Intelligent mediation case analysis method and system based on AI large model

Abstract

The invention belongs to the technical field of artificial intelligence, and particularly relates to an intelligent conciliation case analysis method and system based on an AI large model, and the method comprises the steps: inputting a text material, a voice transcription text and a picture evidence of a conciliation case into a preset multi-mode fusion large model; generating a high-dimensional vector representation; condensing the case core law semantic feature vector; calculating a flow failure risk index; judging whether the flow failure risk index exceeds a preset risk threshold value or not; a brand new workflow topological structure and node tasks are generated; solving a flow adaptability index; judging whether the flow adaptability index is lower than a preset adaptability threshold value or not; when the flow adaptability index is lower than a preset adaptability threshold value, a brand new workflow topological structure and a node task are generated again; and when the flow adaptability index is not lower than a preset adaptability threshold value, ending the flow. According to the method, the accuracy and comprehensiveness of the case core semantic features are improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically a method and system for intelligent analysis of mediation cases based on a large AI model. Background Technology

[0002] Case processing often involves multi-source, heterogeneous data, including written materials submitted by the parties, transcripts of recorded communications, and various types of photographic evidence. This data collectively constitutes a comprehensive description of the facts of the case, but its multimodal and unstructured nature presents significant challenges to traditional processing methods. Traditional mediation case management systems and methods primarily rely on fixed, experience-driven process templates and mainly process structured text information; however, these methods have the following limitations:

[0003] Limitations of data processing: Traditional systems struggle to effectively integrate and process heterogeneous data sources such as text, voice, and images; they lack the ability to analyze unstructured evidence such as voice and images, leading to a one-sided understanding of the facts of the case, an inability to construct a complete chain of evidence, and an easy oversight of key case details hidden in non-textual information;

[0004] Limitations of semantic understanding: Existing analysis methods mostly remain at the level of keyword matching or shallow text classification, and cannot conduct in-depth legal semantic understanding of case materials; these methods are difficult to accurately capture the complex relationships between multiple pieces of evidence, and even more difficult to identify potential key evidence that could overturn the course of a case, leading to deviations in the judgment of the core nature of the case;

[0005] Limitations of process management: Current case management systems generally use preset, static workflow templates; such workflows lack the ability to perceive dynamic changes in the case. When important new evidence emerges, they cannot assess the applicability and compliance risks of existing processes, nor do they have the ability to dynamically adjust and self-optimize. They can only rely on manual judgment for intervention, which is slow to respond and prone to errors.

[0006] In recent years, significant breakthroughs have been made in artificial intelligence large-scale model technology characterized by multimodal fusion. By mapping data from different sources to a unified high-dimensional semantic space, this technology can achieve deep fusion and comprehensive understanding of case information, providing new possibilities for solving the above problems. Multimodal fusion large-scale models can extract the core legal semantic features that represent the essence of a case from a global perspective, providing a solid foundation for dynamic evaluation and decision-making.

[0007] In summary, existing technologies, particularly traditional case management systems, have significant shortcomings in processing multimodal data, deep semantic understanding, and dynamic process adaptability. No research has yet combined the deep semantic understanding capabilities of a multimodal fusion model with workflow risk perception, dynamic reconstruction, and closed-loop correction mechanisms to construct an intelligent, adaptive mediation case handling system capable of proactively adapting to changes in case circumstances. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention discloses a method and system for intelligent analysis of mediation cases based on an AI large-scale model. Specifically, the technical solution of this invention is as follows:

[0009] The intelligent analysis method for mediation cases based on AI large-scale models includes the following steps:

[0010] S1. Input the text materials, speech-to-text transcripts, and image evidence of the mediation case into the preset multimodal fusion model;

[0011] S2. Through a multimodal fusion model, features are extracted from textual materials, speech-to-text transcriptions, and image evidence to generate high-dimensional vector representations.

[0012] S3. Deeply integrate and analyze the high-dimensional vector representation to extract the core legal semantic feature vector of the case;

[0013] S4. Real-time monitoring of the core legal semantic feature vector of the case, combined with the latest legal semantic feature vector of the case containing new evidence, which is parsed by the multimodal fusion model, to calculate the process failure risk index;

[0014] S5. Determine whether the process failure risk index exceeds the preset risk threshold;

[0015] S6. When the process failure risk index exceeds the preset risk threshold, a brand-new workflow topology and node tasks are generated based on the latest case legal semantic feature vector and the preset set of legal compliance rules.

[0016] S7. Continuously monitor the execution status of the new workflow topology and node tasks, and calculate the process adaptability index;

[0017] S8. Determine whether the process adaptability index is lower than the preset adaptability threshold.

[0018] S9. When the process adaptability index is lower than the preset adaptability threshold, a completely new workflow topology and node tasks are generated again to form a closed-loop correction.

[0019] S10. When the process adaptability index is not lower than the preset adaptability threshold, the process ends.

[0020] Preferably, S2 specifically includes:

[0021] The multimodal fusion big model extracts and embeds features from each modality of data in the case’s textual materials, speech-to-text, and image evidence, generating a high-dimensional vector representation.

[0022] Preferably, S4 specifically includes:

[0023] The cosine similarity between the latest legal semantic feature vector of the case containing new evidence and the core legal semantic feature vector of the case is calculated, and combined with the quantitative value of the impact of new evidence on the process, the process failure risk index is calculated.

[0024] Preferably, S6 specifically includes:

[0025] The process reengineering model, built on a multimodal fusion big model, automatically generates or recommends new workflow topologies and node tasks based on the latest case legal semantic feature vectors and a pre-set set of legal compliance rules.

[0026] Preferably, S7 specifically includes:

[0027] The quantitative value of process execution effectiveness is assessed by monitoring intermediate results such as task completion rate and legal risk reports to evaluate the degree of matching with the semantics of the latest cases. The process adaptability index is calculated by combining the quantitative value of process execution effectiveness with the risk index that may be generated during the reconstruction of the new workflow topology and node tasks.

[0028] Preferably, S9 specifically includes:

[0029] The perception-reconstruction cycle is triggered again to fine-tune or reconstruct the process until the process adaptability index reaches the preset adaptability threshold.

[0030] The AI-based intelligent analysis system for mediation cases includes:

[0031] The data input module is used to input the case's text materials, speech-to-text transcription, and image evidence into the multimodal fusion model;

[0032] The feature extraction module is used to extract features from text materials, speech-to-text, and image evidence through a multimodal fusion model, and generate high-dimensional vector representations.

[0033] The semantic condensation module is used to deeply fuse and analyze high-dimensional vector representations to condense the core legal semantic feature vectors of a case.

[0034] The risk assessment module is used to monitor the core legal semantic feature vector of a case, and combine it with the latest legal semantic feature vector of the case containing new evidence, which is parsed from the multimodal fusion model, to calculate the process failure risk index.

[0035] The process refactoring module is used to generate a new workflow topology and node tasks based on the latest case legal semantic feature vector and a set of preset legal compliance rules when the process failure risk index exceeds the preset risk threshold.

[0036] The performance evaluation module is used to continuously monitor the execution status of the new workflow topology and node tasks, and calculate the process adaptability index.

[0037] The closed-loop correction module is used to generate a completely new workflow topology and node tasks when the process adaptability index is lower than the preset adaptability threshold, so as to form a closed-loop correction.

[0038] Preferred intelligent analysis systems for mediation cases based on large AI models include:

[0039] Processor, memory, communication interface; memory stores program code, and the processor executes the program code.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] 1. This invention achieves comprehensive collection and deep semantic understanding of multi-source heterogeneous data in cases by inputting written materials, speech-to-text transcription, and image evidence into a multimodal fusion model. The method first performs independent feature extraction and embedding for each modality of data, and then performs deep fusion analysis to condense the core legal semantic feature vector of the case. This strategy of first separating and extracting, and then deeply fusing, preserves the unique semantic dimensions of various types of data, and can capture key non-textual information in oral statements, tone, and image evidence in addition to written materials, thereby improving the accuracy and comprehensiveness of the core semantic features of the case and enhancing the system's ability to identify subversive evidence hidden in non-textual data.

[0042] 2. This invention establishes a dynamic risk warning and process refactoring mechanism. The system monitors the core legal semantic feature vector of a case in real time. When new evidence appears, it calculates the semantic similarity between the latest case legal semantic feature vector containing the new evidence and the original core vector, and combines this with the quantitative value of the impact of the new evidence on the process to jointly calculate a process failure risk index. This risk measurement method upgrades the fuzzy difference judgment to a comprehensive assessment of the importance and magnitude of the change, which can accurately identify the key case changes that truly need to trigger process refactoring. When the risk exceeds the threshold, the process refactoring model based on the large model will automatically generate a new workflow topology and node tasks according to the latest case profile and preset legal compliance rules. This transforms the previous process design work that relied on human experience into automated generation driven by AI models, significantly improving the system's speed of response to sudden changes in case conditions and the quality of decision-making, and upgrading process management from rigid task execution to dynamic self-refactoring.

[0043] 3. This invention achieves continuous self-optimization of the process through closed-loop correction. After the execution of a new workflow, the system continuously monitors its task completion status and legal risks. Combining the risk index that may arise from the process reconstruction itself, a process adaptability index is calculated to quantitatively evaluate the actual effectiveness of the new process. This index comprehensively weighs the execution benefits and potential costs of the new process. When the adaptability is lower than a preset threshold, the system will trigger the perception and reconstruction cycle again, fine-tune or reconstruct the workflow, and learn from the ineffective reconstruction until the process adaptability meets the requirements. This iterative optimization closed-loop mechanism ensures that the system will not remain at a suboptimal solution. It ensures that no matter how complex the case is, the system can always evolve itself and eventually converge to an efficient, compliant, and practice-tested optimal workflow, significantly improving the mediation success rate and processing quality of difficult cases. Attached Figure Description

[0044] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0045] Figure 1 This is a flowchart of the method of the present invention.

[0046] Figure 2 This is a flowchart of the system of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0048] Example 1:

[0049] Please see Figure 1The intelligent analysis method for mediation cases based on AI big data models includes the following steps:

[0050] S1. Input the text materials, speech-to-text transcripts, and image evidence of the mediation case into the preset multimodal fusion model;

[0051] S2. Through a multimodal fusion model, features are extracted from textual materials, speech-to-text transcriptions, and image evidence to generate high-dimensional vector representations.

[0052] S3. Deeply integrate and analyze the high-dimensional vector representation to extract the core legal semantic feature vector of the case;

[0053] S4. Real-time monitoring of the core legal semantic feature vector of the case, combined with the latest legal semantic feature vector of the case containing new evidence, which is parsed by the multimodal fusion model, to calculate the process failure risk index;

[0054] S5. Determine whether the process failure risk index exceeds the preset risk threshold;

[0055] S6. When the process failure risk index exceeds the preset risk threshold, a brand-new workflow topology and node tasks are generated based on the latest case legal semantic feature vector and the preset set of legal compliance rules.

[0056] S7. Continuously monitor the execution status of the new workflow topology and node tasks, and calculate the process adaptability index;

[0057] S8. Determine whether the process adaptability index is lower than the preset adaptability threshold.

[0058] S9. When the process adaptability index is lower than the preset adaptability threshold, a completely new workflow topology and node tasks are generated again to form a closed-loop correction.

[0059] S10. When the process adaptability index is not lower than the preset adaptability threshold, the process ends.

[0060] This invention provides an intelligent analysis method for mediation cases based on a large AI model. This method constitutes a complete and self-consistent technical closed loop, and its specific implementation is as follows:

[0061] S1. Input the text materials, speech-to-text transcripts, and image evidence of the mediation case into the preset multimodal fusion model;

[0062] The purpose of this step is to achieve comprehensive collection of multi-source heterogeneous data in the case. In a specific mediation case scenario, written materials refer to structured and unstructured texts such as complaints, answers, and lists of evidence submitted by the parties, which serve to provide the basic facts and legal claims of the case. Speech-to-text refers to text generated by automatic speech recognition (ASR) of court recordings and telephone communication recordings during the mediation process, which serves to capture oral statements, tone, and key information beyond written materials. Image evidence refers to image data such as photos, scans, and screenshots, such as scanned copies of contracts and photographs of physical evidence, which serve to provide intuitive, non-textual evidence support. These three types of data together constitute a holographic description of the case. Inputting this data into a pre-set multimodal fusion model is the foundation for subsequent deep semantic analysis. The pre-set multimodal fusion model refers to a deep learning model pre-trained with massive amounts of legal data, capable of understanding and processing multiple data modalities, and its core function is to serve as a unified semantic understanding engine for all subsequent intelligent analyses.

[0063] S2. Through a multimodal fusion model, features are extracted from textual materials, speech-to-text transcriptions, and image evidence to generate high-dimensional vector representations.

[0064] The purpose of this step is to transform unstructured multimodal data into a machine-understandable mathematical representation. The large multimodal fusion model integrates encoders for different data types; for example, it uses models like BERT as text encoders to process written materials and speech-to-text transcription, and models like CNN or ViT as image encoders to process image evidence. These encoders map the raw data into a high-dimensional vector space, generating a high-dimensional vector representation. Specifically, written materials are converted into vectors. The speech-to-text is converted into a vector. The visual evidence was converted into vectors. These vectors are not only a compact mathematical representation of the original data, but more importantly, their positions and orientations in the vector space contain deep semantic information about the data.

[0065] S3. Deeply integrate and analyze the high-dimensional vector representation to extract the core legal semantic feature vector of the case;

[0066] The purpose of this step is to extract the core semantics that have a decisive impact on the legal characterization of the case from the scattered multimodal information; the multimodal fusion big model uses attention mechanisms or other fusion strategies to represent high-dimensional vectors from different modalities ( The model performs weighted fusion and in-depth analysis; this process is not a simple vector concatenation, but rather simulates the process by which human legal experts comprehensively analyze evidence from different sources, identifying and amplifying key information while suppressing redundant or contradictory information; ultimately, the model condenses and outputs a single core legal semantic feature vector for the case. ; Case core legal semantic feature vector ( This refers to a high-dimensional vector that can comprehensively and accurately represent the legal attributes of the current case. Its function is to serve as the benchmark and starting point for the entire subsequent process analysis and decision-making; its calculation process can be formally represented as: ,in, This represents the fusion and analysis function of a large-scale multimodal fusion model, and its input is the high-dimensional vector representation of each modality generated by S2; this vector... It can capture potentially subversive evidence or related case details that are difficult to discover using traditional methods;

[0067] S4. Real-time monitoring of the core legal semantic feature vector of the case, combined with the latest legal semantic feature vector of the case containing new evidence, which is parsed by the multimodal fusion model, to calculate the process failure risk index;

[0068] The purpose of this step is to quantify the impact of new evidence on the stability of the existing mediation process. During the mediation process, any newly submitted evidence will be input into the system and analyzed in real time by a multimodal fusion model to generate a new legal semantic feature vector of the case that includes the new evidence. The system compares the new vector. Compared with the original case core legal semantic feature vector The difference between them is used to calculate the process failure risk index. Process failure risk index ( This refers to an indicator used to measure the degree of risk that the current mediation process may become inapplicable or illegal due to significant changes in the case. Its role is to serve as a basis for decision-making to trigger process restructuring.

[0069] S5. Determine whether the process failure risk index exceeds the preset risk threshold;

[0070] This step is a decision-triggered stage; the system will use the process failure risk index calculated in the previous step. With a preset risk threshold Comparison; preset risk thresholds ( This refers to a threshold value set based on historical case data and legal expert experience; to clarify its setting logic, this threshold can be represented by a risk index arising from the supplementation of routine evidence in a large number of historical cases. The value is determined by statistical analysis and the 99th percentile of its probability distribution. Its purpose is to strike a balance between ensuring that the system can respond promptly to major changes in the case and avoiding frequent process reconstruction due to minor fluctuations.

[0071] S6. When the process failure risk index exceeds the preset risk threshold, a brand-new workflow topology and node tasks are generated based on the latest case legal semantic feature vector and the preset set of legal compliance rules.

[0072] when When the system determines that the current process has a legality or validity crisis and must be restructured, the system will invoke a process restructuring model, which uses the latest case's legal semantic feature vector. As input, it represents the most comprehensive state of the case; at the same time, the model must also comply with a pre-defined set of legal compliance rules ( The pre-defined set of legal compliance rules refers to a rule base that encodes relevant laws, regulations, judicial interpretations, and procedural requirements, such as the provisions on mediation procedures in the Civil Procedure Law. Its function is to ensure that newly generated workflows are legally compliant. The process refactoring model is based on... The understanding, and limited by Given the constraints, output a completely new workflow topology and node tasks. ;

[0073] For example, if new evidence indicates that the case involves corporate bankruptcy, the new workflow may automatically add necessary steps such as notifying the bankruptcy administrator and reviewing creditor claims. This new workflow topology and node tasks... It is then loaded into the system's workflow execution engine, which is responsible for parsing the workflow and scheduling resources to execute the various node tasks it contains;

[0074] S7. Continuously monitor the execution status of the new workflow topology and node tasks, and calculate the process adaptability index;

[0075] The purpose of this step is to evaluate the actual performance of the newly generated workflow and establish a feedback mechanism; this involves the new workflow topology and node tasks. During execution, the system continuously monitors its key performance indicators, such as the completion rate of each node and whether there are any legal risk warnings, and calculates a process adaptability index based on this monitoring data. Process adaptability index It refers to a comprehensive indicator that quantitatively evaluates the degree of matching between the new workflow and the current complex case and the effectiveness of its execution. Its value range is usually between [0,1]. Its function is to determine whether the restructured process has achieved the expected results.

[0076] S8. Determine whether the process adaptability index is lower than the preset adaptability threshold.

[0077] Similar to risk assessment, this step is a closed-loop correction decision trigger; the system will use the process adaptability index calculated in the previous step. With a preset adaptive threshold Comparison; preset adaptive threshold This refers to a critical value that represents the minimum acceptable process efficiency. It is set based on the analysis of the process adaptability index distribution of historical successful mediation cases, and a reasonable value is set that can both ensure the basic effectiveness of the process and avoid endless fine-tuning in pursuit of perfection, thereby maintaining the stability of the system.

[0078] S9. When the process adaptability index is lower than the preset adaptability threshold, a completely new workflow topology and node tasks are generated again to form a closed-loop correction.

[0079] when When the refactored process fails to perform well in actual operation and fails to effectively solve the problem, the system will trigger the perception-refactoring loop again, that is, repeat the operation of step S6 and thereafter to further fine-tune or completely refactor the workflow. This process iterates until the adaptability index of the newly generated workflow meets the requirements, thus forming a closed-loop correction.

[0080] S10. When the process adaptability index is not lower than the preset adaptability threshold, the process ends.

[0081] when When the system deems the current workflow effective, adaptive, and stable, the correction cycle terminates, and the system continues to execute the current workflow until the mediation ends.

[0082] Through the above steps, this invention constructs a complete intelligent processing flow from multimodal data understanding, risk perception, dynamic reconstruction to closed-loop correction; it transforms the system from a rigid task executor into an adaptive process management core capable of proactively sensing changes in case conditions, providing early warnings of process failure risks, and dynamically adjusting its own structure to adapt to new legal realities, greatly improving the intelligence level, compliance, and agility in handling complex emergencies in mediation case processing.

[0083] Example 2:

[0084] S2 specifically includes:

[0085] The multimodal fusion big model extracts and embeds features from each modality of data in the case’s textual materials, speech-to-text, and image evidence, generating a high-dimensional vector representation.

[0086] This implementation method further defines the technical details of step S2 based on Example 1. The multimodal fusion model extracts and embeds features from each modality of data in the case's textual materials, speech-to-text, and image evidence, generating high-dimensional vector representations. Specifically, this process ensures that the unique information of each data source is fully preserved and deeply mined. For example, for textual materials, the model not only focuses on keywords but also understands the legal logic through contextual analysis. For speech-to-text, the model can capture the semantic information implied by interjections such as hesitation and emphasis. For image evidence, the model can identify non-textual information such as the authenticity of seals and alterations to key content. By independently generating optimal vector representations for each modality of data before subsequent fusion, semantic loss in the early stages of information conversion can be minimized, facilitating the subsequent refinement of high-quality core legal semantic feature vectors for the case. It laid a solid foundation;

[0087] This strategy of first separating and extracting, and then deeply fusing, results in more refined feature extraction. Compared with the method of simply textualizing all data and then processing it, this invention can retain the unique semantic dimension of each modality of data, thereby improving the accuracy and comprehensiveness of semantic feature vectors and enhancing the system's ability to identify subversive evidence hidden in non-textual data.

[0088] Example 3:

[0089] S4 specifically includes:

[0090] The cosine similarity between the latest legal semantic feature vector of the case containing new evidence and the core legal semantic feature vector of the case is calculated, and combined with the quantitative value of the impact of new evidence on the process, the process failure risk index is calculated.

[0091] This implementation method is based on Example 1, and adjusts the process failure risk index in step S4. The calculation method was specified; the cosine similarity between the latest case legal semantic feature vector containing new evidence and the core case legal semantic feature vector was calculated, and combined with the quantitative value of the impact of new evidence on the process, the process failure risk index was calculated; this calculation follows the following formula: ;

[0092] in: The process failure risk index is a quantitative representation of the degree of deviation between the changes in the case due to the introduction of new evidence and the original case basis. It is a dimensionless scalar value and its source is the calculation result of this formula.

[0093] The quantitative value of the impact of new evidence on the process, in technical terms, is the importance or disruptive level of the evidence itself, and is a normalized scalar value in the range of [0,1]. Its source is based on preset rules or AI model evaluation. To further clarify, the preset rules are a knowledge base formed by mapping specific types of evidence with high-impact quantitative values ​​based on the experience of legal experts.

[0094] The cosine similarity between two vectors, technically speaking, represents the degree of correlation between old and new cases in the semantic space. Its value ranges from [0,1], and it is derived by calculating the cosine similarity between the vectors. and The dot product is obtained by normalizing the values; the closer the value is to 1, the more similar the old and new cases are; the closer it is to 0, the greater the difference.

[0095] The latest legal semantic feature vector of the case containing new evidence, which is parsed by the AI ​​big model, comes from the first half of step S4 and is the real-time processing result of the model on the new evidence.

[0096] The core legal semantic feature vector of the case on which the current process is based originates from step S3;

[0097] A very small positive number, its function is to prevent the denominator from being zero and to ensure the stability of the calculation. Its source is a system-preset constant, such as... ;

[0098] Based on the above definition, the technical motivation of this formula is that the extent of the change depends more on the importance of the new evidence itself; a new piece of evidence of great importance may trigger huge procedural risks even if the overall case does not change much.

[0099] The application of this formula brings more accurate risk measurement capabilities; it upgrades risk assessment from a vague difference judgment to a two-dimensional assessment that comprehensively considers the importance of the direction of change and the magnitude of change, thereby more accurately identifying high-value case changes that truly need to trigger process restructuring, avoiding overreaction to non-critical changes, and improving the system's decision-making quality and operational efficiency.

[0100] Example 4:

[0101] S6 specifically includes:

[0102] The process reengineering model, built on a multimodal fusion big model, automatically generates or recommends new workflow topologies and node tasks based on the latest case legal semantic feature vectors and a pre-set set of legal compliance rules.

[0103] This implementation method, based on Example 1, specifies the method for generating a new workflow in step S6. Based on a multimodal fusion model, the process reconstruction model automatically generates or recommends a new workflow topology and node tasks according to the latest case legal semantic feature vectors and a preset set of legal compliance rules. Its workflow can be formally represented as follows: ;

[0104] in: The AI ​​model generates a new workflow topology and node tasks based on new circumstances. Its data type can be a directed acyclic graph or a structured list of tasks, and its source is the output of this model.

[0105] The process reconstruction model, built upon a large AI model, serves as the core generation engine for new processes. To illustrate its construction process, the model is trained on a labeled dataset containing a large number of historical cases. The variables in this dataset are not input variables for the model's runtime, but rather data pairs consisting of the 'final case feature vector' of each historical case and the 'optimal process template' corresponding to that case. By learning from these data pairs, the model grasps the mapping relationship from case details to process strategies.

[0106] The latest case legal semantic feature vector serves to provide the model with a complete and accurate profile of the current case and is the input for the model's operation.

[0107] The pre-defined set of legal compliance rules serves to constrain the output space of the model, ensuring that no generated process violates mandatory legal provisions.

[0108] This implementation method achieves intelligent and automated process refactoring; it transforms the process design work that previously relied on human experience and manual adjustments into one that is automatically generated by an AI model based on a deep understanding of the essence of the case; this not only greatly improves the response speed, but more importantly, it can generate solutions that surpass human experience and are legally superior, upgrading the process engine from an unintelligent executor to an intelligent process core capable of dynamic self-refactoring.

[0109] Example 5:

[0110] S7 specifically includes:

[0111] The quantitative value of process execution effectiveness is assessed by monitoring intermediate results such as task completion rate and legal risk reports to evaluate the degree of matching with the semantics of the latest cases. The process adaptability index is calculated by combining the quantitative value of process execution effectiveness with the risk index that may be generated during the reconstruction of the new workflow topology and node tasks.

[0112] This implementation method is based on Example 1, and adjusts the process adaptability index in step S7. The calculation method has been specified; the quantitative value of process execution effect is evaluated by monitoring intermediate results such as task completion rate and legal risk reports to assess the degree of matching with the semantics of the latest cases; the process adaptability index is calculated by combining the quantitative value of process execution effect and the risk index that may be generated during the reconstruction of the new workflow topology and node tasks; the calculation follows the following formula: ;

[0113] in: The process adaptability index, technically speaking, is a comprehensive trade-off between the benefits and costs of a new process. It is a scalar value normalized to the [0,1] interval and is derived from the calculation results of this formula.

[0114] The quantitative value of process execution effect is technically defined as the degree to which the new process matches the expected goals in actual operation. It is a scalar value normalized to the range of [0,1]. It is derived by continuously monitoring the execution status of each node in the process, the feedback from mediators, and whether any new legal risk warnings are triggered, and is calculated by a preset weighting function.

[0115] The risk index, which represents the potential risks arising from the restructuring of the new process, is technically an estimate of the potential problems caused by the complexity and uncertainty of the new process itself. It is a scalar value normalized to the [0,1] interval; its source is the new process. After generation and before formal execution, another AI evaluation model simulates and extrapolates the process to assess potential risks such as task allocation conflicts and omissions in legal clauses. This AI evaluation model itself is obtained through supervised learning of the process structure characteristics and actual failure mode data of historical reconstruction failure cases.

[0116] This formula ensures that the evaluation of the new process is comprehensive, taking into account both how many old problems it solves. We also need to consider whether it introduces new problems. This allows for closed-loop control of process quality.

[0117] This implementation method endows the system with the ability to provide early warning of system risks and to self-evolve; by conducting a simulation assessment of the risks of the new process before reconstruction ( The system can provide managers with decision-making insights, avoiding the hasty adoption of a new process with potential flaws; and through comprehensive evaluation of execution effectiveness ( The system can determine whether the reconstruction is successful and use this as a basis for the next round of optimization and iteration; this makes the system no longer a simple automation tool, but an intelligent platform that can handle unknown complexities and continuously improve itself when handling difficult cases.

[0118] Example 6:

[0119] S9 specifically includes:

[0120] The perception-reconstruction cycle is triggered again to fine-tune or reconstruct the process until the process adaptability index reaches the preset adaptability threshold.

[0121] This implementation further clarifies the closed-loop correction mechanism in step S9 of Example 1; it triggers the perception-reconstruction cycle again to fine-tune or reconstruct the process until the process adaptability index reaches the preset adaptability threshold; this means that when In such cases, the system will not simply revert to the original process, but will instead implement the current, ineffective new process. and its corresponding fitness index As new input information, the process refactoring model is restarted. The model learns why the refactoring failed and avoids similar problems in the next process generation. This cycle continues, sometimes involving a completely new refactoring, sometimes just minor tweaks to existing process nodes, until a final version of the workflow is generated. Adaptability index satisfy until;

[0122] This implementation method ensures the convergence and final effect of process optimization; by establishing a closed loop of continuous feedback and iterative optimization, it guarantees that the system will not remain at any suboptimal solution; this correction mechanism ensures that no matter how complex and changeable the case is, the system can always evolve itself and eventually converge to an efficient, compliant and practice-tested optimal workflow, significantly improving the mediation success rate and handling quality of difficult cases.

[0123] Example 7:

[0124] Please see Figure 2 The intelligent analysis system for mediation cases based on AI big data models includes:

[0125] The data input module is used to input the case's text materials, speech-to-text transcription, and image evidence into the multimodal fusion model;

[0126] The feature extraction module is used to extract features from text materials, speech-to-text, and image evidence through a multimodal fusion model, and generate high-dimensional vector representations.

[0127] The semantic condensation module is used to deeply fuse and analyze high-dimensional vector representations to condense the core legal semantic feature vectors of a case.

[0128] The risk assessment module is used to monitor the core legal semantic feature vector of a case, and combine it with the latest legal semantic feature vector of the case containing new evidence, which is parsed from the multimodal fusion model, to calculate the process failure risk index.

[0129] The process refactoring module is used to generate a new workflow topology and node tasks based on the latest case legal semantic feature vector and a set of preset legal compliance rules when the process failure risk index exceeds the preset risk threshold.

[0130] The performance evaluation module is used to continuously monitor the execution status of the new workflow topology and node tasks, and calculate the process adaptability index.

[0131] The closed-loop correction module is used to generate a completely new workflow topology and node tasks when the process adaptability index is lower than the preset adaptability threshold, so as to form a closed-loop correction.

[0132] This invention also provides an intelligent analysis system for mediation cases based on an AI large-scale model. This system serves as the hardware carrier and functional implementation of the aforementioned method; the system includes:

[0133] The data input module is used to collect all the information of the case. In this embodiment, it is responsible for receiving and preprocessing the text materials, speech-to-text, and image evidence uploaded by the user, and transmitting these multimodal data to the subsequent modules.

[0134] The feature extraction module is used to transform raw data into machine-understandable semantic vectors. In this embodiment, it has built-in encoders of various multimodal fusion large models to perform deep feature extraction on different types of data and generate high-dimensional vector representations.

[0135] The semantic condensation module is used to form a unified and core understanding of the case. In this embodiment, it utilizes the fusion mechanism of a multimodal fusion model to deeply fuse and analyze high-dimensional vector representations, condensing the core legal semantic feature vectors of the case. ;

[0136] The risk assessment module is used to detect procedural risks arising from changes in case details in real time. In this embodiment, it monitors the core legal semantic feature vector of the case. And when new evidence is input, it is combined with the newly analyzed vector. The process failure risk index was calculated. ;

[0137] The process refactoring module is used to dynamically generate mediation processes adapted to new case circumstances. In this embodiment, when the process failure risk index exceeds a preset risk threshold, the internal process refactoring model will be invoked. Based on the latest case legal semantic feature vector and the pre-set set of legal compliance rules Generate a completely new workflow topology and node tasks. ;

[0138] The performance evaluation module is used to verify the actual effectiveness of the new process. In this embodiment, it continuously monitors the execution of the new workflow and calculates the process adaptability index according to the formula. ;

[0139] The closed-loop correction module is used to realize the self-optimization and evolution of the system. In this embodiment, when the process adaptability index is lower than the preset adaptability threshold, the process reconstruction module will be instructed to generate a brand new workflow to form a closed-loop correction.

[0140] The model validation module ensures the robustness of the system. In this embodiment, this module is responsible for stress testing the core model using an independent test set containing historical challenging cases, edge cases, and simulated adversarial samples before its formal deployment. It verifies whether the model's behavior when handling extreme input values ​​conforms to logical expectations and continuously monitors the online model's output to prevent performance degradation caused by data drift or malicious input, thereby ensuring the safe and stable operation of the entire system under various complex and even adversarial scenarios.

[0141] Through the organic combination and collaborative work of the above modules, this system physically realizes an end-to-end intelligent case handling capability; it deeply couples functions such as data collection, semantic understanding, risk assessment, decision generation and effect feedback to form a highly efficient whole, providing a complete system-level solution for realizing intelligent and adaptive management of the entire process of mediation cases.

[0142] Example 8:

[0143] The AI-based intelligent analysis system for mediation cases includes:

[0144] The system includes a processor, a memory, and a communication interface. The memory stores program code, and the processor executes the program code. This implementation provides a specific example of the hardware implementation of the system based on Embodiment 7. The system can be a server, a computer, or a distributed computing cluster, and its core hardware includes at least: a processor, a memory, and a communication interface.

[0145] The processor is the computing core of the system, responsible for executing program instructions and running the algorithm logic of the above-mentioned major modules, such as performing inference calculations for the multimodal fusion model, solving risk indices and adaptability indices, etc. The memory is used to store program code, that is, software programs that solidify the above-mentioned method flow and module functions, and is also used to temporarily store case data, intermediate calculation results, and model parameters, etc. The communication interface is used to interact with external devices to realize data input and result output. The processor is used to execute program code, load and run the instructions in the memory, thereby driving the entire system to realize all the functions described in Examples 1 to 7.

[0146] This embodiment clarifies the feasibility of the technical solution of the present invention, provides a specific hardware architecture, and ensures that all the aforementioned methods and module functions have their physical execution carriers, enabling the technical solution of the present invention to be transformed from theory into practical application products, and possessing the foundation for commercialization and engineering.

[0147] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention; any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0148] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for intelligent analysis of mediation cases based on AI large-scale models, characterized in that, Includes the following steps: S1. Input the text materials, speech-to-text transcripts, and image evidence of the mediation case into the preset multimodal fusion model; S2. Through a multimodal fusion model, features are extracted from textual materials, speech-to-text transcriptions, and image evidence to generate high-dimensional vector representations. S3. Deeply integrate and analyze the high-dimensional vector representation to extract the core legal semantic feature vector of the case; S4. Real-time monitoring of the core legal semantic feature vector of the case, combined with the latest legal semantic feature vector of the case containing new evidence parsed by the multimodal fusion model, to calculate the process failure risk index; S5. Determine whether the process failure risk index exceeds the preset risk threshold; S6. When the process failure risk index exceeds the preset risk threshold, a brand-new workflow topology and node tasks are generated based on the latest case legal semantic feature vector and the preset set of legal compliance rules. S7. Continuously monitor the execution status of the new workflow topology and node tasks, and calculate the process adaptability index; S8. Determine whether the process adaptability index is lower than the preset adaptability threshold. S9. When the process adaptability index is lower than the preset adaptability threshold, a completely new workflow topology and node tasks are generated again to form a closed-loop correction. S10. When the process adaptability index is not lower than the preset adaptability threshold, the process ends.

2. The intelligent analysis method for mediation cases based on AI large model according to claim 1, characterized in that, S2 specifically includes: The multimodal fusion big model extracts and embeds features from each modality of data in the case’s textual materials, speech-to-text, and image evidence, generating a high-dimensional vector representation.

3. The intelligent analysis method for mediation cases based on AI large-scale models according to claim 1, characterized in that, S4 specifically includes: The cosine similarity between the latest legal semantic feature vector of the case containing new evidence and the core legal semantic feature vector of the case is calculated, and combined with the quantitative value of the impact of new evidence on the process, the process failure risk index is calculated.

4. The intelligent analysis method for mediation cases based on AI large model according to claim 1, characterized in that, S6 specifically includes: The process reengineering model, built on a multimodal fusion big model, automatically generates or recommends new workflow topologies and node tasks based on the latest case legal semantic feature vectors and a pre-set set of legal compliance rules.

5. The intelligent analysis method for mediation cases based on AI large model according to claim 1, characterized in that, S7 specifically includes: The quantitative value of process execution effectiveness is assessed by monitoring intermediate results such as task completion rate and legal risk reports to evaluate the degree of matching with the semantics of the latest cases. The process adaptability index is calculated by combining the quantitative value of process execution effectiveness with the risk index that may be generated during the reconstruction of the new workflow topology and node tasks.

6. The intelligent analysis method for mediation cases based on AI large model according to claim 1, characterized in that, S9 specifically includes: The perception-reconstruction cycle is triggered again to fine-tune or reconstruct the process until the process adaptability index reaches the preset adaptability threshold.

7. A mediation case intelligent analysis system based on an AI large-scale model, applied to the mediation case intelligent analysis method based on an AI large-scale model as described in any one of claims 1 to 6, characterized in that, include: The data input module is used to input the case's text materials, speech-to-text transcription, and image evidence into the multimodal fusion model; The feature extraction module is used to extract features from text materials, speech-to-text, and image evidence through a multimodal fusion model, and generate high-dimensional vector representations. The semantic condensation module is used to deeply fuse and analyze high-dimensional vector representations to condense the core legal semantic feature vectors of a case. The risk assessment module is used to monitor the core legal semantic feature vector of a case, and combine it with the latest legal semantic feature vector of the case containing new evidence, which is parsed from the multimodal fusion model, to calculate the process failure risk index. The process refactoring module is used to generate a new workflow topology and node tasks based on the latest case legal semantic feature vector and a set of preset legal compliance rules when the process failure risk index exceeds the preset risk threshold. The performance evaluation module is used to continuously monitor the execution status of the new workflow topology and node tasks, and calculate the process adaptability index. The closed-loop correction module is used to generate a completely new workflow topology and node tasks when the process adaptability index is lower than the preset adaptability threshold, so as to form a closed-loop correction.

8. The intelligent analysis system for mediation cases based on an AI large model according to claim 7, characterized in that, include: Processor, memory, communication interface; memory stores program code, and the processor executes the program code.

Images