Multi-agent coordination-based labor dispute whole-process intelligent arbitration method and system

Abstract

The application discloses a labor dispute whole-process intelligent arbitration method and system based on multi-agent coordination, relates to the technical field of intelligent arbitration of labor disputes, constructs a special intelligent agent cluster, each intelligent agent has dispute identification, decision-making and execution capability, resolves case filing, case division, court trial, document generation whole-process subtasks through a collaborative scheduler and distribution, and arbitration intelligent agents solve operation conflicts, realize whole-process collaboration and information flow. The application realizes intelligent collaborative disposal of the whole process of labor dispute arbitration, greatly improves arbitration efficiency, shortens the case handling period, and alleviates the contradiction between more cases and fewer people. Through unified data lake construction and standardized modeling, efficient integration and reuse of multi-source data are realized, and solid data support is provided for intelligent disposal.

Description

Technical Field

[0001] This invention relates to the field of intelligent arbitration technology for labor disputes, and in particular to an intelligent arbitration method and system for the entire process of labor disputes based on multi-agent coordination. Background Technology

[0002] The current labor dispute arbitration field faces prominent problems such as a large number of cases and a shortage of personnel, long trial periods, and insufficient consistency between arbitration and adjudication. The traditional arbitration model, centered on manual processing, relies on the professional experience of arbitrators to carry out the entire process from case filing and assignment to hearings and document generation, which has many limitations. Multi-source labor dispute data is stored in a scattered manner, lacking unified data integration and standardized modeling, resulting in low data utilization and difficulty in supporting intelligent processing throughout the entire process. The extraction of core characteristics of disputed events is not comprehensive enough, failing to accurately depict the essence of the dispute and affecting the accuracy of case matching.

[0003] Existing intelligent methods are mostly limited to assisting a single link, lacking the ability to coordinate the entire process. The links are not well connected, which can easily lead to operational conflicts. Furthermore, there is a lack of effective rollback mechanisms during task execution, resulting in poor fault tolerance.

[0004] Case matching often relies on traditional keyword retrieval methods, which are prone to missed detections due to differences in wording, making accurate matching difficult and failing to effectively support the goal of consistency between arbitration and adjudication. Document generation is mostly template-based and lacks intelligence, requiring arbitrators to invest significant time in revisions and improvements, further increasing the workload of arbitration.

[0005] Meanwhile, the existing system lacks a continuous optimization mechanism, cannot iterate and upgrade based on feedback data from actual arbitration scenarios, and is difficult to adapt to the complex and ever-changing needs of labor dispute scenarios, thus hindering the efficient and standardized development of labor dispute arbitration work. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method and system for intelligent arbitration of labor disputes throughout the entire process based on multi-agent coordination. The technical solution adopted is as follows: A multi-agent coordinated intelligent arbitration method for labor disputes throughout the entire process includes the following steps: Step 1: Collect multi-source labor dispute related data, construct a unified labor dispute data lake, and perform standardized modeling of dispute events; Step 2: Extract the six-dimensional semantic features of the disputed event, including the key points of the dispute, the claims of the parties, the type of evidence, the subject matter of the dispute, the duration of the dispute, and the semantic text, and construct a six-dimensional semantic vector of the dispute. Step 3: Construct a dedicated intelligent agent cluster. Each intelligent agent has the ability to identify disputes, make decisions and execute them. The entire process of case filing, case assignment, court hearing and document generation is broken down and assigned by the collaborative scheduler. The arbitration intelligent agent resolves operational conflicts and realizes full-process collaboration and information flow. Step 4: Using a multi-dimensional semantic alignment algorithm, calculate the weighted cosine similarity between the current dispute and historical similar cases, match and push high-similarity historical similar cases for arbitrators to refer to, and improve the consistency between adjudication and arbitration. Step 5: Generate a directed acyclic graph of the tasks to be processed, which is then executed by the distributed execution engine, supporting automatic rollback in case of task failure. Step 6: Combine the knowledge enhancement model with the labor dispute document template to generate an L4 level document, which will then be reviewed and modified by the arbitrator.

[0007] Optional, multi-source labor dispute related data includes data such as labor dispute parties' information, arbitration requests, statements of facts, evidence materials, historical case rulings, legal provisions, arbitration process records, court hearing audio and video recordings, and party feedback; evidence materials include labor contracts, wage payment vouchers, attendance records, work permits, colleague testimonies, notices of termination of labor contracts, and overtime application forms, which meet the requirements of the labor arbitration evidence rules; The expression for the unified dispute event model is: ,in As the unique identifier for the disputed event, For the precise timestamp of the submission or occurrence of the disputed event, The source of the event specifically includes at least one of the following: arbitration institution, parties involved, lawyers, or courts. The core content of the dispute can be divided into structured fields or unstructured text. This is a list of context event IDs associated with the current disputed event, used to trace the context of disputed events and support information exchange in the process of adjudication and litigation.

[0008] Optionally, in step 2, the specific extraction methods for the six-dimensional semantic features are as follows: Key Controversial Point O: A deep learning-based entity recognition and key point extraction model is used to extract core controversial points from the content of labor dispute events. These core controversial points refer to the core focus of the labor dispute. An embedding algorithm is then used to transform the extracted controversial points into vectors. ; The parties' claims (T) are categorized and analyzed using a lightweight classification model. This categorizes and clarifies the parties' core demands and factual basis, and the results are then transformed into vectors through embedding. ; Evidence Type S: Using an evidence classification algorithm, various types of evidence submitted by the parties are categorized. The classification results are then vectorized based on the weight of each piece of evidence's probative value to obtain a vector. It is compatible with the requirements of the evidence list in labor arbitration; Disputed Subject Matter X: The amount and type of the subject matter of the labor dispute are quantified and coded. Combined with the relevant legal basis for the subject matter, the quantification and coding results are vectorized to obtain a vector. ; Dispute Duration C: Based on the collected dispute process records, calculate the duration from the occurrence of the dispute to its submission to arbitration, and the current progress time of the arbitration process. Vectorize the duration information to obtain a vector. This provides support for process scheduling; Semantic Text D: A BERT-like pre-trained model is used to perform semantic embedding on unstructured text, which includes statements of disputed facts, courtroom testimonies, and descriptions of evidence. Semantic relevance features are extracted from the text to obtain vectors. ; Six-dimensional controversial semantic vector The expression is: The dimensions of all vectors are unified.

[0009] Optionally, in step 3, the intelligent agent cluster includes intelligent agents for confirming labor relations, intelligent agents for pursuing unpaid wages, intelligent agents for illegal termination of labor contracts, intelligent agents for disputes over overtime pay, and intelligent agents for work-related injury compensation. The dispute identification module adopts a hybrid architecture that combines labor dispute domain rules with a lightweight classification model. Domain rules are used to match known typical labor dispute scenarios, evidence rules and adjudication norms, while the lightweight classification model is used to identify complex and atypical labor disputes, together achieving accurate judgment of dispute type and key points. The collaborative scheduler uses a directed acyclic graph to model the entire process of labor dispute resolution, clarifying the dependencies between sub-tasks such as case filing, case assignment, court hearing, and document generation. Arbitration agents are used to resolve operational conflicts arising from multiple agents dealing with the same disputed case or the same process node. The conflict resolution method involves the arbitration agent calculating the impact score of each conflicting operation. , ;in , , For preset weighting coefficients, As for operation priority, To determine the scope of cases and process progress affected by the operation, Based on the estimated operation time, the operation with the highest impact score I is selected for execution, and the remaining conflicting operations are placed in a waiting queue. They are then evaluated and executed in turn after the current operation is completed.

[0010] Optionally, in step 4, each historical case in the historical case library... All are represented as: ; where Vei is the six-dimensional semantic vector of the dispute corresponding to the historical case, Solutioni is the specific adjudication approach and handling process of the corresponding historical case, including the filing standards, the basis for case assignment, the focus of the trial, the basis for the adjudication, and the application of legal provisions, and Resulti is the handling result of the corresponding historical case, including the adjudication result, the appeal situation, the consistency between the adjudication and the judgment, and the satisfaction score of the parties. Current Labor Dispute Case F and Historical Cases The formula for calculating similarity is: ,in For cosine similarity, Let F be the vector of the current labor dispute event in dimension d. For historical cases A vector in dimension d; The rules for matching and recommending similar cases are as follows: Select the top-K similar cases with the highest similarity. If there are similar cases with a similarity exceeding a preset threshold, prioritize recommending such cases and their adjudication reasoning and applicable legal basis, and focus on recommending cases where the adjudication and judgment are consistent. If the similarity of all similar cases is lower than the threshold θ, then integrate the adjudication reasoning and handling process of the top-K similar cases to generate a reference solution suitable for the current labor dispute and recommend it to the arbitrators for reference.

[0011] Optionally, in step 5, the nodes of the directed acyclic graph of the labor dispute resolution task are specifically defined as follows: The isolation node is used to temporarily isolate cases with dispute risks or incomplete materials, and notify the parties to supplement the materials; the case filing review node is used to automatically review the arbitration applications and evidence materials submitted by the parties, determine whether they meet the filing conditions, and if the review is successful, proceed to the case assignment stage; if the review fails, supplementary opinions are provided to ensure compliance with labor arbitration case filing regulations; the case assignment node is used to automatically assign cases to the corresponding arbitrators based on the dispute type, disputed subject matter, and arbitrator's professional area; the trial assistance node is used to assist arbitrators in completing trial preparation, trial recording, evidence presentation, and cross-examination guidance, and to synchronize trial information to relevant intelligent agents; the document generation node is used to automatically generate relevant documents based on the trial results and similar case references; the verification node is used to check the content and process compliance of the documents and determine whether the resolution process is complete; each node includes specific operation instructions, timeout time, and rollback instructions. When a node times out or fails to execute, a rollback instruction is triggered.

[0012] Optionally, in step 6, the L4 level labor dispute documents conform to the standards for labor arbitration documents.

[0013] Optionally, step 7 is also included, which involves collecting explicit and implicit feedback data to optimize semantic alignment weights and the overall model. The feedback data includes: explicit feedback such as the arbitrator's acceptance or rejection of the case recommendations in step 4, the score for the accuracy of case matching, the content of the modifications made to the documents generated in step 6, and the satisfaction score; and implicit feedback such as the duration of labor dispute resolution, the case appeal rate, the consistency rate between arbitration and adjudication, the adoption rate of the recommended case scheme, and the proportion of document modifications. Special attention is paid to the consistency rate between arbitration and adjudication and the appeal rate to optimize the model.

[0014] A multi-agent coordination-based intelligent arbitration system for the entire process of labor disputes is used to realize an intelligent arbitration method for the entire process of labor disputes based on multi-agent coordination. It is characterized by including a data acquisition and fusion module, a semantic feature extraction module, an agent cluster module, a collaborative scheduling and arbitration module, a semantic alignment and case matching module, a distributed disposal and execution module, a document autonomous generation module, and a feedback optimization module. The data acquisition and fusion module constructs a unified labor dispute data lake and completes the modeling of disputed events; the semantic feature extraction module generates a six-dimensional dispute semantic vector; the intelligent agent cluster module provides dedicated intelligent agents and multi-agent collaborative capabilities; the semantic alignment and case matching module realizes high-precision matching and push of similar cases; the distributed disposal execution module schedules task execution and supports rollback; the document autonomous generation module generates L4 level labor dispute documents; and the feedback optimization module realizes continuous optimization of system parameters and models.

[0015] In summary, the present invention has at least one of the following beneficial technical effects: This invention provides a method and system for intelligent arbitration of labor disputes throughout the entire process based on multi-agent coordination. This enables intelligent collaborative handling of the entire labor dispute arbitration process, significantly improving arbitration efficiency, shortening case handling cycles, and alleviating the contradiction of too many cases and too few personnel. Through the construction of a unified data lake and standardized modeling, it achieves efficient integration and reuse of multi-source data, providing solid data support for intelligent handling.

[0016] Six-dimensional semantic feature extraction and semantic vector construction improve the accuracy of characterizing controversial events. Combined with multi-dimensional semantic alignment algorithms, it achieves high-precision matching of historical similar cases, effectively improving the consistency of adjudication and reducing inconsistent rulings in similar cases. A dedicated intelligent agent cluster and collaborative scheduling mechanism enable the reasonable decomposition and allocation of sub-tasks throughout the entire process, resolving conflicts in the operation of multiple intelligent agents and ensuring smooth process continuity.

[0017] The task-oriented acyclic graph and distributed execution engine enable ordered task scheduling and automatic rollback in case of failure, improving system fault tolerance and stability. Level 4 document autonomous generation significantly reduces the workload of arbitrators in drafting documents, improving document generation efficiency and standardization.

[0018] The feedback optimization mechanism enables continuous iteration of system parameters and models, adapting to complex labor dispute scenarios, further improving the quality and intelligence of arbitration, while also facilitating the connection between arbitration and adjudication and enhancing the credibility of arbitration. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the intelligent arbitration method for the entire labor dispute process based on multi-agent coordination, as proposed in this invention. Figure 2 This is a schematic diagram of the first page of a Level L4 labor dispute document according to a specific embodiment of the present invention; Figure 3 This is a schematic diagram of the second page of a Level L4 labor dispute document according to a specific embodiment of the present invention; Figure 4 This is a schematic diagram of the third page of a Level L4 labor dispute document according to a specific embodiment of the present invention; Figure 5 This is a schematic diagram of the fourth page of the L4 level labor dispute document in a specific embodiment of the present invention. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to the accompanying drawings.

[0021] This invention discloses a method and system for intelligent arbitration of labor disputes throughout the entire process based on multi-agent coordination.

[0022] Reference Figures 1-5 Example 1, a full-process intelligent arbitration method for labor disputes based on multi-agent coordination, includes the following steps: Step 1: Collect multi-source labor dispute related data, construct a unified labor dispute data lake, and perform standardized modeling of dispute events; Step 2: Extract the six-dimensional semantic features of the disputed event, including the key points of the dispute, the claims of the parties, the type of evidence, the subject matter of the dispute, the duration of the dispute, and the semantic text, and construct a six-dimensional semantic vector of the dispute. Step 3: Construct a dedicated intelligent agent cluster. Each intelligent agent has the ability to identify disputes, make decisions and execute them. The entire process of case filing, case assignment, court hearing and document generation is broken down and assigned by the collaborative scheduler. The arbitration intelligent agent resolves operational conflicts and realizes full-process collaboration and information flow. Step 4: Using a multi-dimensional semantic alignment algorithm, calculate the weighted cosine similarity between the current dispute and historical similar cases, match and push high-similarity historical similar cases for arbitrators to refer to, and improve the consistency between adjudication and arbitration. Step 5: Generate a directed acyclic graph of the tasks to be processed, which is then executed by the distributed execution engine, supporting automatic rollback in case of task failure. Step 6: Combine the knowledge enhancement model with the labor dispute document template to generate an L4 level document, which will then be reviewed and modified by the arbitrator.

[0023] By adopting the above technical solution, the fragmented barriers of labor dispute-related data are first broken down through multi-source data collection and integration, constructing a unified data lake. Simultaneously, standardized modeling of disputed events provides a standardized and unified data foundation for all subsequent technical stages, ensuring data traceability and reusability. Next, six-dimensional semantic features of disputed events are extracted and semantic vectors are constructed, transforming unstructured and semi-structured dispute information into a computable and comparable vector form, providing data support for case matching and intelligent decision-making. Then, a dedicated intelligent agent cluster is built, combined with a collaborative scheduler to decompose and allocate sub-tasks throughout the arbitration process. Arbitration intelligent agents resolve operational conflicts, ensuring efficient collaboration and seamless information flow across all stages. A multi-dimensional semantic alignment algorithm is used to achieve accurate matching between current disputes and historical similar cases, providing reference for arbitrators and improving consistency between adjudication and arbitration. Through a directed acyclic graph of tasks and a distributed execution engine, orderly task scheduling and automatic rollback in case of failure are achieved, ensuring process stability. Finally, by combining the knowledge-enhanced big data model with standardized document templates, the system can autonomously generate L4-level documents that meet the standards, reducing the repetitive workload of arbitrators while retaining the step of review and modification by arbitrators, thus balancing automation and professionalism.

[0024] Example 2: The multi-source labor dispute related data includes information on the parties to the labor dispute, arbitration requests, statements of facts, evidentiary materials, historical case rulings, legal provisions, arbitration process records, court hearing audio and video recordings, and feedback from the parties; the evidentiary materials include labor contracts, wage payment vouchers, attendance records, work permits, colleague testimonies, notices of termination of labor contracts, and overtime application forms, which meet the requirements of the labor arbitration evidence rules; The expression for the unified dispute event model is: ,in As the unique identifier for the disputed event, For the precise timestamp of the submission or occurrence of the disputed event, The source of the event specifically includes at least one of the following: arbitration institution, parties involved, lawyers, or courts. The core content of the dispute can be divided into structured fields or unstructured text. This is a list of context event IDs associated with the current disputed event, used to trace the context of disputed events and support information exchange in the process of adjudication and litigation.

[0025] By adopting the above technical solutions, multi-source data covers various types of information related to the entire process of labor dispute resolution. This includes core claims such as basic information of the parties and arbitration requests, as well as supporting data such as evidence materials, historical similar cases, and laws and regulations. It also covers procedural data such as court hearing audio and video recordings and process records. Among them, the evidence materials strictly adhere to the rules of evidence presentation in labor arbitration, ensuring the legality and validity of the data and providing a high-quality data source for subsequent semantic extraction and similar case matching. The unified dispute event model, by setting key elements such as unique identifiers, timestamps, event sources, core content, and context-related ID lists, structurally integrates scattered dispute event information, achieving unique identification, accurate tracing, and contextual association of dispute events. This solves the problems of disorganized and low-relevance information in traditional dispute events, providing technical support for information exchange in the arbitration-adjudication process and ensuring that the dispute information obtained by different parties at different stages is consistent and complete.

[0026] Example 3, in step 2, the specific extraction methods for the six-dimensional semantic features are as follows: Key Controversial Point O: A deep learning-based entity recognition and key point extraction model is used to extract core controversial points from the content of labor dispute events. These core controversial points refer to the core focus of the labor dispute. An embedding algorithm is then used to transform the extracted controversial points into vectors. ; The parties' claims (T) are categorized and analyzed using a lightweight classification model. This categorizes and clarifies the parties' core demands and factual basis, and the results are then transformed into vectors through embedding. ; Evidence Type S: Using an evidence classification algorithm, various types of evidence submitted by the parties are categorized. The classification results are then vectorized based on the weight of each piece of evidence's probative value to obtain a vector. It is compatible with the requirements of the evidence list in labor arbitration; Disputed Subject Matter X: The amount and type of the subject matter of the labor dispute are quantified and coded. Combined with the relevant legal basis for the subject matter, the quantification and coding results are vectorized to obtain a vector. ; Dispute Duration C: Based on the collected dispute process records, calculate the duration from the occurrence of the dispute to its submission to arbitration, and the current progress time of the arbitration process. Vectorize the duration information to obtain a vector. This provides support for process scheduling; Semantic Text D: A BERT-like pre-trained model is used to perform semantic embedding on unstructured text, which includes statements of disputed facts, courtroom testimonies, and descriptions of evidence. Semantic relevance features are extracted from the text to obtain vectors. ; Six-dimensional controversial semantic vector The expression is: The dimensions of all vectors are unified.

[0027] By adopting the above technical solutions, for the dimension of disputed points, a deep learning-driven entity recognition and key point extraction model is used to accurately capture the core focus of labor disputes. Then, an embedding algorithm transforms the extracted key points into standardized vectors, ensuring the computability of the disputed points. For the dimension of party claims, a lightweight classification model is used to categorize and organize the parties' arbitration requests and factual claims, clarifying the core demands and factual basis. Then, vector embedding is used to quantify the claim information. For the dimension of evidence type, an evidence classification algorithm is used to standardize the classification of various types of evidence. Combined with the weight of evidentiary value, vectorization is performed to adapt to the requirements of the labor arbitration evidence list, ensuring the standardization and validity of evidentiary information. For the dimension of disputed subject matter, the amount and type of subject matter are quantitatively encoded and vectorized based on relevant legal grounds, achieving accurate quantification of subject matter information. For the dimension of dispute duration, arbitration process records are linked to calculate the dual duration from the occurrence of the dispute to the submission of arbitration and the current progress of the arbitration process. Vectorization provides a time-dimensional reference for process scheduling. For the semantic text dimension, a BERT-like pre-trained model is used to semantically embed various types of unstructured text, extracting semantic association features from the text and converting them into semantic vectors to capture deep information in unstructured text. Finally, the vectors of the six dimensions are standardized and unified to construct a complete six-dimensional semantic vector of disputes, ensuring that the vectors of each dimension can collaboratively participate in subsequent calculations.

[0028] Example 4, in step 3, the intelligent agent cluster includes intelligent agents for confirming labor relations, intelligent agents for pursuing labor remuneration, intelligent agents for illegal termination of labor contracts, intelligent agents for disputes over overtime pay, and intelligent agents for work injury compensation; The dispute identification module adopts a hybrid architecture that combines labor dispute domain rules with a lightweight classification model. Domain rules are used to match known typical labor dispute scenarios, evidence rules and adjudication norms, while the lightweight classification model is used to identify complex and atypical labor disputes, together achieving accurate judgment of dispute type and key points. The collaborative scheduler uses a directed acyclic graph to model the entire process of labor dispute resolution, clarifying the dependencies between sub-tasks such as case filing, case assignment, court hearing, and document generation. Arbitration agents are used to resolve operational conflicts arising from multiple agents dealing with the same disputed case or the same process node. The conflict resolution method involves the arbitration agent calculating the impact score of each conflicting operation. , ;in , , For preset weighting coefficients, As for operation priority, To determine the scope of cases and process progress affected by the operation, Based on the estimated operation time, the operation with the highest impact score I is selected for execution, and the remaining conflicting operations are placed in a waiting queue. They are then evaluated and executed in turn after the current operation is completed.

[0029] By adopting the above technical solution, the intelligent agent cluster is professionally divided according to common types of labor disputes. Each intelligent agent focuses on the handling of a specific type of dispute and has corresponding dispute identification, decision-making, and execution capabilities, achieving professional and refined dispute handling. The dispute identification module adopts a hybrid architecture combining domain rules and a lightweight classification model. Domain rules are used to quickly match known typical labor dispute scenarios, evidence rules, and adjudication norms to improve identification efficiency; the lightweight classification model is used to identify complex and atypical labor disputes, making up for the limitations of domain rules. The two work together to achieve accurate judgment of dispute type and key points.

[0030] The collaborative scheduler uses a directed acyclic graph to model the entire process, clearly defining the dependencies between sub-tasks such as case filing, case assignment, court hearings, and document generation, ensuring the rationality of task allocation and the orderly progress of the process. The arbitration agent, as the core of conflict coordination, addresses operational conflicts arising from multiple agents in the same case and at the same process node. By constructing an impact score calculation model, it comprehensively considers factors such as operation priority, case impact scope, process progress, and operation time consumption to quantify the impact of each conflicting operation. The operation with the highest impact score is selected for priority execution, while other conflicting operations are placed in a waiting queue and executed sequentially after the current operation is completed. This ensures the orderly progress of the multi-agent collaboration process and avoids process delays or errors caused by operational conflicts.

[0031] Example 5, in step 4, each historical case in the historical case library All are represented as: ; where Vei is the six-dimensional semantic vector of the dispute corresponding to the historical case, Solutioni is the specific adjudication approach and handling process of the corresponding historical case, including the filing standards, the basis for case assignment, the focus of the trial, the basis for the adjudication, and the application of legal provisions, and Resulti is the handling result of the corresponding historical case, including the adjudication result, the appeal situation, the consistency between the adjudication and the judgment, and the satisfaction score of the parties. Current Labor Dispute Case F and Historical Cases The formula for calculating similarity is: ,in For cosine similarity, Let F be the vector of the current labor dispute event in dimension d. For historical cases A vector in dimension d; The rules for matching and recommending similar cases are as follows: Select the top-K similar cases with the highest similarity. If there are similar cases with a similarity exceeding a preset threshold, prioritize recommending such cases and their adjudication reasoning and applicable legal basis, and focus on recommending cases where the adjudication and judgment are consistent. If the similarity of all similar cases is lower than the threshold θ, then integrate the adjudication reasoning and handling process of the top-K similar cases to generate a reference solution suitable for the current labor dispute and recommend it to the arbitrators for reference.

[0032] By adopting the above technical solution, each case in the historical case database is first structured and modeled, transforming each historical case into standardized data containing a six-dimensional semantic vector of the dispute, the adjudication reasoning and handling process, and the handling result. This ensures the completeness and comparability of the historical case information. The adjudication reasoning and handling process cover core aspects such as case filing standards and case assignment basis, while the handling result includes key information such as the adjudication result and appeal status, providing comprehensive support for subsequent reference. Similarity calculation employs a multi-dimensional semantic alignment algorithm. By assigning weights to the six-dimensional semantic features, the cosine similarity between the current dispute and each historical case in each dimension is calculated. Then, a weighted summation is used to obtain the overall similarity, taking into account the importance of each dimension's features and improving the accuracy of similarity calculation. Case matching and recommendation follow differentiated rules, prioritizing the selection of the Top-K cases with the highest similarity. If there are cases with similarity exceeding a preset threshold, these cases, along with their adjudication reasoning and applicable legal basis, are recommended for recommendation, with a focus on cases where the adjudication and judgment are consistent to ensure reference value. If the similarity of all cases is below the threshold, the core adjudication reasoning and handling process of the Top-K cases are integrated to generate a personalized reference solution adapted to the current dispute. This not only solves the problem of lacking highly similar cases but also provides arbitrators with scientific and reasonable decision-making support.

[0033] In Example 6, step 5, the nodes of the directed acyclic graph of the labor dispute resolution task are specifically defined as follows: The isolation node is used to temporarily isolate cases with dispute risks or incomplete materials, and notify the parties to supplement the materials; the case filing review node is used to automatically review the arbitration applications and evidence materials submitted by the parties, determine whether they meet the filing conditions, and if the review is successful, proceed to the case assignment stage; if the review fails, supplementary opinions are provided to ensure compliance with labor arbitration case filing regulations; the case assignment node is used to automatically assign cases to corresponding arbitrators based on the dispute type, dispute subject matter, and arbitrator's professional area; the trial assistance node is used to assist arbitrators in completing trial preparation, trial recording, evidence presentation, and cross-examination guidance, and to synchronize trial information to relevant intelligent agents; the document generation node is used to automatically generate relevant documents based on the trial results and similar case references; the verification node is used to check the content and process compliance of the documents and determine whether the resolution process is complete; each node includes specific operation instructions, timeout time, and rollback instructions. When a node times out or fails to execute, a rollback instruction is triggered.

[0034] By adopting the above technical solution, each node of the directed acyclic graph corresponds to a key stage in the entire arbitration process. Each node has a clearly defined functional positioning and operational specifications, aligning with the requirements of the labor arbitration process. The isolation node is used to screen cases with potential disputes or incomplete materials. By temporarily isolating these cases and notifying the parties to supplement their materials, it ensures that cases entering subsequent stages have complete materials and no obvious risks. The case filing review node automates the review of arbitration applications and evidence materials, strictly adhering to the labor arbitration case filing regulations to determine whether they meet the filing conditions. If the review is successful, the case proceeds to the case assignment stage; if it fails, specific supplementary opinions are provided, improving the efficiency and standardization of the case filing review. The case assignment node automatically assigns cases based on the type of dispute, the subject matter of the dispute, and the arbitrator's area of ​​expertise, ensuring the rationality and professionalism of case assignment and improving processing efficiency. The trial assistance node provides arbitrators with full-process trial support, covering trial preparation, recording, evidence presentation, and cross-examination guidance, while simultaneously synchronizing trial information to relevant agents to ensure information synchronization. The document generation node automatically connects to the document generation module based on the trial results and similar case references to complete the initial generation of relevant documents. The verification node is responsible for checking the accuracy of document content, the compliance of processes, and determining whether the handling process is complete, ensuring that the entire handling process complies with regulations. Each node is pre-set with specific operation instructions, timeout periods, and rollback instructions. When a node times out or fails to execute, the rollback instruction is automatically triggered, returning to the previous normal node for re-execution, thus avoiding process interruption or non-compliant handling.

[0035] In Example 7, step 6, the L4 level labor dispute document conforms to the standard for labor arbitration documents.

[0036] By adopting the above technical solution, the L4 level labor dispute documents, as the core output generated autonomously by the system, must strictly adhere to the norms of labor arbitration documents, covering all aspects such as document format, content, and language, to ensure the legality, standardization, and rigor of the documents.

[0037] By integrating knowledge in the field of labor disputes with document norms through a knowledge-enhancing big data model, and combining standardized document templates, the system automatically integrates relevant information such as court hearing results, similar case references, and legal basis to generate documents that are complete in content, standardized in format, and professional in language. At the same time, it reserves space for arbitrators to review and modify the documents, thereby automating document generation and ensuring that the quality of the documents meets the requirements of arbitration work, avoiding process delays or disputes caused by non-standard documents.

[0038] Example 8 also includes step 7, which involves collecting explicit and implicit feedback data to optimize semantic alignment weights and the overall model. The feedback data includes: explicit feedback such as the arbitrator's acceptance or rejection of the case recommendations in step 4, the score for the accuracy of case matching, the content of the modifications made to the documents generated in step 6, and the satisfaction score; and implicit feedback such as the duration of labor dispute resolution, the case appeal rate, the consistency rate between arbitration and adjudication, the adoption rate of the recommended case scheme, and the proportion of document modifications. The focus is on the consistency rate between arbitration and adjudication and the appeal rate to optimize the model.

[0039] By adopting the above technical solutions, the feedback data encompasses two main categories: explicit and implicit feedback, comprehensively capturing key information such as arbitrators' operational preferences and system performance. Explicit feedback comes directly from arbitrators' proactive operations and evaluations, including their acceptance or rejection of similar case recommendations, ratings of similar case matching accuracy, and ratings of the content and satisfaction with the generated documents. This directly reflects the effectiveness and shortcomings of each stage of the system. Implicit feedback comes from objective data during system operation, including labor dispute resolution time, case appeal rate, arbitration-adjudication consistency rate, similar case recommendation adoption rate, and document modification ratio. Among these, the arbitration-adjudication consistency rate and appeal rate serve as core optimization indicators, directly reflecting the system's support effect on arbitration and adjudication work. Through the analysis and mining of this feedback data, the weight parameters of the semantic alignment algorithm are adjusted, and the training data and model structure of the knowledge augmentation model are optimized to continuously improve the accuracy of similar case matching, the quality of document generation, and the efficiency of process scheduling, ensuring that the system can adapt to the ever-changing needs of labor dispute resolution and arbitration and adjudication standards.

[0040] Example 9: A full-process intelligent arbitration system for labor disputes based on multi-agent coordination, used to realize a full-process intelligent arbitration method for labor disputes based on multi-agent coordination, characterized by including a data acquisition and fusion module, a semantic feature extraction module, an agent cluster module, a collaborative scheduling and arbitration module, a semantic alignment and case matching module, a distributed disposal and execution module, a document autonomous generation module, and a feedback optimization module; The data acquisition and fusion module constructs a unified labor dispute data lake and completes the modeling of disputed events; the semantic feature extraction module generates a six-dimensional dispute semantic vector; the intelligent agent cluster module provides dedicated intelligent agents and multi-agent collaborative capabilities; the semantic alignment and case matching module realizes high-precision matching and push of similar cases; the distributed disposal execution module schedules task execution and supports rollback; the document autonomous generation module generates L4 level labor dispute documents; and the feedback optimization module realizes continuous optimization of system parameters and models.

[0041] By adopting the above technical solutions, each module of the system performs its respective function and works in concert to form a complete intelligent arbitration system for the entire labor dispute process. The data acquisition and fusion module is responsible for collecting, integrating, and standardizing multi-source labor dispute data, constructing a unified labor dispute data lake, and completing standardized modeling of disputed events, providing data support for the entire system. The semantic feature extraction module receives the output from the data acquisition and fusion module, generating standardized six-dimensional dispute semantic vectors according to the extraction rules of six-dimensional semantic features, providing core data for case matching and intelligent decision-making. The intelligent agent cluster module is equipped with various specialized intelligent agents, providing professional dispute identification, decision-making, and execution capabilities, while supporting collaborative operation between multiple intelligent agents. The collaborative scheduling and arbitration module is responsible for the decomposition, allocation, and conflict resolution of sub-tasks throughout the process, achieving orderly task scheduling through a collaborative scheduler and resolving operational conflicts through arbitration intelligent agents. The semantic alignment and case matching module uses a multi-dimensional semantic alignment algorithm to achieve accurate matching and scientific recommendation of current disputes and historical similar cases, improving the consistency of adjudication and arbitration. The distributed execution module, based on a directed acyclic graph of tasks, schedules agents and modules to execute relevant tasks, while providing automatic rollback functionality for task failures to ensure process stability. The document autonomous generation module, combining a knowledge-enhanced model with document templates, autonomously generates L4-level labor dispute documents, supporting automated document generation. The feedback optimization module collects explicit and implicit feedback data to continuously optimize system parameters, semantic alignment weights, and the large model, ensuring continuous improvement in system performance and enabling the efficient and stable implementation of intelligent arbitration methods.

[0042] The following specific embodiments illustrate the implementation principle of the present invention: This embodiment uses a labor and personnel dispute arbitration committee in a prefecture-level city as the application subject. It fully implements a multi-agent coordinated intelligent arbitration method and system for the entire labor dispute process, addressing labor remuneration disputes between enterprises and workers within its jurisdiction. This achieves intelligent handling of the entire process from case submission to document effectiveness, improving arbitration efficiency and ensuring consistency between arbitration and adjudication. The specific implementation process is as follows: A smart arbitration system for the entire labor dispute process based on multi-agent coordination was built, including a data collection and fusion module, a semantic feature extraction module, an agent cluster module, a collaborative scheduling and arbitration module, a semantic alignment and case matching module, a distributed disposal and execution module, a document autonomous generation module, and a feedback and optimization module. All modules work together to ensure a closed loop of functions throughout the entire process.

[0043] Multi-source labor dispute data is collected to construct a unified labor dispute data lake. The data scope includes: information on parties involved in labor disputes within the jurisdiction over the past five years, arbitration requests, statements of fact, evidentiary materials, historical case rulings, labor laws and regulations, arbitration process records, court hearing audio and video recordings, and party feedback. Simultaneously, structured modeling is performed on all historical cases, initializing a six-dimensional dispute semantic vector extraction model, a multi-dimensional semantic alignment algorithm, a knowledge-enhanced large-scale model, and intelligent agent cluster parameters, and pre-setting conflict resolution weight coefficients for arbitration intelligent agents. =0.4、 =0.3、 =0.3, similarity threshold θ=0.8, Top-K value is 5.

[0044] Intelligent agent cluster configuration: Build a dedicated intelligent agent cluster, including intelligent agents for confirming labor relations, intelligent agents for recovering unpaid wages, intelligent agents for illegal termination of labor contracts, intelligent agents for disputes over overtime pay, and intelligent agents for work injury compensation. Among them, the intelligent agent for recovering unpaid wages is the core intelligent agent in this embodiment. Its dispute identification, decision-making and execution capabilities are optimized to meet the needs of handling disputes related to recovering unpaid wages.

[0045] Step 1: Multi-source data collection and standardized modeling of controversial events; Worker Zhang submitted an arbitration application to the arbitration committee, claiming that the technology company he worked for owed him wages for January to March 2026, totaling 15,000 yuan. He also submitted his employment contract, bank statements showing his wages from October to December 2025, attendance records, and WeChat communication records with the company's HR department as evidence.

[0046] The data acquisition and fusion module automatically collects the relevant data from the arbitration application, and simultaneously retrieves the technology company's business registration information, past labor dispute records, and relevant labor laws and regulations, integrating them into a unified labor dispute data lake. It also performs standardized modeling of the dispute event, generating a unique identifier for the dispute event. =20260327001, timestamp =2026-03-27 14:30:00, Source of the event =The parties involved, core content =A technology company owes Mr. Zhang 15,000 yuan in wages for January to March 2026. He requests payment of the owed wages. (Context-related ID list) =[], to complete the standardized modeling of the disputed event and provide a standardized data foundation for subsequent steps.

[0047] Step 2: Extraction of six-dimensional controversial semantic features and construction of semantic vectors; The semantic feature extraction module extracts six-dimensional semantic features and constructs a six-dimensional semantic vector for this controversial event. The specific extraction process is as follows: Point of contention O: Using a deep learning-based entity recognition and key point extraction model, the core point of contention extracted from the arbitration application and evidence materials is that a technology company owes Mr. Zhang 15,000 yuan in wages for January to March 2026, and Mr. Zhang requests payment of the owed wages. This is then transformed into a vector using an embedding algorithm. .

[0048] The party concerned, T, employed a lightweight classification model to categorize and analyze Zhang's arbitration requests and factual statements, clarifying that his core claim was for payment of 15,000 yuan in unpaid wages. The factual basis for this was the existence of an employment relationship between the two parties and the company's failure to pay wages for January to March 2026 on time. The analysis results were then transformed into vectors through embedding. .

[0049] Evidence Type S: The evidence submitted by Zhang is classified using an evidence classification algorithm. The labor contract, bank statements of salary, and attendance records are the core evidence, while WeChat communication records are the auxiliary evidence. Combining the probative weight of each type of evidence, the evidence is vectorized to obtain vector VS, which is adapted to the requirements of the labor arbitration evidence list.

[0050] Disputed Subject Matter X: The disputed amount of 15,000 yuan is quantified and coded. The type of subject matter is labor remuneration recovery. Combining the relevant legal basis in the Labor Contract Law regarding wage arrears, the quantification and coding results are vectorized to obtain a vector. .

[0051] Dispute Duration C: Based on the collected dispute process records, the duration from the occurrence of the dispute to the submission of arbitration is calculated to be 7 days. The current arbitration process progress time is 0. The duration information is vectorized to obtain a vector. This provides a time-based reference for process scheduling.

[0052] Semantic Text D: Using a BERT-like pre-trained model, semantic embedding is performed on Zhang's factual statements, WeChat communication records, and other unstructured texts to extract semantic association features from the text, resulting in a vector. .

[0053] By standardizing and unifying the vectors of the above six dimensions, a complete six-dimensional semantic vector of controversy is constructed. This ensures that vectors of each dimension can work together to participate in subsequent case matching and intelligent decision-making.

[0054] Step 3: Intelligent agent cluster collaboration and full-process sub-task allocation; The collaborative scheduler uses a directed acyclic graph to model the entire process of this labor remuneration claim dispute, clarifying the dependencies between the sub-tasks of case filing, case assignment, court hearing, and document generation, and then decomposes the sub-tasks and assigns them to the corresponding intelligent agents: Dispute Identification: The dispute identification module adopts a hybrid architecture that combines labor dispute domain rules with a lightweight classification model. Domain rules match typical scenarios of wage arrears, while the lightweight classification model further confirms that the dispute is a simple wage recovery dispute without complex details. The wage recovery agent takes the lead in handling the dispute, while the labor relationship confirmation agent assists in verifying the authenticity of the labor relationship.

[0055] Subtask allocation: The collaborative scheduler allocates the case filing review task to the labor remuneration recovery agent, the case assignment task to the labor remuneration recovery agent, the court hearing assistance task to the labor remuneration recovery agent and the court hearing assistance related sub-agents, and the document generation task to the labor remuneration recovery agent and the document generation sub-agent.

[0056] Conflict Resolution: During task allocation, the agent for pursuing unpaid wages and the agent for confirming employment relationships simultaneously request access to the technology company's employment relationship registration information, resulting in an operational conflict. The arbitration agent initiates the conflict resolution mechanism, calculating the impact score I of the two conflicting operations: the priority of accessing employment relationship registration information P=0.8, the scope of the affected case R=1, and the estimated time T=5 minutes. =0.4×0.8+0.3×1+0.3×5=2.02; Priority P=0.6 for temporarily not retrieving and verifying through existing labor contracts, impacting the scope of the case R=1, estimated time T=2 minutes. =0.4×0.6+0.3×1+0.3×2=1.14. The arbitration agent selects the operation with the higher impact score to execute, while the other operation enters the waiting queue. After the current operation is completed, a supplementary verification process is executed based on the retrieved labor relations filing information to ensure no operational conflicts and achieve full-process collaboration and information flow.

[0057] Step 4: Case matching and recommendation; The semantic alignment and case matching module uses a multi-dimensional semantic alignment algorithm to calculate the weighted cosine similarity between the current dispute and all labor remuneration recovery cases in the historical case database.

[0058] Each historical case in the historical case library All are represented as , where Vei is the six-dimensional semantic vector of the historical case, Solutioni is the adjudication approach and handling process, and Resulti is the handling result.

[0059] The formula for calculating similarity is: Where λd is the weight of each dimension, Let d be the cosine similarity between the current controversy and historical similar cases.

[0060] After calculation, the top 5 historical cases with the highest similarity were selected. One of these historical cases had a similarity of 0.85, exceeding the preset threshold of 0.8. This case involved a labor remuneration dispute between Mr. Li and a trading company. The ruling was based on the premise that the labor relationship was legal and that the company was in arrears with wages. According to Article 30 of the Labor Contract Law, the court ruled that the company should pay the arrears. The outcome was that the company paid the arrears, and neither party appealed. The ruling was consistent with the judgment, and the parties' satisfaction score was 90%.

[0061] According to the case matching and recommendation rules, the highly similar historical cases, their adjudication reasoning, and the legal basis for application are recommended first. At the same time, the other four Top-K cases are recommended as supplementary references for arbitrators to review and refer to, so as to improve the consistency between adjudication and arbitration.

[0062] Step 5: Generating the directed acyclic graph for the task and executing the task; Based on the dispute resolution requirements, the distributed resolution execution module generates a directed acyclic graph of labor dispute resolution tasks. The definitions and execution processes of each node are as follows: Isolation Node: Conduct a risk assessment of the dispute, confirm that the evidence submitted by Zhang is complete and there is no obvious risk of dispute, and no temporary isolation is required. Proceed directly to the next node. Case filing review stage: The intelligent agent for recovering unpaid wages automatically reviews the arbitration application and evidence materials submitted by Zhang. By comparing them with the labor arbitration case filing norms, it confirms that the applicant is qualified, the respondent is clear, the arbitration request is specific, there are clear facts and reasons, and the evidence is sufficient. Once the review is passed, the case will proceed to the case assignment stage.

[0063] Case assignment node: The intelligent agent for recovering unpaid wages automatically assigns the case to arbitrator Wang based on the type of dispute, the subject matter of the dispute, and the arbitrator's area of ​​expertise.

[0064] Trial Assistance Node: The trial assistance sub-agent assists arbitrator Wang in completing trial preparation, trial recording, evidence presentation, and cross-examination guidance. At the same time, it synchronizes the trial information to the labor remuneration recovery agent, semantic alignment, and case matching module to ensure information synchronization.

[0065] Document generation node: The document generation sub-agent automatically connects to the document generation module based on the court hearing results and similar case references to prepare for the generation of the judgment.

[0066] Verification Node: The verification node checks the accuracy of the document content and the compliance of the process, confirms the completeness of the processing procedure, and then proceeds to the final stage of document generation.

[0067] Throughout the entire node execution process, each node has pre-set specific operation instructions, timeout time, and rollback instructions. During this implementation, all nodes executed normally without any timeouts or execution failures, and there was no need to trigger rollback instructions.

[0068] Step 6: Autonomous generation, review, and modification of Level 4 documents; The document self-generation module combines a knowledge-enhanced big data model with a labor dispute arbitration award template, integrating relevant information such as court hearing results, similar case references, and legal basis to autonomously generate an L4 level labor dispute arbitration award. This document strictly follows the norms of labor arbitration documents, covering core contents such as basic information of the parties, arbitration request, facts and reasons, evidence determination, basis for the award, award result, and the time limit for filing a lawsuit if dissatisfied with the award. It clearly awards a technology company to pay Mr. Zhang 15,000 yuan in back wages within 10 days from the date the award takes effect.

[0069] Arbitrator Wang reviewed the generated award, confirming that the document format was standardized, the content was accurate, and the legal provisions were correctly applied. He only modified some wording details. After the review was approved, a formal award was generated and served on both parties.

[0070] Step 7: Feedback data collection and system optimization; 1. Explicit Feedback: Arbitrator Wang fully adopted the similar case recommendations, scored 9 points for the accuracy of similar case matching, modified 5% of the generated documents, and scored 9 points for satisfaction; Party Zhang was satisfied with the ruling and had no objection.

[0071] 2. Implicit feedback: The case was processed in 7 working days, there was no appeal, the consistency rate of the ruling and the trial was 100%, the adoption rate of the recommended scheme for similar cases was 100%, and the document modification rate was 5%.

[0072] The feedback optimization module analyzes and mines the above feedback data, appropriately adjusts the weights of disputed points and evidence types in the semantic alignment algorithm, optimizes the training data of the knowledge-enhanced big data model, further improves the accuracy of case matching and the quality of document generation, and achieves continuous system optimization.

[0073] This implementation, through a multi-agent coordinated intelligent arbitration method and system for the entire labor dispute process, achieved intelligent handling of labor remuneration claims disputes throughout the entire process. The case handling time was shortened from an average of 15 working days to 7 working days, improving handling efficiency by 53%. The accuracy of similar case matching reached 85%, and the consistency rate of arbitration and adjudication reached 100%, effectively avoiding inconsistent rulings for similar cases. The automation rate of document generation reached 95%, significantly reducing the repetitive workload of arbitrators. At the same time, the satisfaction rate of the parties reached over 90%, achieving a dual improvement in efficiency and quality, fully demonstrating the practicality and superiority of this method and system in labor dispute arbitration.

[0074] The intelligent arbitration system adopted in this implementation utilizes a collaborative approach across its various modules: the data acquisition and fusion module ensures the integration and standardization of multi-source data; the semantic feature extraction module enables the quantitative transformation of disputed information; the intelligent agent cluster module provides professional and collaborative handling capabilities; the collaborative scheduling and arbitration module ensures the rationality of task allocation and the effectiveness of conflict resolution; the semantic alignment and case matching module enhances consistency between adjudication and arbitration; the distributed handling and execution module ensures the stability of the process; the document autonomous generation module enables automated document generation; and the feedback optimization module enables continuous system iteration. The organic integration of these modules ensures the successful implementation of the full-process intelligent arbitration method.

[0075] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A fully intelligent arbitration method for labor disputes based on multi-agent coordination, characterized in that: Includes the following steps: Step 1: Collect multi-source labor dispute related data, construct a unified labor dispute data lake, and perform standardized modeling of dispute events; Step 2: Extract the six-dimensional semantic features of the disputed event, including the key points of the dispute, the claims of the parties, the type of evidence, the subject matter of the dispute, the duration of the dispute, and the semantic text, and construct a six-dimensional semantic vector of the dispute. Step 3: Construct a dedicated intelligent agent cluster. Each intelligent agent has the ability to identify disputes, make decisions and execute them. The entire process of case filing, case assignment, court hearing and document generation is broken down and assigned by the collaborative scheduler. The arbitration intelligent agent resolves operational conflicts and realizes full-process collaboration and information flow. Step 4: Using a multi-dimensional semantic alignment algorithm, calculate the weighted cosine similarity between the current dispute and historical similar cases, match and push high-similarity historical similar cases for arbitrators to refer to, and improve the consistency between adjudication and arbitration. Step 5: Generate a directed acyclic graph of the tasks to be processed, which is then executed by the distributed execution engine, supporting automatic rollback in case of task failure. Step 6: Combine the knowledge enhancement model with the labor dispute document template to generate an L4 level document, which will then be reviewed and modified by the arbitrator.

2. The intelligent arbitration method for the entire labor dispute process based on multi-agent coordination as described in claim 1, characterized in that, Data related to labor disputes from multiple sources includes information on the parties involved in the labor dispute, arbitration requests, statements of facts, evidentiary materials, historical case rulings, legal provisions, arbitration process records, audio and video recordings of court hearings, and feedback from the parties. Evidence includes labor contracts, wage payment vouchers, attendance records, work permits, colleague testimonies, notices of termination of labor contracts, and overtime application forms, which meet the requirements of the rules of evidence for labor arbitration. The expression for the unified dispute event model is: ,in As the unique identifier of the disputed event, For the precise timestamp of the submission or occurrence of the disputed event, The source of the event specifically includes at least one of the following: arbitration institution, parties involved, lawyers, or courts. The core content of the dispute can be divided into structured fields or unstructured text. This is a list of context event IDs associated with the current disputed event, used to trace the context of disputed events and support information exchange in the process of adjudication and litigation.

3. The intelligent arbitration method for the entire labor dispute process based on multi-agent coordination as described in claim 2, characterized in that, In step 2, the specific extraction methods for the six-dimensional semantic features are as follows: Key Controversial Point O: A deep learning-based entity recognition and key point extraction model is used to extract core controversial points from the content of labor dispute events. These core controversial points refer to the core focus of the labor dispute. An embedding algorithm is then used to transform the extracted controversial points into vectors. ; The parties' claims (T) are categorized and analyzed using a lightweight classification model. This categorizes and clarifies the parties' core demands and factual basis, and the results are then transformed into vectors through embedding. ; Evidence Type S: Using an evidence classification algorithm, various types of evidence submitted by the parties are categorized. The classification results are then vectorized based on the weight of each piece of evidence's probative value to obtain a vector. It is compatible with the requirements of the evidence list in labor arbitration; Disputed Subject Matter X: The amount and type of the subject matter of the labor dispute are quantified and coded. Combined with the relevant legal basis for the subject matter, the quantification and coding results are vectorized to obtain a vector. ; Dispute Duration C: Based on the collected dispute process records, calculate the duration from the occurrence of the dispute to its submission to arbitration, and the current progress time of the arbitration process. Vectorize the duration information to obtain a vector. This provides support for process scheduling; Semantic Text D: A BERT-like pre-trained model is used to perform semantic embedding on unstructured text, which includes statements of disputed facts, courtroom testimonies, and descriptions of evidence. Semantic relevance features are extracted from the text to obtain vectors. ; Six-dimensional controversial semantic vector The expression is: The dimensions of all vectors are unified.

4. The intelligent arbitration method for the entire labor dispute process based on multi-agent coordination as described in claim 3, characterized in that, In step 3, the intelligent agent cluster includes intelligent agents for confirming labor relations, intelligent agents for pursuing unpaid wages, intelligent agents for illegal termination of labor contracts, intelligent agents for disputes over overtime pay, and intelligent agents for work injury compensation. The dispute identification module adopts a hybrid architecture that combines labor dispute domain rules with a lightweight classification model. Domain rules are used to match known typical labor dispute scenarios, evidence rules and adjudication norms, while the lightweight classification model is used to identify complex and atypical labor disputes, together achieving accurate judgment of dispute type and key points. The collaborative scheduler uses a directed acyclic graph to model the entire process of labor dispute resolution, clarifying the dependencies between sub-tasks such as case filing, case assignment, court hearing, and document generation. Arbitration agents are used to resolve operational conflicts arising from multiple agents dealing with the same disputed case or the same process node. The conflict resolution method involves the arbitration agent calculating the impact score of each conflicting operation. , ;in , , For preset weighting coefficients, As for operation priority, To determine the scope of cases and process progress affected by the operation, Based on the estimated operation time, the operation with the highest impact score I is selected for execution, and the remaining conflicting operations are placed in a waiting queue. They are then evaluated and executed in turn after the current operation is completed.

5. The intelligent arbitration method for the entire labor dispute process based on multi-agent coordination as described in claim 4, characterized in that, In step 4, each historical case in the historical case library All are represented as: ; where Vei is the six-dimensional semantic vector of the dispute corresponding to the historical case, Solutioni is the specific adjudication approach and handling process of the corresponding historical case, including the filing standards, the basis for case assignment, the focus of the trial, the basis for the adjudication, and the application of legal provisions, and Resulti is the handling result of the corresponding historical case, including the adjudication result, the appeal situation, the consistency between the adjudication and the judgment, and the satisfaction score of the parties. Current Labor Dispute Case F and Historical Cases The formula for calculating similarity is: ,in For cosine similarity, Let F be the vector of the current labor dispute event in dimension d. For historical cases A vector in dimension d; The rules for matching and recommending similar cases are as follows: Select the top-K similar cases with the highest similarity. If there are similar cases with a similarity exceeding a preset threshold, prioritize recommending such cases and their adjudication reasoning and applicable legal basis, and focus on recommending cases where the adjudication and judgment are consistent. If the similarity of all similar cases is lower than the threshold θ, then integrate the adjudication reasoning and handling process of the top-K similar cases to generate a reference solution suitable for the current labor dispute and recommend it to the arbitrators for reference.

6. The intelligent arbitration method for the entire labor dispute process based on multi-agent coordination as described in claim 5, characterized in that, In step 5, the nodes of the directed acyclic graph of the labor dispute resolution task are specifically defined as follows: The isolation node is used to temporarily isolate cases with dispute risks or incomplete materials, and notify the parties to supplement the materials; the case filing review node is used to automatically review the arbitration applications and evidence materials submitted by the parties, determine whether they meet the filing conditions, and if the review is successful, proceed to the case assignment stage; if the review fails, supplementary opinions are provided, in accordance with the labor arbitration case filing norms; the case assignment node is used to automatically assign cases to the corresponding arbitrators based on the dispute type, disputed subject matter, and arbitrator's professional area; the trial assistance node is used to assist arbitrators in completing trial preparation, trial recording, evidence presentation, and cross-examination guidance, and to synchronize trial information to relevant agents. The document generation node is used to automatically generate relevant documents based on the court hearing results and similar cases; the verification node is used to check the content of the documents and the compliance of the process, and to determine whether the handling process is complete; each node contains specific operation instructions, timeout time, and rollback instructions. When a node times out or fails to execute, the rollback instruction is triggered.

7. The intelligent arbitration method for the entire labor dispute process based on multi-agent coordination as described in claim 6, characterized in that, In step 6, the L4 level labor dispute documents conform to the standards for labor arbitration documents.

8. The intelligent arbitration method for the entire labor dispute process based on multi-agent coordination as described in claim 7, characterized in that, The process also includes step 7, which involves collecting explicit and implicit feedback data to optimize semantic alignment weights and the overall model. The feedback data includes: explicit feedback such as the arbitrator's acceptance or rejection of the case recommendations in step 4, the score for the accuracy of case matching, the content of the modifications made to the documents generated in step 6, and the satisfaction score; and implicit feedback such as the duration of labor dispute resolution, the case appeal rate, the consistency rate between arbitration and adjudication, the adoption rate of the recommended case scheme, and the proportion of document modifications. Special attention is paid to the consistency rate between arbitration and adjudication and the appeal rate to optimize the model.

9. A multi-agent coordination-based intelligent arbitration system for the entire labor dispute process, used to implement the multi-agent coordination-based intelligent arbitration method for the entire labor dispute process as described in claim 8, characterized in that, It includes a data acquisition and fusion module, a semantic feature extraction module, an intelligent agent cluster module, a collaborative scheduling and arbitration module, a semantic alignment and case matching module, a distributed disposal and execution module, a document autonomous generation module, and a feedback optimization module; The data acquisition and fusion module constructs a unified labor dispute data lake and completes the modeling of disputed events; the semantic feature extraction module generates a six-dimensional dispute semantic vector; the intelligent agent cluster module provides dedicated intelligent agents and multi-agent collaborative capabilities; the semantic alignment and case matching module realizes high-precision matching and push of similar cases; the distributed disposal execution module schedules task execution and supports rollback; the document autonomous generation module generates L4 level labor dispute documents; and the feedback optimization module realizes continuous optimization of system parameters and models.

Images