification and task assignment.

[0022] Preferably, the alliance value function constructed in S3 is:

[0023]

[0024] Among them, C l : indicates the lth alliance (collaboration group); v(C l ):League C l The synergistic benefit value; R: the total number of dimensions of the multidimensional semantic factor; φ r : The weight coefficient of the r-th dimension semantic factor (reflecting the importance of this dimension); Node v i Contribution value under the r-th dimension semantic factor;

[0025] The optimization function is defined as follows:

[0026]

[0027] in, The set of all possible alliances; Subset to be selected (coalition combination to be optimized); v(C l ):League C l synergistic benefit value.

[0028] Preferably, in S4, an encoder is used to characterize the Chinese and English inputs respectively to obtain a vector h en and h zh , and adopt dynamic gating mechanism to merge into:

[0029] g=σ(W g [h en ;h zh ]+b g );

[0030] h fuse =g·h en +(1-g)·h zh ;

[0031] Where g is the gating factor; W g is the fusion weight matrix; σ is the Sigmoid activation function.

[0032] Preferably, after the semantic fusion, the topological entropy is calculated:

[0033]

[0034] And set the structure adjustment threshold:

[0035] θ=αH t +β||g||2+γ·rank(M);

[0036] Among them, π(p) is the sampling probability of process path p; λ controls the smoothness of entropy change; α, β, γ are adjustment parameters; and g is the gating factor.

[0037] Preferably, S5 constructs a set of variation values ​​ΔD by monitoring the policy dimension changes between consecutive time points, and projects it into the current parameter space to perform weight adjustment, thereby achieving dynamic response and update of the system model to the new policy.

[0038] Preferably, the workflow structure output in S6 is a set of agent-task-confidence triples, and the compliance of the process execution is verified based on the set rules to avoid illegal task allocation and behavioral paths that violate regulations.

[0039] A multi-agent collaborative workflow construction system for English customs consultation, including:

[0040] Data fusion and hypergraph construction module, used to collect multi-source data and construct a four-dimensional dynamic hypergraph;

[0041] Tensor decomposition and role identification module, used to perform tensor decomposition on the hypergraph structure and identify key agent roles;

[0042] Alliance optimization and conflict coordination module, used to achieve optimal resource allocation based on the kernel solution method and fuzzy constraints;

[0043] Cross-modal semantic fusion module, used for semantic alignment and structural feedback adjustment of Chinese and English information;

[0044] Policy synchronization and incremental learning module, used to detect policy changes and adjust parameters;

[0045] The process generation and compliance verification module is used to output collaborative processes and verify their compliance.

[0046] The present invention provides a multi-agent collaborative workflow construction method for English customs consultation. It has the following beneficial effects:

[0047] 1. This invention utilizes a multi-agent collaboration mechanism and dynamic hypergraph construction technology to achieve real-time information integration in English customs consulting. This technology significantly improves the accuracy of policy interpretation. Compared with existing information access and processing technologies based on static databases, it solves the problem of information silos and enables timely response to customer needs.

[0048] 2. This invention combines tensor decomposition with a non-negative optimization algorithm to generate an efficient agent role identification model. Compared with traditional methods, this invention offers greater flexibility in role assignment, overcomes the drawbacks of relying on fixed rules, and enables real-time analysis and interpretation of complex laws and regulations.

[0049] 3. This invention enhances cross-language information processing capabilities by integrating Chinese and English semantic representations through the introduction of a dynamic gating mechanism. This technical solution, while addressing the shortcomings of existing static translation systems, enables a better understanding of policy clauses in multicultural contexts, achieving higher semantic accuracy and applicability, and effectively improving the quality of consulting services.

[0050] 4. This invention combines policy change monitoring with system parameter optimization, enabling the system to adjust response strategies in a timely manner. Compared to existing passive systems, this invention addresses the issue of delayed response due to policy changes, enabling real-time updates and flexible responses, ensuring continuity and compliance of customer service. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 Schematic diagram of the method flow of the present invention;

[0052] Figure 2 It is the system framework structure diagram of the present invention. DETAILED DESCRIPTION

[0053] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0054] Please see the attached Figure 1 The embodiment of the present invention provides a multi-agent collaborative workflow construction method for English customs consultation, including:

[0055] S1 collects and integrates English and Chinese customs policy texts, historical customs declaration case data, and real-time announcement API data to construct a four-dimensional dynamic hypergraph consisting of entity nodes, hyperedge relationships, relationship weights, and dynamic policy dimensions;

[0056] In this embodiment, first, the system includes a data fusion and hypergraph construction unit, which is connected to a multi-source data acquisition interface, including an English customs policy capture terminal, a Chinese policy database interface, a historical customs declaration case storage module, and a real-time announcement API service port.

[0057] This unit synchronously accesses various data sources to obtain policy semantics and relevant rule instances at different points in time. It then uses a unified word segmentation and entity recognition model to perform preliminary analysis of the Chinese and English texts. The entity recognition results are used to construct the node structure of the hypergraph, and the semantic relationships are classified to generate the hyperedge structure.

[0058] In the initial graph construction phase, the entity node set is defined as:

[0059]

[0060] Among them, v i Represents the i-th node with policy semantics, such as “tariff”, “HS code”, “regulatory conditions”, etc.

[0061] At the same time, build a set of hyperedges:

[0062]

[0063] Among them, e j Represents the jth semantic or logical relationship edge, which is used to connect two or more nodes and describe their co-occurrence relationship, causal relationship, or cross-language mapping relationship.

[0064] In order to characterize the strength of semantic linkage between nodes, the weight matrix between hyperedges is further defined:

[0065]

[0066] Among them, W ij Represents the hyperedge e i With e j The strength of semantic dependency in a joint context is calculated based on co-occurrence frequency and semantic similarity.

[0067] Based on the temporal evolution characteristics of policy texts, a dynamic policy dimension set is introduced:

[0068]

[0069] Among them, d k (t) represents the kth policy dimension at time point t, such as "number of declaration elements" or "change in customs declaration code category." Changes in policy dimensions are detected by comparing the TF-IDF index over consecutive time slices to determine the significance of fluctuations. Changes are determined using the following formula:

[0070]

[0071] in, represents the TF-IDF change amplitude of dimension k, τ is the set significance threshold, Δd k Update the delta for the policy dimension.

[0072] The updated dimension set is used to expand the original graph structure to form a four-dimensional dynamic hypergraph structure with a time-evolving dimension:

[0073]

[0074] The construction process of this hypergraph structure is as follows: first, entity nodes are extracted, then multi-dimensional relationships are established, and relationship strength and dynamic dimension elements are introduced, finally generating a heterogeneous graph structure suitable for subsequent tensor modeling and dynamic evolution analysis.

[0075] After the above structure is output by the construction module, it is further processed by the tensor modeling module within the system to achieve unified encoding of cross-language semantic data and preparation for intelligent agent role modeling.

[0076] The four-dimensional dynamic hypergraph constructed in the above way can realize the cross-language and cross-data source integrated expression between policy semantic entities, has dynamic scalability, and provides a complete semantic foundation and structural input for subsequent tensor decomposition and multi-agent collaborative process generation.

[0077] S2, converting the four-dimensional dynamic hypergraph into a tensor representation and performing tensor decomposition to identify key agent role nodes in high-weight dimensions;

[0078] In this embodiment, first, the system includes a tensor decomposition and role recognition unit, which receives the four-dimensional dynamic hypergraph structure constructed in S1. And perform structural mapping and tensor construction processing on it. Based on the data structure of multimodal features, the hypergraph is embedded into a fourth-order tensor as an example:

[0079]

[0080] Where: N represents the number of entity nodes, M represents the number of hyperedges, K represents the number of policy dimensions, and T represents the number of time steps.

[0081] Each tensor dimension corresponds to a node, relationship, dimension, and time axis, respectively, achieving unified encoding of structural information and dynamic changes.

[0082] During tensor construction, the relationship between nodes and hyperedges is first established. To this end, the system maps the relationships between nodes in the hypergraph at each moment, representing the semantic relationship corresponding to each hyperedge in binary form to form a basic second-order relationship tensor. Then, through the weight matrix W in the policy dimension set and the hyperedge weight matrix E, the dynamic changes in the relationship and policy under the time dimension are integrated, expanding the representation into a multidimensional tensor. In this way, each dimension of the tensor can encode different semantic information and temporal evolution characteristics.

[0083] After the tensor is constructed, the system further performs tensor decomposition processing and uses the non-negative tensor decomposition method (NMF) to decompose the tensor. The goal is to extract potential semantic roles and their weights from the multidimensional tensor. Specifically, the decomposition model is iteratively optimized using the following formula:

[0084]

[0085] Among them, λ r Represents the weight of each factor, A r ,B r ,C r ,D r They represent the factor matrices associated with entity nodes, hyperedges, policy dimensions, and time steps, respectively. R is the decomposition rank, which determines the number of potential factors.

[0086] Through this decomposition, the system can extract a set of factor vectors representing each role. These factor vectors can effectively represent the relationship between policies, tasks, and agent roles, providing data support for subsequent role identification and task allocation.

[0087] In the factor vector, the dimension factor c r Indicates the importance of each policy dimension, node factor a r Then it is mapped to the candidate set of agents, and the key agent role nodes are identified by sorting the node factors under the high-weight dimension channel. The key roles are selected according to the following judgment rules:

[0088]

[0089] in, Node v in layer r i The attention weight θ r : The preset threshold of the rth layer; Logical implication symbol (if the left side holds, then the right side holds); A collection of key nodes.

[0090] This processing method has the advantages of strong interpretability and adaptability to multi-dimensional dynamic data structures, and can achieve efficient identification of key role nodes in cross-language policy knowledge graphs.

[0091] S3, based on the key agent role nodes, construct the alliance value function, introduce fuzzy constraints and solve the improved kernel solution optimization model to achieve resource conflict coordination;

[0092] In this embodiment, first, the system includes an alliance modeling and optimization decision unit, which is connected to the key role node set output in S2. And receive the corresponding semantic factor representation result.

[0093] The unit first models the semantic collaborative relationship between key nodes, constructs a coalition candidate set, and defines the coalition structure set as:

[0094]

[0095] in,

[0096] represents the l-th candidate alliance, L is the number of alliances, and satisfies 1≤l≤L.

[0097] Each node in the alliance is associated under multi-dimensional semantic factors. To measure the synergy effect, the alliance value function is introduced:

[0098]

[0099] Among them, C l : indicates the lth alliance (collaboration group); v(C l ):League C l The synergistic benefit value; R: the total number of dimensions of the multidimensional semantic factor; φ r : The weight coefficient of the r-th dimension semantic factor (reflecting the importance of this dimension); Node v i Contribution value under the r-th dimension semantic factor.

[0100] The synergy benefits of the alliance are obtained by weighted summation of contributions from each dimension. The specific calculation process is as follows:

[0101] For each dimension r, calculate the sum of the contribution values ​​of all nodes in the alliance in that dimension

[0102] Multiply the result of each dimension by the corresponding weight φ.

[0103] Through weighted summation, the collective contribution value of key nodes under different semantics is quantified to reflect their overall performance in the alliance.

[0104] Based on the definition of the alliance value function, the system further performs combinatorial optimization on all alliance candidate sets, with the goal of selecting an alliance configuration solution that maximizes the overall value. The optimization objective function is defined as follows:

[0105]

[0106] in, The set of all possible alliances; Subset to be selected (coalition combination to be optimized); v(C l ):League C l synergistic benefit value.

[0107] The constraint is that the alliances do not overlap:

[0108]

[0109] The above optimization problem is formally a weighted subset partitioning problem. The system uses a greedy iterative algorithm to solve it. The specific steps include:

[0110] First, sort by descending value of a single alliance, then check each one to see if it conflicts with the selected alliance. If there is no overlap, add it to the final result set until no more alliances can be added or the alliance pool is exhausted. The final output is the optimal alliance set:

[0111]

[0112] This process completes the full-chain modeling from key role identification to alliance configuration, and ensures that there is no cross-interference between alliances. It has the characteristics of clear structure, reasonable logic, and strong computability.

[0113] The alliance modeling and optimization module is physically connected to the aforementioned tensor decomposition unit, and its output results will serve as the basic input for the system's subsequent strategy decision-making or resource allocation modules to realize the automated reasoning of the multi-agent task collaboration mechanism.

[0114] Through this implementation method, it is possible to realize the value modeling of the intelligent agent alliance based on multi-dimensional policy factors and role participation, and complete the search for the optimal task configuration combination, ensuring that the system has high adaptability and execution efficiency under dynamic policy changes.

[0115] S4 uses an encoder to represent the Chinese and English inputs respectively, generates a semantic representation vector, and combines it with the output of the alliance value function to fuse the semantic information through a dynamic gating mechanism to obtain a structured semantic fusion result;

[0116] In this embodiment, first, the system receives Chinese and English text as input. For each language, a pre-trained encoder (such as Transformer or BERT) is used to convert the text into a semantic representation vector. Through this encoder, the English text is converted into a semantic vector h en , Chinese text is converted into semantic vector h zh These two semantic vectors contain the customs semantic information of their respective languages ​​and are ready for subsequent semantic fusion.

[0117] Next, the Chinese and English semantic vectors h en and h zh The vectors are concatenated into a large vector and fed into the dynamic gating mechanism. The gating mechanism calculates the gating factor g, which determines the weighted ratio of English and Chinese semantics in the final fusion result. The formula for calculating the gating factor g is as follows:

[0118] g=σ(W g [h en ;h zh ]+b g );

[0119] Among them, σ is the Sigmoid activation function, the formula is:

[0120]

[0121] Used to calculate the gating factor g, whose value range is between 0 and 1. Through this factor, the gating mechanism can determine the weighted ratio of English semantic vectors and Chinese semantic vectors when fusion. Weight matrix W g and the bias term b g It is used to learn how to dynamically adjust weights based on the concatenation of Chinese and English semantics.

[0122] Then, the obtained gating factor g is used to perform weighted fusion of the Chinese and English semantic vectors, and finally the fused semantic vector h is generated. fuse , and its calculation formula is:

[0123] h fuse =g·h en +(1-g)·h zh ;

[0124] At this time, h fuse This represents the weighted fusion of Chinese and English semantic information and serves as the input for subsequent multi-agent collaborative workflow generation. This fusion vector contains comprehensive features of Chinese and English semantics, providing comprehensive information for subsequent task allocation and decision-making.

[0125] In order to further improve the performance of multilingual collaborative workflow, this embodiment introduces a structural adjustment mechanism based on topological entropy. After semantic fusion, the system calculates the topological entropy H t , which is used to measure the complexity of the workflow structure. The topological entropy is calculated as follows:

[0126]

[0127] Among them, π(p) is the sampling probability of process path p, which represents the probability of occurrence of each path at a given time point; It represents the gradient of the topological entropy at the previous moment, and λ controls the smoothness of the entropy change.

[0128] According to the calculated topological entropy H t , the system further sets a structural adjustment threshold θ to optimize the structure of the workflow and adjust the task execution order of each agent. The calculation formula of the structural adjustment threshold is as follows:

[0129] θ=αH t +β||g||2+γ·rank(M);

[0130] Among them, α, β, and γ are adjustment factors used to balance the impact of different items on structural adjustment; ||g||2 is the L2 norm of the gating factor g, which measures the strength of the semantic fusion of Chinese and English; rank(M) is the rank of the matrix M, which represents the complexity of the agent allocation in the workflow.

[0131] In the subsequent processing, the final generated semantic vector h fuse This is passed as input to the task assignment and decision-making modules. Based on this fusion result, the agents make appropriate task assignments and resource scheduling to ensure efficient workflow execution. This process allows the system to flexibly adjust resource allocation based on the semantic information of the Chinese and English inputs, ensuring the efficiency and accuracy of multi-agent collaborative workflows.

[0132] This embodiment performs weighted fusion of Chinese and English inputs through a dynamic gating mechanism, and optimizes the workflow structure by combining topological entropy and structural adjustment thresholds, thereby ensuring efficient integration of semantic information in different languages.

[0133] S5, based on the policy dimension semantic channel in the structured semantic fusion result, monitoring the policy dimension changes at consecutive time points, and using the projection operator to incrementally update the system weight parameters to achieve model synchronization of dynamic policy content;

[0134] In this embodiment, first, based on the structured semantic fusion result h obtained in step S4, fuse,t , extracting the policy-related semantic channel information and constructing the policy dimension semantic representation in the form of time series. This representation is achieved through channel selection operation and is expressed as:

[0135] p t =Π policy (h fuse,t );

[0136] in:

[0137] The semantic representation vector of the policy dimension at the t-th time point;

[0138] Π policy (·): channel selection mapping function, extracting the subvector corresponding to the policy semantic channel in the structured semantic vector;

[0139] h fuse,t : The Chinese and English fusion semantic vector at the t-th time point comes from the aforementioned gated weighted fusion process.

[0140] Then, within the continuous time window [t-Δt, t], the difference between the policy semantic vectors at any two adjacent time points t-1 and t is calculated to obtain the policy change set ΔP, whose elements are defined as follows:

[0141] δ t =p t -p t-1 ;

[0142] in:

[0143] Represents the difference in policy semantics between time points t and t-1;

[0144] The differences indicate the direction and magnitude of policy change trends and are the basis for subsequent model parameter adjustments.

[0145] In order to respond to policy changes and achieve synchronous updates of model parameters, this implementation constructs a linear projection transformation operator at the structural level. Used to change policy semantics δ t Project to the model parameter space to obtain parameter increments:

[0146]

[0147] in:

[0148] Incremental adjustment matrix of system parameters;

[0149] Projection weight matrix, which plays a mapping role in linear transformation;

[0150] Bias vector, used to perform translation correction on the mapping result;

[0151] Projection operator, parameter set θ={W p ,b p} are trainable model parameters.

[0152] The mapping process can be trained and optimized using a standard back-propagation algorithm, where the objective function includes the model prediction error and the parameter adjustment regularization term, as detailed below.

[0153] The model execution parameter synchronization update process adopts the following incremental update formula:

[0154] Θ t =Θ t-1 +η·ΔΘ t ;

[0155] in:

[0156] Represents the complete parameter matrix of the model at time t

[0157] η∈(0,1]: learning rate, used to control the speed of weight update;

[0158] Θ t-1 : The parameter matrix of the system that has been trained at the previous time point.

[0159] In order to avoid model oscillation in scenarios with frequent policy changes, a regularization constraint term is further introduced to control the weight adjustment range and ensure the continuity and stability of system parameters. The regularization term is calculated as follows:

[0160]

[0161] in:

[0162] Parameter update regularization loss;

[0163] Regularization coefficient, used to balance model sensitivity and robustness;

[0164] ||·||2: L2 norm, which measures the magnitude of the change in the parameter matrix.

[0165] This regularization term and the main loss function work together in the training process to minimize the comprehensive loss function. Achieve stable synchronization.

[0166] In the above process, the system continuously monitors the dynamic changes of the policy dimensions in the fusion semantics in a time series manner. After detecting the policy change trend, it explicitly models its differential vector and maps it to the model parameter space to achieve structured response to the latest policy content and automatic synchronization of the model.

[0167] S6, based on the output of the alliance value function, the structured semantic fusion results and the updated system parameters, generates an executable process including agents, tasks and confidence levels, and performs compliance verification based on predefined rules.

[0168] In this implementation, we first establish an agent-task joint modeling structure. The model input includes: the structured semantic fusion result h fuse,t , the weight parameter matrix Θ after the system is updated t , and the alliance value function output v(C l ), the output is a set of structured process triples

[0169] in:

[0170] is the fused semantic vector; is the system parameter after dynamic update in step S5; v(C l ): alliance value function; The execution process structure generated at the current time point is generated by agent a i , task τ j , confidence c ij Triple composition.

[0171] To generate the above triplet set, first calculate the agent-task matching score matrix Each of its elements is given by the following formula:

[0172] s ij =φ(a i ,τ j ;h fuse,t ,Θ t )·V(a i ,τ j );

[0173] in:

[0174] φ(·): represents the matching function between task semantics and agent features, which is modeled by a bidirectional attention mechanism, where the semantic vector input is h fuse,t , the model parameters are given by Θ t Provide; V(a i ,τ j ): The cooperative benefit value between the agent and the task at the current time point given by the alliance value function.

[0175] By applying a normalization function (such as Softmax) to the score matrix S, we can get the confidence matrix C = {c ij}, where each c ij Represents agent a i Undertake task τ under the current strategy j The execution credibility is calculated as follows:

[0176]

[0177] When constructing the process structure, the triples with confidence greater than the set threshold δ are taken as valid allocations, and the process output set is defined as:

[0178]

[0179] in:

[0180] δ: The set task allocation confidence threshold to ensure that only highly reliable tasks are scheduled for execution;

[0181] The system outputs a set of agent-task-confidence triples at time t for processing by subsequent execution modules.

[0182] In order to ensure the standardization of the task execution process, the process compliance verification is further performed based on the rule engine. The rule set is defined as Each rule r k Represents a set of legal constraints, such as task ownership, execution order, and compliance restrictions.

[0183] For any triple If it violates the constraints in the rule set, it is marked as an illegal path and removed from the executable set. The verification formula is as follows:

[0184]

[0185] in:

[0186] The legal process structure that is ultimately verified by the rules;

[0187] r k (a i ,τ j ): Boolean function, True if the current allocation satisfies the rule.

[0188] This implementation enables task matching that integrates multidimensional semantic information, system parameters, and value assessment results, outputting structured process results with confidence and credibility, and filtering out illegal task paths. This method enables automated task generation and compliance assurance in a multilingual, multi-agent environment, achieving clear process structure, controllable execution, and traceable decision-making.

[0189] The multi-agent collaborative workflow construction system for English customs consultation described below and the multi-agent collaborative workflow construction method for English customs consultation described above can be referenced to each other.

[0190] See also Figure 2 The present invention also provides a multi-agent collaborative workflow construction system for English customs consultation, including:

[0191] Data fusion and hypergraph construction module, used to collect multi-source data and construct a four-dimensional dynamic hypergraph;

[0192] Tensor decomposition and role identification module, used to perform tensor decomposition on the hypergraph structure and identify key agent roles;

[0193] Alliance optimization and conflict coordination module, used to achieve optimal resource allocation based on the kernel solution method and fuzzy constraints;

[0194] Cross-modal semantic fusion module, used for semantic alignment and structural feedback adjustment of Chinese and English information;

[0195] Policy synchronization and incremental learning module, used to detect policy changes and adjust parameters;

[0196] The process generation and compliance verification module is used to output collaborative processes and verify their compliance.

[0197] The system of this embodiment can be used to execute the above method embodiments, and its principles and technical effects are similar, so they will not be repeated here.

[0198] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A multi-agent collaborative workflow construction method for English customs consultation, characterized by: The following steps are involved: S1 collects and integrates English and Chinese customs policy texts, historical customs declaration case data, and real-time announcement API data to construct a four-dimensional dynamic hypergraph consisting of entity nodes, hyperedge relationships, relationship weights, and dynamic policy dimensions; S2, converting the four-dimensional dynamic hypergraph into a tensor representation and performing tensor decomposition to identify key agent role nodes in high-weight dimensions; S3, builds the alliance value function based on the identified key agent role nodes and performs optimization solution; S4 uses an encoder to represent the Chinese and English inputs respectively, generates a semantic representation vector, and combines it with the output of the alliance value function to fuse the semantic information through a dynamic gating mechanism to obtain a structured semantic fusion result; S5, based on the policy dimension semantic channel in the structured semantic fusion result, monitoring the policy dimension changes at consecutive time points, and using the projection operator to incrementally update the system weight parameters to achieve model synchronization of dynamic policy content; S6, based on the output of the alliance value function, the structured semantic fusion results and the updated system parameters, generates an executable process including agents, tasks and confidence levels, and performs compliance verification based on predefined rules.

2. The method for constructing a multi-agent collaborative workflow for English customs consultation according to claim 1 is characterized in that: The four-dimensional dynamic hypergraph includes: Entity node collection Represents different customs semantic entities; Hyperedge set Used to express multiple semantic or logical relationships between entities; Weight Matrix Used to characterize the strength of the association between hyperedges; Dynamic policy dimension set It is dynamically updated at different time points based on the TF-IDF anomaly detection rules.

3. The method for constructing a multi-agent collaborative workflow for English customs consultation according to claim 2 is characterized in that: The dynamic policy dimension set D is updated according to the following rules: in, represents the TF-IDF change of the kth dimension; τ is the threshold; Δd k Incremental for updates.

4. The method for constructing a multi-agent collaborative workflow for English customs consultation according to claim 1 is characterized in that: In S2, the four-dimensional hypergraph is tensor-constructed and decomposed. The four-dimensional hypergraph tensor is: The factor vectors representing different roles are extracted through the non-negative tensor decomposition method for subsequent role identification and task assignment.

5. The method for constructing a multi-agent collaborative workflow for English customs consultation according to claim 1 is characterized in that: The alliance value function constructed in S3 is: Among them, C l : indicates the lth alliance (collaboration group); v(C l ):League C l The synergistic benefit value; R: the total number of dimensions of the multidimensional semantic factor; φ r : The weight coefficient of the r-th dimension semantic factor (reflecting the importance of this dimension); Node v i Contribution value under the r-th dimension semantic factor; The optimization function is defined as follows: in, The set of all possible alliances; Subset to be selected (coalition combination to be optimized); v(C l ):League C l synergistic benefit value.

6. The method for constructing a multi-agent collaborative workflow for English customs consultation according to claim 1 is characterized in that: In S4, an encoder is used to represent the Chinese and English inputs respectively to obtain the vector h en and h zh , and adopt dynamic gating mechanism to merge into: g=σ(W g [h en ;h zh ]+b g ); h fuse =g·h en +(1-g)·h zh ; Where g is the gating factor; W g is the fusion weight matrix; σ is the Sigmoid activation function.

7. The method for constructing a multi-agent collaborative workflow for English customs consultation according to claim 6 is characterized in that: After the semantic fusion, the topological entropy is calculated: And set the structure adjustment threshold: θ=αH t +β||g||2+γ·rank(M); Among them, π(p) is the sampling probability of process path p; λ controls the smoothness of entropy change; α, β, γ are adjustment parameters; and g is the gating factor.

8. The method for constructing a multi-agent collaborative workflow for English customs consultation according to claim 1 is characterized in that: The S5 constructs a set of variation ΔD by monitoring the policy dimension changes between consecutive time points, and projects it into the current parameter space to perform weight adjustment, thereby achieving dynamic response and update of the system model to the new policy.

9. The method for constructing a multi-agent collaborative workflow for English customs consultation according to claim 1 is characterized in that: The workflow structure output in S6 is a set of agent-task-confidence triples, and the compliance of the process execution is verified based on the set rules to avoid illegal task allocation and behavioral paths that violate the norms.

10. A multi-agent collaborative workflow construction system for English customs consultation, applied to the multi-agent collaborative workflow construction method for English customs consultation according to any one of claims 1 to 9, characterized in that: include: Data fusion and hypergraph construction module, used to collect multi-source data and construct a four-dimensional dynamic hypergraph; Tensor decomposition and role identification module, used to perform tensor decomposition on the hypergraph structure and identify key agent roles; Alliance optimization and conflict coordination module, used to achieve optimal resource allocation based on the kernel solution method and fuzzy constraints; Cross-modal semantic fusion module, used for semantic alignment and structural feedback adjustment of Chinese and English information; Policy synchronization and incremental learning module, used to detect policy changes and adjust parameters; The process generation and compliance verification module is used to output collaborative processes and verify their compliance.

Images