An experience-driven multi-agent negotiation contract construction method and device

Abstract

The application provides an experience-driven multi-agent negotiation contract construction method and device, and relates to the technical field of contract negotiation. The method generates an individualized initial clause scheme according to the initial appeal description of a user through a pre-trained experience prediction model. In subsequent rounds of negotiation, a multi-agent module containing a decision agent and a coordination agent is used to promote the negotiation process, wherein the decision agent is responsible for generating a candidate scheme, the coordination agent is responsible for analyzing user feedback intentions, evaluating candidate schemes and controlling the interaction process. After the negotiation is completed, the effectiveness of the pending scheme reached is checked, and the final consensus scheme that passes the verification is stored in the experience library for subsequent incremental updating of the experience prediction model. The application realizes the automation and intelligence of contract negotiation through experience prediction and multi-agent collaboration, and has the significant advantages of high negotiation efficiency, stable scheme quality and continuous self-enhancement with experience accumulation.

Description

Technical Field

[0001] This invention relates to the field of contract negotiation technology, and specifically to an experience-driven method and apparatus for constructing multi-agent negotiation clause contracts. Background Technology

[0002] As large language models demonstrate increasingly powerful capabilities in natural language understanding, knowledge reasoning, and complex task processing, building intelligent agents based on large language models for negotiation and dialogue scenarios has become an important research direction in the field of dialogue systems. Among these, contract negotiation agents for business negotiations are expected to assist business personnel in reviewing terms, assessing risks, and making negotiation decisions, while supporting natural language interaction between the parties. From a technological evolution perspective, AI-driven contract negotiation systems have evolved from early rule-based and utility function-based automated negotiation agents to intelligent negotiation systems centered on large language models in recent years. However, existing methods still face several technical challenges in building contract negotiation agents with continuous negotiation capabilities.

[0003] Based on the construction method of the decision engine, existing contract negotiation and bargaining agents can be roughly divided into three categories: rule-based and utility function-driven methods, large language model-based direct-driven methods, and hybrid methods. Rule-based and utility function-driven methods do not rely on large language models for decision-making. Instead, they predefine negotiation topics, preference weights, and utility functions, and calculate the price adjustment range through a preset algorithm. The large language model is only responsible for converting the calculation results into natural language expressions, acting as a "translator" rather than a "decision-maker." This technical approach can be traced back to early foundational work, such as a series of automated negotiation agent patents filed by HP between 2003 and 2004, which established an electronic contract negotiation paradigm based on multi-perspective risk and trust assessment. IBM's 2014 patent application on financial risk analysis of IT service contracts introduced a technical solution that compares new contracts with historical precedents based on similarity scoring. Large language model-based direct-driven methods use the large language model as the core decision engine, writing negotiation strategies, counterparty preferences, and historical dialogues into prompts, with the large language model generating the next round of offers or responses end-to-end. For example, existing patent applications have proposed technical solutions using chatbots driven by large language models to negotiate contract terms. Some systems can already complete contract negotiations between agents in a completely autonomous manner. The advantages of this type of method lie in its flexibility and language expression capabilities, but the decision-making process lacks interpretability and controllability. Hybrid-driven methods, on the other hand, involve a negotiation engine evaluating the utility of offers based on preset preferences and calculating adjustment ranges, followed by a large language model converting the quantified results into natural language offers. The core idea is a division of labor between "computational decision-making and language expression." In terms of knowledge and experience utilization, all three types of methods mainly enhance decision-making capabilities through legal and regulatory databases, external searches, historical dialogue memories, and case databases. Existing research has applied search-enhanced generation to contract term generation and bias detection, or to automated contract revision through a multi-agent framework.

[0004] However, the existing methods mentioned above still have many shortcomings in contract negotiation scenarios.

[0005] First, decision-making methods are significantly limited. Rule-driven methods rely on pre-defined utility functions and cannot adapt to the dynamic changes in the other party's demands during negotiations. While hybrid-driven methods partially solve the problem of natural language interaction, their decision-making logic is still constrained by pre-defined rules, lacking flexibility and strategic depth. Although methods directly driven by large language models possess flexible semantic understanding capabilities, they lack systematic planning for balancing long-term goals and issues.

[0006] Second, cross-round information processing primarily relies on simple context splicing. Existing methods in multi-round negotiations mainly depend on splicing historical dialogues into the current context window. However, negotiations involve multiple rounds of interaction, and the dialogue history grows rapidly, easily exceeding the context window limit of large language models. Even within the window, early key information is diluted due to the inherent positional bias of large language models—the "Lost in the Middle" effect. Existing methods lack a hierarchical understanding and structured extraction of interaction history, leading to the easy forgetting or dilution of early key information.

[0007] Third, the reliance on experience, primarily based on similarity retrieval, lacks inductive and generalization capabilities. Existing methods mainly rely on explicit feature matching to retrieve similar samples from a case library for reference. Essentially, they depend on the superficial similarity between cases, rather than learning and abstracting the underlying patterns of experience. When facing negotiating partners whose demands lack highly similar samples in historical cases, the effectiveness of the retrieval results is difficult to guarantee. There is a significant discrepancy between the initial proposal and the other party's expectations, requiring multiple rounds of exploratory interactions to gradually narrow the differences.

[0008] Fourth, existing methods rely solely on historical records pieced together from the current input and context for each round of decision-making. They lack mechanisms to proactively identify information patterns from multiple rounds of interaction, resulting in a lack of multi-path comparison and filtering in decision-making. Consequently, the output quality is highly dependent on the randomness of a single inference. The other party's feedback in multiple rounds of negotiation often exhibits phenomena such as repetition, gradual correction, and shifting focus. This contains structured information about their core demands and acceptable boundaries, but existing methods struggle to extract comprehensive judgments that can support strategic decision-making from the scattered natural language feedback across multiple rounds.

[0009] Furthermore, existing systems lack an effective closed-loop mechanism for experience accumulation—while successful negotiation experiences may be stored in a case library for later retrieval, they are not used to continuously optimize the predictive capabilities of the decision-making model itself. The performance of the decision-making model remains relatively fixed after deployment and cannot continuously improve with increased usage.

[0010] In view of the above, this application is hereby submitted. Summary of the Invention

[0011] This invention provides an experience-driven method and apparatus for constructing multi-agent negotiation terms contracts, which can at least partially improve the above-mentioned problems.

[0012] To achieve the above objectives, the present invention adopts the following technical solution:

[0013] An experience-driven approach to constructing multi-agent negotiation terms contracts includes: Obtain the user's initial request description, call the pre-trained experience prediction model to predict the initial request description, and obtain a complete initial terms and conditions scheme. The multi-agent negotiation support module is used to negotiate the initial terms and conditions, generating pending solutions. The multi-agent negotiation support module includes a decision agent module and a coordination agent module. The decision agent module is used to generate candidate solutions that respond to the user's adjustment intentions, and the coordination agent module is used to parse the user's feedback information, evaluate the candidate solutions generated by the decision agent module, and control the progress of the negotiation process. A final validity check is performed on the proposed solution. When the check is deemed successful, a final consensus clause proposal is generated and added to the experience base.

[0014] Preferably, the user's initial request description is obtained, and a pre-trained experience prediction model is used to predict and process the initial request description to obtain a complete initial terms and conditions scheme, specifically as follows: Obtain the user's initial request description and the negotiation dimension list configured according to the current negotiation scenario. Based on the request parsing prompts and the negotiation dimension list, guide the preset large language model to extract the relative importance weights of each dimension from the initial request description. Based on relative importance weights, the initial request descriptions are transformed into standardized text with uniform format and clearly defined fields; The BERT encoder is used to map the canonical text to obtain a high-dimensional semantic vector, and the prediction layer of the empirical prediction model is used to process the high-dimensional semantic vector to obtain a multi-dimensional vector. According to the preset dimension mapping rules, the multidimensional vector is converted into specific clause values ​​and filled into the placeholders corresponding to the preset initial template to generate a complete initial clause scheme.

[0015] Preferably, the multi-agent negotiation support module is used to negotiate the initial terms and conditions, generating a pending solution, specifically: For the t-th consultation, obtain the latest feedback information from the user and determine whether all consultation dimensions in the latest user feedback information remain unchanged; If so, stop the negotiations and treat the best terms from the (t-1)th negotiation as a pending option; If not, the coordination agent module is called to parse and process the latest feedback information from the user to obtain the parsing result of the t-th negotiation. The parsing process includes: matching modifiers in the latest feedback information from the user based on a predefined sentiment lexicon and degree adverb dictionary, using the implicit reasoning ability of the big language model in combination with contextual semantics to judge the urgency of the other party's request, and normalizing the intensity into three discrete levels: high, medium, and low. The decision-making agent module is invoked to process the current terms and conditions, negotiation strategy, and the parsing result of the t-th negotiation, generate the optimal terms and conditions for the t-th negotiation, and send it to the user. When t equals 1, the current terms and conditions are the initial terms and conditions. When t does not equal 1, the current terms and conditions are the optimal terms and conditions for the (t-1)-th negotiation. Repeat the above steps until a solution is obtained.

[0016] Preferably, the decision-making agent module is invoked to process the current terms and conditions, negotiation strategy, and the analysis results of the t-th negotiation, and to generate the optimal terms and conditions for the t-th negotiation. Specifically: Based on the current terms and conditions, the negotiation strategy, and the analysis results of the t-th negotiation, a candidate set containing multiple candidate solutions is generated according to the core principles and decision constraints defined by the third decision prompt of the decision agent module. A pre-defined multi-dimensional comprehensive evaluation model is used to comprehensively score each candidate solution in the candidate set, and the control coordination agent module selects the candidate solution with the highest comprehensive score as the optimal terms solution for the t-th negotiation.

[0017] Preferred options also include: The number of negotiations is recorded using a periodic summary start value S, which is initially 0. Each time a negotiation is conducted, the periodic summary start value S is incremented by 1. When it is determined that the periodic summary initiation value S is equal to the periodic summary initiation threshold, the periodic summary initiation value S is reassigned to 0. The decision agent module triggers a periodic summary mechanism. Based on the second decision prompt of the decision agent module, the sequence of the other party's adjustment intentions in the previous periodic summary initiation threshold negotiations is analyzed, and the negotiation dimensions that the other party continues to focus on are identified, and the negotiation strategy is updated.

[0018] Preferably, a final validity check is performed on the proposed solution. When the check is deemed successful, a final consensus clause proposal is generated and added to the experience base. Specifically: The proposed solutions are filtered to exclude those that have reached a consensus in form but fail to meet the basic constraints of either party in substance. After the filtering process, the remaining pending solutions are recorded and standardized in format to generate the final consensus terms. The final consensus terms will be added to the experience base as a new negotiation sample.

[0019] When the number of new negotiation samples in the experience base reaches the preset update threshold, the experience prediction model is updated. The current training set is constructed based on the preset update threshold of the experience base, the number of newly added consultation samples, and a subset randomly sampled from the historical experience base. The experience prediction model is then updated based on the current training set.

[0020] The present invention also provides an experience-driven multi-agent negotiation terms contract construction apparatus, comprising: The initial terms and conditions generation unit is used to obtain the user's initial request description, call the pre-trained experience prediction model to predict the initial request description, and obtain the complete initial terms and conditions. The multi-agent negotiation unit is used to negotiate the initial terms and conditions using the multi-agent negotiation support module and generate pending solutions. The multi-agent negotiation support module includes a decision agent module and a coordination agent module. The decision agent module is used to generate candidate solutions in response to the user's adjustment intentions, and the coordination agent module is used to parse the user's feedback information, evaluate the candidate solutions generated by the decision agent module, and control the progress of the negotiation process. The final consensus clause generation unit is used to perform a final validity check on the proposed clause. When the check is passed, the final consensus clause is generated and added to the experience base.

[0021] In summary, the core of this method lies in mapping the user's initial demand description into personalized initial terms and conditions through an experience-based prediction model, serving as the starting point for multiple rounds of negotiation. During these rounds, a dual-agent architecture is employed, with a decision agent and a coordinating agent working collaboratively. The decision agent is responsible for generating multiple candidate solutions in response to the user's adjustment intentions, while the coordinating agent is responsible for fine-grained intent analysis of user feedback, comprehensive evaluation of each candidate solution, and selection of the optimal solution to present to the user. Simultaneously, a periodic summary mechanism is used to identify the user's continuous focus dimensions across rounds of interaction to dynamically adjust the concession strategy. Once a consensus is reached, the validity of the proposed solution is checked, and the verified final consensus terms and conditions are stored in an experience base for subsequent incremental updates to the experience-based prediction model.

[0022] Compared with existing technologies, this invention has the following beneficial effects: 1. It realizes a paradigm shift from experience retrieval to pattern prediction, so that the generation of initial solutions is no longer limited by the coverage of the case library; 2. It realizes the comprehensive utilization of cross-round interactive information through a periodic summary mechanism, giving multi-round consultations clear direction and coherence; 3. Through the multi-candidate solution generation and comprehensive evaluation mechanism, it effectively avoids the instability of single-round reasoning and significantly improves the reliability of consultation decisions; 4. Through the continuous accumulation of experience base and the incremental update of prediction model, it constructs a continuous evolutionary closed loop of "experience-model-experience", so that consultation capabilities continuously improve with the number of uses; 5. It provides efficient consensus-building assistance for both parties in consultation, significantly reducing the number of exploratory interaction rounds, while retaining the human right to final confirmation of the contract content. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the experience-driven multi-agent negotiation terms contract construction method provided in the first embodiment of the present invention.

[0024] Figure 2 This is a flowchart of the experience-driven multi-agent negotiation terms contract construction method provided in the first embodiment of the present invention.

[0025] Figure 3 This is a schematic diagram of a module of an experience-driven multi-agent negotiation terms contract construction device provided in the second embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0027] refer to Figure 1 , Figure 2 As shown, the first embodiment of the present invention discloses an experience-driven multi-agent negotiation terms contract construction method, which can be executed by an experience-driven multi-agent negotiation terms contract construction device (hereinafter referred to as the construction device), specifically, by one or more processors within the construction device, to implement the following method: S1, Obtain the user's initial request description and call the pre-trained empirical prediction model M. ED The initial demand description is processed to obtain a complete initial clause scheme; Specifically, step S1 further includes: obtaining the user's initial request description P counterparty Based on the negotiation dimension list configured in the current negotiation scenario, and based on the request parsing prompts and the negotiation dimension list, guide the preset large language model to extract the relative importance weights of each dimension from the initial request description; Based on relative importance weights, the initial request descriptions are transformed into standardized text with uniform format and clearly defined fields; The BERT encoder is used to map the canonical text to obtain a high-dimensional semantic vector. This high-dimensional semantic vector is then processed by the prediction layer of an empirical prediction model to obtain a multi-dimensional vector X. init ; According to the preset dimension mapping rules, the multidimensional vector is converted into specific clause values, and these values ​​are filled into the placeholders corresponding to the preset initial template to generate a complete initial clause scheme S. init .

[0028] In this embodiment, when a new contract negotiation process is initiated, the system first obtains the user's initial request description in natural language. This request description is a free-text expression of the user's needs, characterized by high unstructured nature and individual variability. To transform this unstructured, diverse free text into a data format that can be stably processed by the model, the system obtains a pre-configured list of negotiation dimensions based on the current negotiation scenario. This list of negotiation dimensions includes various clause dimensions involved in the current contract type. For example, in a goods procurement negotiation scenario, it can be configured as dimensions such as price, delivery cycle, quality standards, and payment terms; in a service contract negotiation scenario, it can be configured as dimensions such as service scope, service level, and charging method. Domain experts can flexibly define the applicable dimension system according to the actual business scenario, and the conversion rules for each dimension are adjusted accordingly.

[0029] Based on the negotiation dimension list and pre-defined request parsing prompts, the system guides a pre-defined large language model to extract the relative importance weights of each dimension from the user's initial request description. Specifically, the request parsing prompts guide the large language model to extract the relative importance weights of each dimension from the user's description, that is, the relative emphasis of each dimension in the user's decision-making. Through this step, the system transforms the original free text request description into a structured expression containing the weight information of each dimension, enabling request information from different users and with different expression methods to be stably encoded and compared within a unified representation framework.

[0030] Subsequently, based on the extracted relative importance weights, the system transforms the initial demand description into a standardized text with a unified format and clearly defined fields. This standardized text, within a unified data framework, structurally expresses the demands and weights of each dimension, eliminating individual differences in expression among different users and ensuring that demand information from different sources can be stably processed under the same representation structure. After obtaining the standardized text, the system uses a BERT encoder to semantically encode it. The BERT encoder maps the semantic information in the standardized text into a high-dimensional semantic vector, which comprehensively expresses the user's demand tendencies and emphases in each negotiation dimension. Specifically, the BERT encoder takes the pooled output of the [CLS] flag as the overall demand semantic vector. Next, the prediction layer of the empirical prediction model performs regression processing on this high-dimensional semantic vector, outputting a multi-dimensional vector. The number of dimensions in this multi-dimensional vector is consistent with the number of negotiation dimensions, and the values ​​of each dimension are all within the range [0,1], reflecting the ideal values ​​of the predicted terms for the current user's demand in each dimension.

[0031] Finally, the system converts the [0,1] values ​​of each dimension in the multi-dimensional vector into specific clause values ​​that conform to business semantics, according to the preset dimension mapping rules for each negotiation dimension. Specifically, for numerical dimensions such as price and delivery cycle, the [0,1] values ​​are converted into actual values ​​based on the preset numerical range of that dimension; for non-numerical dimensions with a clear hierarchical order, such as quality standards, the [0,1] values ​​are mapped to the corresponding level descriptions according to the level list configured for that dimension. After the conversion is completed, the system fills these specific clause values ​​into the placeholders corresponding to the preset initial template, generating a complete initial clause scheme to present to the user.

[0032] Specifically, in this embodiment, to address the problems of existing methods relying solely on explicit feature matching for case retrieval, lacking the ability to summarize decision-making patterns from historical experience, and resulting in initial clause proposals failing to effectively align with the other party's demands, this invention designs an experience-based prediction model (denoted as M) within the contract negotiation agent. ED This model uses the BERT pre-trained language model as its core architecture and learns the mapping relationship from the other party's demand description to the successful contract terms by fine-tuning it based on historical successful contract experience.

[0033] The demands expressed by the other party in natural language are highly unstructured and exhibit significant individual variability. The same demand may be expressed in drastically different ways (e.g., "We hope for a more favorable price" versus "Our budget is limited, so we need to negotiate the price further"). This diversity in expression can interfere with BERT's stable understanding of the demand semantics, thus affecting the reliability of predictions. To address this issue, this invention designs a demand parsing module as an intermediate layer. Leveraging a large language model and prompt words, it transforms the free-text demand from the other party into a structured, standardized text representation. Specifically, the system first obtains a list of negotiation dimensions configured according to the current negotiation scenario (for example, in a goods procurement negotiation scenario, this could be configured as price, delivery cycle, quality standards, payment terms, etc.). The demand parsing prompt words guide the large language model to extract the relative importance weights (i.e., the relative emphasis of each dimension in the other party's decision-making) of each dimension from the other party's description (i.e., the initial demand description). This transforms the original free text (i.e., the initial demand description) into standardized text with a unified format and clearly defined fields, enabling demand information from different parties and expressed in different ways to be stably encoded and compared by BERT within a unified representation framework.

[0034] The structured request (i.e., canonical text) is represented by a BERT encoder, which maps it into a high-dimensional semantic vector. The prediction layer then outputs a multi-dimensional vector X. init The value is a continuous numerical value in the range [0,1] for each negotiation dimension. In the specific implementation, the BERT encoder takes the pooled output (768 dimensions) of the [CLS] marker as the overall appeal semantic vector, followed by a two-layer fully connected regression head, with the dimension changing from 768→256→n (where n is the number of negotiation dimensions). The output layer uses the Sigmoid activation function to ensure that the predicted values ​​of each dimension are compressed within the range [0,1]. During the model training phase, the terms (target values) in historical successful contracts are transformed into numerical vectors within the range [0,1] according to the preset transformation rules for each dimension: for numerical dimensions such as price and delivery cycle, the specific values ​​are normalized to the range [0,1] according to the preset numerical range of the dimension; for non-numerical dimensions with a clear order of grade such as quality standards, they are mapped sequentially to equally spaced positions within the range [0,1] according to the grade list configured for the dimension. During training, preprocessed data (structured claim representation, normalized clause schemes) is batch-input into the model. After BERT encoding and forward propagation by the regression head, the predicted scheme vector is obtained. The mean squared error loss between the predicted scheme vector and the target scheme vector is calculated. Then, backpropagation is performed through the Adam optimizer to update all parameters of the BERT encoder and regression head. The learning rate is set to 1×10. -5The batch size was set to 16, and the maximum sequence length was set to 192 tokens. Each training round iterated through all training data, for a total of 100 rounds. During training, the loss was monitored on the validation set, and the model parameters with the lowest validation loss were selected as the final training result and saved. Through this training process, the model learns the implicit mapping pattern between demands and solutions from a large amount of historical experience.

[0035] The system pre-sets a contract template containing placeholders for each negotiation dimension. Each dimension's value range and mapping rules are pre-configured by domain experts (e.g., price range of 0-200 yuan, quality standard level list, etc.). The system will then... init The [0,1] values ​​of each dimension are converted into specific clause values ​​according to the mapping rules of the corresponding dimensions (e.g., 0.60 in the price dimension maps to 120 yuan, and 1.00 in the quality standard dimension maps to "superior quality"), and filled into the corresponding placeholders in the template to generate a complete initial clause scheme S. init Presented to the user.

[0036] Unlike existing methods that rely on similar case retrieval, the M of this invention ED By learning the demand-solution mapping patterns inherent in a large amount of historical experience, the generation of initial clause solutions is no longer limited by the coverage of the historical case library. Even if the current demand pattern lacks highly similar precedents in historical cases, the model can still output reasonable initial predictions based on the learned mapping relationships, thereby effectively improving the generalization ability of experience utilization. Through continuous expansion of the experience library and incremental updates to the model, M ED It can provide a high-quality initial terms that are more aligned with the other party's demands for subsequent multi-agency negotiations, significantly narrowing the initial position gap between the two parties, reducing ineffective exploratory rounds of interaction, and laying an efficient foundation for the entire negotiation process.

[0037] S2, the initial terms and conditions are negotiated using the multi-agent negotiation support module to generate pending solutions. This multi-agent negotiation support module includes a decision-making agent module A. D Coordination Agent Module A C The decision agent module is used to generate candidate solutions in response to the user's adjustment intentions, while the coordination agent module is used to parse the user's feedback information, evaluate the candidate solutions generated by the decision agent module, and control the progress of the consultation process. Specifically, step S2 further includes: for the t-th consultation, obtaining the latest feedback information from the user, and determining whether all consultation dimensions of the latest user feedback information remain unchanged; If so, stop the negotiations and treat the best terms from the (t-1)th negotiation as a pending option; If not, the coordination agent module is invoked to parse and process the latest user feedback information to obtain the parsing result I of the t-th negotiation. t The parsing process includes: matching modifiers in the latest user feedback information based on a predefined sentiment lexicon and degree adverb dictionary; using the implicit reasoning ability of the big language model in conjunction with contextual semantics to judge the urgency of the other party's request; and normalizing the intensity into three discrete levels: high, medium, and low. The decision-making agent module is invoked to process the current terms and conditions, negotiation strategy, and the parsing result of the t-th negotiation, generate the optimal terms and conditions for the t-th negotiation, and send it to the user. When t equals 1, the current terms and conditions are the initial terms and conditions. When t does not equal 1, the current terms and conditions are the optimal terms and conditions for the (t-1)-th negotiation. Repeat the above steps until a solution is obtained.

[0038] The decision-making agent module is invoked to process the current terms and conditions, negotiation strategy, and the analysis results of the t-th negotiation, and to generate the optimal terms and conditions for the t-th negotiation. Specifically: Based on the current terms and conditions, the negotiation strategy, and the analysis results of the t-th negotiation, a candidate set containing multiple candidate solutions is generated according to the core principles and decision constraints defined by the third decision prompt of the decision agent module. A pre-defined multi-dimensional comprehensive evaluation model is used to comprehensively score each candidate solution in the candidate set, and the control coordination agent module selects the candidate solution with the highest comprehensive score as the optimal terms solution for the t-th negotiation.

[0039] It also includes: using a periodic summary start value S to record the number of consultations, with an initial value of 0, wherein the periodic summary start value S is incremented by 1 for each consultation; When it is determined that the periodic summary initiation value S equals the periodic summary initiation threshold S int At that time, the periodic summary start value S will be reassigned to 0; The decision agent module triggers a periodic summary mechanism. Based on the second decision prompt of the decision agent module, the sequence of the other party's adjustment intentions in the previous periodic summary initiation threshold negotiations is analyzed, and the negotiation dimensions that the other party continues to focus on are identified, and the negotiation strategy is updated.

[0040] In this embodiment, after generating the initial terms scheme in step S1, the system activates the multi-agent negotiation support module to advance the multi-round negotiation process through the collaborative work of decision agents and coordination agents. The system uses a periodic summary initiation value S to record the number of negotiations, with an initial value of 0. Each time a negotiation is conducted, the system increments the periodic summary initiation value S by 1. When the periodic summary initiation value S equals a preset periodic summary initiation threshold, the periodic summary initiation value S is reset to 0, and the periodic summary mechanism is triggered.

[0041] For the t-th negotiation, the system first obtains the user's latest feedback information and determines whether all negotiation dimensions in the user's latest feedback information remain unchanged. If the adjustment intentions of all dimensions remain unchanged, the system determines that a consensus has been reached, stops the negotiation, and outputs the optimal terms from the (t-1)-th negotiation as a pending solution. If the adjustment intention of any dimension changes, the system continues the negotiation.

[0042] In the absence of consensus, the system invokes the coordination agent module to parse and process the latest user feedback, obtaining the parsing result of the t-th negotiation. Specifically, the coordination agent module matches modifiers in the latest user feedback based on a predefined sentiment lexicon and degree adverb dictionary. For example, words like "must" and "very" are mapped to high intensity, while words like "slightly" and "somewhat" are mapped to low intensity. Simultaneously, the coordination agent module combines contextual semantics with the implicit reasoning capabilities of a large language model to determine the urgency of the other party's request, normalizing the intensity into three discrete levels: high, medium, and low. This intensity level is used to guide the decision agent module in generating candidate solutions with differentiated magnitudes.

[0043] Subsequently, the system invokes the decision-making agent module to process the current terms and conditions, negotiation strategy, and the analysis result of the t-th negotiation. Specifically, when t equals 1, the current terms and conditions are the initial terms and conditions generated in step S1; when t does not equal 1, the current terms and conditions are the optimal terms and conditions obtained in the (t-1)-th negotiation.

[0044] During the decision-making process, the decision-making agent module first generates a candidate set containing multiple candidate solutions based on the current terms and conditions, negotiation strategy, and the analysis results of the t-th negotiation, according to the core principles and decision constraints defined in the third decision prompt of the decision-making agent module. These core principles and decision constraints include hard commercial limitations such as the client's cost floor, delivery capability, and minimum quality standards, ensuring that the generated candidate solutions do not deviate from the client's acceptable range. The decision-making agent module generates differentiated candidate solutions by adjusting the focus and the extent of concessions. Some solutions strictly adhere to the established negotiation strategy, some slightly explore non-core dimensions, and some respond to high-intensity demands by moderately breaking some constraints, thus maintaining the ability to explore potential consensus space within a strategy-oriented framework.

[0045] After candidate solutions are generated, the system uses a pre-set multi-dimensional comprehensive evaluation model to comprehensively score each candidate solution in the candidate set. The evaluation model assesses each candidate solution from dimensions such as demand matching degree, business feasibility, and strategy consistency. Demand matching degree measures whether the candidate solution matches the user's expressed adjustment intention well; business feasibility measures whether the various adjustments in the candidate solution are in line with business logic and do not deviate from the acceptable range of the user; and strategy consistency measures whether the candidate solution follows the established negotiation strategy. By comprehensively evaluating the results of each dimension, the coordination agent module selects the candidate solution with the highest comprehensive score as the optimal terms for the t-th negotiation and sends it to the user. This mechanism, through the generation and comparison of multiple candidate solutions, effectively avoids the quality instability problem of single-path decision-making and significantly improves the reliability of negotiation decisions.

[0046] Furthermore, when the periodic summary initiation value S equals the periodic summary initiation threshold, the system triggers the periodic summary mechanism using the decision agent module. Specifically, based on its second decision prompt, the decision agent module analyzes the sequence of the other party's adjustment intentions in the previous rounds of negotiations before the periodic summary initiation threshold, identifying the negotiation dimensions that the user consistently focuses on in multiple rounds of interaction—the more times the same dimension is mentioned, the higher the average adjustment intensity, and the stronger the persistence of recurrence in multiple rounds, the greater the likelihood that the dimension is judged as a dimension of continuous focus. Based on this analysis, the decision agent module updates the negotiation strategy, giving higher adjustment weights to dimensions that the user consistently focuses on, while adopting a limited concession strategy for dimensions with low focus. Through this mechanism, the system can extract comprehensive judgments that can support strategy decisions from information scattered across multiple rounds of interaction, giving multi-round negotiations a clear direction and coherence, helping both parties to focus on key points of disagreement more quickly, and improving the efficiency of consensus-building. In this way, the system repeats the above steps, that is, after obtaining user feedback each time, it performs intent analysis to determine whether a consensus has been reached. If no consensus has been reached, it generates candidate solutions, evaluates and filters them, and presents the optimal solution until the adjustment intent of all dimensions remains unchanged, then stops the negotiation and takes the optimal terms solution of the (t-1)th negotiation as the pending solution. Specifically, after the empirical prediction model generates the initial terms and conditions, the contract negotiation agent initiates a multi-agent negotiation support mechanism. This mechanism establishes two collaborative modules within the contract negotiation agent: a decision-making agent and a coordination agent. These two modules work together through multiple iterations to adjust the terms and conditions until a consensus is reached. This dual-agent structure is adopted because of the differences in functional objectives between terms generation and negotiation process management. Terms generation focuses on producing reasonable terms and conditions that align with our interests under given constraints, while negotiation management focuses on understanding the other party's intentions, evaluating the quality of the terms and conditions, and controlling the interaction flow. Assigning these two functions to different agents, and instantiating them using an LLM with complementary capabilities, allows the model to leverage its strengths in both generation and inference dimensions, while enhancing the structural clarity of the negotiation mechanism. In this architecture, the roles and behaviors of the agents are determined by task allocation and internally predefined prompts. The technical details of each sub-module are described below according to the interaction flow: To address the challenge of existing methods failing to identify stable demands from the other party's natural language feedback during cross-round interactions, this invention designs a fine-grained intent parsing mechanism executed by a coordinating agent within the contract negotiation agent. The core of this mechanism lies in not only identifying the other party's demands for each negotiation dimension (a i For the i-th negotiation dimension, the adjustment direction (increase / decrease / remain unchanged) is further inferred and quantified from the semantic and emotional intensity of the text to determine discrete adjustment intensity levels (e.g., "high," "medium," "low"). Specifically, the first prompt P1 of the coordination agent module requires A... C For each dimension, the following steps are performed: First, based on a predefined sentiment lexicon and a dictionary of degree adverbs (e.g., "must" and "very" are mapped to high intensity, and "slightly" and "somewhat" are mapped to low intensity), modifiers in the other party's feedback are matched; second, combined with contextual semantics (e.g., "the price must be lowered further" versus "it would be great if the price could be more favorable"), the implicit reasoning ability of LLM is used to determine the urgency of the other party's request; finally, the intensity is normalized into three discrete levels: high (urgent and strong demand), medium (demanding but negotiable), and low (slight inclination or not strong). This parsing process is as follows: As shown, I t For the analysis result of the t-th negotiation, F t This represents the latest user feedback information obtained during the t-th consultation.

[0047] In this invention, the intensity level is calculated independently during the intent resolution process, and its design purpose is to inform subsequent decision-making by agent A. D The generation of quantified concession levels provides a basis: high intensity corresponds to larger concession levels (i.e., significant adjustments to the corresponding dimension in the generated candidate solutions), while low intensity corresponds to smaller concession levels. For example, if the other party expresses high intensity dissatisfaction with the "price" dimension, A... DThe price index will be significantly reduced in the next round of candidate solutions; if the dissatisfaction is only low-intensity, only minor adjustments will be made. This "intensity → magnitude" mapping rule is a core component of the negotiation strategy. Through the above design, this invention transforms unstructured counterpart feedback into structured, quantifiable strategy-driven parameters, providing key inputs for subsequent precise and differentiated negotiation decisions.

[0048] Based on the above analysis results, the contract negotiation agent executes the consensus decision if and only if all a i When all parties' adjustment intentions remain "unchanged," the negotiation stops. At this point, a verification step will be performed to conduct a final validity check on the current solution, confirming whether it is within the preset commercially reasonable and compliance boundaries. If the check passes, the relevant data will be structured and stored in the contract consensus experience base E; otherwise, this round of consensus will be marked as invalid and will not be stored in the experience base. If no consensus is reached, the other party's adjustment intention will be... The proposed solutions provide key guidance, enabling them to develop more targeted concession strategies and drive the negotiations toward consensus.

[0049] When the consultation dimensions are not all constant, to address the issues of quality instability and lack of systematic guarantees in existing single-path decision-making methods, this invention designs a closed-loop decision-making mechanism of "generation-evaluation-screening." In the solution generation phase, the decision agent A... D It does not generate just one solution, but rather a solution based on the current solution (in natural language text form) and the negotiation strategy. And the structured intent (i.e., the parsing result I of the t-th negotiation) t Based on the core principles and decision constraints defined by the third decision prompt P3 of the decision agent module, N is generated. cand A set of N candidate solutions (N) cand (This refers to the number of candidate solutions in the candidate set). Decision constraints include hard business limitations such as the company's cost floor, delivery capability, and minimum quality standards, ensuring that the generated candidate solutions do not deviate from the company's acceptable range. Unlike general random sampling, the candidate generation process of this invention explicitly utilizes... and I t To guide diversity: P3 requires A D Candidate options that differentiate themselves by adjusting focus and the extent of concessions; for example, some options may strictly adhere to Some may tentatively explore non-core dimensions, while others may break through certain constraints in response to strong demands. This controlled diversity allows candidate solutions to still possess the ability to explore potential consensus space within a strategy-oriented framework. This process is as follows: As shown, S t For decision-making agent module A DThe "current solution" used when generating candidate solutions in round t (i.e., the t-th consultation).

[0050] When the negotiation strategy is generated through a periodic summary mechanism, it will serve as a key input to guide the generation of the scheme; otherwise, its value will be empty. At this point, the diversity of solutions is generated solely based on the other party's intent. Subsequently, in the solution evaluation phase, the coordination agent module does not simply sort by preference, but instead employs a multi-dimensional comprehensive evaluation model for each candidate solution. A comprehensive score is given, such as As shown, D t Let d be the candidate set for the t-th consultation. i Let i be the i-th candidate solution in the candidate set.

[0051] This assessment model is guided by the fourth cue, P4, and considers the following three dimensions: the degree of relevance to the request, i.e., d i Does it align well with the adjustment intentions just expressed by the other party? Commercial feasibility, i.e., d i Do the various adjustments in the process conform to business logic, and are the reasons for the adjustments self-consistent? Is the strategy consistent, i.e., consistent with the already generated... In the case of d i Whether the strategy was followed (e.g., making concessions on high-attention dimensions and maintaining low-attention dimensions). When In this case, the strategy consistency dimension is not included in the scoring (i.e., its weight is zero), and the overall score is determined solely by the relevance of the appeal and commercial feasibility. By integrating these evaluation results, A... C Select the optimal terms and conditions, such as As shown, This is the optimal terms and conditions.

[0052] Ultimately, the coordination agent module will This is presented to the other party, forming a complete decision-making loop. The other party then interprets the received information based on their own perspective. Generate feedback F t+1 The feedback was passed to A C A new round of intent analysis is conducted to initiate the next round of negotiations or terminate the entire cycle. When negotiations terminate, the framework outputs a consensus state and the final terms S. final This information, once confirmed by both parties, serves as the basis for the formal contract terms. This "generation-evaluation-screening" mechanism integrates high-level strategies and fine-grained intentions into the candidate generation and screening process, achieving a structured guarantee of the quality of consultation and decision-making.

[0053] Furthermore, in this embodiment, to address the problem that existing methods primarily rely on single-wheel drive and lack the ability to comprehensively utilize information across multiple rounds, this invention designs a periodic summary mechanism in the decision-making agent of the contract negotiation intelligent agent. Each time the negotiation has gone through S... intAfter the wheel, A D This mechanism will be triggered, and under the guidance of the second prompt P2, a deep analysis of near S will be performed. int The sequence of adjustments made by the opposing party during the rounds of negotiations identifies the negotiation dimensions that the opposing party continues to focus on. Based on this in-depth analysis, A D The system dynamically generates corresponding negotiation strategies. The core of this strategy is to output a differentiated weight vector, giving higher adjustment weights to the "negotiation dimensions that the other party is continuously focusing on," while adopting a limited concession strategy (i.e., maintaining the original terms or making only minor adjustments) for dimensions with low focus. This process is as follows: As shown.

[0054] The determination of "continuous attention" does not rely on a preset counting threshold. Instead, the decision agent, guided by prompt P2, performs semantic reasoning by comprehensively considering the frequency of mentions of each dimension within the window, the level of adjustment intensity, and the continuity across rounds: the more times the same dimension is mentioned, the higher the average adjustment intensity, and the stronger the continuity of repeated appearances in multiple rounds, the greater the likelihood that the dimension will be judged as continuous attention.

[0055] Taking a goods procurement negotiation scenario as an example, let the negotiation dimensions be price, delivery cycle, quality standards, and payment terms, S int =5. The feedback and adjustment intentions from the other party in the last 5 rounds are as follows: In the first round, the other party proposed "a more favorable price and a shorter delivery cycle," which was interpreted as a price reduction (medium intensity) and a shorter delivery cycle (medium intensity); In the second round, "the price is still too high, and the quality standard must be guaranteed," which was interpreted as a price reduction (high intensity) and no change in quality standards; In the third round, "if the price is lowered further, the payment terms can be negotiated," which was interpreted as a price reduction (high intensity) and negotiable payment terms (low intensity); In the fourth round, "the main issue is the price," which was interpreted as a price reduction (high intensity); In the fifth round, "if the price is suitable, it would be best if the delivery cycle remained as planned," which was interpreted as a price reduction (high intensity). After analyzing the above sequence, the decision-making agent identified that the "price" dimension was mentioned in all 5 rounds and had a high intensity demand in multiple rounds, and was determined to be a dimension of continuous concern; the "delivery cycle" was mentioned only in the first and fifth rounds and its intensity was weakened, and was determined to be a dimension of non-continuous concern; the "quality standard" and "payment terms" were mentioned only in a single round and were determined to be dimensions of non-continuous concern. Based on this, A DThe output negotiation strategy is described in natural language, for example: "The other party has shown sustained and high-intensity attention to the price dimension in the last 5 rounds of negotiations, and subsequent solutions should make significant concessions in the price dimension; although the delivery cycle was mentioned, it was not consistently emphasized, and only minor adjustments were made; the quality standards and payment terms can be considered to remain unchanged." The expressions "significant concessions," "minor adjustments," and "maintaining the original solution" in the above natural language strategy are specific instructions on the degree of differentiated adjustment for different dimensions. The decision-making agent will implement the corresponding adjustments when generating subsequent candidate solutions, thereby achieving a differentiated strategy of "significant concessions for high-concession dimensions and limited adjustments for low-concession dimensions."

[0056] Unlike the general strategy reconstruction in existing methods, the periodic summary mechanism of this invention has two specific characteristics: First, its goal is not a generalized "belief calibration," but rather to explicitly identify the negotiation dimensions repeatedly mentioned by the other party in multiple rounds. This identification result is the direct basis for subsequent differentiated concessions. Second, its output is a quantifiable distribution of dimension weights, rather than a vague strategy direction. This allows the generation of subsequent candidate solutions to precisely execute "significant adjustments to high-priority dimensions and limited concessions to low-priority dimensions." Without periodic summaries, the contract negotiation agent can only react in isolation to each round of feedback from the other party, lacking long-term strategic planning and easily leading to negotiations getting bogged down in local adjustments and deviating from the overall goal. With the introduction of this mechanism, the system can dynamically identify changes in the negotiation focus and proactively adjust the concession strategy, thereby improving the efficiency of consensus-building and the cooperative benefits of the final solution.

[0057] S3 performs a final validity check on the proposed solution. When the check is passed, a final consensus clause scheme is generated and added to the experience base.

[0058] Specifically, step S3 further includes: filtering the proposed solutions to exclude those that have reached a consensus in form but fail to meet the basic constraints of either party in substance; After the filtering process, the remaining pending solutions are recorded and standardized in format to generate the final consensus terms. The final consensus terms will be added to the experience base as a new negotiation sample.

[0059] When the number of new negotiation samples in the experience base reaches the preset update threshold, the experience prediction model is updated. The current training set is constructed based on the preset update threshold of the experience base, the number of newly added consultation samples, and a subset randomly sampled from the historical experience base. The experience prediction model is then updated based on the current training set.

[0060] In this embodiment, after reaching a consensus and outputting pending solutions in step S2 through multiple rounds of consultation, the system performs a final validity check on the pending solutions. Specifically, the system filters the pending solutions to exclude those that have reached a consensus in form but fail to meet the basic constraints of either party in substance, such as reaching a price consensus but the price being lower than one party's cost floor. Through this filtering process, the system ensures that only solutions that truly meet the commercially reasonable scope and compliance boundaries of both parties can proceed to subsequent processing stages, thereby preventing low-quality data from polluting the experience base.

[0061] After filtering, the system records the remaining pending solutions and standardizes their format. Specifically, based on the dimensional configuration of the current negotiation scenario, the system converts the specific clause values ​​in the pending solutions into numerical vectors of a unified format according to preset transformation rules. Numerical dimensions are normalized according to preset numerical ranges, while hierarchical dimensions are mapped sequentially to their corresponding numerical positions according to a configured hierarchical list. After standardization, the system generates the final consensus clause scheme. Through the above filtering and standardization processes, the system transforms the original negotiation results into clean, structured training instances compatible with the input format of the prediction model.

[0062] Subsequently, the system adds the final consensus terms to the experience base as a new negotiation sample. The experience base uses persistent storage, allowing accumulated knowledge to be retained across sessions. In addition, the system stores metadata information for this negotiation to support subsequent quality analysis and cross-record comparisons. For records with the same or similar request configurations, the system retains only those with higher overall quality, ensuring that the experience base continuously stores high-quality results under each request profile.

[0063] When the system determines that the number of newly added negotiation samples in the experience base has reached a preset update threshold, it triggers an update of the experience prediction model. Specifically, the system constructs the current training set based on the preset update threshold number of newly added negotiation samples in the experience base and a subset randomly sampled from the historical experience base. Subsequently, the system incrementally updates the experience prediction model based on the current training set, enabling the model to adapt to the current new consensus mode while retaining previously learned knowledge, effectively mitigating the catastrophic forgetting problem. Through the above-mentioned experience evolution mechanism, this invention constructs a continuous evolutionary closed loop from experience to model and then from model to experience, enabling the decision-making ability of the contract negotiation agent to continuously improve with the accumulation of experience, and the quality of future initial solution predictions to continuously improve accordingly.

[0064] Specifically, once a valid consensus is reached through multi-agent negotiations, the negotiation results undergo a series of processing steps before being incorporated into the experience base for model updates. This processing flow transforms the original negotiation records into clean, structured training instances, making them compatible with the input format of the prediction model and suitable for incremental learning. The processing flow first filters the negotiation results. Not all consensus results meet the inclusion criteria; only records where the final terms simultaneously satisfy the basic business constraints and compliance requirements of both parties are retained. This criterion excludes cases where a consensus is reached in form but fails to meet the basic constraints of either party in substance (e.g., a price consensus is reached, but the price is below one party's cost floor). Failed negotiation cases and invalid results are discarded directly to prevent low-quality data from polluting the experience base.

[0065] The filtered records are then standardized in format. The opposing party's demands are saved in the same text format as the prediction model input (including the relative importance weights of each negotiation dimension), ensuring that the semantic content of the demands is fully preserved in subsequent model fine-tuning. The final consensus terms are automatically converted into a numerical vector (i.e., X) within the [0,1] interval by the system according to the dimension configuration of the current negotiation scenario and preset conversion rules. final For numerical dimensions such as price and delivery cycle, the specific values ​​are normalized to the [0,1] interval according to the preset numerical range of the dimension; for non-numerical dimensions with a clear order of priority, such as quality standards, the values ​​are mapped sequentially to equally spaced positions within the [0,1] interval according to the level list configured for that dimension. After the conversion, the complete record is ( , Yes, add it to the experience base: The experience base E employs persistent storage, enabling accumulated knowledge to be retained across sessions. Each record additionally stores metadata information about the negotiation (such as the time the negotiation was reached) to support quality analysis and cross-record comparisons. For records with the same or similar counterparty request configurations, only those with higher overall quality are retained, ensuring that the experience base continuously stores high-quality results under each request profile.

[0066] Furthermore, this method incorporates an experience evolution module within the contract negotiation agent, comprising two sub-modules: experience data processing and incremental model updating. This module continuously feeds back the consensus reached during negotiations to the experience prediction model, enabling the contract negotiation agent to continuously learn from experience and achieve self-enhancement.

[0067] To make M ED To continuously update and mitigate the catastrophic forgetting of learned knowledge, this invention employs a hybrid training strategy inspired by a replay mechanism. A model update is triggered after accumulating a certain number (e.g., 50) of new consultation samples. The current training set consists of newly added data and a subset randomly sampled from the historical experience database. ,in This represents newly accumulated training data. This represents historical data in the experience base. To control the proportion of historical sampling, this hybrid training strategy balances adaptation to new consensus patterns with retention of learned knowledge, mitigating the catastrophic forgetting problem without requiring access to the complete historical dataset. The value of is dynamically set based on the number of newly accumulated samples, ensuring that the number of historical records sampled in each update is roughly equivalent to the number of new samples. Each incremental update uses the same fine-tuning configuration (including learning rate, batch size, and number of training epochs) to maintain consistent optimization dynamics and avoid performance fluctuations caused by arbitrarily changing parameters.

[0068] Specifically, the training process in the incremental update stage remains consistent with the initial training in terms of data processing and model optimization. Each training sample is a data pair (structured claim description, normalized clause scheme), including newly added data. Data derived from recent successful negotiations and consensus records, and historical sampling data. Historical consensus records already existing in the experience base. During training, Samples are entered into M in batches. ED After BERT encoding and forward propagation with the regression head, the predicted scheme vector is obtained. The mean squared error loss between the predicted scheme vector and the target scheme vector is calculated. Then, backpropagation is performed through the Adam optimizer to update all parameters of the BERT encoder and regression head. Similar to the initial training, the learning rate is set to 1×10. -5 The batch size was set to 16, the maximum sequence length was set to 192 tokens, and the number of training rounds was set to 100. During training, the loss was monitored on the validation set, and the model parameters with the lowest validation loss were selected as the final result of this incremental update and saved.

[0069] As the experience base continues to expand, M ED By gradually learning a more stable mapping relationship between the other party's demands and the successful consensus terms, the quality of future initial proposal predictions also continuously improves. This dynamic adaptive mechanism enables the contract negotiation agent to respond to changes in the market environment or the evolution of business strategies.

[0070] It should be noted that this invention uses BERT as the basic architecture of the prediction model. In other embodiments, other pre-trained language models can be used instead, such as RoBERTa, ALBERT, or LegalBERT pre-trained in the legal field. These models also have the ability to map natural language into semantic vectors, requiring only corresponding adjustments to the model structure and training parameters. This invention uses a large language model in conjunction with prompt words to transform the other party's natural language request into a structured, standardized text representation. In other embodiments, other methods can also be used: for example, training a dedicated natural language understanding model (such as a BERT-based sequence labeling or text classification model) to complete the extraction and structuring of the request information; or using a combination of rule-based keyword matching and template filling to achieve basic request parsing. In specific embodiments, this invention uses price, delivery cycle, and quality standards as examples of negotiation dimensions. In other embodiments, however, the specific type and number of negotiation dimensions can be flexibly configured according to different contract types. For example, in service contract negotiations, dimensions such as service scope, service level, and charging method can be configured; in lease contract negotiations, dimensions such as rent and lease term can be configured. Domain experts can define applicable dimensional systems based on actual business scenarios, and the conversion rules for each dimension (numerical normalization methods, level mapping methods, etc.) will be adjusted accordingly without affecting the operation of the core mechanism of this invention. No limitations are made here, but all these solutions are within the protection scope of this invention.

[0071] In summary, this invention aims to address the following issues: First, how to generalize decision-making patterns from historical successful contracts and apply them to predict personalized initial terms. Existing methods can only retrieve similar samples from a case library based on explicit feature matching, essentially relying on superficial similarities between cases rather than learning and abstracting the underlying patterns. When facing negotiating partners whose demands lack highly similar samples in historical cases, the effectiveness of the retrieval results is difficult to guarantee, and there is a significant deviation between the initial proposal and the other party's expectations. Both parties need to go through multiple rounds of exploratory interactions to gradually narrow the differences. Second, how to comprehensively understand the other party's stable demands across multiple rounds of negotiations and make systematic strategy adjustments accordingly to alleviate the quality instability problem caused by single-path decision-making. Existing methods rely solely on the current input and historical records pieced together from the context for each round of decision-making, lacking a mechanism to actively identify information patterns from multiple rounds of interactions. Decision-making lacks multi-path comparison and filtering, resulting in output quality highly dependent on the randomness of a single inference. The other party's feedback in multiple rounds of consultations often involves repetition, gradual revision, and shifts in focus. This contains structured information about their core demands and acceptable boundaries, but existing methods struggle to extract comprehensive judgments that can support strategic decisions from the scattered natural language feedback across multiple rounds.

[0072] Specifically, this method involves an experience-driven contract negotiation agent, which comprises two collaborative modules: a decision-making agent and a coordination agent. These two modules work together to complete the entire process from the initial generation of the contract terms to multiple rounds of negotiation and adjustment. The workflow of this contract negotiation agent (i.e., the workflow of this method) is as follows: First, an empirical prediction model is trained using historical successful contract data to map the other party's demands, described in natural language, into personalized initial contract terms, serving as the starting point for negotiations. By learning the demand-term mapping patterns inherent in historical experience, this model ensures that the generation of initial solutions is no longer limited by the coverage of the historical case library. Even if the current demand pattern lacks highly similar precedents in historical cases, the model can still output reasonable initial predictions based on the learned mapping relationships.

[0073] After entering the multi-round consultation stage, the coordinating agent is responsible for fine-grained analysis of the other party's natural language feedback, identifying the direction of adjustment and the intensity of their demands for each contract issue dimension (distinguishing between different levels of urgency such as "must be adjusted" and "suggested adjustment"); the decision-making agent generates multiple candidate clause schemes based on the analysis results; the coordinating agent further conducts a comprehensive evaluation of the candidate schemes from dimensions such as demand matching degree, commercial feasibility and consistency of consultation strategy, and selects the optimal scheme to present to the other party.

[0074] Every few rounds, the decision-making agent will also periodically summarize the historical demand sequence, identify the contract issues that the other party has been consistently focused on in multiple rounds of interaction, and generate differentiated concession strategies accordingly—giving greater adjustment weight to the dimensions of sustained focus and only making minor adjustments to the dimensions of low focus, thereby guiding the focus of negotiations to concentrate on key points of disagreement.

[0075] Once a consensus is reached through consultation, the successful solution and the corresponding demands of the other party are stored in the experience base. The prediction model is periodically updated through an incremental learning strategy of mixed sample replay, enabling the contract negotiation agent to continuously learn from experience and enhance itself. Its decision-making level can be continuously improved with the increase of usage.

[0076] Please participate in the reading. Figure 2In simple terms, this contract negotiation agent consists of three interconnected modules: experience prediction, multi-agent negotiation support, and experience evolution. These correspond to the three core stages of negotiation initialization, negotiation advancement, and experience accumulation, respectively. The experience prediction module is responsible for summarizing decision-making patterns from historical successful contract experiences to generate personalized initial clause proposals. The multi-agent negotiation support module is responsible for advancing multiple rounds of negotiation, achieving solution generation, intent parsing, and strategy adjustment through the collaborative work of decision-making agents and coordinating agents. The experience evolution module is responsible for continuously feeding back the consensus solutions generated from successful negotiations into the prediction model, enabling the contract negotiation agent to continuously learn from experience and achieve self-enhancement.

[0077] Compared with existing technologies, this method has the following advantages: (1) A paradigm shift from experience retrieval to pattern prediction has been achieved: This invention learns the mapping pattern between claims and terms directly from historical successful contract data through an experience prediction model, so that the generation of initial solutions is no longer limited by the coverage of the historical case library. Even if the current claim pattern of the other party lacks highly similar precedents in historical cases, the model can still output a reasonable initial prediction based on the learned mapping relationship, which effectively improves the generalization ability of existing methods in the use of experience.

[0078] (2) Comprehensive utilization of cross-round interactive information is achieved, significantly improving the adaptability and coherence of multi-round negotiations: This invention, by designing a periodic summary mechanism in the decision-making agent, enables the contract negotiation agent to identify the other party's continuous focus dimensions from the information scattered across multiple rounds of interaction, and generate differentiated negotiation strategies accordingly. Compared with the passive response mode of existing methods that only rely on context splicing, the proactive strategy planning of this invention gives multi-round negotiations clear directionality and coherence.

[0079] (3) A high-quality multi-candidate decision-making assurance mechanism has been established: This invention adopts a multi-candidate solution generation and a three-dimensional comprehensive evaluation mechanism of "demand matching degree, commercial feasibility, and strategy consistency", integrating high-level consultation strategies and fine-grained intentions into the candidate solution generation and evaluation process. Compared with the single-path generation mode of existing methods, this invention effectively avoids the instability of single-model reasoning through the closed-loop mechanism of "generation-evaluation-screening", and significantly improves the reliability, rationality and consistency of consultation decisions.

[0080] (4) A continuous evolutionary closed loop of "experience-model-experience" is constructed: This invention continuously feeds back the consensus solution generated by successful negotiation to the prediction model through the experience evolution module, and adopts an incremental learning strategy of mixed sample replay to effectively alleviate catastrophic forgetting while absorbing new knowledge. This design enables the decision-making ability of the contract negotiation agent to be continuously enhanced with the accumulation of experience, overcoming the limitation of existing methods where experience is only used for retrieval reference and cannot be fed back to the decision model.

[0081] (5) Provides efficient consensus-building assistance for both parties in the negotiation: This invention is positioned to assist both parties in the negotiation process to efficiently advance the negotiation of contract terms, rather than to replace the final contract confirmation and signing. By generating personalized initial solutions through experience prediction models, recommending strategies in multiple rounds of negotiation, and evaluating multiple candidate solutions, the number of exploratory interaction rounds is significantly reduced, helping both parties to focus on key points of disagreement more quickly; the final output solution still needs to be manually confirmed by both parties before it can be signed and take effect, thus retaining the human final decision-making power and compliance review power over the contract content, which meets the objective requirement in commercial practice that the signing of contracts must be finally confirmed by the authorized party.

[0082] Please see Figure 3 A second embodiment of the present invention provides an experience-driven multi-agent negotiation terms contract construction apparatus, comprising: The initial terms and conditions generation unit 101 is used to obtain the user's initial demand description, call the pre-trained experience prediction model to predict the initial demand description, and obtain a complete initial terms and conditions scheme. The multi-agent negotiation unit 102 is used to negotiate the initial terms and conditions using the multi-agent negotiation support module to generate pending solutions. The multi-agent negotiation support module includes a decision agent module and a coordination agent module. The decision agent module is used to generate candidate solutions that respond to the user's adjustment intentions, and the coordination agent module is used to parse the user's feedback information, evaluate the candidate solutions generated by the decision agent module, and control the progress of the negotiation process. The final consensus clause scheme generation unit 103 is used to perform a final validity check on the proposed scheme. When the check is passed, the final consensus clause scheme is generated and added to the experience base.

[0083] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. An experience-driven method for constructing multi-agent negotiation clause contracts, characterized in that, include: Obtain the user's initial request description, call the pre-trained experience prediction model to predict the initial request description, and obtain a complete initial terms and conditions scheme. The multi-agent negotiation support module is used to negotiate the initial terms and conditions, generating pending solutions. The multi-agent negotiation support module includes a decision agent module and a coordination agent module. The decision agent module is used to generate candidate solutions that respond to the user's adjustment intentions, and the coordination agent module is used to parse the user's feedback information, evaluate the candidate solutions generated by the decision agent module, and control the progress of the negotiation process. A final validity check is performed on the proposed solution. When the check is deemed successful, a final consensus clause proposal is generated and added to the experience base.

2. The experience-driven multi-agent negotiation clause contract construction method according to claim 1, characterized in that, Obtain the user's initial request description, call the pre-trained empirical prediction model to predict the initial request description, and obtain a complete initial terms and conditions proposal, specifically: Obtain the user's initial request description and the negotiation dimension list configured according to the current negotiation scenario. Based on the request parsing prompts and the negotiation dimension list, guide the preset large language model to extract the relative importance weights of each dimension from the initial request description. Based on relative importance weights, the initial request descriptions are transformed into standardized text with uniform format and clearly defined fields; The BERT encoder is used to map the canonical text to obtain a high-dimensional semantic vector, and the prediction layer of the empirical prediction model is used to process the high-dimensional semantic vector to obtain a multi-dimensional vector. According to the preset dimension mapping rules, the multidimensional vector is converted into specific clause values ​​and filled into the placeholders corresponding to the preset initial template to generate a complete initial clause scheme.

3. The experience-driven multi-agent negotiation clause contract construction method according to claim 1, characterized in that, The initial terms and conditions are negotiated using the multi-agent negotiation support module to generate a pending solution, specifically: For the t-th consultation, obtain the latest feedback information from the user and determine whether all consultation dimensions in the latest user feedback information remain unchanged; If so, stop the negotiations and treat the best terms from the (t-1)th negotiation as a pending option; If not, the coordination agent module is called to parse and process the latest feedback information from the user to obtain the parsing result of the t-th negotiation. The parsing process includes: matching modifiers in the latest feedback information from the user based on a predefined sentiment lexicon and degree adverb dictionary, using the implicit reasoning ability of the big language model in combination with contextual semantics to judge the urgency of the other party's request, and normalizing the intensity into three discrete levels: high, medium, and low. The decision-making agent module is invoked to process the current terms and conditions, negotiation strategy, and the parsing result of the t-th negotiation, generate the optimal terms and conditions for the t-th negotiation, and send it to the user. When t equals 1, the current terms and conditions are the initial terms and conditions. When t does not equal 1, the current terms and conditions are the optimal terms and conditions for the (t-1)-th negotiation. Repeat the above steps until a solution is obtained.

4. The experience-driven multi-agent negotiation clause contract construction method according to claim 3, characterized in that, The decision-making agent module is invoked to process the current terms and conditions, negotiation strategy, and the analysis results of the t-th negotiation, and to generate the optimal terms and conditions for the t-th negotiation. Specifically: Based on the current terms and conditions, the negotiation strategy, and the analysis results of the t-th negotiation, a candidate set containing multiple candidate solutions is generated according to the core principles and decision constraints defined by the third decision prompt of the decision agent module. A pre-defined multi-dimensional comprehensive evaluation model is used to comprehensively score each candidate solution in the candidate set, and the control coordination agent module selects the candidate solution with the highest comprehensive score as the optimal terms solution for the t-th negotiation.

5. The experience-driven multi-agent negotiation clause contract construction method according to claim 4, characterized in that, Also includes: The number of negotiations is recorded using a periodic summary start value S, which is initially 0. Each time a negotiation is conducted, the periodic summary start value S is incremented by 1. When it is determined that the periodic summary initiation value S is equal to the periodic summary initiation threshold, the periodic summary initiation value S is reassigned to 0. The decision agent module triggers a periodic summary mechanism. Based on the second decision prompt of the decision agent module, the sequence of the other party's adjustment intentions in the previous periodic summary initiation threshold negotiations is analyzed, and the negotiation dimensions that the other party continues to focus on are identified, and the negotiation strategy is updated.

6. The experience-driven multi-agent negotiation clause contract construction method according to claim 3, characterized in that, A final validity check is performed on the proposed solution. If the check passes, a final consensus clause is generated and added to the experience base. Specifically: The proposed solutions are filtered to exclude those that have reached a consensus in form but fail to meet the basic constraints of either party in substance. After the filtering process, the remaining pending solutions are recorded and standardized in format to generate the final consensus terms. The final consensus terms will be added to the experience base as a new negotiation sample.

7. The experience-driven multi-agent negotiation clause contract construction method according to claim 6, characterized in that, Also includes: When the number of new negotiation samples in the experience base reaches the preset update threshold, the experience prediction model is updated. The current training set is constructed based on the preset update threshold of the experience base, the number of newly added consultation samples, and a subset randomly sampled from the historical experience base. The experience prediction model is then updated based on the current training set.

8. An experience-driven multi-agent negotiation terms contract construction device, characterized in that, include: The initial terms and conditions generation unit is used to obtain the user's initial request description, call the pre-trained experience prediction model to predict the initial request description, and obtain the complete initial terms and conditions. The multi-agent negotiation unit is used to negotiate the initial terms and conditions using the multi-agent negotiation support module and generate pending solutions. The multi-agent negotiation support module includes a decision agent module and a coordination agent module. The decision agent module is used to generate candidate solutions in response to the user's adjustment intentions, and the coordination agent module is used to parse the user's feedback information, evaluate the candidate solutions generated by the decision agent module, and control the progress of the negotiation process. The final consensus clause generation unit is used to perform a final validity check on the proposed clause. When the check is passed, the final consensus clause is generated and added to the experience base.

Images