Dispute mediation key information extraction system based on user feedback

Abstract

The invention provides a dispute mediation key information extraction system based on user feedback. The dispute mediation key information extraction system comprises a multi-source data acquisition module, a legal semantic analysis module, an intelligent decision support module and a user feedback analysis module. The system can collect mediation voice, extract legal information, generate a mediation scheme, perform risk assessment, collect user satisfaction feedback after mediation is completed, automatically optimize the scheme to generate parameters, form a self-adaptive optimization closed loop of a mediation strategy, and effectively improve the mediation effect and the user satisfaction.

Description

Technical Field

[0001] The present invention relates to the cross-application field of artificial intelligence and legal technology, and specifically to a dispute mediation key information extraction system based on user feedback, belonging to the field of intelligent judicial assistance technology. Background Art

[0002] In current dispute mediation practices, the collection and processing of voice data has become a crucial tool for improving mediation efficiency and quality. Typically, voice information generated during mediation is collected using recording equipment and subsequently converted into text using speech recognition technology. However, due to the diversity of mediation environments and the presence of background noise, the quality of this voice data can vary, hindering subsequent information extraction and analysis.

[0003] In terms of legal semantic analysis, existing technologies mainly rely on natural language processing (NLP) methods to analyze texts to identify legal terms and related elements. For example, the existing patent CN113011185A proposes a legal text analysis and recognition method based on the RoBERTa model, which aims to improve the accuracy of legal terminology recognition and assist in case semantic understanding and legal application judgment. In addition, constructing a legal knowledge graph is also a common practice. Legal entities and their relationships are represented by a graph structure to improve the organizational efficiency and automated processing capabilities of legal information. However, most of the related technologies still remain at the static graph construction level, making it difficult to respond to dynamic changes in specific mediation scenarios in real time.

[0004] In terms of intelligent decision support, some existing systems attempt to combine intelligent recommendations with legal adaptation logic to assist in the development of mediation strategies. Patent CN109815467B proposes a mediation decision support system based on judicial document analysis. This system generates case-by-case reasoning and reference conclusions based on historical cases, and generates legal basis recommendations based on legal provisions. However, when handling new or complex cases, the adaptability of the model and the transparency of the reasoning path still need to be improved, especially in cases involving multiple parties and complex legal relationships.

[0005] Regarding the utilization of user feedback, current mediation systems rarely systematically collect and analyze feedback from mediators and parties. While research in other fields, such as customer service, has proposed intelligent customer query solution recommendation systems that optimize service quality by analyzing user feedback, similar feedback mechanisms are not yet widely used in legal mediation.

[0006] In summary, the shortcomings of existing dispute mediation-related technologies limit the mediation system's ability to handle complex cases and improve user satisfaction. There is an urgent need to propose a comprehensive solution that better integrates multimodal processing and human-computer feedback mechanisms. Summary of the Invention

[0007] In order to solve the above problems in the prior art, the present invention proposes a dispute mediation key information extraction system based on user feedback, the extraction system comprising:

[0008] Multi-source data collection module, used to collect dispute mediation voice data;

[0009] Legal semantic parsing module, used to extract legal information from the text converted from mediation recordings;

[0010] An intelligent decision support module for generating mediation plans and conducting risk assessments based on the legal information;

[0011] A user feedback analysis module is used to collect user evaluation data after mediation is completed and optimize the mediation strategy generation logic, wherein the user feedback analysis module includes:

[0012] Satisfaction collection unit, used to collect satisfaction feedback from mediators and parties on the mediation plan;

[0013] A feedback analysis unit, configured to perform text analysis and index extraction on the feedback data;

[0014] A parameter optimization unit is used to adjust the parameter configuration of the mediation solution generation engine according to the analysis results to form a feedback loop for strategy optimization.

[0015] The multi-source data acquisition module includes:

[0016] Cloud call system for multi-channel voice communication recording;

[0017] The adaptive noise reduction unit is used to perform noise suppression processing on the collected voice data.

[0018] The legal semantic analysis module includes:

[0019] Legal entity identification unit, used to identify legal terms and relevant legal elements in mediation statements;

[0020] A dynamic legal knowledge graph construction unit is used to establish and update the legal relationship graph related to mediation cases.

[0021] The intelligent decision support module includes:

[0022] A mediation solution generation engine, used to generate multiple candidate solutions oriented towards mediation goals;

[0023] The risk identification unit is used to identify risk points in the legal feasibility and enforceability of mediation strategies.

[0024] The mediation solution generation engine receives information on dispute elements, parties' positions, conflicting points of claims, and applicable legal provisions in the legal knowledge graph;

[0025] generating a mediation plan based on at least one mediation objective, wherein the mediation objective includes mediation cost, mediation time, or satisfaction;

[0026] Match the feature vectors of the current case with those of historical cases to obtain several historical mediation paths as candidate strategies;

[0027] The mediation schemes are organized in a tree structure, each mediation scheme includes at least one mediation direction and its corresponding execution suggestion, and multiple sets of optional mediation strategies are output.

[0028] The risk identification unit includes:

[0029] Conduct legal feasibility assessments on candidate mediation proposals, using a legal database and case conflict judgment model to identify legal application issues in the mediation proposals;

[0030] Performing an execution feasibility assessment on the candidate mediation solutions, wherein the assessment quantifies the execution risks of different solutions based on execution-related information;

[0031] Set multiple risk factors and generate a risk score for each candidate plan through weighted calculation;

[0032] The candidate mediation solutions are ranked according to the risk scores for mediation strategy recommendation.

[0033] The parameter optimization unit makes targeted adjustments to multiple control parameters in the mediation solution generation engine based on the structured feedback indicators output by the feedback analysis unit;

[0034] The feedback indicators include user subjective evaluation results and semantic sentiment analysis results;

[0035] The control parameters include mediation problem modeling template selection weight, historical case recall strategy weight, mediation tone and wording control parameters.

[0036] The parameter optimization unit supports incremental learning and periodic weight updates, and uses a sliding window mechanism to record the feedback results of the latest several rounds.

[0037] In each optimization cycle, the following weight evolution function is executed to calculate the strategy parameter θ k The incremental update value Δθ k :

[0038] Δθ k =η·(S k ·(1-β·V k )+γ·Ck ),

[0039] in:

[0040] η is the basic learning rate,

[0041] S k is the feedback guidance strength,

[0042] V k is the historical weight volatility variance,

[0043] β is the fluctuation suppression factor,

[0044] C k is the semantic concentration index,

[0045] γ is the semantic convergence adjustment factor.

[0046] The optimized parameters are stored in the parameter configuration library and called by the mediation scheme generation engine in the next round of mediation tasks to achieve closed-loop adaptive optimization of the mediation scheme.

[0047] Beneficial effects:

[0048] This invention incorporates a user feedback analysis module to achieve closed-loop optimization of the mediation solution generation process, significantly improving the system's adaptability to changes in user satisfaction. After the mediation is completed, the system collects feedback from mediators and parties, automatically performing text analysis and parameter adjustments, thereby continuously optimizing the strategy generation logic and improving the fit and feasibility of mediation solutions. Furthermore, by integrating voice acquisition with legal semantic analysis, the system accurately extracts dispute information, improving mediation efficiency and intelligent decision-making overall. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application, but do not constitute an improper limitation of the present invention. In the drawings:

[0050] Figure 1 Shows the overall structural diagram of the dispute mediation key information extraction system.

[0051] Figure 2 A functional flow diagram of the user feedback analysis module is shown. DETAILED DESCRIPTION

[0052] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The exemplary embodiments and descriptions are only used to explain the present invention but are not intended to limit the present invention.

[0053] Example 1: Dispute Mediation Key Information Extraction System Based on User Feedback

[0054] like Figure 1 As shown, this embodiment provides a system for extracting key information for dispute mediation based on user feedback. The system includes a multi-source data acquisition module 101, a legal semantics analysis module 102, an intelligent decision support module 103, and a user feedback analysis module 104. These modules are interconnected via internal data links, forming a closed-loop processing system for data collection, semantic recognition, decision generation, and feedback optimization during the dispute mediation process.

[0055] The multi-source data acquisition module 101 is used to collect voice conversation data, including that between the mediator and the parties, during the mediation process. This module integrates a cloud call system 1011 and an adaptive noise reduction unit 1012. Deployed at the front end of the mediation platform, the cloud call system 1011 supports multi-channel voice communication, synchronously recording the conversation content between the mediator and the parties, and identifying the call channels for subsequent semantic separation. The adaptive noise reduction unit 1012 receives the raw voice stream from the cloud call system 1011, dynamically removes ambient noise using spectral smoothing and echo suppression algorithms, and outputs a clear voice signal to improve recognition accuracy.

[0056] The Legal Semantic Parsing Module 102 converts audio data into text and extracts legal information relevant to the mediation matter. First, the mediation recording is converted into editable text using a speech recognition model. Subsequently, the Legal Entity Identification Unit 1021 structures the text, identifying elements such as legal terminology, points of dispute, and subject relationships. The Dynamic Legal Knowledge Graph Construction Unit 1022 dynamically updates the knowledge graph nodes and edge weights based on these elements to represent the rights and obligations formed during the mediation process and their evolutionary paths. This graph is stored in real time in a graph database, with open interfaces for access by other modules.

[0057] The intelligent decision support module 103 generates feasible mediation plans based on semantic analysis results and assesses their execution risks. The mediation plan generation engine 1031 combines graph information, historical cases, and mediation objectives to construct multiple alternative plans, targeting different options such as reconciliation, concessions, compensation, and third-party intervention. The risk identification unit 1032 verifies the legal feasibility and analyzes the execution costs of each candidate plan, identifying risks such as potential legal conflicts and insufficient precedent adaptability, and generates a risk score for each plan.

[0058] like Figure 2 As shown, the user feedback analysis module 104 includes a satisfaction collection unit 1041, a feedback analysis unit 1042, and a parameter optimization unit 1043, forming a self-learning closed loop for the system. After the mediation concludes, the system pushes a structured evaluation questionnaire or an open comment interface to the mediator and parties via a mobile device or web interface to collect feedback. This feedback covers the rationality of the mediation plan, its feasibility, and the clarity of the legal interpretation.

[0059] Feedback analysis unit 1042, using natural language processing technology and sentiment recognition algorithms, extracts keywords, sentiment, and ratings from feedback texts to form a structured feedback index set. This index set is fed into parameter optimization unit 1043, which fine-tunes the weighting parameters in mediation solution generation engine 1031 based on various indicators, including but not limited to the weighting for problem modeling template selection, historical case retrieval strategies, and mediation tone and wording parameters, to improve solution fit and satisfaction.

[0060] Through the above-mentioned structural configuration, the system realizes a closed loop of the entire process from data collection, legal semantic recognition, solution generation to user feedback optimization, thereby improving the intelligence level and user acceptance of the dispute mediation process. A modular deployment structure is adopted between each module, which can be flexibly integrated into the existing judicial mediation platform or social conflict mediation system. This embodiment is only for illustrative purposes and does not constitute a limitation on the scope of protection of the present invention. The implementation of each specific module is specifically shown in the following Examples 2-4.

[0061] Example 2 Multi-source data acquisition module

[0062] Multi-source data acquisition module 101 is used to collect voice conversation data, including that between the mediator and the parties, during the mediation process, ensuring complete recording of the entire mediation process and the quality of the voice for subsequent processing. In this embodiment, multi-source data acquisition module 101 includes a cloud call system 1011 and an adaptive noise reduction unit 1012, which work together to achieve real-time collection and preprocessing of high-fidelity voice data.

[0063] The cloud call system 1011 is deployed at the front-end node of the mediation system and supports two-way voice communication with multiple call terminals. The cloud call system adopts a communication architecture based on SIP (Session Initiation Protocol) or WebRTC protocol, which can establish multiple concurrent audio channels and assign a unique call channel identifier to each call through the audio data marking mechanism. The system automatically generates a session record ID when mediation is initiated, and binds the voice streams of both parties to achieve parallel recording and channel attribution management of the audio during the entire conversation. The recording data is collected in a frame-by-frame manner, and the sampling rate is preferably set to 16kHz and the bit depth is 16bit to ensure the clarity of voice restoration. The collected data is temporarily cached in the RAM cache in PCM format, and is synchronously written to the voice stream buffer for subsequent module calls.

[0064] The adaptive noise reduction unit 1012 is used to perform noise suppression processing on the original voice stream from the cloud call system 1011. In this embodiment, the noise reduction unit adopts a multi-channel noise suppression strategy based on a combination of spectral subtraction and Wiener filtering. First, a short-time Fourier transform (STFT) is performed on the voice signal to extract the amplitude spectrum and phase spectrum of each frame; after estimating the noise power spectrum, the spectral subtraction algorithm is used to subtract the noise estimate from the input voice spectrum to generate a preliminary denoised voice spectrum; then, the Wiener filter is used to further smooth the spectrum to reduce residual noise. For situations where there is background continuous noise or burst noise interference during a call, the minimum mean square error (MMSE) enhancement criterion is introduced to dynamically adjust the noise estimate to improve the adaptability of the noise reduction process in different call scenarios.

[0065] After noise reduction, the voice data is restored to a time-domain voice signal through an inverse STFT transform and uniformly encoded into a compressed format such as FLAC or Opus before being sent to the subsequent semantic recognition module. In this way, the system can effectively filter out environmental noise and device echoes without affecting the integrity of the voice content, thereby improving the accuracy of voice transcription and legal semantic extraction. The audio processing flow in this embodiment is all completed at the edge node, avoiding delays or packet loss during transmission, and improving the real-time responsiveness and stability of the system.

[0066] Example 3 Legal Semantic Analysis Module

[0067] The Legal Semantic Parsing Module 102 converts audio data collected during the mediation process into structured legal information, providing a semantic foundation for subsequent mediation solution generation and risk assessment. This module primarily includes a speech recognition processing flow, a legal entity identification unit 1021, and a dynamic legal knowledge graph construction unit 1022. These units collaborate to accurately extract and graphically represent legal semantics.

[0068] First, the input audio data comes from the noise reduction processing results of the front-end multi-source data acquisition module. The system uses an end-to-end speech recognition model to transcribe the speech data. The model is based on a deep neural network structure, such as the Transformer or Conformer architecture. After supervised training based on legal mediation scenario corpus, it can achieve highly robust automatic speech recognition (ASR) in situations with multiple speakers, multiple accents, and irregular punctuation. The recognition process extracts frame-level features from the speech data, including input features such as Mel-frequency cepstral coefficients (MFCC) or logarithmic Mel-spectrograms, and generates editable text output with timestamps after decoding by the model.

[0069] The text output enters the legal entity recognition unit 1021, which uses named entity recognition (NER) technology to perform structured analysis on the text and extract multiple types of legal entities, including legal terms, dispute focus, claims, party identities, and rights and obligations. Specifically, the system introduces a model structure based on bidirectional encoding representation (such as BiLSTM-CRF, BERT+CRF), combined with labeled data in the legal field for training, to distinguish entity categories and mark their start and end positions. This recognition process simultaneously considers the contextual information and syntactic dependencies of words to improve the recognition accuracy of nested legal entities in complex and long sentences.

[0070] Taking structured text data from legal entity identification as input, the dynamic legal knowledge graph construction unit 1022 organizes various entities and their logical relationships into a graph structure. The graph takes the form of a directed graph, with entity nodes representing case elements such as parties, contract terms, legal basis, and claims. Edges represent the relationship types between entities, such as "filing a request," "obligation," and "according to a clause." The system extracts edge types and weights using rule templates and dependency syntactic relationships, and then attaches relationship semantic labels and contextual sentence information to the edges.

[0071] The constructed legal knowledge graph is dynamically stored in a graph database, such as Neo4j or JanusGraph, and accessible to other modules within the system through standard interfaces provided by graph database query languages ​​(such as Cypher). The graph supports real-time updates. As new information enters or existing relationships evolve during the mediation process, the graph structure is automatically updated incrementally, ensuring the real-time and consistent knowledge representation. Each node and edge is bound to a timestamp and source information, enabling case tracing and visualization of mediation paths.

[0072] The design and deployment of this legal semantic analysis module enables deep semantic understanding of unstructured voice data, storing and accessing it in a structured, graphical format. This significantly improves the efficiency of semantic analysis and the accuracy of solution support during mediation. This module demonstrates stability, scalability, and cross-module interface compatibility during execution, making it suitable for integration and expansion into various mediation service platforms.

[0073] Example 4 Intelligent Decision Support Module

[0074] The intelligent decision support module 103 generates feasible mediation solutions based on the structured information output by the legal semantics analysis module and systematically assesses the execution risks of each solution. This module includes a mediation solution generation engine 1031 and a risk identification unit 1032, which collaborate to translate semantics into mediation strategies, enabling multi-strategy responses to complex dispute situations.

[0075] The mediation solution generation engine 1031 first receives the map information such as dispute elements, party positions, conflicting points of claims, and scope of application of relevant laws expressed in the legal knowledge map, and constructs a solution in combination with the mediation goal. In this embodiment, the mediation goal is set by the system or manually by the mediator, including objective functions such as minimizing mediation costs, shortening mediation time, and maximizing satisfaction. During the solution generation process, the system calls the historical case library, matches the feature vectors of the current case with the archived cases based on the characteristics, and extracts relevant historical mediation paths as candidate strategy references through similarity calculation. The above matching adopts a weighted vector similarity model, and the vector dimensions include features such as dispute type, claim structure, applicable legal clause number, number of mediation entities, and case complexity.

[0076] During the proposal development phase, the system pre-determines potential outcomes, including but not limited to: reaching a settlement, partial concessions, financial compensation, or recommending the involvement of a third party (such as a professional organization or notary). The proposal is presented as a structured tree, with the root node representing the mediation goal, child nodes representing alternative paths, and leaf nodes representing recommended actions. Each path undergoes strategic evaluation and ranking based on historical data, allowing the system to generate multiple sets of alternatives for subsequent screening by the evaluation module.

[0077] Risk identification unit 1032 is responsible for evaluating the legal feasibility and enforceability of the aforementioned mediation proposals. This legal feasibility assessment, based on the binding rules and case conflict judgment model in the legal database, compares the applicable legal provisions in the mediation proposal with the actual case facts, identifying potential legality flaws or inconsistencies. The enforcement cost analysis, based on information such as the regional legal enforcement environment, the parties' ability to perform, and historical enforcement cases, quantifies the actual obstacles that different mediation paths may encounter in terms of enforcement.

[0078] The system sets multi-dimensional risk factors, including: precedent suitability score, clause conflict risk, execution difficulty assessment and performance expectation deviation. Each plan is assigned a risk score, which is calculated by multi-factor weighting. The scoring model (RiskScore) can be set as:

[0079] RiskScore=w1·F1+w2·F2+w3·F3+…+w n ·F n ,

[0080] Among them, F1 to F n Represent the risk factor values, w1 to w n The final output list of solutions will be ranked based on the risk score and the preset mediation objective function, with low-risk and high-satisfaction strategies being recommended for mediators to choose.

[0081] Through the construction of the above-mentioned intelligent decision support module, this embodiment can realize an automated closed loop in dispute mediation from case semantic understanding to solution generation and risk assessment, greatly improving mediation efficiency and strategic scientificity, and effectively reducing the risk of performance failure or secondary disputes caused by improper solution design.

[0082] Example 5 User Feedback Analysis Module

[0083] The user feedback analysis module 104 is used to quantitatively assess user acceptance of mediation solutions after the mediation process concludes. Based on this feedback, it optimizes the core parameter configuration of the mediation solution generation engine 1031, thereby forming a closed-loop strategy improvement mechanism. This module comprises a satisfaction collection unit 1041, a feedback analysis unit 1042, and a parameter optimization unit 1043, all of which operate in tandem, enabling efficient feedback collection and response capabilities.

[0084] After the mediation is completed, the satisfaction collection unit 1041 sends feedback invitations to the mediator and all parties through the system's built-in feedback push mechanism. Feedback comes in two forms: a structured evaluation questionnaire and an open natural language input interface. The structured questionnaire sets questions in multiple dimensions, covering the rationality of the mediation plan (such as whether it is fair and whether it fits the focus of the dispute), the feasibility of implementation (such as whether the terms of the agreement are operational), and the clarity of legal interpretation (such as whether the applicable laws and precedents are understood). The system uses a five-level scale to record the questionnaire scoring results. For open text input, users can fill in suggestions, emotional tendencies or dissatisfaction content on their own, and the feedback information will retain the original expression in full.

[0085] The feedback analysis unit 1042 receives the feedback content from the satisfaction collection unit 1041 and converts it into structured indicators. For the structured questionnaire part, the scale score is directly extracted and normalized to a unified indicator space. For the open text part, the system uses natural language processing technology, including word segmentation, part-of-speech tagging, named entity recognition and dependency syntax analysis. Keyword extraction is performed through the dual strategy of TF-IDF and TextRank, and emotional tendency recognition adopts a deep sentiment classification model based on BiGRU-Attention to output emotional polarity labels and confidence. The model performs transfer learning on an emotional dataset containing legal mediation corpus to improve the accuracy of emotion recognition in specific scenarios.

[0086] The analysis results are output as a structured feedback metrics set, including but not limited to: overall satisfaction score, suggested content keyword set, positive / negative sentiment ratio, expression intensity distribution, content focus, etc. The feedback analysis unit also features an anomaly detection module that automatically identifies feedback scores below a set threshold or instances of extreme sentiment fluctuations, prompting manual review to mitigate systemic biases.

[0087] Parameter Optimization Unit 1043 is used to fine-tune the core control parameters in Mediation Solution Generation Engine 1031 based on the structured indicators output by Feedback Analysis Unit 1042, thereby continuously optimizing mediation strategies and iteratively improving user satisfaction. This optimization process, informed by user subjective evaluations and semantic sentiment analysis, is highly sensitive and adaptable, supporting personalized parameter adjustments across a wide range of mediation scenarios.

[0088] In terms of the selection of mediation problem modeling templates, the system presets a variety of semantic modeling structure templates, such as a three-segment template based on conflict classification, a comparative template based on interest reconstruction, and a cross-type template based on demand coordination. Each template has a set of structural weight parameters, which represent its applicable priority in different situations. The parameter optimization unit 1043 dynamically updates the selection weight of each template in the template set based on the scores of dimensions such as "logical clarity of the solution" or "degree of focus on the dispute" in the feedback. For example, if the feedback indicates that the current mediation solution is vague in clarifying the key points of the dispute, the system will increase the weight of the template suitable for structured segmented expression and reduce the priority of the template that tends to be emotionally soothing, thereby guiding the subsequent mediation generation logic to focus more on the core of the dispute.

[0089] In order to optimize the historical case retrieval strategy, the system uses multi-dimensional vectors to represent the content characteristics of mediation cases, including case type, dispute attributes, applicable laws, mediation result form, etc. In the initial stage of solution generation, vector recall uses weighted cosine similarity to match similar cases. After receiving indicators such as "irrelevant cases" or "unrepresentative cases" output by the feedback analysis unit, the parameter optimization unit 1043 adjusts the weight matrix in the recall strategy. For example, when a user expresses dissatisfaction with the reference nature of the mediation result, the system will increase the weight of the vector dimension related to the "mediation path evolution pattern" and weaken the weight of the case background information to improve the reference value of the recalled cases in the solution structure.

[0090] To adjust mediation tone and wording parameters, the mediation solution generation engine 1031 includes a language generation layer whose control parameters include language style, emotion regulation factors, and tone moderation. The parameter optimization unit 1043 adjusts the activation function and word vector weights of the emotion embedding layer in the tone generation template based on the distribution of feedback emotion labels. For example, if the system identifies user feedback that the mediation language is "stiff" or "unfriendly," the system will increase the sampling probability of positive-leaning words in the emotion vocabulary during generation, reduce the use of calm terms with strong logical constraints, and dynamically adjust the language generation temperature value to improve the naturalness and acceptability of the language.

[0091] The parameter optimization process supports incremental learning and periodic weight updates. The system uses a sliding window mechanism to record recent feedback results and executes weight regression and gradient update algorithms within each optimization cycle to maintain the stability and goal orientation of parameter changes. To achieve a more adaptive and semantically convergent parameter update effect, this embodiment proposes a new feedback-driven mediation weight evolution function to replace the conventional weighted average or SGD update method. This function comprehensively considers the feedback sentiment fluctuation range, historical parameter fluctuation frequency, and semantic concentration trend, and its form is as follows:

[0092] Δθ k =η·(S k ·(1-β·V k )+γ·C k )

[0093] This formula is used to calculate the update increment Δθ of the kth mediation strategy control parameter in the current optimization cycle k This update strategy combines the guiding strength of user feedback, the inhibitory effect of historical parameter fluctuations, and the regulatory ability of semantic central tendency on the generation strategy, forming an optimization driving mechanism with multi-factor trade-offs.

[0094] The meaning of each parameter is as follows:

[0095] Δθ k : represents the update amount of the kth adjustment strategy control parameter (such as template weight, wording factor, etc.). It determines in which direction and by how much the parameter should be adjusted in the current optimization cycle.

[0096] η (eta): represents the basic learning rate, a global coefficient used to adjust the overall update step size. Its value is usually set to a small positive number (such as 0.01-0.1) to ensure the stability and gradualness of parameter updates.

[0097] S k : Indicates the strength of the positive feedback on the kth parameter in user feedback. This is determined by the feedback analysis module by normalizing the structured questionnaire scores or sentiment quantification results. Its value range is [0, 1], reflecting the degree to which the current parameter directly impacts user satisfaction.

[0098] V k : represents the parameter θ k The weight variance over the last n optimization cycles measures the volatility of the parameter over its historical evolution. Greater volatility indicates unstable parameter adjustments, and the current update amplitude should be appropriately suppressed to prevent strategy drift.

[0099] β: Variance inhibition factor, whose value range is [0,1], is used to control V k The suppression effect on the update amplitude. The larger the value, the stronger the fluctuation suppression effect on the optimization. It is often used to ensure the stability of the overall system adjustment strategy.

[0100] C k : represents the concentration index of the mediation semantic vector on the kth parameter control dimension. k This parameter measures the linguistic or logical cohesion of the current semantic output (e.g., the language of the mediation proposal) under the influence of this parameter. A larger value indicates a more focused and consistent semantic expression, which helps improve the standardization and persuasiveness of the mediation proposal.

[0101] γ: semantic convergence coefficient, a positive real number used to adjust C k The weight of its role in the overall update volume. The value of γ should be set according to the scenario requirements. If the mediation strategy tends to be standardized, it can be set higher; if it encourages strategy diversity, it can be set lower.

[0102] This formula introduces a multi-dimensional control factor when updating the policy parameters to dynamically integrate user feedback (S k ), parameter history behavior (V k ) and semantic consistency (C k ) to achieve an optimal balance between accuracy, adaptability, and stability. Unlike existing optimization mechanisms driven by loss derivatives or single feedback scores, this approach integrates feedback and semantic constraints while ensuring the stability and diversity of the mediation language, achieving a higher-dimensional personalized solution generation and control strategy.

[0103] The optimized parameters are stored in a parameter configuration library and invoked by the mediation solution generation engine in the next mediation round, enabling feedback-driven closed-loop adaptive evolution of the policy. This mechanism improves the mediation system's responsiveness to user evaluation results and significantly enhances the contextual fit and linguistic expressiveness of the generated solutions. It also avoids policy drift caused by overfitting to short-term feedback during the evolution process, ensuring the system's long-term stability and adaptability.

[0104] Through the closed-loop mechanism of the user feedback analysis module 104, the system can achieve adaptive mediation strategy evolution, improving the intelligent mediation system's ability to perceive and respond to user needs, thereby continuously enhancing the acceptance and social recognition of mediation results. This module is highly scalable in implementation and is suitable for feedback-driven strategy optimization tasks across different mediation platforms and dispute types.

[0105] The above description is only a preferred embodiment of the present invention. Therefore, any equivalent changes or modifications made according to the structure, characteristics and principles described in the scope of the patent application of the present invention are included in the scope of the patent application of the present invention.

Claims

1. A dispute mediation key information extraction system based on user feedback, characterized by: The extraction system comprises: Multi-source data collection module, used to collect dispute mediation voice data; Legal semantic parsing module, used to extract legal information from the text converted from mediation recordings; An intelligent decision support module for generating mediation plans and conducting risk assessments based on the legal information; A user feedback analysis module is used to collect user evaluation data after mediation is completed and optimize the mediation strategy generation logic, wherein the user feedback analysis module includes: Satisfaction collection unit, used to collect satisfaction feedback from mediators and parties on the mediation plan; A feedback analysis unit, configured to perform text analysis and index extraction on the feedback data; A parameter optimization unit is used to adjust the parameter configuration of the mediation solution generation engine according to the analysis results to form a feedback loop for strategy optimization.

2. The system for extracting key dispute mediation information based on user feedback according to claim 1, characterized in that: The multi-source data acquisition module includes: Cloud call system for multi-channel voice communication recording; The adaptive noise reduction unit is used to perform noise suppression processing on the collected voice data.

3. The system for extracting key dispute mediation information based on user feedback according to claim 1, characterized in that: The legal semantic analysis module includes: Legal entity identification unit, used to identify legal terms and relevant legal elements in mediation statements; A dynamic legal knowledge graph construction unit is used to establish and update the legal relationship graph related to mediation cases.

4. The system for extracting key dispute mediation information based on user feedback according to claim 1, characterized in that: The intelligent decision support module includes: A mediation solution generation engine, used to generate multiple candidate solutions oriented towards mediation goals; The risk identification unit is used to identify risk points in the legal feasibility and enforceability of mediation strategies.

5. The system for extracting key dispute mediation information based on user feedback according to claim 4, characterized in that: The mediation solution generation engine receives information on dispute elements, parties' positions, conflicting points of claims, and applicable legal provisions in the legal knowledge graph; generating a mediation plan based on at least one mediation objective, wherein the mediation objective includes mediation cost, mediation time, or satisfaction; Match the feature vectors of the current case with those of historical cases to obtain several historical mediation paths as candidate strategies; The mediation schemes are organized in a tree structure, each mediation scheme includes at least one mediation direction and its corresponding execution suggestion, and multiple sets of optional mediation strategies are output.

6. The system for extracting key dispute mediation information based on user feedback according to claim 4, characterized in that: The risk identification unit includes: Conduct legal feasibility assessments on candidate mediation proposals, using a legal database and case conflict judgment model to identify legal application issues in the mediation proposals; Performing an execution feasibility assessment on the candidate mediation solutions, wherein the assessment quantifies the execution risks of different solutions based on execution-related information; Set multiple risk factors and generate a risk score for each candidate plan through weighted calculation; The candidate mediation solutions are ranked according to the risk scores for mediation strategy recommendation.

7. The system for extracting key dispute mediation information based on user feedback according to claim 1, characterized in that: The parameter optimization unit makes targeted adjustments to multiple control parameters in the mediation solution generation engine based on the structured feedback indicators output by the feedback analysis unit; The feedback indicators include user subjective evaluation results and semantic sentiment analysis results; The control parameters include mediation problem modeling template selection weight, historical case recall strategy weight, mediation tone and wording control parameters.

8. The system for extracting key dispute mediation information based on user feedback according to claim 1, characterized in that: The parameter optimization unit supports incremental learning and periodic weight updates, and uses a sliding window mechanism to record the feedback results of the latest several rounds.

9. The system for extracting key dispute mediation information based on user feedback according to claim 8, characterized in that: In each optimization cycle, the following weight evolution function is executed to calculate the incremental update value Δθk of the policy parameter θk: Δθk=η·(Sk·(1-β·Vk)+γ·Ck), in: η is the basic learning rate, Sk is the feedback guidance strength, Vk is the historical weight volatility variance, β is the fluctuation suppression factor, Ck is the semantic concentration index, γ is the semantic convergence adjustment factor.

10. The system for extracting key dispute mediation information based on user feedback according to claim 9, characterized in that: The optimized parameters are stored in the parameter configuration library and called by the mediation scheme generation engine in the next round of mediation tasks to achieve closed-loop adaptive optimization of the mediation scheme.

Images