Multimodal big language model-based contradictory dispute real-time mediation agent method

Abstract

The invention discloses a contradictory dispute real-time mediation agent method based on a multi-mode large language model. The method comprises the following steps: firstly, acquiring voice information of two parties of a contradictory dispute and converting the voice information into text information; calculating an emotional state score based on the voice information and the text information, and predicting intentions of both parties of a contradictory dispute; then extracting facts, appeals and dispute focuses of the two parties of contradictory disputes, and quantifying the conflict degree of the dispute focuses at the same time; thirdly, distributing weights for the demands of the two parties with contradictory disputes to generate an optimal compromise point, generating a mediation strategy, and carrying out dynamic optimization by utilizing reinforcement learning; then, mediation terms are generated, and the terms are updated based on the emotional state scores and feedback of the two parties of the contradictory dispute; and finally, generating a protocol hash value according to the updated mediation terms, synchronizing the protocol hash value to the block chain for evidence storage, and automatically executing the updated mediation terms. According to the method, the mediation verbal skill and strategy can be dynamically adjusted according to the real-time context, the mediation pertinence and flexibility are improved, and the method is suitable for diversified and complicated mediation scenes.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence and legal technology, and specifically relates to an intelligent agent method for real-time mediation of conflicts and disputes based on a multimodal large language model. Background Art

[0002] As social conflicts and disputes become increasingly diverse and complex, traditional mediation methods face significant challenges in terms of efficiency, accuracy, and legal compliance. Most current mediation technologies rely on single-modal data (such as text or voice), making it difficult to fully capture the parties' true intentions and emotional states. For example, when one party verbally expresses a willingness to compromise but reveals resistance through body language (such as crossed arms, averted eye contact) or micro-expressions (such as frequent frowning), single-modal systems can easily misjudge their true intentions, causing mediation strategies to deviate from actual needs.

[0003] In recent years, multimodal emotion recognition technology has gained increasing attention, aiming to improve the accuracy and robustness of emotion recognition by integrating multiple data modalities, such as text and speech. However, existing research has primarily focused on fields such as education and healthcare, and has yet to be fully applied to legal mediation scenarios. Practical applications also face challenges such as complex data fusion and difficult model training.

[0004] At the same time, existing legal reasoning tools are mostly based on static rule engines and lack the ability to dynamically link legal provisions with dispute points. In complex scenarios such as cross-border contract disputes, conflicts between clauses in different jurisdictions (such as the Civil Code and the United Nations Convention on International Commercial Contracts) are difficult for the system to identify. The resulting mediation agreements carry legal risks and have low judicial acceptance rates. Despite the recent emergence of some large language models in the legal field (such as Lawformer and DISC-LawLLM), their practical applications still face challenges such as insufficient model generalization and lagging legal knowledge updates.

[0005] Furthermore, traditional mediation strategies rely on predefined rules, making them unable to adjust in real time to the dynamics of disputes (such as sudden shifts in sentiment and interests). Furthermore, they lack a trustworthy mechanism for agreement execution, leading to frequent breach of contract disputes. Despite the potential of blockchain technology for data storage and automated execution, its application in mediation agreement execution is still in its exploratory stages, lacking mature solutions.

[0006] Therefore, the industry urgently needs an intelligent mediation solution that can integrate multimodal information, enhance legal reasoning, support dynamic optimization and ensure execution transparency, so as to improve mediation efficiency and the legal effect of the agreement. Summary of the Invention

[0007] The main purpose of the present invention is to overcome the shortcomings and deficiencies of the prior art and to provide a real-time intelligent mediation method for conflicts and disputes based on a multimodal large language model.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions: The method for real-time mediation of conflicts and disputes based on a multimodal large language model includes the following steps: Obtain voice information from both parties of the dispute and convert it into text information; Calculate emotional state scores based on voice and text information, and use a large language model to predict the intentions of both parties in the dispute; these intentions include insisting on the request and accepting mediation; Perform semantic analysis on text information to extract the facts, demands, and focus of the dispute between the two parties, and perform contradiction detection to quantify the degree of conflict in the focus of the dispute; The optimal compromise point is generated by assigning weights to the demands of both parties in the dispute. A mediation strategy is generated based on the optimal compromise point and the focus of the dispute, and the maximum satisfaction of both parties in the dispute is used as the reward function. The mediation strategy is dynamically optimized using reinforcement learning. Generate mediation terms based on the optimized mediation strategy and update the terms based on the emotional state scores and feedback from both parties involved in the dispute; The agreement hash value generated based on the updated mediation terms is synchronized to the blockchain for evidence storage, and the updated mediation terms are automatically executed.

[0009] As a preferred technical solution, the voice information is converted into text information using automatic speech recognition technology.

[0010] As a preferred technical solution, the emotional state score is calculated based on the emotional characteristics of voice information and text information in the following way: Based on the pre-trained BERT sentiment analysis model, the sentiment tendency of the text information is scored to obtain the text sentiment score; The emotional features of speech information are extracted and scored based on the Mel-frequency cepstral coefficients to obtain the speech emotion score; The emotional state score is obtained by weighted summing of the text emotion score and the voice emotion score; The use of a large language model to predict the intentions of both parties in a conflict or dispute is specifically as follows: Fuse the acoustic features corresponding to text information and speech information to form a multimodal input feature vector; The multimodal input feature vector is fed into a pre-trained large language model fine-tuned with legal domain data for semantic understanding and reasoning, analyzing the language expressions and emotional tendencies of both parties in the dispute, and predicting the intentions of both parties in the dispute. A confidence assessment is performed on the intent predicted by the large language model. If the confidence is lower than the set confidence threshold, feature fusion and model inference are re-performed, or a comprehensive judgment is made in combination with other modal information; the other modal information includes visual information.

[0011] As a preferred technical solution, a legally enhanced large language model is used to perform semantic analysis on text information, extract the facts, demands, and focus of the dispute between the two parties, and map the facts and demands of the two parties to the focus of the dispute; The cosine similarity is used to quantify the focus of the dispute between the two parties in the conflict and obtain the degree of conflict.

[0012] As a preferred technical solution, the optimal compromise point is generated by allocating weights to the demands of both parties in the conflict, specifically: The interest balance algorithm is used to allocate the weights of the demands of the two parties in the conflict based on the Shapley value; The Pareto optimal solution is calculated based on the demands and weights of the two parties in the conflict and dispute, and the optimal compromise point is obtained.

[0013] As a preferred technical solution, the interest balancing algorithm is used to allocate the weights of the claims of both parties in the conflict based on the Shapley value, specifically: Collect all possible combinations of demands and their corresponding value functions of both parties in the conflict; Initialize the weight of each claim of both parties in the conflict to 0; For all demand combinations that do not include a certain demand, calculate the marginal contribution of the demand combination that does not include the certain demand after adding the demand; The marginal contribution of the demand is added to all demand combinations that do not contain the demand, and the average is used to obtain the Shapley value of the demand and update it as the weight of the demand; The Pareto optimal solution is calculated based on the demands and weights of the two parties in the conflict, specifically: The demands of both parties in the conflict and dispute and their corresponding weights are used as variables, and corresponding constraints are set based on actual conditions; Taking maximizing the weighted satisfaction of both parties in the conflict as the objective function, a Pareto model is established; Mathematical programming methods or intelligent optimization algorithms are used to iteratively solve the Pareto model to obtain the optimal compromise point.

[0014] As a preferred technical solution, the mediation clauses generated based on the optimal compromise point and the focus of the dispute are specifically as follows: Based on the optimal compromise point and the dispute focus of the two parties in the conflict, combined with the legal clause template library, a matching algorithm is used to select the clause template that best matches the current dispute focus and the optimal compromise point; Fill the facts and demands of both parties in the dispute into the corresponding positions of the clause template, use natural language generation technology to semantically optimize and integrate the clause content, and generate preliminary mediation clauses; Input the preliminary mediation clauses into the multimodal large language model to further optimize the legal terms and logical structure of the preliminary mediation clauses and generate the final mediation clauses; The matching algorithm calculates the matching degree with each legal clause template in the legal clause template library according to the characteristics of the current dispute focus and the optimal compromise point, and selects the template with the highest matching degree as the clause template.

[0015] As a preferred technical solution, the maximum satisfaction of both parties in the conflict is used as the reward function, and the mediation strategy is dynamically optimized using reinforcement learning, specifically: 1) Initialize the reinforcement learning model: Select the reinforcement learning model and initialize the model parameters; 2) Define the state space and action space: The state space includes the emotional state of the parties in the dispute, the focus of the dispute, and the current mediation progress; the action space includes various mediation strategies; 3) Strategy Evaluation and Update: In each mediation step, a mediation strategy is selected as an action based on the current state, and a reward value is calculated based on the feedback from both parties. The parameters of the reinforcement learning model are updated based on the reward value, and the quality of the current mediation strategy is evaluated and adjusted. 4) Training and Optimization: Repeat the strategy evaluation and update steps, continuously allowing the reinforcement learning model to interact with the mediation environment and gradually optimize the mediation strategy. The exploration-exploitation strategy is adopted during the training process, that is, different mediation strategies are explored and tried in the early stages of training, and the learned excellent mediation strategies are utilized in the later stages of training. 5) Strategy output and application: After training, the reinforcement learning model outputs the optimal mediation strategy for the current conflict and dispute situation.

[0016] As a preferred technical solution, the updating clauses based on the emotional state scores and feedback of the two parties in the conflict or dispute are specifically as follows: Obtain the emotional status scores of both parties in the dispute in real time, and collect feedback from both parties on the mediation terms; Based on the monitored emotional state scores and feedback information, refer to the predefined terms adjustment rules to clarify the adjustment direction and amount; Using natural language processing technology, the mediation clauses are partially modified according to the determined adjustment amount to generate updated mediation clauses; the updated mediation clauses are input into the multimodal large language model for review and optimization to obtain the final updated mediation clauses.

[0017] As a preferred technical solution, the agreement hash value is generated using a hash algorithm based on the content of the updated mediation clause; The blockchain adopts Hyperledger Fabric chain.

[0018] Compared with the prior art, the present invention has the following advantages and beneficial effects: 1. Fusion of multimodal inputs to achieve dual recognition of emotions and legal intent: This invention utilizes multimodal data such as speech and text as input, combined with the capabilities of large language models for semantic understanding and emotion recognition, effectively capturing the true expression intentions and emotional states of both parties to the dispute, overcoming the problem of insufficient recognition accuracy in traditional mediation systems due to the single information dimension.

[0019] 2. Introducing a legally enhanced large language model to achieve refined understanding and reasoning of legal semantics: The legally enhanced large language model used in this invention (based on LLaMA fine-tuning and combined with legal corpus training) can accurately extract dispute facts, demands and focus of disputes, and supports reasoning and semantic matching in dynamic and diverse legal semantic scenarios, which is significantly superior to the static reasoning method of traditional rule engines; it has stronger generalization and adaptability in extracting legal relationships, mapping legal provisions and avoiding conflicting clauses.

[0020] 3. Extraction of dispute focus and dynamic optimization of mediation strategies: By extracting keywords, focus topics and core conflicts from user input, the present invention can dynamically adjust mediation language and strategies according to real-time context, thereby improving the pertinence and flexibility of mediation and adapting to diverse and complex mediation scenarios.

[0021] 4. Strengthen the credibility of agreement execution and ensure the implementation of mediation results: This invention introduces blockchain technology to store mediation results and combines it with the smart contract mechanism to realize the automatic execution of the agreement, effectively preventing the agreement from being tampered with or refused to perform, thereby enhancing the legal effect and judicial acceptance rate of the mediation agreement.

[0022] 5. Adaptable to multiple scenarios, with good practicality and promotion value: This invention is applicable to various mediation scenarios such as pre-trial mediation in courts, community conflict resolution, and e-commerce platform disputes. It has cross-scenario deployment capabilities and helps to build a standardized and intelligent modern dispute mediation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0024] Figure 1 This is a flow chart of an intelligent agent method for real-time mediation of conflicts and disputes based on a multimodal large language model in an embodiment of the present invention.

[0025] Figure 2 This is a flowchart of the controversial focus and quantification in an embodiment of the present invention.

[0026] Figure 3 This is a flow chart of generating the optimal compromise point and optimizing the mediation strategy in an embodiment of the present invention.

[0027] Figure 4 This is a flowchart of blockchain evidence storage and automatic execution in an embodiment of the present invention. DETAILED DESCRIPTION

[0028] In order to enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0029] References to "embodiments" in this application mean that a particular feature, structure, or characteristic described in connection with the embodiment may be included in at least one embodiment of the application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described in this application may be combined with other embodiments.

[0030] like Figure 1 As shown, this embodiment discloses a method for real-time mediation of conflicts and disputes based on a multimodal large language model, comprising the following steps: S1. Acquire voice information of both parties of the conflict and convert the voice information into text information.

[0031] In the present invention, the voice information of the two parties in the dispute is received and converted into text information using Automatic Speech Recognition (ASR) technology. The process is described as follows: T = ASR(V), where T is the text information and V is the voice information.

[0032] In this embodiment, the Whisper model is used for automatic speech recognition.

[0033] S2. Calculate emotional state scores based on voice and text information, and use a large language model to predict the intentions of both parties in a conflict; intentions include insisting on demands and accepting mediation.

[0034] Furthermore, sentiment analysis is first performed to extract the emotional state score based on the emotional features of voice information and text information in the following way: Based on the pre-trained BERT sentiment analysis model, the sentiment tendency of the text information is scored to obtain the text sentiment score; The emotional features of speech information are extracted and scored based on the Mel-frequency cepstral coefficients to obtain the speech emotion score; The emotional state score is obtained by weighted summing of the text emotion score and the voice emotion score. The process is described as follows: Ec = α·SentimentAnalysis (T) + β·AudioFeature(V), Where Ec is the emotional state score, α and β are weight coefficients with α + β = 1, SentimentAnalysis() is a function that scores the emotional tendency of text information based on the pre-trained BERT sentiment analysis model, T is the text information, AudioFeature() is a function that extracts and scores the emotional features (pitch, volume, and speaking speed, etc.) of speech information based on Mel-frequency cepstral coefficients (MFCC), and V is the speech information.

[0035] Next, we use the large language model to predict the intentions of both parties involved in the conflict, specifically: The acoustic features corresponding to the text information T and the voice information V are fused to form a multimodal input feature vector; The multimodal input feature vector is fed into a pre-trained large language model fine-tuned with legal domain data for semantic understanding and reasoning, analyzing the language expressions and emotional tendencies of both parties in the dispute, and predicting the intentions of both parties in the dispute. Perform a confidence assessment on the intention predicted by the large language model. If the confidence is lower than the set confidence threshold, re-perform feature fusion and model inference, or combine other modal information (such as visual information that may be obtained later) for comprehensive judgment.

[0036] S3. Perform semantic analysis on text information to extract the facts, demands, and focus of dispute of both parties in the conflict and dispute, and perform conflict detection to quantify the degree of conflict of the focus of dispute.

[0037] The purpose of extracting facts, demands and dispute focus is to fully and accurately understand the core content of the conflict and the dispute between the two parties, so as to provide an accurate information basis for the subsequent formulation of mediation strategies. By mapping facts and demands to dispute focus and quantifying the dispute focus, the degree and focus of the dispute between the two parties can be intuitively presented, providing a quantitative basis for the formulation of mediation strategies. Figure 2 As shown, the present invention uses a legally enhanced large language model (legal enhanced LLM) to perform semantic analysis on text information and extract the facts of both parties in the conflict. F i , demands R iand controversial issues C i The facts and demands of both parties in the conflict and dispute are mapped to the focus of the dispute. The mapping method is: C i = LLM( F i , R i ), i∈{1,2}, in, C i For conflicts and disputes i The focus of the dispute, F i For conflicts and disputes i The fact that R i For conflicts and disputes i The party’s demands.

[0038] Then, the cosine similarity is used to quantify the focus of the dispute between the two parties in the conflict and obtain the degree of conflict: , in, C 1 is the focus of dispute on one side of the conflict or dispute, C 2 is the focus of the dispute on the other side of the conflict. C 1i The focus of the dispute C 1st i The value of the dimension, C 2i To indicate the focus of the other party in the conflict or dispute C 2nd i The value of the dimension, n is the number of dimensions used to represent the focus of the dispute, that is, the total number of characteristic dimensions used to represent the focus of the dispute in the legal semantic space.

[0039] The legally enhanced large language model used in this example is built using Llama as the base model and LoRA fine-tuning technology. It is a large language model specifically designed for processing legal-related texts and tasks. Its construction process is as follows: 1) Data Collection: Selenium is used to collect a large amount of legal case texts, legal provisions, and other information from multiple public channels such as the national case database, legal provisions database, and related system databases as raw data for model training.

[0040] 2) Data preprocessing: Using the Python programming language, the collected raw data is cleaned, organized, and converted. This includes removing irrelevant characters, unifying the text encoding format, and performing word segmentation, converting the data into a structured, regular format that meets the input requirements for model training.

[0041] 3) Model training: The LLAMA-FACTORY open-source framework is used, and the LoRA fine-tuning method is used to train and optimize the model based on different scenario prompts. This allows the model to learn the semantic features, professional terminology, and logical structure of legal texts, adapting to the professional characteristics of the legal field and the requirements of mediation tasks.

[0042] S4. Assign weights to the demands of both parties in the conflict and dispute to generate the optimal compromise point. Generate a mediation strategy based on the optimal compromise point and the focus of the dispute, and use the maximum satisfaction of both parties in the conflict and dispute as the reward function. Use reinforcement learning to dynamically optimize the mediation strategy.

[0043] S4.1. In the process of conflict and dispute mediation, the demands of both parties often have different levels of importance and urgency. Traditional mediation methods usually have difficulty finding a balance between the demands of both parties that can maximize the interests of both parties while ensuring the feasibility of the mediation plan. Therefore, this invention assigns weights to the demands of both parties in the conflict and dispute, quantifies the relative importance of each demand in the overall conflict, and thus reasonably balances the interests of both parties in the mediation plan, such as Figure 3 As shown, specifically: S4.1.1. Use the interest balance algorithm based on Shapley value to allocate the weights of the claims of both parties in the dispute W ij By quantifying the importance of the demands of both parties, we ensure that the mediation plan maximizes the balance of interests and acceptance of both parties on the basis of fairness and rationality, avoids mediation failure due to over-emphasis or neglect of a certain demand, and improves the success rate of mediation and the satisfaction of both parties.

[0044] Furthermore, the specific steps for assigning weights are: 1) Input data: Collect all possible combinations of demands and their corresponding value functions of both parties in the conflict v ( S ),in Represents a combination of demands, N ={ R 11 , R 12 ,..., R 2n} represents the set of all demands of both parties in the conflict. n is the total number of requests.

[0045] 2) Initialization weight: i Fangdi j demands R ij , initialize its weight W ij =0.

[0046] 3) Calculate marginal contribution: For all R ij Combination of demands , calculate the joining demands R ij hour S’ Marginal contribution: , that is, in the combination of demands S’ Add demands R ij The increase in value afterward.

[0047] 4) Calculate Shapley value: Add appeal R ij Combination of demands S’ The marginal contribution of the average is the demand R ij The Shapley value is updated to the demand R ij The weight of , that is: , in, For all those who do not include demands R ij The number of demands in the demand combination.

[0048] S4.1.2. Calculate the Pareto optimal solution based on the demands and weights of the two parties in the conflict and obtain the optimal compromise point Be.

[0049] Furthermore, the specific method for calculating the Pareto optimal solution is: 1) Define variables: The demands of both parties in the dispute R ij and its weight W ij As a variable, corresponding constraints are set based on actual conditions. For example, the degree of satisfaction of the demands cannot exceed the upper limit proposed, nor can it be lower than a certain lower limit.

[0050] 2) Establish a Pareto model: Taking maximizing the weighted satisfaction of both parties in a conflict as the objective function, establish a Pareto model: , in, f ( R ij , B ), i ∈{1,2} represents the demand R ij At the point of compromise B The matching function is used to measure its satisfaction. ni For the i The number of demands from the parties.

[0051] 3) Solving the model: Use mathematical programming methods (such as linear programming or nonlinear programming) and intelligent optimization algorithms (such as genetic algorithms or particle swarm algorithms) to iteratively solve the Pareto model for the specific optimal compromise point Be.

[0052] S4.2. Further, based on the optimal compromise point and the focus of the dispute, mediation clauses are generated, specifically: S4.2.1. Based on the optimal compromise point Be and the focus of the dispute between the two parties C i , combined with the legal clause template library, select the clause template that best meets the current dispute focus and the optimal compromise point through a matching algorithm T m .

[0053] Furthermore, the matching algorithm is based on the current focus of the dispute. C i and the characteristics of the optimal compromise point Be, calculate the matching degree with each legal clause template in the legal clause template library, and select the template with the highest matching degree as the clause template T m , and the matching degree calculation formula is: , in, Indicates the focus of the dispute C i With the j Legal clause templates The similarity of the corresponding focus is obtained through semantic similarity calculation methods, such as cosine similarity, Jaccrad similarity, etc. For the j Legal clause templates The weight of the legal clause template is determined based on factors such as the universality and frequency of application of the legal clause template.

[0054] S4.2.2. Fill in the clause template with the facts and demands of both parties in the dispute T m The corresponding position of the clause is semantically optimized and integrated using natural language generation technology to generate preliminary mediation clauses. T k .

[0055] S4.2.3. Preliminary Mediation Clauses T kInput a multimodal large language model, combine the model's understanding of legal norms and analysis of similar cases, further optimize the legal terms and logical structure of the preliminary mediation clauses, ensure the legality and enforceability of the clauses, and generate the final mediation clauses. .

[0056] S4.3. Furthermore, reinforcement learning can achieve real-time strategy adjustment to adapt to information changes during the mediation process; with maximum satisfaction as the optimization goal, the quality of the mediation strategy and the satisfaction of both parties can be improved; and through real-time feedback optimization strategy, the mediation success rate can be improved. Therefore, the present invention uses the maximum satisfaction of both parties in the conflict as the reward function and uses reinforcement learning to dynamically optimize the mediation strategy. Specifically, S4.3.1. Initialize the reinforcement learning model: Select an appropriate reinforcement learning model (such as Q-learning, Deep Q-Network, etc.) and initialize the model parameters, such as the learning rate and discount factor.

[0057] S4.3.2. Define state space and action space: The state space includes information such as the emotional state of the parties to the conflict, the focus of the dispute, the current mediation progress, etc.; the action space includes various possible mediation strategies, such as adjusting the proportion of responsibility division, changing the compensation amount, introducing third-party evaluation, etc.

[0058] S4.3.3. Strategy evaluation and update: In each mediation step, a mediation strategy is selected as an action based on the current state, and then the reward value is calculated based on the feedback from both parties to the conflict (such as acceptance level, emotional changes, etc.); the parameters of the reinforcement learning model are updated based on the reward value, the quality of the current strategy is evaluated, and the mediation strategy is continuously adjusted to maximize the long-term cumulative reward.

[0059] S4.3.4, Training and Optimization: Repeat the above strategy evaluation and update steps, continuously let the reinforcement learning model interact with the mediation environment, accumulate experience, and gradually optimize the mediation strategy; during the training process, an exploration-exploitation strategy can be adopted, that is, in the early stage, more exploration and trying different strategies are carried out, and in the later stage, more use is made of the excellent strategies that have been learned.

[0060] S4.3.5. Strategy output and application: After sufficient training, the reinforcement learning model outputs the optimal mediation strategy for the current state of conflict and dispute for execution by the mediation agent, thereby dynamically optimizing the mediation process and improving the mediation success rate and satisfaction of both parties.

[0061] S5. Generate mediation clauses based on the optimized mediation strategy, and update the clauses based on the emotional state scores and feedback of the two parties in the dispute.

[0062] Furthermore, in order to make the mediation clauses better adapt to the emotional changes and real-time feedback of the parties, improve the acceptability and adaptability of the clauses, promote the smooth progress of the mediation process and the successful achievement of mediation results, the present invention updates the clauses based on the emotional state scores and feedback of the parties to the conflict, specifically: S5.1. Emotion and feedback monitoring: Obtain the emotional state scores Ec of both parties to the conflict in real time, and collect feedback information on the mediation terms from both parties to the conflict (such as acceptance of verbal expression and affirmative or negative gestures).

[0063] S5.2. Application of adjustment rules: Based on the emotional state score Ec and feedback information, refer to the predefined terms adjustment rules to determine the adjustment direction and adjustment amount .

[0064] S5.3, Clause Update: Using natural language processing technology, based on the determined adjustment amount Mediation Clauses Make partial modifications to generate updated mediation clauses When updating terms, natural language processing technology primarily utilizes operations such as text replacement, insertion, and deletion, combined with grammatical checking and semantic analysis, to ensure that the updated mediation terms maintain the original intent while better adapting to the parties’ emotional changes and implementation feedback.

[0065] S5.4. Updated Mediation Clauses The mediation terms are finally updated after being input into the multimodal large language model for review and optimization to ensure their legal compliance and logical rigor.

[0066] S6. Generate an agreement hash value based on the updated mediation terms and synchronize it to the blockchain for evidence storage, and automatically execute the updated mediation terms.

[0067] Further, such as Figure 4 As shown, a hash algorithm is used to generate a protocol hash value based on the updated mediation clause: H = Hash (updated mediation clause). This hash value is then synchronized to the Hyperledger Fabric chain for storage. The updated mediation clause is automatically executed. Failure to fully or incompletely execute the clause triggers a breach of contract (e.g., late payment triggers liquidated damages).

[0068] Example 2 To facilitate understanding of the method of the present invention, this embodiment illustrates the steps of the method of the present invention through a specific example. Assume that a conflict dispute involving noise disturbing residents caused by a shop occurs in a community conflict mediation center, involving shop owner A and resident B. The mediation agent is enabled with the method of the present invention to assist staff in conducting real-time mediation. The specific steps are: Step S1: Get voice and convert it into text First, the mediation agent collects on-site voice statements from shop owners and residents, and uses the Whisper automatic speech recognition model to convert these voice statements into text information T. For example, if a resident says, "Playing music outside after 10 o'clock every night really disturbs my rest," this speech is converted into corresponding text content and passed to the subsequent module for processing.

[0069] Step S2: Calculate emotional state scores and predict intentions The mediation agent first extracts emotional features from the voice and text information of residents and shop owners. The text is analyzed using the BERT sentiment analysis model to extract sentiment scores, while the speech is analyzed by extracting MFCC acoustic features and inputting them into a CNN model for sentiment scoring. The resulting emotional state score is recorded as Ec.

[0070] Next, the text information T and the emotional state score Ec are fed into a Transformer-based legal language model, which outputs a probability distribution of the intentions of both parties. For example, the model identifies a resident's intention to "insist on their demands" with a probability of 0.82, while the probability of "accepting mediation" is 0.18, indicating that their emotions are intense and they are not likely to accept the mediation proposal.

[0071] Step S3: Extract facts, claims, and dispute points and quantify conflicts The legally enhanced large language model performs semantic understanding of the text content and extracts the facts about store A. F 1 is "comply with business regulations and use legal audio equipment", R 1 is "continue normal business"; the facts of residents F 2 is "sound affects sleep", appeal R 2 means "No playing of music outside after 10pm".

[0072] Then, the agent F 1. R 1 and F 2, R 2 Mapping out the focus of their respective disputes C 1. C 2 and calculated the cosine similarity, and found that the conflict degree between the two on the controversial point of "audio playback time limit" is 0.85, indicating that this issue is highly controversial.

[0073] Step S4: Generate compromise points and optimize mediation strategies The agent quantifies the demands of both parties and assigns weights based on Shapley values. For example, a resident's demand for "no noise after 10 pm" is weighted 0.6, while a shopkeeper's demand for "normal business hours are a legal right" is weighted 0.4.

[0074] Step S5: Generate mediation terms and negotiate updates Based on the weights of each party's claims, the mediation agent combines a corpus of historical similar cases with the current focus of the conflict and, using a legally enhanced large language model, generates a mediation strategy S that aligns with the wishes of all parties. In this example, the mediation agent, combining the dispute focus of "audio playback time limits" with the weighted claims, generates the following mediation strategy: "Shops are allowed to play music indoors during legal business hours, but are prohibited from playing any music outdoors after 10:00 PM daily."

[0075] Subsequently, the agent automatically generates a structured mediation clause P based on the mediation strategy, which includes standard elements such as mediation matters, obligations of each party, enforcement methods, and consequences of breach of contract. An example mediation clause is as follows: "Party A (the shop owner) shall, from now on, cease using audio equipment to play music outdoors after 10:00 PM daily. Party B (the resident) shall refrain from making new complaints regarding noise issues if Party A complies with the above obligations." If the emotional state or intention of any party fluctuates during the mediation process, the intelligent agent can recalculate the emotional state score Ec and the intention probability distribution, and automatically iterate and update the mediation strategy and terms to achieve dynamic optimization of the mediation suggestions.

[0076] Step S6: Confirmation of mediation and electronic evidence storage after reaching an agreement When the mediation clause P is confirmed and accepted by both parties, the mediation agent integrates the mediation results with the structured data of the entire process to generate an electronic mediation agreement document D, which includes but is not limited to: Basic information of both parties and the time and location of mediation; automatic identification and extraction of facts F1, F2, demands R1, R2; analysis of the focus of dispute C and the degree of conflict; emotional state score Ec and probability distribution of intention; final mediation terms P and source annotation of the generated strategy; Speech-to-text conversion and a summary of important mediation records are generated. Agreement Document D is encrypted and stored in the backend mediation filing system, equipped with an electronic signature function. This can be used for subsequent filing with regulatory agencies, judicial confirmation, or data analysis. This automatically completes the mediation process.

[0077] It should be noted that, for the sake of convenience, the aforementioned method embodiments are all expressed as a series of action combinations, but those skilled in the art should know that the present invention is not limited to the described order of actions, because according to the present invention, certain steps can be performed in other orders or simultaneously.

[0078] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0079] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A real-time intelligent agent method for conflict and dispute mediation based on a multimodal large language model, characterized by: The steps include: Obtain voice information from both parties of the dispute and convert it into text information; Calculate emotional state scores based on voice and text information, and use a large language model to predict the intentions of both parties in the dispute; these intentions include insisting on the request and accepting mediation; Perform semantic analysis on text information to extract the facts, demands, and focus of the dispute between the two parties, and perform contradiction detection to quantify the degree of conflict in the focus of the dispute; The optimal compromise point is generated by assigning weights to the demands of both parties in the dispute. A mediation strategy is generated based on the optimal compromise point and the focus of the dispute, and the maximum satisfaction of both parties in the dispute is used as the reward function. The mediation strategy is dynamically optimized using reinforcement learning. Generate mediation terms based on the optimized mediation strategy and update the terms based on the emotional state scores and feedback from both parties involved in the dispute; The agreement hash value generated based on the updated mediation terms is synchronized to the blockchain for evidence storage, and the updated mediation terms are automatically executed.

2. The method for real-time mediation of conflicts and disputes based on a multimodal large language model according to claim 1 is characterized in that: The voice information is converted into text information using automatic voice recognition technology.

3. The method for real-time mediation of conflicts and disputes based on a multimodal large language model according to claim 1 is characterized in that: The emotional state score is calculated based on the emotional features of voice information and text information as follows: Based on the pre-trained BERT sentiment analysis model, the sentiment tendency of the text information is scored to obtain the text sentiment score; The emotional features of speech information are extracted and scored based on the Mel-frequency cepstral coefficients to obtain the speech emotion score; The emotional state score is obtained by weighted summing of the text emotion score and the voice emotion score; The use of a large language model to predict the intentions of both parties in a conflict or dispute is specifically as follows: Fuse the acoustic features corresponding to text information and speech information to form a multimodal input feature vector; The multimodal input feature vector is fed into a pre-trained large language model fine-tuned with legal domain data for semantic understanding and reasoning, analyzing the language expressions and emotional tendencies of both parties in the dispute, and predicting the intentions of both parties in the dispute. A confidence assessment is performed on the intent predicted by the large language model. If the confidence is lower than the set confidence threshold, feature fusion and model inference are re-performed, or a comprehensive judgment is made in combination with other modal information; the other modal information includes visual information.

4. The method for real-time mediation of conflicts and disputes based on a multimodal large language model according to claim 1 is characterized in that: Use a legally enhanced large language model to perform semantic analysis on text information, extract the facts, demands, and focus of the dispute between the two parties, and map the facts and demands of the two parties to the focus of the dispute; The cosine similarity is used to quantify the focus of the dispute between the two parties in the conflict and obtain the degree of conflict.

5. The method for real-time mediation of conflicts and disputes based on a multimodal large language model according to claim 1 is characterized in that: The optimal compromise point is generated by allocating weights to the demands of both parties in the conflict, specifically: The interest balance algorithm is used to allocate the weights of the demands of the two parties in the conflict based on the Shapley value; The Pareto optimal solution is calculated based on the demands and weights of the two parties in the conflict and dispute, and the optimal compromise point is obtained.

6. The method for real-time mediation of conflicts and disputes based on a multimodal large language model according to claim 5 is characterized in that: The interest balancing algorithm is used to allocate the weights of the claims of both parties in the conflict based on the Shapley value, specifically: Collect all possible combinations of demands and their corresponding value functions of both parties in the conflict; Initialize the weight of each claim of both parties in the conflict to 0; For all demand combinations that do not include a certain demand, calculate the marginal contribution of the demand combination that does not include the certain demand after adding the demand; The marginal contribution of the demand is added to all demand combinations that do not contain the demand, and the average is used to obtain the Shapley value of the demand and update it as the weight of the demand; The Pareto optimal solution is calculated based on the demands and weights of the two parties in the conflict, specifically: The demands of both parties in the conflict and dispute and their corresponding weights are used as variables, and corresponding constraints are set based on actual conditions; Taking maximizing the weighted satisfaction of both parties in the conflict as the objective function, a Pareto model is established; Mathematical programming methods or intelligent optimization algorithms are used to iteratively solve the Pareto model to obtain the optimal compromise point.

7. The method for real-time mediation of conflicts and disputes based on a multimodal large language model according to claim 1 is characterized in that: The mediation clauses generated based on the optimal compromise point and the focus of the dispute are specifically as follows: Based on the optimal compromise point and the dispute focus of the two parties in the conflict, combined with the legal clause template library, a matching algorithm is used to select the clause template that best matches the current dispute focus and the optimal compromise point; Fill the facts and demands of both parties in the dispute into the corresponding positions of the clause template, use natural language generation technology to semantically optimize and integrate the clause content, and generate preliminary mediation clauses; Input the preliminary mediation clauses into the multimodal large language model to further optimize the legal terms and logical structure of the preliminary mediation clauses and generate the final mediation clauses; The matching algorithm calculates the matching degree with each legal clause template in the legal clause template library according to the characteristics of the current dispute focus and the optimal compromise point, and selects the template with the highest matching degree as the clause template.

8. The method for real-time mediation of conflicts and disputes based on a multimodal large language model according to claim 1 is characterized in that: The maximum satisfaction of both parties in the conflict is used as the reward function, and the mediation strategy is dynamically optimized using reinforcement learning. Specifically: 1) Initialize the reinforcement learning model: Select the reinforcement learning model and initialize the model parameters; 2) Define the state space and action space: The state space includes the emotional state of the parties in the dispute, the focus of the dispute, and the current mediation progress; the action space includes various mediation strategies; 3) Strategy Evaluation and Update: In each mediation step, a mediation strategy is selected as an action based on the current state, and a reward value is calculated based on the feedback from both parties. The parameters of the reinforcement learning model are updated based on the reward value, and the quality of the current mediation strategy is evaluated and adjusted. 4) Training and Optimization: Repeat the strategy evaluation and update steps, continuously allowing the reinforcement learning model to interact with the mediation environment and gradually optimize the mediation strategy; An exploration-exploitation strategy is adopted during the training process, that is, different mediation strategies are explored and tried in the early stages of training, and the learned excellent mediation strategies are utilized in the later stages of training; 5) Strategy output and application: After training, the reinforcement learning model outputs the optimal mediation strategy for the current conflict and dispute situation.

9. The method for real-time mediation of conflicts and disputes based on a multimodal large language model according to claim 1 is characterized in that: The update terms based on the emotional state scores and feedback of both parties in the conflict are specifically as follows: Obtain the emotional status scores of both parties in the dispute in real time, and collect feedback from both parties on the mediation terms; Based on the monitored emotional state scores and feedback information, refer to the predefined terms adjustment rules to clarify the adjustment direction and amount; Using natural language processing technology, the mediation clauses are partially modified according to the determined adjustment amount to generate updated mediation clauses; the updated mediation clauses are input into the multimodal large language model for review and optimization to obtain the final updated mediation clauses.

10. The method for real-time mediation of conflicts and disputes based on a multimodal large language model according to claim 1 is characterized in that: The agreement hash value is generated by using a hash algorithm based on the content of the updated mediation clause; The blockchain adopts Hyperledger Fabric chain.

Images