An intelligent information extraction system based on multimodal adaptive analysis

Abstract

This invention provides an intelligent information extraction system based on multimodal adaptive analysis, comprising a multi-source data acquisition module, a legal semantic parsing module, and an intelligent decision support module. The system collects voice data and performs noise reduction processing to extract legal semantic information, constructs a dynamic legal knowledge graph, generates mediation plans based on case characteristics, and assesses risks. This achieves intelligent and adaptive optimization of mediation strategies, improving the efficiency and effectiveness of dispute resolution.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary applications of artificial intelligence and legal technology, specifically an intelligent information extraction system based on multimodal adaptive analysis. Background Technology

[0002] In existing technologies, some dispute resolution platforms combine voice recording with text transcription to preserve the statements of both parties as searchable electronic archives. These platforms typically acquire recordings via telephone or internet calls and then rely on manual or general speech recognition models for transcription. Given the potential for noise interference from the on-site environment or network conditions, some systems may still have room for improvement in terms of voice clarity and subsequent recognition accuracy.

[0003] In the processing of text and voice information, some technical solutions attempt to extract key points and points of contention from dialogue using rule-based or shallow semantic natural language processing methods. However, when faced with highly specialized legal terminology or emotional expressions from parties involved, such solutions often struggle to comprehensively identify and summarize the crucial information of a case. Furthermore, some systems analyze data based on accumulated case databases or general knowledge graphs, but if the legal relationships involved in a case change, or if relevant regulations are frequently updated, the existing structured knowledge base cannot respond to this dynamic change in a timely manner, requiring manual correction in subsequent processing.

[0004] Regarding the generation of mediation plans, many dispute resolution systems provide feasible suggestions to parties based on established rules or limited case templates. Some plans may roughly assess potential risks during this process, but these often rely on subjective judgment or a single indicator, making it difficult to consider the multidimensional factors under different types of disputes. Especially when the case type is complex and the relationships or claims of the parties are diverse, the approach of generating mediation strategies with uniform parameter settings may be difficult to adapt to the actual situation, thereby affecting the expected results of the plan at the implementation level.

[0005] Regarding feedback and adjustments in mediation execution, current practices mostly rely on manual retrospective analysis to optimize parameters or improve handling strategies after the fact. Without sufficient data collection on the subsequent execution results and influencing factors of various cases, the mediation system's capacity for sustainable optimization is relatively limited. This, to some extent, renders the existing system insufficiently adaptable, making it difficult to continuously improve its mediation efficiency and relevance in a context-specific manner. Summary of the Invention

[0006] To address the aforementioned problems in the prior art, this invention proposes an intelligent information extraction system for dispute mediation based on multimodal adaptive analysis, comprising:

[0007] A multi-source data acquisition module is used to collect voice data for dispute mediation.

[0008] The legal semantic parsing module is used to extract legal information from mediation texts;

[0009] An intelligent decision support module is used to generate a mediation plan and conduct a risk assessment based on the legal information. The intelligent decision support module includes:

[0010] A mediation proposal generation engine is used to generate a set of candidate proposals that include mediation suggestions.

[0011] The risk identification unit is used to assess the potential risks of mediation schemes at the legal and enforcement levels.

[0012] The dispute type adaptive unit is used to dynamically adjust the parameters for generating the mediation plan based on the characteristics of the case.

[0013] The multi-source data acquisition module includes:

[0014] The cloud call system is used to collect mediation voice data through PSTN or VoIP channels and upload the voice data to the server;

[0015] An adaptive noise reduction unit is used to preprocess the speech data and employs a noise suppression algorithm based on a deep neural network to preserve the original audio and the noise-reduced audio.

[0016] The dialect recognition unit is used to label the speech data with dialect types.

[0017] The legal semantic parsing module includes:

[0018] The legal entity identification unit is used to receive the mediation text transcribed by the speech recognition engine and identify the legal terms, party identity information and disputed elements in the text;

[0019] A dynamic legal knowledge graph construction unit is used to construct a knowledge graph structure based on the legal entity identification results. The knowledge graph structure includes nodes and edges. The nodes represent legal subjects, disputed objects, and legal clauses, and the edges represent the legal application relationship or dispute affiliation relationship between the nodes.

[0020] The mediation scheme generation engine further includes:

[0021] Call the interfaces of legal clause nodes and similar case nodes that match the disputed elements identified in the current case;

[0022] The rules engine is used to generate mediation suggestions based on preset rules;

[0023] The case reasoning model is used to identify reference cases and generate mediation paths based on the semantic features of the current case and the similarity between them and historical cases.

[0024] The risk identification unit includes a risk scoring module for evaluating candidate mediation schemes in terms of execution difficulty, legal compliance, and historical execution feedback. The evaluation is based on a risk scoring matrix constructed from a risk characteristic indicator system.

[0025] A risk ranking module is used to calculate the risk score matrix according to a set weighted scoring model to output the risk ranking results of candidate mediation schemes.

[0026] The dispute type adaptive unit includes:

[0027] The case feature extraction module is used to extract case features, including case type, relationship between parties, and intensity of claims.

[0028] The parameter configuration module is used to match the corresponding dispute type in the preset dispute case classification database based on the extracted case features, and to adjust the parameter weight configuration in the mediation scheme generation engine.

[0029] The parameter update module is used to adaptively fine-tune the mediation parameters based on the mediation execution results and mediation effect feedback.

[0030] Feedback indicators include agreement signing rate, performance rate, mediation period, and complaint rate. These indicators are normalized to form a mediation performance vector.

[0031]

[0032] in:

[0033] Let represent the mediation performance vector formed by the j-th mediation strategy in the t-th feedback cycle;

[0034] Let represent the score of mediation strategy j under the i-th specific evaluation indicator dimension in the t-th feedback cycle; subscript j: indicates the mediation strategy number, used to distinguish different mediation suggestions generated internally by the system; superscript (t): indicates that the current feedback cycle is t, corresponding to the result record after a complete mediation execution; subscript i: indicates the i-th dimension indicator for evaluating the mediation effect, with a total of m indicators.

[0035] The parameter update module calculates the update amount of the i-th parameter based on the following formula.

[0036]

[0037] Where, η i For learning rate, and Let be the historical mean and standard deviation of the i-th indicator within period t, respectively, and let ∈ be a constant. γ is used to score the confidence level of the indicator. i This is the confidence suppression coefficient.

[0038] The system constructs an incremental vector of adjustment parameters based on the update amounts of each dimension.

[0039]

[0040] The parameters are updated based on the following formula:

[0041]

[0042] in:

[0043] This represents the current parameter vector of the j-th mediation strategy at the beginning of the t-th feedback cycle;

[0044] This represents the new parameter vector obtained after the feedback adjustment in this round;

[0045] The system adjusts the weights generated by the mediation strategies for subsequent similar cases based on the updated parameter vector.

[0046] Beneficial effects:

[0047] This invention significantly improves the efficiency of information extraction and the accuracy of strategy matching in dispute mediation by integrating multimodal voice acquisition, legal semantic analysis, and intelligent mediation decision-making. The system can adapt to different dispute types, dynamically optimize mediation parameter configuration, and achieve intelligent generation and risk assessment of mediation plans, reducing mediation costs and increasing mediation success rates and party satisfaction. It has good practicality and promotional value. Attached Figure Description

[0048] The accompanying drawings, which are provided to further illustrate the invention and form part of this application, are not intended to unduly limit the invention. In the drawings:

[0049] Figure 1 A schematic diagram of a key information extraction system for dispute mediation based on multimodal adaptive analysis is shown. Detailed Implementation

[0050] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The illustrative embodiments and descriptions are only used to explain the present invention and are not intended to limit the present invention.

[0051] Example 1: Intelligent Extraction System for Key Information in Dispute Mediation

[0052] like Figure 1 As shown, this embodiment provides an intelligent system for extracting key information in dispute mediation based on multimodal adaptive analysis. The system includes a multi-source data acquisition module, a multimodal semantic parsing module, and an intelligent decision support module. These three modules work collaboratively to achieve comprehensive extraction, intelligent processing, and strategy generation of key legal information in dispute mediation.

[0053] The multi-source data acquisition module is used to collect voice data during the mediation process, ensuring the authenticity and completeness of the data sources for subsequent analysis. This module includes a cloud calling system and an adaptive noise reduction unit. The cloud calling system can record voice calls during the mediation process in real time via PSTN or VoIP lines, and audio data is collected synchronously through local caching and cloud transmission. The adaptive noise reduction unit processes the acquired voice signals, employing an environmental noise suppression algorithm based on deep neural networks to improve voice clarity and ensure the accuracy of subsequent speech recognition and semantic extraction.

[0054] The legal semantic parsing module transcribes the denoised speech signal into text and performs legal semantic analysis on the text. This module includes a legal entity recognition unit and a dynamic legal knowledge graph construction unit. The legal entity recognition unit uses a BiLSTM-CRF model to label and classify legal terms, party identities, and disputed elements in the text. The dynamic legal knowledge graph construction unit models the legal clauses, factual relationships, and evidence chains involved in the dispute based on the identified entities and relationships, forming a queryable and scalable legal knowledge structure to support the automated reasoning and generation of subsequent mediation plans.

[0055] The intelligent decision support module generates multiple alternative mediation solutions based on semantic parsing results and assesses the applicability and risks of each solution. This module includes a mediation solution generation engine, a risk identification unit, and a dispute type adaptive unit. Supported by a dynamic knowledge graph, the mediation solution generation engine outputs a set of mediation suggestions by combining case reasoning and rule engines. The risk identification unit quantitatively analyzes the potential risks of each solution based on dimensions such as execution difficulty, legal compliance, and historical data feedback. The dispute type adaptive unit dynamically adjusts the generation parameters of the mediation solution based on the specific case's characteristics. The case feature extraction module extracts multi-dimensional feature vectors, including case type, party relationship structure, and the intensity of claims. The parameter configuration module matches the corresponding dispute type in a pre-defined dispute case classification library based on the features and adjusts the parameter weights in the generation engine. The parameter update module continuously fine-tunes the parameters based on historical mediation results and execution feedback to improve the system's adaptability to complex and diverse dispute scenarios.

[0056] Through the integration and collaboration of the above modules, this invention realizes a closed-loop processing mechanism for voice data collection, legal information extraction, dispute structure modeling, and mediation strategy generation throughout the entire dispute mediation process, which helps to improve mediation efficiency and the professionalism and enforceability of the results.

[0057] Example 2: Multi-source data acquisition module

[0058] The multi-source data acquisition module includes a cloud call system and an adaptive noise reduction unit. The cloud call system is deployed within the mediation platform and supports dual-channel access via the Public Switched Telephone Network (PSTN) and Voice over IP (VoIP). The system connects a telephone box and terminal devices to achieve real-time recording of mediation calls. The telephone box uses hardware interception to acquire audio signals and obtains the recording stream through local devices using C language to access computer permissions, then uploads it to the system server via HTTP. Local caching and cloud uploading operate synchronously to ensure data continuity and integrity even under network fluctuations.

[0059] After the recording data acquisition is completed, the adaptive noise reduction unit initiates the audio preprocessing process. This invention preferably employs an environmental noise suppression algorithm based on a deep neural network (DNN), such as the Wave-U-Net architecture, to address potential background interference (such as environmental noise and overlapping human voices) at the mediation site. This algorithm can stably maintain a signal-to-noise ratio (SNR) of no less than 35dB at a 48kHz sampling rate, ensuring the clarity and stability of the speech signal. During noise suppression, the system simultaneously retains both the original audio and the noise-reduced data to support subsequent semantic processing modules in comparing and analyzing specific modal data.

[0060] To accommodate the diverse dialects encountered in mediation scenarios across different regions, this module includes a pre-defined dialect feature annotation interface. During data collection, the system can utilize a dialect classification model to identify the language of audio segments and perform feature modeling and annotation for specific regions (such as Wenzhou dialect and Cantonese), assisting the subsequent ASR engine in improving recognition accuracy under dialect conditions. Validation results for this part of the model show a dialect recognition accuracy of 98.7%, significantly reducing semantic recognition bias caused by language differences.

[0061] Through the aforementioned technical means, the multi-source data acquisition module not only achieves high-fidelity acquisition and processing of dispute mediation voice data, but also provides subsequent modules with a complete, high-quality, and semantically modelable source of original information, thereby improving the overall mediation assistance capability and information extraction accuracy of the system. This module also supports redundant data storage, timestamp synchronization, and concurrent multi-channel recording, demonstrating good engineering feasibility and deployment adaptability.

[0062] Example 3: Legal Semantic Analysis Module

[0063] The role of the legal semantic parsing module is to transcribe the noise-reduced mediation voice data into text, and on this basis, to complete the structured analysis and knowledge modeling of legal semantics, providing semantic support for the subsequent generation of mediation plans and risk assessment.

[0064] In this embodiment, the legal semantic parsing module consists of a legal entity recognition unit and a dynamic legal knowledge graph construction unit. These two units work together to transform speech text into a legal semantic structure. First, the speech-to-text process is completed by the system's Automatic Speech Recognition (ASR) engine, and its output is input into the model via the entity recognition unit for further processing. This legal entity recognition unit uses a BiLSTM-CRF (Bidirectional Long Short-Term Memory-Conditional Random Field) model to perform sequence labeling on the input text, extracting key legal terms, party identification information, disputed elements, and other content from the text.

[0065] The BiLSTM layer is used to capture the semantics of the text's context, enhancing the ability to identify keywords in the context through bidirectional semantic awareness. The CRF layer performs a global optimal annotation path search on the BiLSTM output, thereby improving the overall accuracy of named entity recognition. This model, trained on the dispute text corpus described in this embodiment, achieves an F1 score of 0.92, demonstrating high accuracy in extracting legal information in most civil mediation cases.

[0066] After entity recognition is completed, the system sends the extracted entity items to the dynamic legal knowledge graph construction unit for modeling. This unit, based on preset legal relationship templates and graph node types, transforms the semantic connections between entities into a graph structure representation. Nodes in the graph include legal subjects (such as plaintiffs, defendants, and guarantors), disputed objects (such as loans and real estate), and legal clauses (such as Article 680 of the Civil Code). Edges represent the relationship types between nodes, such as "applicable clause," "subject to dispute," and "dependent on evidence." The system employs an incremental update strategy, allowing for real-time supplementation of the graph structure based on new input cases, maintaining its timeliness and scalability.

[0067] In its implementation, the system can load different sub-graph templates based on case type to adapt to the legal logic structures of different fields such as marriage and family, private lending, and property services. For example, in private lending cases, the graph construction prioritizes identifying core elements such as "loan amount," "repayment period," "interest rate," and "guarantee liability" to ensure that the constructed graph has strong problem relevance and factual coverage. After the graph is generated, the system can perform further reasoning driven by rules or neural networks to assist in identifying points of contention, constructing chains of evidence, or determining applicable legal provisions.

[0068] In summary, this legal semantic parsing module, by combining a high-precision legal entity recognition model with a flexible and scalable knowledge graph modeling mechanism, achieves structured extraction and semantic reconstruction of legal elements in mediation voice text. It can effectively support downstream modules in understanding legal disputes and generating mediation strategies, significantly improving the legal semantic processing capabilities and intelligence level of the mediation system.

[0069] Example 4 Intelligent Decision Support Module

[0070] In this embodiment, the intelligent decision support module generates multiple candidate mediation schemes based on the output of the legal semantic parsing module. It then quantitatively or semi-quantitatively evaluates each scheme in terms of legal applicability, enforceability, and deviation from historical similar cases, thereby providing intelligent scheme recommendations for mediators or auxiliary systems. This module mainly consists of a mediation scheme generation engine, a risk identification unit, and a dispute type adaptive unit. These components form an organic linkage mechanism to jointly realize the generation and dynamic optimization of mediation strategies.

[0071] Supported by a dynamic legal knowledge graph, the mediation scheme generation engine first invokes legal clause nodes and typical case nodes that match the disputed elements identified in the current case to achieve semantic connections in the application of law. Within a baseline legal framework, this engine, combined with a rule engine and a case-based reasoning (CBR) model, generates multiple structured mediation suggestion sets. The rule engine includes, but is not limited to, rules for determining mediation priority, rules for recommending amount and interest rate ranges, and rules for combining installment payment conditions. The case reasoning part calculates the similarity between the semantic features of the current case and existing cases (e.g., vector space similarity or structured distance based on graph structure), identifies reference cases with a deviation of less than a preset threshold (e.g., 15%), and generates transferable mediation paths.

[0072] The risk identification unit performs multi-dimensional risk analysis on the generated candidate mediation solutions, evaluating aspects such as the difficulty of implementation, legal compliance, and success rate indicators from past implementation feedback. This unit establishes a risk scoring matrix based on a risk characteristic indicator system. These dimensions may include, but are not limited to: whether complex property division is involved, whether there is reliance on third-party performance, the agreement performance rate of similar historical solutions, and whether it violates provisions of laws and regulations such as the Civil Procedure Law and the Mediation Law. In its implementation, each solution is assigned a set of quantitative risk scores, and a final risk ranking is output based on a predefined weighted scoring model.

[0073] The dispute type adaptive unit achieves context adaptation and continuous optimization in the mediation strategy generation process through parameter configuration and parameter update mechanisms, thereby enhancing the system's adaptability and intelligence level to complex and ever-changing dispute scenarios.

[0074] The dispute type adaptive unit includes a case feature extraction module, a parameter configuration module, and a parameter update module. The case feature extraction module extracts multi-dimensional case features, including case type, party relationship structure, and claim intensity level, by calling the output of the semantic analysis model. Case type can be represented as discrete category variables, such as "private lending," "marriage and family," and "property services"; the party relationship structure is constructed as a structural vector based on the equality of rights and obligations and the degree of economic dependence; the claim intensity is obtained by normalizing factors such as semantic sentiment recognition values ​​and the frequency of emotional expression words to obtain an sentiment index. These variables are combined to form the case feature vector F:

[0075] F = {f1, f2, ... f i ,f n}

[0076] Where: f n This represents the nth feature component in the case feature vector, which is the last dimension (or the nth dimension) in the entire case feature set.

[0077] In the case feature extraction module, the system extracts several key features that influence mediation strategies based on the semantic analysis of the mediation voice. These features are arranged in order to form a vector F. Each f... i Each is an independent feature variable, while f n This is the last feature in this vector, such as "whether the dispute involves parties from different regions" or "the number of times mediation has failed in the past".

[0078] n: represents the total number of case feature dimensions, which is predefined by the system, such as 10, 12, or 20 dimensions, depending on the complexity of the model design.

[0079] The parameter configuration module matches the aforementioned case feature vectors with existing types in the system's preset dispute case classification database, selects the closest dispute type label, and loads the mediation parameter template corresponding to that type. This template is used to initialize the parameter vector θ in the mediation engine. j :

[0080] θ j =[w1,w2,...w i ,w m ]

[0081] Where: w m This represents the m-th parameter in the mediation strategy parameter vector, and is the last weight term in the mediation parameter set. Each w... i These all control the priority of a certain optimization objective in the mediation engine, such as the relative focus on "shortest mediation time," "strongest agreement enforcement," or "highest party satisfaction." mThis is the last item, and it may represent subsequent optimization dimensions such as "mediation cost control weight" or "negotiation round limit".

[0082] m: represents the total number of dimensions of the mediation parameter vector, which is the number of optimization objectives, and is usually set when modeling the mediation strategy.

[0083] To enable dynamic adjustment and self-adaptation of this parameter configuration, this invention further implements feedback-driven parameter optimization through a parameter update module. After each round of mediation, the system automatically records feedback indicators, including completion rate, fulfillment rate, mediation time cost, and complaint ratio. To support the feedback-driven mechanism of the parameter update module, after each round of mediation, the system collects and organizes data based on the actual execution situation to form a mediation performance vector, represented as follows:

[0084]

[0085] in:

[0086] Let represent the mediation performance vector formed by the j-th mediation strategy in the t-th feedback cycle, which is used to quantify the actual performance of the strategy across multiple evaluation dimensions;

[0087] Let be the score of mediation strategy j under the i-th specific evaluation index dimension in the t-th feedback cycle. This score can be normalized based on the original mediation data, and its value is usually normalized to an interval.

[0088] [0,1] reflects the relative merits of this indicator;

[0089] Subscript j: Indicates the mediation strategy number, used to distinguish different mediation proposals generated internally by the system;

[0090] Superscript (t): Indicates that the current feedback cycle is t, corresponding to the result record after a complete mediation execution;

[0091] Subscript i: indicates the i-th dimension indicator of the mediation effect evaluation, with a total of m indicators, such as:

[0092] It can represent the completion rate, that is, whether the proposed mediation plan is ultimately accepted and signed by both parties;

[0093] It can represent the fulfillment rate, which refers to whether both parties have fulfilled their obligations under the agreement after mediation;

[0094] It can represent the mediation cycle (Time Cost), which is an inverse indicator of the time taken from when the system generates a solution to when an agreement is reached (the shorter the cycle, the higher the score);

[0095] It can represent the complaint ratio, which is the degree of objection or dissatisfaction of the parties with the mediation result. It is usually a negative indicator and needs to be reversed during normalization.

[0096] Other dimensions (such as) (etc.) can be expanded to include system satisfaction, mediator workload, post-intervention frequency, etc.

[0097] The vector It forms the basis for achieving multidimensional evaluation, used to compare with historical averages. Standard deviation By combining statistical measures, the various parameters w of the mediation strategy are analyzed. i Differential adjustments are made. Through this structure, the system can track the actual effect of the mediation strategy dimension by dimension and drive the precise operation of the adaptive parameter fine-tuning mechanism.

[0098] To further achieve stepwise feedback learning and fine-tuning, the following adjustment parameter vector update formula is adopted:

[0099]

[0100] in:

[0101] This represents the update amount of the i-th adjustment parameter;

[0102] This represents the score for the i-th indicator in the current period.

[0103] These are the historical average and standard deviation of the indicator for this dimension, respectively.

[0104] The system is given a confidence score for this indicator;

[0105] η i The adjustment range is controlled for the learning rate of this dimension;

[0106] γ i The confidence level suppression coefficient;

[0107] ∈ is a small constant to prevent division by zero.

[0108] The system calculates the update values ​​for all parameter dimensions sequentially according to the above formula, thus obtaining the adjustment parameter increment vector:

[0109]

[0110] And update the adjustment parameters accordingly:

[0111]

[0112] in:

[0113] This represents the current parameter vector of the j-th mediation strategy at the beginning of the t-th feedback cycle, with its initial values ​​coming from the mediation parameter template matching the dispute type.

[0114] This indicates that the new parameter vector obtained after this round of feedback adjustments will be used to generate mediation plans in similar cases in the future;

[0115] For example, if the system detects that the "one-time performance + third-party guarantee" combination solution performs best in terms of performance rate and satisfaction in cases with "high demand intensity + unstable family structure", then the relevant parameters in the mediation engine will automatically receive higher weight in subsequent similar cases, thereby realizing the evolution and optimization of strategy priority ranking.

[0116] In summary, this embodiment establishes a complete optimization closed loop for dispute mediation strategy generation through case feature-driven initial parameter configuration and a dynamic parameter update mechanism with vectorized feedback. This demonstrates significant intelligent adaptive capabilities and practical application value. The system can continuously improve mediation efficiency and increase the success rate of execution in various dispute scenarios, exhibiting good engineering feasibility and innovation.

[0117] Through the above methods, the intelligent decision support module can not only automatically generate mediation suggestions based on fixed rules and knowledge graphs, but also adapt parameters through feature-driven adjustments in diverse case scenarios, thereby improving the pertinence, enforceability, and overall mediation efficiency of the mediation plan and meeting the actual needs for intelligent and personalized mediation strategy generation in real-world mediation scenarios.

[0118] The above description is only a preferred embodiment of the present invention. Therefore, all equivalent changes or modifications made to the structure, features and principles described in the claims of this patent application are included in the scope of this patent application.

Claims

1. An information intelligent extraction system based on multi-modal adaptive analysis, characterized in that: include: A multi-source data acquisition module is used to collect voice data for dispute mediation. The legal semantic parsing module is used to extract legal information from mediation texts; An intelligent decision support module is used to generate a mediation plan and conduct a risk assessment based on the legal information. The intelligent decision support module includes: A mediation proposal generation engine is used to generate a set of candidate proposals that include mediation suggestions. The risk identification unit is used to assess the potential risks of mediation schemes at the legal and enforcement levels. The dispute type adaptive unit is used to dynamically adjust the parameters for generating the mediation plan based on the characteristics of the case. The dispute type adaptive unit includes: The case feature extraction module is used to extract case features, including case type, relationship between parties, and intensity of claims. The parameter configuration module is used to match the corresponding dispute type in the preset dispute case classification database based on the extracted case features, and to adjust the parameter weight configuration in the mediation scheme generation engine. The parameter update module is used to adaptively fine-tune the mediation parameters based on the mediation execution results and mediation effect feedback. After each round of mediation, the system automatically records feedback indicators, including agreement signing rate, performance rate, mediation period, and complaint rate. These indicators are normalized to form a mediation performance vector. in: : represents the mediation policy at the th mediation policy iteration; and the mediation performance vector formed in the th feedback cycle. Mediation strategy In the In the feedback cycle, the first Scores under specific evaluation indicator dimensions; subscript : Indicates the mediation strategy number, used to distinguish different mediation proposals generated internally by the system; superscript : Indicates the current position is the th Each feedback cycle corresponds to the result record after a complete mediation execution; subscript : Indicates the first step in the evaluation of mediation effectiveness Dimensional indicators, totaling The parameter update module calculates the first one based on the following formula; Update amount of each parameter : in, For learning rate, and The first Item in cycle The historical mean and standard deviation within the range, It is a constant. Score the confidence level of the indicator. The confidence level suppression coefficient; The system constructs an incremental vector of adjustment parameters based on the update amounts of each dimension. : The parameters are updated based on the following formula: in: : indicates the first The mediation strategy in the first The current parameter vector at the start of the next feedback cycle; : This represents the new parameter vector obtained after the update following this round of feedback adjustments.

2. The information intelligent extraction system based on multimodal adaptive analysis as described in claim 1, characterized in that: The multi-source data acquisition module includes: The cloud call system is used to collect mediation voice data through PSTN or VoIP channels and upload the voice data to the server; An adaptive noise reduction unit is used to preprocess the speech data and employs a noise suppression algorithm based on a deep neural network to preserve the original audio and the noise-reduced audio. The dialect recognition unit is used to label the speech data with dialect types.

3. The information intelligent extraction system based on multimodal adaptive analysis as described in claim 1, characterized in that: The legal semantic parsing module includes: The legal entity identification unit is used to receive the mediation text transcribed by the speech recognition engine and identify the legal terms, party identity information and disputed elements in the text; A dynamic legal knowledge graph construction unit is used to construct a knowledge graph structure based on the legal entity identification results. The knowledge graph structure includes nodes and edges. The nodes represent legal subjects, disputed objects, and legal clauses, and the edges represent the legal application relationship or dispute affiliation relationship between the nodes.

4. The information intelligent extraction system based on multimodal adaptive analysis as described in claim 1, characterized in that: The mediation scheme generation engine further includes: Call the interfaces of legal clause nodes and similar case nodes that match the disputed elements identified in the current case; The rules engine is used to generate mediation suggestions based on preset rules; The case reasoning model is used to identify reference cases and generate mediation paths based on the semantic features of the current case and the similarity between them and historical cases.

5. The information intelligent extraction system based on multimodal adaptive analysis as described in claim 1, characterized in that: The risk identification unit includes a risk scoring module for evaluating candidate mediation schemes in terms of execution difficulty, legal compliance, and historical execution feedback. The evaluation is based on a risk scoring matrix constructed from a risk characteristic indicator system. A risk ranking module is used to calculate the risk score matrix according to a set weighted scoring model to output the risk ranking results of candidate mediation schemes.

6. The information intelligent extraction system based on multimodal adaptive analysis as described in claim 1, characterized in that: The system adjusts the weights generated by the mediation strategies for subsequent similar cases based on the updated parameter vector.

Images