Intelligent evaluation method and device based on dispute point matching and dynamic weight fusion

Abstract

The embodiment of the invention provides an intelligent appraisal method and device based on dispute point matching and dynamic weight fusion, and effective establishment of association is realized through innovatively designing a dispute analysis system and through feature extraction and evidence matching. And a scheduling distribution mechanism is constructed, and a reliable evaluation strategy is established in combination with adaptation degree evaluation and weight calculation. And result fusion is introduced, and the accuracy of a conclusion is ensured through deviation processing and confidence evaluation. According to the method, the defects of the traditional technology in the aspects of dispute analysis, scheduling distribution, result fusion and the like are effectively overcome, and technical guarantee is provided for intelligent evaluation.

Description

Technical Field

[0001] This application relates to the field of data processing, specifically to an intelligent adjudication method and apparatus based on dispute point matching and dynamic weight fusion. Background Technology

[0002] Existing intelligent adjudication methods have significant shortcomings. Traditional systems perform poorly in analyzing points of contention and matching evidence, failing to effectively achieve feature correlation and thus affecting the accuracy of adjudication.

[0003] Furthermore, existing technologies suffer from bottlenecks in expert scheduling and weight allocation. Most systems lack robust adaptation assessment mechanisms and allocation strategies, leading to inadequate adjudication.

[0004] Existing systems have technical shortcomings in result fusion. They lack in-depth analysis of scoring biases and struggle to achieve efficient anomaly handling through weight optimization, thus affecting the credibility of conclusions. Solving these problems is crucial for improving intelligent adjudication capabilities. Summary of the Invention

[0005] To address the problems in existing technologies, this application provides an intelligent adjudication method and apparatus based on dispute point matching and dynamic weight fusion, which can effectively solve the shortcomings of traditional technologies in dispute analysis, scheduling and allocation, and result fusion, and provide technical support for intelligent adjudication.

[0006] To solve at least one of the above problems, this application provides the following technical solution: Firstly, this application provides an intelligent adjudication method based on dispute point matching and dynamic weight fusion, including: The system receives the initial dispute point data stream output by the large model, extracts the dispute point feature vectors, constructs a calculation index matrix, imports the influence data items, evidence association parameters, and rule matching factors into a preset calculation unit, generates a dispute point graded dataset, constructs a feature mapping matrix based on the graded dataset, performs data parsing and format conversion on the input evidence files, performs vector operations on the evidence feature data and the mapping matrix, generates an evidence annotation data stream based on the matching degree threshold, and outputs a data card containing dispute level identifiers and evidence association parameters. Based on the data card, expert processing parameters are obtained, and the adjudication records in the expert historical database are read. The quantitative indicators of dispute type processing and the result pass rate data are numerically calculated according to the configuration parameters to generate domain fit data. Based on the fit data and historical accuracy parameters, a weight allocation model is constructed. According to the allocation model, a task scheduling sequence and a weight coefficient matrix are generated, and a task data table containing allocation parameters and scheduling strategies is output. Based on the task data table, a scoring data stream is collected, and a standard deviation calculation unit is executed. Data items whose scoring deviation values ​​exceed a set threshold are marked as abnormal nodes. The weight coefficients of the abnormal nodes are corrected. Based on the corrected weight matrix, a data fusion operation is performed to generate conclusion confidence data. The task reassignment mechanism is triggered based on the confidence data, and a result dataset containing the fusion conclusion and confidence coefficient is output.

[0007] Furthermore, it also includes: obtaining large model output data from the data access terminal, parsing the data format type according to the data header identifier, establishing dispute point parsing rules, extracting dispute point text attributes according to a preset feature encoding scheme, performing word segmentation on the dispute point text, constructing a feature mapping table, and generating a dispute point feature set containing topic word vectors and semantic structures according to the feature mapping table; Based on the feature set of disputed points, an index calculation rule is constructed. The influence parameter is discretely encoded according to a preset quantification standard. An evaluation score is generated according to the evidence correlation standard. Weight coefficients are configured according to the rule matching degree. The score and the weight coefficients are substituted into the calculation matrix. The disputed points are classified according to the calculation results to generate a disputed point dataset containing graded identifiers.

[0008] Furthermore, it also includes: constructing feature extraction rules based on hierarchical datasets, performing format recognition and type classification on evidence files, establishing a text preprocessing module, performing data format conversion according to a preset standardized process, performing word segmentation and vectorization processing on evidence content, constructing a feature dimension table, and generating a feature data stream containing evidence subject words and semantic vectors based on the feature dimension table; Based on the feature data stream, a mapping calculation rule is constructed, the evidence feature vector is normalized according to a preset standard, matching degree calculation parameters are generated according to the dispute level identifier, the labeling rule is configured according to the preset threshold condition, the calculation parameters are substituted into the mapping matrix, the evidence correlation degree is quantified and labeled according to the calculation result, and a data card containing the dispute level identifier and evidence correlation parameters is generated.

[0009] Furthermore, it also includes: constructing expert retrieval rules based on data cards, verifying access permissions to the expert historical database, establishing a data reading channel, obtaining adjudication records according to preset filtering conditions, classifying and statistically analyzing the record content, constructing a processing quantitative indicator table, and generating an expert basic dataset containing processing frequency and pass rate based on the quantitative indicator table. Based on the aforementioned expert dataset, a fitting degree calculation rule is constructed. The processing quantitative indicators are normalized according to a preset standard. Weight coefficients are generated based on the pass rate parameter. Calculation templates are configured according to the dispute type. The weight coefficients are substituted into the calculation unit. The expert fitting degree is quantitatively evaluated based on the calculation results, and fitting degree data containing domain matching degree and processing capability is generated.

[0010] Furthermore, it also includes: constructing weight calculation rules based on fit data, performing data standardization processing on historical accuracy parameters, establishing a parameter fusion module, combining indicators according to preset priority standards, validating the fusion results, constructing a weight calculation matrix, and generating an allocation model containing expert rating weights and task priorities based on the weight calculation matrix; Based on the allocation model, a scheduling rule set is constructed, the task sequence is divided according to a preset time window, a scheduling priority is generated according to the system load parameters, an allocation strategy is configured according to the task type, the priority is substituted into the scheduling matrix, and the task is configured with resources according to the allocation result, generating a task data table containing allocation parameters and scheduling strategies.

[0011] Furthermore, it also includes: constructing data collection rules based on the task data table, collecting and formatting expert scoring data in real time, establishing a deviation calculation module, performing standard deviation calculation according to preset statistical standards, comparing the scoring results with thresholds, constructing an anomaly marking matrix, and generating a dataset containing scoring deviation values ​​and anomaly nodes based on the anomaly marking matrix; Based on the dataset, a weight correction rule is constructed, the weight coefficients of abnormal nodes are adjusted according to a preset ratio, a correction matrix is ​​generated according to the adjustment parameters, a calculation template is configured according to the fusion calculation standard, the correction matrix is ​​substituted into the fusion unit, and the confidence of the conclusion is quantified according to the calculation result to generate a dataset containing the fusion conclusion and confidence.

[0012] Furthermore, it also includes: constructing redistribution triggering rules based on confidence data, determining the threshold of confidence values, establishing a task allocation controller, screening experts according to preset supplementary standards, performing status detection on allocable resources, constructing a task allocation matrix, and generating a redistribution scheme containing an expert supplementary list and allocation strategy based on the task allocation matrix. Based on the redistribution scheme, result synthesis rules are constructed, supplementary scoring data are fused according to preset standards, weight parameters are generated based on confidence coefficients, synthesis templates are configured according to conclusion types, the weight parameters are substituted into the calculation unit, the fusion conclusion is finally confirmed based on the calculation results, and a result dataset containing conclusion data and confidence coefficients is generated.

[0013] Secondly, this application provides an intelligent adjudication device based on dispute point matching and dynamic weight fusion, comprising: The dispute matching module is used to receive the initial dispute point data stream output by the large model, extract the dispute point feature vectors, construct the calculation index matrix, import the influence data items, evidence association parameters, and rule matching factors into the preset calculation unit, generate a dispute point hierarchical dataset, construct a feature mapping matrix based on the hierarchical dataset, perform data parsing and format conversion on the input evidence files, perform vector operations on the evidence feature data and the mapping matrix, generate an evidence annotation data stream according to the matching degree threshold, and output a data card containing dispute level identifiers and evidence association parameters. The weight fusion module is used to obtain expert processing parameters based on the data card, read the adjudication records in the expert historical database, perform numerical calculations on the dispute type processing quantitative indicators and result pass rate data according to the configuration parameters, generate domain fit data, construct a weight allocation model based on the fit data and historical accuracy parameters, generate a task scheduling sequence and weight coefficient matrix according to the allocation model, and output a task data table containing allocation parameters and scheduling strategies. The intelligent assessment module is used to collect scoring data streams based on the task data table, execute the standard deviation calculation unit, mark data items whose scoring deviation values ​​exceed a set threshold as abnormal nodes, perform parameter correction on the weight coefficients of the abnormal nodes, perform data fusion operations based on the corrected weight matrix, generate conclusion confidence data, trigger the task reassignment mechanism based on the confidence data, and output a result dataset containing the fusion conclusion and confidence coefficients.

[0014] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the intelligent adjudication method based on dispute point matching and dynamic weight fusion.

[0015] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the intelligent adjudication method based on dispute point matching and dynamic weight fusion.

[0016] Fifthly, this application provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the intelligent adjudication method based on dispute point matching and dynamic weight fusion.

[0017] As can be seen from the above technical solution, this application provides an intelligent adjudication method and apparatus based on dispute point matching and dynamic weight fusion. Through innovative design of the dispute analysis system, it achieves effective establishment of associations through feature extraction and evidence matching. A scheduling and allocation mechanism is constructed, combined with fit assessment and weight calculation, to establish a reliable adjudication strategy. Result fusion is introduced, and through bias processing and confidence assessment, the accuracy of the conclusions is ensured. This method effectively solves the shortcomings of traditional technologies in dispute analysis, scheduling and allocation, and result fusion, providing technical support for intelligent adjudication. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the intelligent adjudication method based on dispute point matching and dynamic weight fusion in the embodiments of this application; Figure 2 This is a structural diagram of the intelligent adjudication device based on dispute point matching and dynamic weight fusion in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the electronic device in the embodiments of this application.

[0020] Figure label: Electronic device 9600, central processing unit 9100, memory 9140, communication module 9110, input unit 9120, audio processor 9130, display 9160, power supply 9170, buffer memory 9141, application / function storage unit 9142, data storage unit 9143, driver storage unit 9144, antenna 9111, speaker 9131, microphone 9132. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] The acquisition, storage, use, and processing of data in this application comply with relevant laws and regulations.

[0023] In view of the problems existing in the prior art, this application provides an intelligent adjudication method and device based on dispute point matching and dynamic weight fusion. Through innovative design of the dispute analysis system, it achieves effective establishment of associations through feature extraction and evidence matching. A scheduling and allocation mechanism is constructed, combined with fit assessment and weight calculation, to establish a reliable adjudication strategy. Result fusion is introduced, and through bias handling and confidence assessment, the accuracy of the conclusions is ensured. This method effectively solves the shortcomings of traditional technologies in dispute analysis, scheduling and allocation, and result fusion, providing technical support for intelligent adjudication.

[0024] To effectively address the shortcomings of traditional technologies in dispute analysis, scheduling and allocation, and result fusion, and to provide technical support for intelligent adjudication, this application provides an embodiment of an intelligent adjudication method based on dispute point matching and dynamic weight fusion. See [link to relevant documentation]. Figure 1 The intelligent adjudication method based on dispute point matching and dynamic weight fusion specifically includes the following: Step S101: Receive the initial dispute point data stream output by the large model, extract the dispute point feature vector, construct the calculation index matrix, import the influence data item, evidence association parameters, and rule matching factors into the preset calculation unit, generate a dispute point graded dataset, construct a feature mapping matrix based on the graded dataset, perform data parsing and format conversion on the input evidence file, perform vector operations on the evidence feature data and the mapping matrix, generate an evidence annotation data stream according to the matching degree threshold, and output a data card containing dispute level identifier and evidence association parameters; First, this embodiment receives the initial dispute point data stream output by the large model, aligns the timestamp and source identifier, parses the text type according to the format identifier in the data header, and extracts the dispute point text field and context fragments. Then, the dispute point text is segmented according to a preset feature encoding scheme and mapped to a topic thesaurus and a rule thesaurus to obtain a dispute point feature vector containing topic vectors and rule trigger bits.

[0025] To ensure consistent calculations in the future, this embodiment assembles the influence data item, evidence association parameter placeholder, and rule matching factor placeholder into vector slots of the same dimension to generate an initial feature block for calculation, which is then read by the index matrix in the next step.

[0026] Based on the initial feature blocks mentioned above, a calculation index matrix is ​​constructed. The column fields of the matrix correspond to three types of indicators: influence, evidence association, and rule matching, while the row fields align to individual instances of disputed points. The calculation index matrix is ​​read by a "grading and rating device," which discretizes the influence based on the quantitative criteria of the rule base and assigns interval-based weights to evidence association and rule matching.

[0027] To constrain the stability of the scoring, the following weighted and penalized calculation formula is introduced: R=α·∑iqi·ui+β·∑j vj·wj-γ·∑k zk^2.

[0028] In the formula, R is the graded score, qi is the discrete value of influence, and ui is the corresponding weight; vj is the evidence relevance score, and wj is the corresponding weight; zk is the rule matching bias term, and α, β, and γ are non-negative adjustment coefficients; the summation indices i, j, and k traverse along their respective dimensions. R is used to determine the grade identifier of the disputed point and write it back to the graded dataset.

[0029] After the graded dataset is generated, a feature mapping matrix is ​​established, which maps the topic words, rule triggering bits and grade identifiers of the disputed points to the topic space of the evidence side.

[0030] To establish the mapping relationship, this embodiment defines an "evidence parser" that performs format recognition and content parsing on the input evidence files. Text-based evidence undergoes word segmentation and vectorization, image-based evidence undergoes text extraction and key area hashing, and form-based evidence undergoes field standardization. The parsing results are uniformly encoded into evidence feature data, with fields including evidence topic word vectors and evidence source identifiers, and aligned with the coordinate system of the feature mapping matrix.

[0031] Once the evidence feature data is ready, the "matching calculation unit" is invoked to perform vector operations. Specifically, the evidence feature vector is normalized, and then compared with the corresponding weight column in the feature mapping matrix by performing a dot product and a cooperating term to obtain an initial matching degree. A reduction factor is then introduced based on source quality and temporal proximity. The matching calculation unit outputs a matching record for each piece of evidence for each point of contention. The record contains the matching degree value and the triggered rule bit, which is used for subsequent threshold determination. To ensure consistency across sources, the vector scales generated by different parsing paths are corrected before the operation to ensure that the matching degrees are in the same metric space.

[0032] After the matched records are formed, threshold determination rules are loaded, and the records are divided according to preset upper and lower threshold intervals. Records with a matching degree below the lower threshold are marked as weakly correlated, records with a matching degree above the upper threshold are marked as strongly correlated, and the remaining records are kept in a pending review state. The threshold interval boundaries are adaptively adjusted by the distribution of the aforementioned graded score R to ensure that the strong correlation threshold is appropriately increased when the level of dispute is high. The threshold determination results are read by the "annotation generator" and assembled into an evidence annotation data stream.

[0033] Based on the aforementioned evidence-annotated data stream, data cards are generated. Each card uses the point of contention as the primary key, aggregating the contention level identifier and evidence-related parameters, and includes the evidence source, timestamp, and rule trigger details. The data cards retain the original matching degree, threshold determination status, and snapshots of the weights involved in the calculation, for subsequent review and traceability. To facilitate subsequent model calls, the cards use a unified structure encoding to ensure consistency with the interface specifications of subsequent steps.

[0034] After the data card is written to the cache, the score R of the "grading evaluator" is read by the scheduler and used as the baseline for the next window threshold; simultaneously, the data card is registered by the task allocation module as input for subsequent steps. For example, in a scenario of "duplicate submission dispute," two pieces of comparative evidence are parsed to form topic vectors, which are then mapped to rows with high-level identifiers to generate high-match records. After threshold determination, these records are entered into the strong association set and finally written to the data card of the dispute point, with the rule trigger bit being registered synchronously in the card.

[0035] Finally, the data card, as the output of this step, is provided to subsequent steps through a unified interface. On the one hand, step S102 can directly read the dispute level identifier and evidence association parameters in the card, calculate the expert domain fit, and generate a weight allocation model. On the other hand, step S103 reads back the weight snapshot and threshold status in the card when collecting scoring data to explain scoring deviations and guide anomaly labeling. Through the above read-write relationship, the grading results, mapping relationships, and annotation information are continuously referenced in subsequent processes, avoiding information gaps.

[0036] Step S102: Obtain expert processing parameters based on the data card, read the adjudication records in the expert historical database, perform numerical calculations on the dispute type processing quantitative indicators and result pass rate data according to the configuration parameters, generate domain fit data, construct a weight allocation model based on the fit data and historical accuracy parameters, generate a task scheduling sequence and weight coefficient matrix according to the allocation model, and output a task data table containing allocation parameters and scheduling strategies. First, after reading the data cards output in step S101, the fields "Dispute Level Identifier, Evidence Association Parameters, and Rule Trigger Details" are validated, and multiple cards with the same dispute point are deduplicated and merged based on timestamps. Then, based on the dispute type in the card, the corresponding adjudication records in the expert historical database are filtered by type, time window, and source domain, and three types of raw statistics are extracted: processing frequency, near-window pass rate, and rejection rate, forming a candidate record set that maps one-to-one with the card type for use in the fit calculation.

[0037] Based on the candidate record set, a "fitness evaluator" is constructed to perform missing data imputation and distribution truncation on each statistic, ensuring that outliers do not dominate the scoring. The fitness evaluator normalizes the processing frequency according to configuration parameters, converts the pass rate into a robust score after shrinking it by confidence interval, and uses the evidence correlation parameters in the cards as correction factors in the weighting, outputting domain fit data specific to the current type of dispute. This data is one-to-one bound to an expert identifier, and a calculation snapshot is retained for verification.

[0038] After the domain fit data is generated, historical accuracy parameters are read and scaled uniformly and time-sensitively verified. To avoid imbalance of a single indicator, this embodiment constructs a "weight allocator" to generate scoring weights and execution priorities by comprehensively considering multiple constraints.

[0039] The weight allocator employs a weighted and regularized computation formula: H=λ·∑p ap·bp+μ·∑q cq·dq-ν·∑r er^2.

[0040] In the formula, H represents the expert comprehensive weight, ap represents the domain fit term, and bp represents the corresponding weight coefficient; cq represents the historical accuracy term, and dq represents the corresponding weight coefficient; er represents the load and conflict penalty term; λ, μ, and ν are non-negative adjustment parameters; the summation indices p, q, and r are traversed along their respective dimensions. H serves as the core input for subsequent scheduling and allocation and is written into the main column of the weight coefficient matrix.

[0041] After the weight coefficient matrix is ​​filled, a task scheduling rule set is constructed. Tasks are sorted according to the dispute level identifier and H value, and resource consumption and time window constraints are introduced to generate execution levels. The rule set explicitly defines the single writer constraint for the same data primary key, ensuring that the same dispute point is assigned to only one responsible expert. At the same time, a cold backup list is generated for the second-highest weighted expert for rapid replacement in case of anomalies. The sorting results are read by the scheduler, which outputs a task scheduling sequence organized by time slot.

[0042] Based on the task scheduling sequence, the scheduler combines system load and expert availability time to refine concurrency and batch size, prioritizing mapping high-level dispute points to high-weight experts.

[0043] The mapping relationship is written back to the allocation view of the weight coefficient matrix, forming resource entries containing "expert identifier, task priority, time slice, and number of concurrent paths". To maintain process traceability, the allocation view retains the decomposition items of H and their generation time.

[0044] Based on the aforementioned allocation view, a task data table is generated. The task data table uses the disputed point as the primary key, and records the target expert, weight coefficient, execution time window, and rollback strategy flag for each entry. The rollback strategy references the aforementioned cold backup list and single-writer constraint, serving as the reassignment entry point when an anomaly occurs in the online channel. The task data table is also bound to the version hash of the data cards, ensuring that the original evidence association parameters and rule trigger details can be reused during the scoring phase.

[0045] Finally, the task data table, as the output of this step, is provided for subsequent steps to read. On the one hand, step S103 directly references the weight coefficient matrix and scheduling sequence in the task data table when collecting the scoring data stream, performing standard deviation calculation and anomaly node marking; on the other hand, step S104 calls the cold backup list and rollback flag in the task data table when reallocation is needed, achieving task migration with minimal scope. Through this read-write relationship, adaptability, weight, and scheduling remain consistent under the same index.

[0046] Step S103: Collect scoring data stream based on the task data table, execute the standard deviation calculation unit, mark data items whose scoring deviation values ​​exceed the set threshold as abnormal nodes, perform parameter correction on the weight coefficients of the abnormal nodes, perform data fusion operation based on the corrected weight matrix, generate conclusion confidence data, trigger the task reassignment mechanism according to the confidence data, and output the result dataset containing the fusion conclusion and confidence coefficient.

[0047] First, after reading the task data table output in step S102, the scoring channels are organized by time slice. A scoring session identifier is generated for the target expert for each point of contention, and the sampling frequency and timeout parameters are configured according to the weight coefficient matrix. Then, the scoring data stream transmitted from the online scoring system is accessed, and three checks are performed on the expert identifier, task primary key, and card version hash. Data that passes the checks are merged into the scoring buffer of the same task, forming a scoring sequence arranged in time for deviation calculation.

[0048] Based on the scoring sequence, the standard deviation calculation unit is invoked to perform denoising and scaling processing on the scores for the same task. Specifically, the scoring intervals of different experts are mapped to a unified metric, and then the mean and standard deviation are calculated using a fixed window sliding method to generate the deviation value of each score.

[0049] The deviation value is mapped to the weight coefficient in the task data table to form a deviation record containing the original score, deviation value and weight snapshot, which is used for anomaly detection.

[0050] After the deviation records are generated, threshold rules are loaded to filter the records. The threshold rules read the dispute level from the task data table and the threshold status from the card, and dynamically set the upper and lower limits.

[0051] For scoring entries whose deviation values ​​exceed the upper limit, they are marked as anomalous nodes, and the reason for the anomalousness and the session context are recorded. For entries near the boundary, they are added to the set to be reviewed and await confirmation by the next batch of data. The set of anomalous nodes is read by the weight corrector, which is ready to perform parameter correction.

[0052] Based on the set of abnormal nodes, a weight corrector is activated to proportionally lower the weight coefficients of the corresponding experts and slightly raise the weight coefficients of experts with stable scores in the same task. To avoid over-correction introducing new biases, this embodiment adopts a piecewise convergence strategy to limit the magnitude and direction of weight changes within a single window. Simultaneously, the corrected weight values ​​are written back to the weight coefficient matrix, and the difference before and after correction is retained for subsequent fusion and auditing.

[0053] Once the modified weight coefficient matrix is ​​ready, data fusion is performed. The fusion unit performs a weighted synthesis of the score vector and weight vector for the same task, and introduces a smoothing term for the nearest time slice to obtain the fusion score and uncertainty estimate for the current window.

[0054] To improve the robustness of the fusion, a weighted and penalized calculation formula is introduced: U=ρ·∑m xm·ym+σ·∑n an·bn-τ·∑h lh^2.

[0055] In the formula, U is the fusion score, xm is the score value, ym is the corrected weight; an is the time smoothing term, bn is the corresponding coefficient; lh is the conflict penalty term; ρ, σ, and τ are non-negative adjustment parameters; and the summation indices m, n, and h traverse along their respective dimensions. U, together with the uncertainty, constitutes the conclusion confidence data and is bound to the task primary key.

[0056] Based on the aforementioned confidence data, a task reassignment mechanism is triggered. The reassignment controller reads the confidence threshold and the cold backup list in the task data table. For tasks with low confidence and a high proportion of abnormal nodes, supplementary scoring is initiated, and these tasks are preferentially assigned to candidate experts, inheriting the original task's session context and card version hash. The reassignment action registers the relationship and time anchor point between the old and new experts in the task data table to ensure that incremental scores can be distinguished during subsequent fusion.

[0057] After the supplementary score is returned, the fusion unit reads the revised weight coefficient matrix and the newly added score again, performs incremental fusion, and updates U and the confidence interval. If the confidence level reaches the threshold, the weight and score set of the task is frozen, and the final conclusion is output; if it still does not reach the threshold, the reason for incompleteness is recorded and the process proceeds to the next window for scrolling processing. Throughout the process, the cumulative information of abnormal nodes is continuously written to the abnormal archive area, providing a basis for subsequent statistics.

[0058] After the final conclusion is determined, a result dataset is generated. The result dataset is indexed by the task primary key, and its fields include the fusion conclusion, U-value, uncertainty, list of outliers, and weight correction trajectory. This result dataset is written back to the monitoring panel for progress display and also provided to subsequent steps for reading through a unified interface. Specifically, step S104 reads the list of outliers and confidence levels from the result dataset for further threshold convergence and strategy adjustment; the scoring end restores the fusion state of the current window based on the weight correction trajectory when needed. Through the above connection, score collection, anomaly identification, weight correction, and redistribution form a closed loop under the same index.

[0059] As described above, the intelligent adjudication method based on dispute point matching and dynamic weight fusion provided in this application can effectively establish associations through innovative design of a dispute analysis system, feature extraction, and evidence matching. It constructs a scheduling and allocation mechanism, combining suitability assessment and weight calculation to establish a reliable adjudication strategy. Furthermore, it introduces result fusion, ensuring the accuracy of conclusions through bias handling and confidence assessment. This method effectively addresses the shortcomings of traditional technologies in dispute analysis, scheduling and allocation, and result fusion, providing technical support for intelligent adjudication.

[0060] In one embodiment of the intelligent adjudication method based on dispute point matching and dynamic weight fusion in this application, it may further include the following: Step S201: Obtain the output data of the large model from the data access terminal, parse the data format type according to the data header identifier, establish the dispute point parsing rules, extract the text attributes of the dispute points according to the preset feature encoding scheme, perform word segmentation on the dispute point text, construct a feature mapping table, and generate a dispute point feature set containing topic word vectors and semantic structures according to the feature mapping table. Step S202: Construct index calculation rules based on the disputed point feature set, discretely encode the influence parameter according to the preset quantification standard, generate evaluation scores according to the evidence correlation standard, configure weight coefficients according to the rule matching degree, substitute the scores and weight coefficients into the calculation matrix, classify the disputed points according to the calculation results, and generate a disputed point dataset containing graded identifiers.

[0061] First, after pulling the output data from the large model from the data access end, data header recognition is performed to verify four fields: source, timestamp, encoding method, and content type. Data is then routed to the text parsing channel based on content type. For text data that passes verification, disputed text attributes are extracted according to a preset feature encoding scheme. These fields cover topic phrases, context windows, rule terms, and entity references. A unique identifier is generated for each disputed point to ensure consistency in subsequent mapping and tracing.

[0062] After the text attributes are extracted, the word segmentation module is invoked to perform fine-grained segmentation of the disputed text, and the proper nouns and rule entries are unified by combining the rule terminology dictionary. The segmentation results are read by the semantic encoder, which generates topic word vectors and dependency structure descriptions, respectively. The topic word vectors are used to measure semantic similarity, and the dependency structures are used to characterize the grammatical relationship between the assertion and the evidence.

[0063] To avoid scale inconsistencies caused by cross-domain vocabularies, this embodiment normalizes the scales of vectors from different sources and outputs a primary vector set with a unified metric space.

[0064] Based on the initial vector set, a feature mapping table is constructed. This mapping table maps topic words, rule terms, and entity references to feature slots of fixed dimensions, and compresses the dependency structure into several relation bits, which are used to trigger rule matching branches in subsequent calculations.

[0065] The mapping table records the bidirectional pointer relationship between lexical units and slots, ensuring that the text at the point of contention can be traced back to a specific slot, and the original text evidence can also be retrieved from the slot. Based on this, a feature set of the points of contention containing topic word vectors and semantic structures is generated, which serves as the direct input for index calculation.

[0066] Once the feature set of disputed points is ready, indicator calculation rules are established. The rules first read the business definition of the impact parameter, then discretize the impact according to a preset quantification standard to form an impact discrete vector. Next, based on the usability of the evidence and the similarity between the evidence and the topic, an evidence association score is calculated, and weak evidence is subject to reduction. Finally, rule matching is triggered based on rule terminology and dependency relationships, generating a rule matching score and recording the trigger details. These three types of quantities are then entered into the same calculation matrix for unified output of graded scores.

[0067] Before the calculation, robustness and constraint terms are loaded to avoid a single source dominating the result.

[0068] This embodiment employs a calculation formula that combines weighted synthesis and penalty in parallel within the "grading evaluator": V=κ·∑a xa·ya+λ·∑b pb·qb-ξ·∑c rc^2.

[0069] In the formula, V represents the graded score, xa represents the impact discrete term, and ya represents the corresponding weight; pb represents the evidence association score, and qb represents the corresponding weight; rc represents the rule matching bias term; κ, λ, and ξ are non-negative adjustment parameters; the summation indices a, b, and c traverse the dimensions of impact, evidence association, and rule bias, respectively. The parameter values ​​are loaded through a configuration file and are not fixed to specific numerical values, allowing for flexible adjustment in different scenarios.

[0070] Based on the aforementioned score V, the rating system generates a rating identifier for each point of contention and backfills the triggered rule details and evidence reduction factors into the record, forming an intermediate entry with a rating identifier. This intermediate entry retains a calculation snapshot, including the weight version and threshold caliber, for easy subsequent review and recalculation. To avoid level boundary fluctuations, the rating system introduces in-window smoothing, using the score distribution of similar points of contention in adjacent time periods as a reference, and delaying the confirmation of samples near the critical value by one window.

[0071] After aggregating the intermediate entries, a dispute point dataset is generated. This dataset uses the unique identifier of the dispute point as the primary key, and the fields include a level identifier, topic word vector, semantic structure summary, impact discrete encoding, evidence association score, and rule matching trigger bit. The dataset also records the source and timestamp index to ensure that subsequent evidence-side matching and task-side scheduling can be time-aligned.

[0072] After the disputed points dataset is established, the version hash of the feature mapping table is written back upstream to record the dictionary and rule set versions used in this mapping, ensuring that the same metric space can be reused for subsequent incremental updates. Downstream, this dataset is referenced in step S101 to construct a feature mapping matrix and perform vector operations on the evidence files, thereby producing data cards; simultaneously, step S102 reads the level identifier and evidence association score when calculating the expert domain fit, as a pre-weight for task allocation. Through this read-write relationship, the disputed points form a closed-loop link from the original text to the classification result, providing a consistent entry point for subsequent matching and scheduling.

[0073] In one embodiment of the intelligent adjudication method based on dispute point matching and dynamic weight fusion in this application, it may further include the following: Step S301: Construct feature extraction rules based on the hierarchical dataset, perform format recognition and type classification on the evidence files, establish a text preprocessing module, perform data format conversion according to the preset standardized process, perform word segmentation and vectorization processing on the evidence content, construct a feature dimension table, and generate a feature data stream containing evidence subject words and semantic vectors based on the feature dimension table; Step S302: Construct mapping calculation rules based on the feature data stream, normalize the evidence feature vector according to a preset standard, generate matching degree calculation parameters according to the dispute level identifier, configure annotation rules according to preset threshold conditions, substitute the calculation parameters into the mapping matrix, quantify the evidence correlation degree according to the calculation results, and generate a data card containing the dispute level identifier and evidence correlation parameters.

[0074] First, after reading the hierarchical dataset output in step S202, version verification is performed on the hierarchical identifiers and topic word vectors, and an evidence access channel is established based on the source domain. After accessing the evidence files, format recognition is performed first, distinguishing them into three types: text, image, and form. Corresponding parsing plugins are then loaded according to the type. Text evidence enters the word segmentation and denoising process; image evidence is extracted and then merged into the text channel; form evidence undergoes key-value standardization according to the field mapping table to ensure that evidence from different sources is comparable under the same structure.

[0075] After the type parsing is completed, the text preprocessing module is started. It performs character set unification, stop word cleanup, and phrase merging according to a preset standardization process, and unifies the referential meaning of fragments involving rule terms. The standardized text output by this module is read by the embedded encoder to generate semantic vectors, while retaining word position information and source timestamps. To ensure consistency with the topic space of the hierarchical dataset, the generated vectors are scale-aligned with the topic word vectors in step S202, and the alignment residuals are recorded as a penalty factor for subsequent matching.

[0076] Based on the alignment results, a feature dimension table is constructed. This dimension table organizes the evidence features into three categories: "topic term slots, rule term slots, and context window indexes," and establishes a unique index for each piece of evidence, mapping it back to the original text position and the parsing plugin version.

[0077] The output contains a feature data stream containing evidence topic terms and semantic vectors; this data stream is aligned with the row keys of the hierarchical dataset, clearly defining the set of disputed points that each piece of evidence may be associated with, and providing input for mapping calculations.

[0078] After the feature data stream is ready, mapping calculation rules are constructed. The rules first normalize the evidence feature vectors so that the vectors of different parsing paths are under the same metric; then, based on the level labels in the hierarchical dataset, matching degree calculation parameters are generated for different levels, including level weights, residual penalty coefficients, and time decay factors.

[0079] The aforementioned calculation parameters and feature data stream are then substituted into the mapping matrix to calculate the matching score between the evidence and each point of contention, and the intermediate matching record containing the score, source weight, and trigger bit is output.

[0080] After the intermediate matching records are generated, annotation rules are loaded for threshold-based data splitting. Based on preset threshold conditions, the annotation rules label records above the upper boundary as strongly correlated, records below the lower boundary as weakly correlated, and the remaining records as pending review. To suppress short-term fluctuations, the threshold boundaries are adaptively adjusted based on the distribution quantile of the previous window.

[0081] Meanwhile, this embodiment employs a weighted and penalized calculation formula in the "matching evaluator" to robusten the score: W=θ·∑u su·tu+φ·∑v gv·hv-ψ·∑w cw^2.

[0082] In the formula, W is the robust matching score, su is the normalized similarity, tu is the rank weight; gv is the temporal proximity term, hv is the corresponding coefficient; cw is the alignment residual term; θ, φ, and ψ are non-negative adjustment parameters; and the summation indices u, v, and w traverse their respective dimensions. The W term replaces the original score in the threshold determination and is then backfilled into the matching records.

[0083] After the threshold-based triage is completed, data cards are generated. Each data card uses the unique identifier of the disputed point as the primary key, aggregates the corresponding list of evidence, and writes a dispute level identifier, evidence association parameters, W value, annotation status, and triggered rule bit for each piece of evidence. Items to be reviewed are accompanied by the minimum necessary context and source fingerprint to facilitate subsequent manual verification. The data card also records the mapping matrix version and threshold caliber to support cross-window recalculation.

[0084] Based on the aforementioned data cards, a write-back relationship is constructed.

[0085] On the one hand, the evidence association parameters in the card are read in step S102 and used as correction factors when calculating the expert domain fit. On the other hand, the labeling status and W value in the card are read in step S103 to explain the scoring deviation and guide the identification of abnormal nodes.

[0086] To ensure data consistency, cards are bound to the version hash of the hierarchical dataset when written, and a read-only view is exposed at the interface layer to avoid data drift caused by online modifications.

[0087] Finally, the data cards are submitted to the task orchestration channel in batches. During the submission process, the system registers the mapping index between the cards and the evidence sources, and records the coverage and proportion of the mapping to be reviewed in this round on the monitoring panel; any subsequent step can be traced back to the feature dimension table and the original evidence fragments by index, ensuring that the matching calculation process is reproducible and verifiable. Through the above link, the transformation of evidence from unstructured input to structured cards is completed, providing a stable entry point for downstream weight allocation and scoring fusion.

[0088] In one embodiment of the intelligent adjudication method based on dispute point matching and dynamic weight fusion in this application, it may further include the following: Step S401: Construct expert retrieval rules based on data cards, verify access permissions to the expert historical database, establish a data reading channel, obtain adjudication records according to preset filtering conditions, classify and statistically analyze the record content, construct a processing quantitative index table, and generate an expert basic dataset containing processing frequency and pass rate based on the quantitative index table. Step S402: Construct adaptation calculation rules based on the expert basic dataset, normalize the processing quantitative indicators according to preset standards, generate weight coefficients according to the pass rate parameter, configure calculation templates according to the dispute type, substitute the weight coefficients into the calculation unit, and perform quantitative evaluation of expert adaptation based on the calculation results to generate adaptation data containing domain matching degree and processing ability.

[0089] First, after reading the data card output in step S302, a search key is established based on the unique identifier of the disputed point. The three types of fields in the card—dispute level identifier, evidence association parameters, and rule trigger details—are parsed and used as components of the search conditions. Then, an access control list is loaded on the expert side, and access permission verification is performed on the expert's historical database. After successful verification, a read-only data reading channel is opened according to the source domain and time window to ensure that subsequent statistics are based on the same criteria. After the channel is established, corresponding adjudication records are filtered based on the dispute type and version hash in the card. Samples with missing primary keys and abnormal timestamps are removed, forming the original record set to be statistically analyzed.

[0090] Based on the original record set, categorized statistics are performed. The statistical process uses a three-dimensional cross-reference of "dispute type, rule entry, and source domain" to count the number of times each expert processes a case, the number of valid scores, and the pass rate of the result separately, and removes duplicate records across domains.

[0091] To mitigate short-term fluctuations, near-window data is aggregated using a sliding window, while simultaneously recording window boundaries and record coverage. Statistical output is renamed and structured into a processing quantification index table, with fields including at least processing frequency, effective sample size, pass rate, and outlier rate, to comprehensively reflect the expert's actual processing capabilities for the current type of dispute.

[0092] After the quantitative indicator table is generated, it is aggregated according to expert identifiers to form an expert basic dataset. This dataset inherits the data card index from step S302, realizes a one-to-many mapping relationship of "disputed points - expert records", and writes version stamps and caliber summaries with each entry for traceability.

[0093] Meanwhile, to ensure the integrity of the link, the dispute level identifier and the mean of the evidence association parameters used for screening are backfilled into the dataset entries as external features for subsequent fit correction.

[0094] After the expert base dataset is ready, fitness calculation rules are constructed. The rules first normalize the processing frequency and effective sample size, using quantile truncation to limit the impact of extreme values. Then, the pass rate is narrowed to a robust score based on confidence intervals, and the anomaly rate is given a reduction coefficient based on threshold intervals. To incorporate evidence-based information, the mean values ​​of evidence-related parameters in the data cards are mapped to correction factors, limiting their influence on the score and ensuring that evidence from different sources does not dominate the results.

[0095] Based on the normalized results, a "domain matching evaluator" is established, and corresponding calculation templates are loaded according to the type of dispute.

[0096] The evaluator employs a weighted synthesis and penalty-based computational approach to quantify the fit: Y=π·∑s ms·ns+χ·∑t ut·vt-ω·∑z dz^2.

[0097] In the formula, Y represents the expert fit score, ms represents the normalized value of the processing frequency, and ns represents the corresponding weight; ut represents the robustness score of the pass rate, and vt represents the corresponding weight; dz represents the penalty term formed by the anomaly rate and evidence correction; π, χ, and ω are non-negative adjustment parameters; the summation indices s, t, and z traverse the processing frequency, pass rate, and penalty dimensions, respectively. Each weight is loaded from the configuration file and updated with the scenario version, and is not fixed to a specific value.

[0098] Based on the aforementioned fit score Y, the evaluator simultaneously calculates two indicators: domain matching degree and processing capacity. Domain matching degree focuses on the alignment between dispute types and rule entries, primarily relying on the robustness score of the pass rate and the rule trigger hit ratio; processing capacity focuses on the carrying capacity within a given time window, primarily relying on the processing frequency and the number of effective samples. These two indicators, along with Y, are written into the fit data entries, and the calculation template and parameter version used are recorded for easy recalculation when needed later.

[0099] After the fit data is generated, it is merged with the expert base dataset to produce a fit view with a two-dimensional key of "expert identifier - dispute type". This view explicitly references the previously written dispute level identifier and the mean of the evidence association parameter, indicating the correction rules and influence range of the two in the scoring. The view also records timestamps and window numbers to ensure that the subsequent weight allocation model can be read by scrolling through the window and maintain consistency.

[0100] Based on the aforementioned adaptation view, the output includes adaptation data containing domain matching degree and processing capability, which serves as the direct input for the next step of weight allocation.

[0101] On the one hand, in step S402, the subsequent weight allocation model reads Y and its decomposition terms to generate a weight coefficient matrix and a task scheduling sequence; On the other hand, step S103 can read back the source of the penalty item dz in the view when a scoring deviation occurs, and explain the basis for weight correction of abnormal nodes.

[0102] Through this read-write relationship, the expert's historical performance and evidence are linked under the same index, providing a reproducible computational link for subsequent allocation and fusion.

[0103] In one embodiment of the intelligent adjudication method based on dispute point matching and dynamic weight fusion in this application, it may further include the following: Step S501: Construct weight calculation rules based on fit data, standardize historical accuracy parameters, establish a parameter fusion module, combine indicators according to preset priority standards, verify the effectiveness of the fusion results, construct a weight calculation matrix, and generate an allocation model containing expert rating weights and task priorities based on the weight calculation matrix. Step S502: Construct a scheduling rule set based on the allocation model, divide the task sequence into preset time windows, generate scheduling priorities according to system load parameters, configure allocation strategies according to task types, substitute the priorities into the scheduling matrix, configure resources for tasks according to the allocation results, and generate a task data table containing allocation parameters and scheduling strategies.

[0104] First, after reading the fit data generated in steps S401 to S402, time window alignment is completed using "expert identifier - dispute type" as the key, and historical accuracy parameters are simultaneously retrieved for caliber verification. Standardization processing is performed on the historical accuracy parameters, including missing data completion, distribution truncation, and scaling, to obtain a comparable accuracy vector. The fit data and accuracy vector are mapped one-to-one in the key space, serving as input pairs for subsequent weight calculations.

[0105] Based on the input pairs, a parameter fusion module is established. This module reads two indicators: domain matching degree and processing capability, and simultaneously incorporates standardized historical accuracy. It combines these indicators according to a preset priority standard, which increases the weight of domain matching degree when the dispute level is high and increases the weight of processing capability when the dispute level is low. During the fusion process, a convergence boundary is introduced to handle short-term anomalies, preventing weight jumps caused by single batches of data. The fusion output forms candidate weight terms and their source annotations for matrix calculation verification.

[0106] After the candidate weights are ready, a weight calculation matrix is ​​constructed. The column domains of the matrix correspond to domain matching degree, processing power, and historical accuracy, respectively, while the row domains align experts and dispute type instances. To balance robustness and interpretability, a weighted and penalized parallel calculation formula is used in the "weight allocator" to complete the comprehensive weight and priority estimation: G=ζ·∑r Ar·Br+eta·∑s Cs·Ds-θ·∑t Ft^2.

[0107] In the formula, G represents the comprehensive weight score, Ar represents the domain matching degree term, and Br represents the corresponding weight; Cs represents the historical accuracy term, and Ds represents the corresponding weight; Ft represents the load and conflict penalty term; ζ, η, and θ are non-negative adjustment parameters; the summation indices r, s, and t traverse the three categories of indicator dimensions respectively. Each symbol is defined only as a dimensionless quantity with consistent dimensions, and the specific values ​​are loaded from the configuration file. G serves as the baseline score for expert scoring weights and is also mapped to the initial value of task priority.

[0108] Based on the aforementioned G value, validity verification is performed. The verification process sets an upper limit on the G difference for the same expert in adjacent windows. Items with a sudden increase in anomaly rate or excessively low sample size are flagged for weight reduction, and the reason for the weight reduction and the time anchor are recorded. Validated weights are written into the main column of the weight calculation matrix, forming an allocation model that includes expert rating weights and task priorities. Items that fail verification are placed in the observation pool and evaluated again after being added to the next window.

[0109] After the allocation model is generated, a scheduling rule set is constructed. The rule set first divides the task sequence into several execution slices according to a preset time window, and then reads system load parameters to generate scheduling priorities, including three constraints: concurrency limit, submission interval, and queue depth. Subsequently, allocation strategies are configured according to task type. For "high-level dispute points," high-weight priority and cold standby follow are enabled; for "low-level dispute points," batch merging and rate-limited execution are enabled, forming a switchable strategy group adaptable to different load states.

[0110] Based on the aforementioned strategy group, priorities are substituted into the scheduling matrix to calculate task allocation within each execution slice. The scheduling matrix reads the expert rating weights and task priorities from the allocation model, combines the available expert time slots with concurrency constraints, and outputs a mapping list of "task-expert-time slice". To avoid write conflicts, a single writer constraint is implemented for the same point of contention, and a cold backup index is registered for the second-highest weighted expert to ensure quick switching in case of online path anomalies.

[0111] After the mapping list is generated, resource configuration is completed. The resource configurator allocates the number of concurrent paths, batch size, and submission rhythm according to the list, and binds each allocation record to the version hash of the source evidence to ensure consistency in the scoring phase. The configuration results are synchronously registered with load fingerprints and weight snapshots for retrospective and auditing in subsequent steps; if the load probe reports congestion exceeding the threshold, the scheduler executes downgraded according to the rule set and records the downgrade interval.

[0112] Finally, a task data table containing allocation parameters and scheduling strategies is output. The task data table uses the unique identifier of the disputed point as the primary key, and records each entry including the target expert, scoring weight G, task priority, time window, number of concurrent paths, and cold backup index, along with the reason for weight reduction, load fingerprint, and parameter version number. This task data table is directly read in step S103 to organize the scoring data stream and standard deviation calculation, while also providing a cold backup entry point for possible reallocation. In the next window, step S501 can read back the weight snapshot and execution fingerprint in this table to perform rolling calibration of the allocation model.

[0113] In one embodiment of the intelligent adjudication method based on dispute point matching and dynamic weight fusion in this application, it may further include the following: Step S601: Construct data collection rules based on the task data table, collect and format expert scoring data in real time, establish a deviation calculation module, perform standard deviation calculation according to preset statistical standards, compare the scoring results with thresholds, construct an anomaly marking matrix, and generate a dataset containing scoring deviation values ​​and anomaly nodes based on the anomaly marking matrix. Step S602: Construct weight correction rules based on the dataset, adjust the weight coefficients of abnormal nodes according to a preset ratio, generate a correction matrix according to the adjustment parameters, configure the operation template according to the fusion calculation standard, substitute the correction matrix into the fusion unit, quantify the confidence of the conclusion according to the operation results, and generate a dataset containing the fusion conclusion and confidence.

[0114] First, after reading the task data table output in step S502, a collection key is established based on the time window and the unique identifier of the dispute point. The four fields "target expert, scoring weight, time slice, and card version hash" are used as the necessary conditions for the collection rules.

[0115] When accessing the score feedback channel, pressing the acquisition button verifies the source and aligns the time of the arriving score packets, removing missing primary keys and duplicate packets, while retaining the original score values, expert identifiers, and session context. The standardized score stream output from this process then undergoes formatting, unifying the units and recording the version fingerprint mapped to the data card for subsequent deviation calculation.

[0116] After the standardized scoring flow is ready, a deviation calculation module is established. This module first standardizes the scale of multi-expert scoring intervals for the same task, and then calculates the mean and standard deviation using a fixed-length sliding window to obtain the deviation value for each score. The deviation value and the weight snapshot in the task data table are summarized into a scoring deviation entry. The entry includes the original score value, deviation value, window number, and expert weight, which is used for threshold comparison.

[0117] To avoid short-term jitter, the module uses weighted attenuation for samples at the window boundaries to reduce the impact of a single outlier score on the result.

[0118] After the scoring deviation entries are generated, threshold rules are loaded for comparison. The threshold rules read the dispute level and historical fluctuation range, set upper and lower limits, and mark entries that cross these limits as anomalies. After marking, the anomaly distribution of each task in the current window is aggregated into an anomaly label matrix. The matrix rows correspond to the task primary key, the columns correspond to the expert identifier, and the elements are the deviation exceeding the limit flag and the deviation magnitude. The anomaly label matrix serves as the core intermediate quantity for weight correction and subsequent fusion.

[0119] Based on the anomaly marker matrix, weight correction rules are constructed. These rules implement segmented downweighting for experts with consecutive anomalies, slight upweighting for experts who are stable and close to the mean in the same task, and set a gradual increase cap for newly added candidate experts. Weight adjustment parameters are generated accordingly, including downweighting percentage, upweighting percentage, and window limit. These parameters are written into a correction matrix, with the same dimension as the weight coefficient matrix, and element values ​​representing the incremental correction for the current window, facilitating item-by-item merging and auditing with historical weights.

[0120] After the correction matrix is ​​generated, the calculation template is configured according to the fusion calculation standard, and a weighted fusion is performed on the scores and corrected weights of the same task. The fusion unit introduces a time smoothing term and a conflict penalty term within the window to reduce the impact of sudden anomalies on the conclusion.

[0121] This embodiment uses the following formula to calculate the fused output: Q=α·∑g yg·zg+β·∑h sh·th-γ·∑k rk^2.

[0122] In the formula, Q is the fusion score, yg is the score value, zg is the corrected weight; sh is the time smoothing term, th is the corresponding coefficient; rk is the conflict penalty term induced by the anomaly labeling matrix; α, β, and γ are non-negative adjustment parameters; the summation indices g, h, and k traverse the scoring, smoothing, and penalty dimensions, respectively. Parameter values ​​are loaded by configuration and constrained by the window strategy, and are not fixed to specific numerical values. Q and its variance information together constitute the fundamental quantity for the reliability of the conclusion.

[0123] Based on the aforementioned Q and variance information, a quantitative calculation of the conclusion confidence level is constructed. The confidence level is reflected by the relative lead and uncertainty of the fusion score. The conclusion category, Q value, and confidence interval are output, and the version of the correction matrix used for calculation and the threshold caliber are also written into the record, forming a dataset containing the fusion conclusion and confidence level. This dataset retains a list of anomalous nodes and correction trajectories for each task, facilitating subsequent verification and playback.

[0124] After the dataset is formed, the abnormal nodes and correction matrices are written back to the extended field of the task data table, recording the "reason for weight change, window number, and candidate expert status". After the write-back is completed, the scheduling controller determines whether to trigger supplementary scoring and reallocation based on the confidence threshold. The results of the supplementary scoring are entered into the next round of fusion using the same acquisition key, thus forming a closed loop of correction-fusion-verification. For tasks that still have insufficient confidence, the reasons for incompleteness are recorded and pushed into the next window observation queue.

[0125] Finally, the dataset containing the fusion conclusions and confidence levels is provided to subsequent steps through a unified interface. On one hand, step S603 reads the Q-value and anomaly distribution when generating a phase report to display the current assessment status; on the other hand, step S701 aggregates the confidence trajectory and anomaly node ratio of this dataset on the monitoring side for progress calculation and subsequent strategy updates. Through this connection, data collection, deviation calculation, weight correction, and fusion output maintain consistency under the same index.

[0126] In one embodiment of the intelligent adjudication method based on dispute point matching and dynamic weight fusion in this application, it may further include the following: Step S701: Construct redistribution trigger rules based on confidence data, determine the threshold of confidence values, establish a task allocation controller, screen experts according to preset supplementary standards, perform status detection on allocable resources, construct a task allocation matrix, and generate a redistribution scheme containing an expert supplementary list and allocation strategy based on the task allocation matrix. Step S702: Based on the redistribution scheme, construct the result synthesis rules, perform fusion calculation on the supplementary scoring data according to the preset standard, generate weight parameters according to the confidence coefficient, configure the synthesis template according to the conclusion type, substitute the weight parameters into the calculation unit, and finally confirm the fusion conclusion according to the calculation result to generate a result dataset containing conclusion data and confidence coefficient.

[0127] First, after reading the dataset containing the fusion conclusion and confidence level output in step S602, align it according to the task primary key and window number, and extract the three fields "confidence level, percentage of abnormal nodes, and status of alternate experts" as the judgment input.

[0128] Next, a redistribution triggering rule is established based on the confidence level value. The threshold judgment result is calculated for each task. The threshold boundary is adaptively adjusted in combination with the confidence level distribution of the two windows. At the same time, the proportion of abnormal nodes and the score coverage are used as auxiliary conditions. The trigger flag and trigger reason list are output for the task allocation controller to read.

[0129] Once the trigger flag is ready, the task allocation controller loads preset supplementary criteria based on the trigger reason and filters candidate experts from the suitability view according to the dispute type and level. During filtering, priority is given to meeting the domain matching requirements of the current dispute type, followed by verifying processing capacity and available time periods, and then checking the timeliness label of historical accuracy. Experts in an abnormal observation state are removed. An exclusive mapping is established between the filtering results and existing responsible experts to avoid concurrent writing of the same task. The candidate set that passes the filtering is labeled with "candidate priority and constraints" and used as input for resource detection.

[0130] Based on the candidate set, allocable resource status checks are performed. The resource detection module reads the system load probe to verify the concurrency limit, queue depth, and local session availability. Simultaneously, it compares the current task usage of candidate experts with their quotas to ensure that it does not exceed the load threshold within the window. Candidates that pass the checks are matched with tasks using Cartesian matching to form preliminary task allocation pairs. Entries that fail the checks are recorded for reasons and retained in the retry list for the next window.

[0131] Based on the initial allocation pairs, a task allocation matrix is ​​constructed. The rows of the matrix represent tasks to be reallocated, the columns represent candidate experts, and the elements represent allocation feasibility, which combines three indicators: domain matching degree, historical accuracy, and workload margin.

[0132] To suppress over-concentration, this embodiment employs a weighted and penalized calculation formula in the "redistribution evaluator": .

[0133] In the formula, S represents the assigned score, ai represents the domain matching degree, and bi represents the corresponding weight; cj represents the historical accuracy, and dj represents the corresponding weight; fk represents the load occupancy and session conflict penalty; δ, ε, The parameters are non-negative adjustment parameters; the summation indices i, j, and k are traversed along the three dimensions respectively. S is used to determine the candidate order for each task and is called in the next segment to generate a redistribution scheme.

[0134] Based on the aforementioned allocation score S, the task allocation controller selects the column with the highest feasibility per row and outputs a redistribution scheme by combining constraints and single-writer constraints. The scheme includes a supplementary expert list, target time slice, number of concurrent paths, and inheritance context markers. It is clearly stated that the supplementary scoring must use the original task's card version hash and weight snapshot to ensure consistency in subsequent fusion. After the scheme is generated, the version number and window number are registered, and it is sent to the scoring side through a message channel to enter the supplementary scoring process.

[0135] After the redistribution scheme is issued, the result synthesis rules are loaded. The synthesis rules define the access order, denoising, and scaling methods for the supplementary scores, and specify the timing and coverage of incremental fusion. When the supplementary score data is transmitted back, the system maps the data to the original task session according to the inheritance marker in the scheme, completing version and time alignment. Then, the confidence coefficient is read, new weight parameters are generated according to the preset segmentation relationship, and a gradual increase limit is applied to the corrected weights to prevent weight jumps caused by a single supplement.

[0136] Once the weight parameters are ready, they are substituted into the computation unit for fusion. The fusion unit performs a weighted synthesis of the original score and the supplementary score, applies smoothing to the time proximity item, and retains conflict penalty terms for experts who have exhibited anomalies. The synthesis template is configured according to the conclusion type; for example, log-odds accumulation is used for binary conclusions, and weighted normalization aggregation is used for multiple-choice conclusions. After fusion, the updated fusion conclusion and new confidence coefficients are output, and the source of the weights and the parameter version are recorded to ensure traceability.

[0137] Based on the aforementioned update results, the final confirmation process is executed. The confirmer jointly determines the lead margin and uncertainty of the conclusion. If the threshold is reached, the weight and score set of the task are frozen and marked as completed; if it is still below the threshold, the task is retained for reallocation and the reason for incompleteness is written back. The confirmation result synchronously updates the exception file and reallocation log for subsequent window strategy fine-tuning.

[0138] After the confirmation action is completed, a result dataset is generated. The result dataset is indexed by the task primary key, and its fields include the latest fusion conclusion, confidence coefficient, list of participating experts, reassigned trajectory, and parameter version fingerprint. This dataset is provided to subsequent steps through a unified interface. In step S702, subsequent stages continue rolling synthesis and judgment by directly referencing the confidence coefficient and trajectory. Simultaneously, the monitoring side reads the reassigned trajectory and failure reasons when generating progress and anomaly distributions, forming a replayable closed-loop evidence chain. Through this connection, triggering, filtering, allocation, and synthesis remain consistent under the same index, supporting steady-state progress across windows.

[0139] To effectively address the shortcomings of traditional technologies in dispute analysis, scheduling and allocation, and result fusion, and to provide technical support for intelligent adjudication, this application provides an embodiment of an intelligent adjudication device based on dispute point matching and dynamic weight fusion for implementing all or part of the aforementioned intelligent adjudication method based on dispute point matching and dynamic weight fusion. See [link to embodiment]. Figure 2 The intelligent adjudication device based on dispute point matching and dynamic weight fusion specifically includes the following components: The dispute matching module 10 is used to receive the initial dispute point data stream output by the large model, extract the dispute point feature vector, construct the calculation index matrix, import the influence data item, evidence association parameters, and rule matching factors into the preset calculation unit, generate a dispute point hierarchical dataset, construct a feature mapping matrix based on the hierarchical dataset, perform data parsing and format conversion on the input evidence file, perform vector operations on the evidence feature data and the mapping matrix, generate an evidence annotation data stream according to the matching degree threshold, and output a data card containing dispute level identifier and evidence association parameters. The weight fusion module 20 is used to obtain expert processing parameters based on the data card, read the adjudication records in the expert historical database, perform numerical calculations on the dispute type processing quantitative indicators and result pass rate data according to the configuration parameters, generate domain adaptability data, construct a weight allocation model based on the adaptability data and historical accuracy parameters, generate a task scheduling sequence and weight coefficient matrix according to the allocation model, and output a task data table containing allocation parameters and scheduling strategies. The intelligent assessment module 30 is used to collect scoring data streams based on the task data table, execute the standard deviation calculation unit, mark data items whose scoring deviation values ​​exceed a set threshold as abnormal nodes, perform parameter correction on the weight coefficients of the abnormal nodes, perform data fusion operations based on the corrected weight matrix, generate conclusion confidence data, trigger the task reassignment mechanism based on the confidence data, and output a result dataset containing the fusion conclusion and confidence coefficient.

[0140] As described above, the intelligent adjudication device based on dispute point matching and dynamic weight fusion provided in this application can effectively establish associations through innovative design of a dispute analysis system, feature extraction, and evidence matching. It constructs a scheduling and allocation mechanism, combining suitability assessment and weight calculation to establish a reliable adjudication strategy. By introducing result fusion and handling bias and assessing confidence, it ensures the accuracy of conclusions. This method effectively addresses the shortcomings of traditional technologies in dispute analysis, scheduling and allocation, and result fusion, providing technical support for intelligent adjudication.

[0141] From a hardware perspective, in order to effectively address the shortcomings of traditional technologies in dispute analysis, scheduling and allocation, and result fusion, and to provide technical support for intelligent adjudication, this application provides an embodiment of an electronic device for implementing all or part of the intelligent adjudication method based on dispute point matching and dynamic weight fusion. The electronic device specifically includes the following components: The system comprises a processor, memory, a communications interface, and a bus; wherein the processor, memory, and communications interface communicate with each other via the bus; the communications interface is used to realize information transmission between the intelligent adjudication device based on dispute point matching and dynamic weight fusion and core business systems, user terminals, and related databases and other related devices; the logic controller can be a desktop computer, tablet computer, or mobile terminal, etc., and this embodiment is not limited to these. In this embodiment, the logic controller can be implemented with reference to the embodiments of the intelligent adjudication method based on dispute point matching and dynamic weight fusion, and the embodiments of the intelligent adjudication device based on dispute point matching and dynamic weight fusion in the embodiments, the content of which is incorporated herein, and repeated parts will not be described again.

[0142] It is understood that the user terminal may include smartphones, tablet computers, network set-top boxes, portable computers, desktop computers, personal digital assistants (PDAs), in-vehicle devices, smart wearable devices, etc. Among these, the smart wearable devices may include smart glasses, smartwatches, smart bracelets, etc.

[0143] In practical applications, the intelligent adjudication method based on dispute point matching and dynamic weight fusion can be partially executed on the electronic device side as described above, or all operations can be completed on the client device. The choice can be made based on the processing power of the client device and the limitations of the user's usage scenario. This application does not impose any limitations on this. If all operations are completed on the client device, the client device may further include a processor.

[0144] The aforementioned client device may have a communication module (i.e., a communication unit) that can communicate with a remote server to achieve data transmission with the server. The server may include a server on the task scheduling center side; in other implementation scenarios, it may also include a server on an intermediate platform, such as a server on a third-party server platform that has a communication link with the task scheduling center server. The server may include a single computer device, a server cluster consisting of multiple servers, or a distributed server structure.

[0145] Figure 3 This is a schematic block diagram illustrating the system configuration of the electronic device 9600 according to an embodiment of this application. Figure 3 As shown, the electronic device 9600 may include a central processing unit 9100 and a memory 9140; the memory 9140 is coupled to the central processing unit 9100. It is worth noting that... Figure 3 This is an example; other types of structures can also be used to supplement or replace this structure to achieve telecommunications functions or other functions.

[0146] In one embodiment, the intelligent adjudication method based on dispute point matching and dynamic weight fusion can be integrated into the central processing unit 9100. The central processing unit 9100 can be configured to perform the following controls: Step S101: Receive the initial dispute point data stream output by the large model, extract the dispute point feature vector, construct the calculation index matrix, import the influence data item, evidence association parameters, and rule matching factors into the preset calculation unit, generate a dispute point graded dataset, construct a feature mapping matrix based on the graded dataset, perform data parsing and format conversion on the input evidence file, perform vector operations on the evidence feature data and the mapping matrix, generate an evidence annotation data stream according to the matching degree threshold, and output a data card containing dispute level identifier and evidence association parameters; Step S102: Obtain expert processing parameters based on the data card, read the adjudication records in the expert historical database, perform numerical calculations on the dispute type processing quantitative indicators and result pass rate data according to the configuration parameters, generate domain fit data, construct a weight allocation model based on the fit data and historical accuracy parameters, generate a task scheduling sequence and weight coefficient matrix according to the allocation model, and output a task data table containing allocation parameters and scheduling strategies. Step S103: Collect scoring data stream based on the task data table, execute the standard deviation calculation unit, mark data items whose scoring deviation values ​​exceed the set threshold as abnormal nodes, perform parameter correction on the weight coefficients of the abnormal nodes, perform data fusion operation based on the corrected weight matrix, generate conclusion confidence data, trigger the task reassignment mechanism according to the confidence data, and output the result dataset containing the fusion conclusion and confidence coefficient.

[0147] As described above, the electronic device provided in this application, through an innovative design of a dispute analysis system, effectively establishes associations through feature extraction and evidence matching. It constructs a scheduling and allocation mechanism, combining suitability assessment and weight calculation to establish a reliable adjudication strategy. Furthermore, it introduces result fusion, ensuring the accuracy of conclusions through bias processing and confidence assessment. This method effectively addresses the shortcomings of traditional technologies in dispute analysis, scheduling and allocation, and result fusion, providing technical support for intelligent adjudication.

[0148] In another embodiment, the intelligent adjudication device based on dispute point matching and dynamic weight fusion can be configured separately from the central processing unit 9100. For example, the intelligent adjudication device based on dispute point matching and dynamic weight fusion can be configured as a chip connected to the central processing unit 9100, and the intelligent adjudication method based on dispute point matching and dynamic weight fusion can be implemented through the control of the central processing unit.

[0149] like Figure 3 As shown, the electronic device 9600 may further include: a communication module 9110, an input unit 9120, an audio processor 9130, a display 9160, and a power supply 9170. It is worth noting that the electronic device 9600 does not necessarily need to include these components. Figure 3 All components shown; in addition, the electronic device 9600 may also include Figure 3 For components not shown, please refer to existing technology.

[0150] like Figure 3 As shown, the central processing unit 9100, sometimes also referred to as a controller or operating control, may include a microprocessor or other processor device and / or logic device, which receives inputs and controls the operation of various components of the electronic device 9600.

[0151] The memory 9140 may be, for example, one or more of a cache, flash memory, hard drive, removable media, volatile memory, non-volatile memory, or other suitable devices. It may store the aforementioned failure-related information, and also store a program for executing that information. The central processing unit 9100 may execute the program stored in the memory 9140 to perform information storage or processing, etc.

[0152] Input unit 9120 provides input to central processing unit 9100. Input unit 9120 may be, for example, a keypad or touch input device. Power supply 9170 provides power to electronic device 9600. Display 9160 displays images and text. Display may be, for example, an LCD display, but is not limited thereto.

[0153] The memory 9140 can be a solid-state memory, such as a read-only memory (ROM), random access memory (RAM), a SIM card, etc. It can also be a memory that retains information even when power is off, can be selectively erased, and contains more data; examples of this type of memory are sometimes referred to as EPROMs. The memory 9140 can also be some other type of device. The memory 9140 includes a buffer memory 9141 (sometimes referred to as a buffer). The memory 9140 may include an application / function storage unit 9142 for storing application programs and function programs or processes for executing the operation of the electronic device 9600 via the central processing unit 9100.

[0154] The memory 9140 may also include a data storage unit 9143 for storing data, such as contacts, digital data, pictures, sounds, and / or any other data used by the electronic device. The driver storage unit 9144 of the memory 9140 may include various drivers for the electronic device for communication functions and / or for performing other functions of the electronic device (such as messaging applications, address book applications, etc.).

[0155] The communication module 9110 is a transmitter / receiver that sends and receives signals via the antenna 9111. The communication module 9110 (transmitter / receiver) is coupled to the central processing unit 9100 to provide input signals and receive output signals, which is the same as in a conventional mobile communication terminal.

[0156] Based on different communication technologies, multiple communication modules 9110 can be configured in the same electronic device, such as cellular network modules, Bluetooth modules, and / or wireless LAN modules. The communication module 9110 (transmitter / receiver) is also coupled to a speaker 9131 and a microphone 9132 via an audio processor 9130 to provide audio output via the speaker 9131 and receive audio input from the microphone 9132, thereby realizing typical telecommunications functions. The audio processor 9130 may include any suitable buffer, decoder, amplifier, etc. Additionally, the audio processor 9130 is coupled to a central processing unit 9100, enabling on-device recording via the microphone 9132 and on-device playback of stored audio via the speaker 9131.

[0157] Embodiments of this application also provide a computer-readable storage medium capable of implementing all steps of the intelligent adjudication method based on dispute point matching and dynamic weight fusion, where the execution subject is a server or client, as described in the above embodiments. The computer-readable storage medium stores a computer program that, when executed by a processor, implements all steps of the intelligent adjudication method based on dispute point matching and dynamic weight fusion, where the execution subject is a server or client, as described in the above embodiments. For example, when the processor executes the computer program, it implements the following steps: Step S101: Receive the initial dispute point data stream output by the large model, extract the dispute point feature vector, construct the calculation index matrix, import the influence data item, evidence association parameters, and rule matching factors into the preset calculation unit, generate a dispute point graded dataset, construct a feature mapping matrix based on the graded dataset, perform data parsing and format conversion on the input evidence file, perform vector operations on the evidence feature data and the mapping matrix, generate an evidence annotation data stream according to the matching degree threshold, and output a data card containing dispute level identifier and evidence association parameters; Step S102: Obtain expert processing parameters based on the data card, read the adjudication records in the expert historical database, perform numerical calculations on the dispute type processing quantitative indicators and result pass rate data according to the configuration parameters, generate domain fit data, construct a weight allocation model based on the fit data and historical accuracy parameters, generate a task scheduling sequence and weight coefficient matrix according to the allocation model, and output a task data table containing allocation parameters and scheduling strategies. Step S103: Collect scoring data stream based on the task data table, execute the standard deviation calculation unit, mark data items whose scoring deviation values ​​exceed the set threshold as abnormal nodes, perform parameter correction on the weight coefficients of the abnormal nodes, perform data fusion operation based on the corrected weight matrix, generate conclusion confidence data, trigger the task reassignment mechanism according to the confidence data, and output the result dataset containing the fusion conclusion and confidence coefficient.

[0158] As described above, the computer-readable storage medium provided in this application, through an innovative design of a dispute analysis system, achieves effective establishment of associations through feature extraction and evidence matching. It constructs a scheduling and allocation mechanism, combining suitability assessment and weight calculation to establish a reliable adjudication strategy. Furthermore, it introduces result fusion, ensuring the accuracy of conclusions through bias processing and confidence assessment. This method effectively addresses the shortcomings of traditional technologies in dispute analysis, scheduling and allocation, and result fusion, providing technical support for intelligent adjudication.

[0159] Embodiments of this application also provide a computer program product capable of implementing all steps of the intelligent adjudication method based on dispute point matching and dynamic weight fusion, where the execution subject is a server or client, as described in the above embodiments. When executed by a processor, this computer program / instruction implements the steps of the intelligent adjudication method based on dispute point matching and dynamic weight fusion. For example, the computer program / instruction implements the following steps: Step S101: Receive the initial dispute point data stream output by the large model, extract the dispute point feature vector, construct the calculation index matrix, import the influence data item, evidence association parameters, and rule matching factors into the preset calculation unit, generate a dispute point graded dataset, construct a feature mapping matrix based on the graded dataset, perform data parsing and format conversion on the input evidence file, perform vector operations on the evidence feature data and the mapping matrix, generate an evidence annotation data stream according to the matching degree threshold, and output a data card containing dispute level identifier and evidence association parameters; Step S102: Obtain expert processing parameters based on the data card, read the adjudication records in the expert historical database, perform numerical calculations on the dispute type processing quantitative indicators and result pass rate data according to the configuration parameters, generate domain fit data, construct a weight allocation model based on the fit data and historical accuracy parameters, generate a task scheduling sequence and weight coefficient matrix according to the allocation model, and output a task data table containing allocation parameters and scheduling strategies. Step S103: Collect scoring data stream based on the task data table, execute the standard deviation calculation unit, mark data items whose scoring deviation values ​​exceed the set threshold as abnormal nodes, perform parameter correction on the weight coefficients of the abnormal nodes, perform data fusion operation based on the corrected weight matrix, generate conclusion confidence data, trigger the task reassignment mechanism according to the confidence data, and output the result dataset containing the fusion conclusion and confidence coefficient.

[0160] As described above, the computer program product provided in this application, through an innovative design of a dispute analysis system, achieves effective establishment of associations through feature extraction and evidence matching. It constructs a scheduling and allocation mechanism, combining suitability assessment and weight calculation to establish a reliable adjudication strategy. Furthermore, it introduces result fusion, ensuring the accuracy of conclusions through bias handling and confidence assessment. This method effectively addresses the shortcomings of traditional technologies in dispute analysis, scheduling and allocation, and result fusion, providing technical support for intelligent adjudication.

[0161] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0162] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0163] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0164] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0165] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. An intelligent adjudication method based on dispute point matching and dynamic weight fusion, characterized in that, The method includes: The system receives the initial dispute point data stream output by the large model, extracts the dispute point feature vectors, constructs a calculation index matrix, imports the influence data items, evidence association parameters, and rule matching factors into a preset calculation unit, generates a dispute point graded dataset, constructs a feature mapping matrix based on the graded dataset, performs data parsing and format conversion on the input evidence files, performs vector operations on the evidence feature data and the mapping matrix, generates an evidence annotation data stream based on the matching degree threshold, and outputs a data card containing dispute level identifiers and evidence association parameters. Based on the data card, expert processing parameters are obtained, and the adjudication records in the expert historical database are read. The quantitative indicators of dispute type processing and the result pass rate data are numerically calculated according to the configuration parameters to generate domain fit data. Based on the fit data and historical accuracy parameters, a weight allocation model is constructed. According to the allocation model, a task scheduling sequence and a weight coefficient matrix are generated, and a task data table containing allocation parameters and scheduling strategies is output. Based on the task data table, a scoring data stream is collected, and a standard deviation calculation unit is executed. Data items whose scoring deviation values ​​exceed a set threshold are marked as abnormal nodes. The weight coefficients of the abnormal nodes are corrected. Based on the corrected weight matrix, a data fusion operation is performed to generate conclusion confidence data. The task reassignment mechanism is triggered based on the confidence data, and a result dataset containing the fusion conclusion and confidence coefficient is output.

2. The intelligent adjudication method based on dispute point matching and dynamic weight fusion according to claim 1, characterized in that, The system receives the initial dispute point data stream output by the large model, extracts the dispute point feature vectors, constructs a calculation index matrix, and imports the influence data items, evidence association parameters, and rule matching factors into a preset calculation unit to generate a dispute point hierarchical dataset, including: The system obtains large model output data from the data access terminal, parses the data format type according to the data header identifier, establishes dispute resolution rules, extracts dispute text attributes according to the preset feature encoding scheme, performs word segmentation on the dispute text, constructs a feature mapping table, and generates a dispute feature set containing topic word vectors and semantic structures based on the feature mapping table. Based on the feature set of disputed points, an index calculation rule is constructed. The influence parameter is discretely encoded according to a preset quantification standard. An evaluation score is generated according to the evidence correlation standard. Weight coefficients are configured according to the rule matching degree. The score and the weight coefficients are substituted into the calculation matrix. The disputed points are classified according to the calculation results to generate a disputed point dataset containing graded identifiers.

3. The intelligent adjudication method based on dispute point matching and dynamic weight fusion according to claim 1, characterized in that, The process involves constructing a feature mapping matrix based on the hierarchical dataset, parsing and converting the input evidence files, performing vector operations on the evidence feature data and the mapping matrix, generating an evidence annotation data stream based on a matching degree threshold, and outputting data cards containing dispute level identifiers and evidence association parameters, including: Based on the hierarchical dataset, feature extraction rules are constructed, the format of evidence files is identified and the types are classified, a text preprocessing module is established, data format conversion is performed according to the preset standardized process, the evidence content is segmented and vectorized, a feature dimension table is constructed, and a feature data stream containing evidence subject words and semantic vectors is generated according to the feature dimension table. Based on the feature data stream, a mapping calculation rule is constructed, the evidence feature vector is normalized according to a preset standard, matching degree calculation parameters are generated according to the dispute level identifier, the labeling rule is configured according to the preset threshold condition, the calculation parameters are substituted into the mapping matrix, the evidence correlation degree is quantified and labeled according to the calculation result, and a data card containing the dispute level identifier and evidence correlation parameters is generated.

4. The intelligent adjudication method based on dispute point matching and dynamic weight fusion according to claim 1, characterized in that, The process of obtaining expert processing parameters based on the data card, reading adjudication records from the expert historical database, and numerically calculating the dispute type processing quantification indicators and result pass rate data according to the configured parameters to generate domain suitability data includes: Based on data cards, expert retrieval rules are constructed, access permissions are verified for the expert historical database, a data reading channel is established, and judgment records are obtained according to preset filtering conditions. The record content is classified and statistically analyzed, and a processing quantitative index table is constructed. Based on the quantitative index table, an expert basic dataset containing processing frequency and pass rate is generated. Based on the aforementioned expert dataset, a fitting degree calculation rule is constructed. The processing quantitative indicators are normalized according to a preset standard. Weight coefficients are generated based on the pass rate parameter. Calculation templates are configured according to the dispute type. The weight coefficients are substituted into the calculation unit. The expert fitting degree is quantitatively evaluated based on the calculation results, and fitting degree data containing domain matching degree and processing capability is generated.

5. The intelligent adjudication method based on dispute point matching and dynamic weight fusion according to claim 1, characterized in that, The weight allocation model is constructed based on the fitness data and historical accuracy parameters. A task scheduling sequence and weight coefficient matrix are generated according to the allocation model, and a task data table containing allocation parameters and scheduling strategies is output, including: Weight calculation rules are constructed based on fit data, historical accuracy parameters are standardized, a parameter fusion module is established, indicators are combined according to preset priority standards, the effectiveness of the fusion results is verified, a weight calculation matrix is ​​constructed, and an allocation model containing expert rating weights and task priorities is generated based on the weight calculation matrix. Based on the allocation model, a scheduling rule set is constructed, the task sequence is divided according to a preset time window, a scheduling priority is generated according to the system load parameters, an allocation strategy is configured according to the task type, the priority is substituted into the scheduling matrix, and the task is configured with resources according to the allocation result, generating a task data table containing allocation parameters and scheduling strategies.

6. The intelligent adjudication method based on dispute point matching and dynamic weight fusion according to claim 1, characterized in that, The process involves collecting scoring data streams based on the task data table, executing a standard deviation calculation unit, marking data items with scoring deviation values ​​exceeding a set threshold as abnormal nodes, correcting the weight coefficients of the abnormal nodes, performing data fusion operations based on the corrected weight matrix, and generating conclusion confidence data, including: Data collection rules are constructed based on the task data table. Expert scoring data is collected and formatted in real time. A deviation calculation module is established. Standard deviation is calculated according to preset statistical standards. Threshold comparison is performed on the scoring results. An anomaly marking matrix is ​​constructed. A dataset containing scoring deviation values ​​and anomaly nodes is generated based on the anomaly marking matrix. Based on the dataset, a weight correction rule is constructed, the weight coefficients of abnormal nodes are adjusted according to a preset ratio, a correction matrix is ​​generated according to the adjustment parameters, a calculation template is configured according to the fusion calculation standard, the correction matrix is ​​substituted into the fusion unit, and the confidence of the conclusion is quantified according to the calculation result to generate a dataset containing the fusion conclusion and confidence.

7. The intelligent adjudication method based on dispute point matching and dynamic weight fusion according to claim 1, characterized in that, The task reassignment mechanism triggered based on the confidence data outputs a result dataset containing fusion conclusions and confidence coefficients, including: Based on confidence data, a redistribution trigger rule is constructed, a threshold determination is made on the confidence value, a task allocation controller is established, experts are screened according to preset supplementary standards, the status of allocable resources is detected, a task allocation matrix is ​​constructed, and a redistribution scheme containing an expert supplementary list and allocation strategy is generated based on the task allocation matrix. Based on the redistribution scheme, result synthesis rules are constructed, supplementary scoring data are fused according to preset standards, weight parameters are generated based on confidence coefficients, synthesis templates are configured according to conclusion types, the weight parameters are substituted into the calculation unit, the fusion conclusion is finally confirmed based on the calculation results, and a result dataset containing conclusion data and confidence coefficients is generated.

8. An intelligent adjudication device based on dispute point matching and dynamic weight fusion, characterized in that, The device includes: The dispute matching module is used to receive the initial dispute point data stream output by the large model, extract the dispute point feature vectors, construct the calculation index matrix, import the influence data items, evidence association parameters, and rule matching factors into the preset calculation unit, generate a dispute point hierarchical dataset, construct a feature mapping matrix based on the hierarchical dataset, perform data parsing and format conversion on the input evidence files, perform vector operations on the evidence feature data and the mapping matrix, generate an evidence annotation data stream according to the matching degree threshold, and output a data card containing dispute level identifiers and evidence association parameters. The weight fusion module is used to obtain expert processing parameters based on the data card, read the adjudication records in the expert historical database, perform numerical calculations on the dispute type processing quantitative indicators and result pass rate data according to the configuration parameters, generate domain fit data, construct a weight allocation model based on the fit data and historical accuracy parameters, generate a task scheduling sequence and weight coefficient matrix according to the allocation model, and output a task data table containing allocation parameters and scheduling strategies. The intelligent assessment module is used to collect scoring data streams based on the task data table, execute the standard deviation calculation unit, mark data items whose scoring deviation values ​​exceed a set threshold as abnormal nodes, perform parameter correction on the weight coefficients of the abnormal nodes, perform data fusion operations based on the corrected weight matrix, generate conclusion confidence data, trigger the task reassignment mechanism based on the confidence data, and output a result dataset containing the fusion conclusion and confidence coefficients.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the intelligent adjudication method based on dispute point matching and dynamic weight fusion as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the intelligent adjudication method based on dispute point matching and dynamic weight fusion as described in any one of claims 1 to 7.

Images