Large model-based digital human purchase negotiation data processing method and system

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a digital human purchase negotiation data processing method and system based on a large model. The method comprises the steps of obtaining negotiation data of an opponent party and a multi-dimensional negotiation target preset by the opponent party, initializing multiple thinking flows in parallel, generating candidate thinking trees, calculating a multi-dimensional negotiation state vector for leaf nodes of each candidate thinking tree, screening out a to-be-selected strategy set, and calculating a multi-dimensional negotiation state vector for a leaf node of each candidate thinking tree; and calling a plurality of opponent portrait game models to carry out adversarial deduction on each to-be-selected strategy, generating a probability distribution matrix, determining an optimal strategy of the round in the to-be-selected strategy set by adopting a minimum-maximum regret criterion, and generating structured negotiation response data based on a thinking path corresponding to the optimal strategy. According to the scheme, the problem that a traditional single thinking chain is prone to falling into local optimum can be solved, one-sided decision making that long-term benefits are sacrificed by only paying attention to short-term prices is avoided, and heavy losses possibly caused by misjudgment of opponent intentions are avoided.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a procurement negotiation data processing method and system based on a large model of a digital human. Background Art

[0002] As a specific application of AI in business interactions, digital humans can autonomously analyze rival bids, develop response strategies, and generate communication scripts, playing a key role in complex procurement negotiations. Early procurement negotiation assistance systems primarily relied on expert systems with pre-set rules or statistical models based on historical data.

[0003] However, while expert systems offer clear logic, they lack the flexibility to adapt to the dynamic dynamics and irrational behavior of real-world negotiations. Traditional machine learning-based statistical models, such as regression analysis or classification algorithms, can learn pricing patterns or counterparty behavior from historical data. However, they often view negotiations as a single-point decision-making problem, lacking a holistic understanding of multi-round, long-term negotiation processes. Furthermore, the decision-making process is often a black box, making it difficult to explain the underlying logic behind the strategy. This is a fatal flaw in high-stakes business negotiations.

[0004] With the rise of large language models, techniques like chain thinking have provided new avenues for simulating the complex human reasoning process. By constructing a chain of reasoning step by step, chain thinking improves the model's logical analysis capabilities and the interpretability of its decisions for complex problems, addressing the black-box problem inherent in traditional models. However, most chain thinking techniques employ a single-threaded linear reasoning model, operating along a single path. If the initial direction is poorly chosen or deviations occur in intermediate links, the entire chain of reasoning can easily become trapped in a local optimum or even go awry. When faced with strategically volatile negotiating opponents, a one-way reasoning process is unable to foresee the opponent's reactions or the evolution of the confrontation over multiple rounds, presenting significant limitations. Summary of the Invention

[0005] The purpose of the present invention is to propose a procurement negotiation data processing method and system based on a large model of digital humans to solve the problems of lack of flexibility in negotiations and the "black box" decision-making process in the existing technology; to this end, the present invention provides solutions in the following two aspects.

[0006] In a first aspect, the present invention provides a procurement negotiation data processing method based on a large model of a digital human, comprising the following steps: Obtain the negotiation data of the opponent in the current negotiation round and the multi-dimensional negotiation goals preset by the party; based on the negotiation data of the opponent and the multi-dimensional negotiation goals, initialize multiple thought flows in parallel in the pre-constructed negotiation knowledge graph, and generate candidate thought trees through hierarchical expansion of each thought flow; during the expansion process, when the confidence of the branch to be expanded of any thought tree is lower than the preset threshold, trigger cross-flow grafting, and merge the nodes with confidence higher than the preset threshold in other thought flows into the current thought tree as new branches; calculate the multi-dimensional negotiation state vector containing the bid advantage, long-term cooperation value and performance reliability for the leaf nodes of each candidate thought tree, and screen out non-dominated leaf nodes as the candidate strategy set; call multiple preset opponent portrait game models, conduct adversarial deductions for each candidate strategy in the next N rounds, and generate a probability distribution matrix containing multiple deduction outcomes; according to the probability distribution matrix, use the minimum and maximum regret criterion to determine the optimal strategy of this round in the candidate strategy set, and generate structured negotiation response data based on the thought path corresponding to the optimal strategy.

[0007] Preferably, based on the negotiation data of the opponent and the multi-dimensional negotiation goals, multiple thought flows are initialized in parallel in a pre-constructed negotiation knowledge graph, and each thought flow generates a candidate thought tree through hierarchical expansion, including: taking the three dimensions of optimal price, shortest delivery cycle, and most stable cooperative relationship as the main goals, initializing three independent thought flows respectively, and setting each main goal as the root node of the corresponding thought flow; for any node in each thought flow, with the negotiation point represented by the node as the center, the related arguments, potential risks and response plans are retrieved in the negotiation knowledge graph as sub-nodes, and breadth-first expansion is performed layer by layer until the expansion depth reaches the preset number of layers to form a candidate thought tree.

[0008] Preferably, when the confidence of the branch to be expanded of any thinking tree is lower than a preset threshold, cross-flow grafting is triggered, and the nodes in other thinking flows with confidence higher than the preset threshold are merged into the current thinking tree as new branches, including: calculating the confidence of the branch to be expanded as the cumulative product of the confidences of all nodes on the path from the root node to the current node to be expanded; when the confidence of the branch to be expanded of thinking tree A is lower than the preset threshold, pausing the expansion of the branch; retrieving all nodes in other thinking trees whose path confidence from the root node to the node is higher than the preset threshold, and selecting the node with the highest path confidence as the grafting node; copying and connecting the grafting node and the complete subtree of the grafting node to the branch to be expanded that triggered the pause in thinking tree A to form a new expanded branch.

[0009] Preferably, the multi-dimensional negotiation state vector including the bid advantage, long-term cooperation value and performance reliability is calculated for each leaf node of the candidate thinking tree, including: calculating the bid advantage , the calculation formula is: ,in, is the bid advantage, is the counterparty’s historical average quote, Quote for the current leaf node strategy; calculate the long-term cooperation value , the historical cooperation years, historical average order amount and supplier rating score are normalized and calculated by weighted summation. The calculation formula is: ,in, For the long-term cooperation value, 、 、 are the normalized values ​​of historical cooperation years, historical average order amount and supplier rating score, 、 、 is the corresponding preset weight coefficient, and Calculate performance reliability , the calculation formula is: ,in, For contract performance reliability, The number of orders delayed in the past year. The total number of orders in the past year.

[0010] Preferably, the method calls multiple preset opponent portrait game models, conducts N rounds of adversarial deductions for each candidate strategy, and generates a probability distribution matrix containing multiple deduction outcomes, including: calling an aggressive price suppression model, a robust value-oriented model, and a compromising relationship maintenance model as multiple opponent portrait game models; taking any candidate strategy as our first-round input, simulating the opponent's response and our subsequent two-round response in the three models respectively, until three rounds of deduction are completed; counting the frequency of occurrence of the three outcomes of reaching a deal, negotiation breakdown, and negotiation stalemate at the end of the deduction under each model, and converting the occurrence frequency into probabilities to obtain a probability distribution matrix of the deduction outcomes.

[0011] Preferably, the minimum-maximum regret criterion is used to determine the optimal strategy for this round in the set of candidate strategies based on the probability distribution matrix, including: presetting benefit values ​​for the three deduction outcomes of reaching a deal, negotiation breakdown, and negotiation stalemate, and combining the probability distribution matrix to calculate the expected benefit value of each candidate strategy under different opponent portrait game models to form a benefit matrix; finding the highest expected benefit value that can be achieved by all candidate strategies under each opponent portrait game model; calculating the regret matrix, wherein the calculation method of each element is: the highest expected benefit value under the opponent portrait game model minus the expected benefit value of a specific strategy under the opponent portrait game model; determining the maximum regret value of each candidate strategy in all opponent portrait game models; and selecting the candidate strategy with the smallest maximum regret value as the optimal strategy for this round.

[0012] In the second aspect, a procurement negotiation data processing system based on a digital human of a large model includes the following modules: a thinking expansion module, which is used to obtain the negotiation data of the opponent in the current negotiation round and the multi-dimensional negotiation goals preset by the party; based on the negotiation data of the opponent and the multi-dimensional negotiation goals, multiple thinking flows are initialized in parallel in a pre-built negotiation knowledge graph, and each thinking flow generates a candidate thinking tree through hierarchical expansion; during the expansion process, when the confidence of the branch to be expanded of any thinking tree is lower than the preset threshold, cross-flow grafting is triggered, and the nodes in other thinking flows with confidence higher than the preset threshold are merged into the current thinking tree as new branches ; A deduction module is used to calculate a multidimensional negotiation state vector including the bid advantage, long-term cooperation value and performance reliability for the leaf nodes of each candidate thinking tree, and screen out non-dominated leaf nodes as the set of candidate strategies; call multiple preset opponent portrait game models to conduct N rounds of adversarial deduction for each candidate strategy in the future, and generate a probability distribution matrix containing multiple deduction outcomes; a negotiation data generation module is used to determine the optimal strategy of this round in the candidate strategy set based on the probability distribution matrix and the minimum and maximum regret criterion, and generate structured negotiation response data based on the thinking path corresponding to the optimal strategy.

[0013] Preferably, based on the negotiation data of the opponent and the multi-dimensional negotiation goals, multiple thought flows are initialized in parallel in a pre-constructed negotiation knowledge graph, and each thought flow generates a candidate thought tree through hierarchical expansion, including: taking the three dimensions of optimal price, shortest delivery cycle, and most stable cooperative relationship as the main goals, initializing three independent thought flows respectively, and setting each main goal as the root node of the corresponding thought flow; for any node in each thought flow, with the negotiation point represented by the node as the center, the related arguments, potential risks and response plans are retrieved in the negotiation knowledge graph as sub-nodes, and breadth-first expansion is performed layer by layer until the expansion depth reaches the preset number of layers to form a candidate thought tree.

[0014] Preferably, when the confidence of the branch to be expanded of any thinking tree is lower than a preset threshold, cross-flow grafting is triggered, and the nodes in other thinking flows with confidence higher than the preset threshold are merged into the current thinking tree as new branches, including: calculating the confidence of the branch to be expanded as the cumulative product of the confidences of all nodes on the path from the root node to the current node to be expanded; when the confidence of the branch to be expanded of thinking tree A is lower than the preset threshold, pausing the expansion of the branch; retrieving all nodes in other thinking trees whose path confidence from the root node to the node is higher than the preset threshold, and selecting the node with the highest path confidence as the grafting node; copying and connecting the grafting node and the complete subtree of the grafting node to the branch to be expanded that triggered the pause in thinking tree A to form a new expanded branch.

[0015] Preferably, the multi-dimensional negotiation state vector including the bid advantage, long-term cooperation value and performance reliability is calculated for each leaf node of the candidate thinking tree, including: calculating the bid advantage , the calculation formula is: ,in, is the bid advantage, is the counterparty’s historical average quote, Quote for the current leaf node strategy; calculate the long-term cooperation value , the historical cooperation years, historical average order amount and supplier rating score are normalized and calculated by weighted summation. The calculation formula is: , where is the long-term cooperation value, 、 、 are the normalized values ​​of historical cooperation years, historical average order amount and supplier rating score, 、 、 is the corresponding preset weight coefficient, and Calculate performance reliability , the calculation formula is: , where is the performance reliability, The number of orders delayed in the past year. The total number of orders in the past year.

[0016] The beneficial effects of the present invention are as follows: the present invention overcomes the problem that traditional single thought chains are prone to falling into local optimality by initializing multiple thought streams in parallel and utilizing a cross-stream grafting mechanism. Moreover, by constructing a multi-dimensional state vector that includes quotation advantages, cooperation value, and performance reliability, it ensures that the selected candidate strategies can achieve an effective balance among multiple conflicting goals, avoiding one-sided decisions that only focus on short-term prices at the expense of long-term interests. In addition, the introduction of multiple rounds of adversarial deduction can foresee the long-term impact and potential risks of strategies, and adopt the minimum maximum regret criterion to make the final decision, so as to select the most reliable solution in a negotiation environment full of uncertainty and avoid significant losses that may result from misjudging the opponent's intentions. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The flowchart schematically shows the steps of the procurement negotiation data processing method based on the large model of digital human in this embodiment. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention.

[0019] like Figure 1As shown, the procurement negotiation data processing method based on the large model of digital human in this embodiment includes the following steps: Step S1, obtain the negotiation data of the opponent in the current negotiation round and the multi-dimensional negotiation goals preset by this party; based on the negotiation data of the opponent and the multi-dimensional negotiation goals, initialize multiple thinking flows in parallel in the pre-built negotiation knowledge graph, and each thinking flow generates a candidate thinking tree through hierarchical expansion; during the expansion process, when the confidence of the branch to be expanded of any thinking tree is lower than the preset threshold, cross-flow grafting is triggered, and the nodes in other thinking flows with confidence higher than the preset threshold are merged into the current thinking tree as new branches.

[0020] The system receives structured data from the counterparty through an API, such as a JSON object containing a quote of 105 yuan, a delivery time of 60 days, and cash on delivery. Simultaneously, it loads its own multi-dimensional negotiation objectives from an internal database. These objectives are quantified as a price target range of 95 to 102 yuan, an optimal delivery time of within 30 days, acceptable payment methods including a 30% down payment, and a desired long-term partnership score of 4.5 or higher out of 5. A pre-built negotiation knowledge graph stores entities and relationships, including supplier performance records, industry cost benchmarks, and market supply and demand. Based on this, three thought streams are initialized in parallel: thought stream A focuses on optimal price, thought stream B on long-term partnerships, and thought stream C on rapid fulfillment. Specifically, when constructing the knowledge graph, the spaCy library is used for basic entity recognition, such as company names and dates, and fine-tuned with a pre-trained model such as BERT to identify negotiation-specific entities, such as negotiation topics and concessions. Relationship extraction is preferably performed using the OpenNRE library to identify relationships between entities. When generating a thought tree, a stateful Actor model is used, and each thought flow is a Ray Actor.

[0021] Based on the opponent's negotiation data and multi-dimensional negotiation goals, multiple thought flows are initialized in parallel in the pre-built negotiation knowledge graph. Each thought flow generates a candidate thought tree through hierarchical expansion, including: taking the three dimensions of optimal price, shortest delivery cycle, and most stable cooperative relationship as the main goals, initializing three independent thought flows respectively, and setting each main goal as the root node of the corresponding thought flow; for any node in each thought flow, with the negotiation point represented by the node as the center, retrieve the related arguments, potential risks and response plans as sub-nodes in the negotiation knowledge graph, and perform breadth-first expansion layer by layer until the expansion depth reaches the preset number of layers to form a candidate thought tree.

[0022] Establish three core negotiation goals: optimal price, shortest delivery cycle, and strongest partnership. Initiate a separate thought process for each goal, using the goal itself as the starting point for a thought tree. Each thought tree is constructed layer by layer. Starting from any key negotiation point, the knowledge graph is searched for supporting arguments, potential risks, and response strategies, adding these as new nodes to the next layer. This process continues in a breadth-first manner until the preset limit on the number of expansion layers is reached, resulting in multiple structured candidate thought trees.

[0023] When the confidence of the branch to be expanded in any thinking tree is lower than the preset threshold, cross-flow grafting is triggered, and the nodes in other thinking flows with confidence higher than the preset threshold are merged into the current thinking tree as new branches, including: calculating the confidence of the branch to be expanded as the cumulative product of the confidences of all nodes on the path from the root node to the current node to be expanded; when the confidence of the branch to be expanded in thinking tree A is lower than the preset threshold, pausing the expansion of the branch; retrieving all nodes in other thinking trees whose path confidence from the root node to the node is higher than the preset threshold, and selecting the node with the highest path confidence as the grafting node; copying and connecting the grafting node and the complete subtree of the grafting node to the branch to be expanded that triggered the pause in thinking tree A to form a new expanded branch.

[0024] When constructing a mind tree, a confidence score is calculated for each branch. This confidence score is preferably calculated by multiplying the confidence scores of all nodes along the path from the root node to the current node. Once the cumulative confidence score of a branch, such as in mind tree A, falls below a preset threshold of 0.3, the growth of that branch is suspended. Other mind trees B and C are searched for candidate nodes with path confidence scores higher than 0.3. The one with the highest confidence score is selected. This high-confidence node and all its substructures are then completely copied and transplanted to the suspended branch in mind tree A, essentially replacing the original weak branch with a more reliable argument path.

[0025] Step S2 calculates a multidimensional negotiation state vector including the bid advantage, long-term cooperation value, and performance reliability for the leaf nodes of each candidate thinking tree, and selects non-dominated leaf nodes as the set of candidate strategies; calls multiple preset opponent portrait game models, conducts N rounds of adversarial deductions for each candidate strategy in the future, and generates a probability distribution matrix containing multiple deduction outcomes.

[0026] Calculate the multi-dimensional negotiation state vector containing the bid advantage, long-term cooperation value and performance reliability for the leaf node of each candidate thinking tree, including: Calculate the bid advantage , the calculation formula is: ,in, is the bid advantage, is the counterparty’s historical average quote, Quote for the current leaf node strategy; calculate the long-term cooperation value , the historical cooperation years, historical average order amount and supplier rating score are normalized and calculated by weighted summation. The calculation formula is: ,in, For the long-term cooperation value, 、 、 are the normalized values ​​of historical cooperation years, historical average order amount and supplier rating score, 、 、 is the corresponding preset weight coefficient, and Calculate performance reliability , the calculation formula is: ,in, For contract performance reliability, The number of orders delayed in the past year. The total number of orders in the past year.

[0027] The above formula calculates bid superiority by subtracting the current strategy's bid from the counterparty's historical average bid, and dividing the result by the historical average bid, thereby quantifying price competitiveness. The long-term partnership value assessment normalizes indicators such as the length of historical partnership, average order value, and supplier rating, and then weights them according to preset weights. Contract performance reliability is calculated by calculating the proportion of orders delivered without delay within the past year to the total number of orders. These three indicators together constitute a comprehensive evaluation system for each strategy, providing comprehensive and objective data support for selecting the optimal strategy and ensuring that decisions strike the optimal balance between cost, value, and risk.

[0028] Calling multiple preset opponent portrait game models, conducting N rounds of adversarial deductions for each candidate strategy, and generating a probability distribution matrix containing multiple deduction outcomes, including: calling an aggressive price suppression model, a robust value-oriented model, and a compromising relationship maintenance model as the multiple opponent portrait game models; using any candidate strategy as our first-round input, simulating the opponent's response and our subsequent two-round response in the three models respectively until three rounds of deduction are completed; counting the frequency of occurrence of the three outcomes of reaching a deal, negotiation breakdown, and negotiation stalemate at the end of the deduction under each model, and converting the occurrence frequency into probabilities to obtain a probability distribution matrix of the deduction outcomes.

[0029] To predict the actual effects of each candidate strategy, we simulated it in a simulated adversarial environment. Three typical counterparty behavior patterns were pre-set: aggressive price suppression, conservative value orientation, and compromising relationship maintenance. The simulations were set to last three rounds. For each candidate strategy, three rounds of interaction were simulated under each of these three counterparty models to observe the dynamic evolution of the results. After the simulations were completed, the number of deals, negotiation breakdowns, and stalemates under each counterparty model was counted. The probability of each outcome was then calculated and summarized to form a probability distribution matrix.

[0030] Step S3: According to the probability distribution matrix, the minimum-maximum-regret criterion is used to determine the optimal strategy for this round in the candidate strategy set, and structured negotiation response data is generated based on the thinking path corresponding to the optimal strategy.

[0031] Based on the probability distribution matrix, a regret matrix is ​​constructed. The specific calculation method is to find the highest achievable probability value among all strategies for each outcome, i.e., each column. Then, subtract the probability value of each strategy in that column from the highest probability value to obtain the regret value for each strategy under that outcome. For each row of the regret matrix, i.e., each strategy, the maximum regret value across all outcomes is found. The strategy with the smallest maximum regret value is selected as the optimal strategy for this round. After determining the optimal strategy, the entire path of the strategy in the mindset tree is traced back from the root node to the leaf node. The logic of each node along the path, such as "We expect price adjustments because we are committed to long-term procurement," is converted into natural language. This is then packaged into a structured JSON object along with the strategy's final quote, delivery date, and other parameters, as the formal negotiation response.

[0032] In an optional embodiment, according to the probability distribution matrix, the minimum-maximum regret criterion is used to determine the optimal strategy for this round in the set of candidate strategies, including: presetting benefit values ​​for the three deduction outcomes of reaching a deal, negotiation breakdown, and negotiation stalemate, and combining the probability distribution matrix to calculate the expected benefit value of each candidate strategy under different opponent portrait game models to form a benefit matrix; finding the highest expected benefit value that can be achieved by all candidate strategies under each opponent portrait game model; calculating the regret matrix, wherein the calculation method of each element is: the highest expected benefit value under the opponent portrait game model minus the expected benefit value of a specific strategy under the opponent portrait game model; determining the maximum regret value of each candidate strategy in all opponent portrait game models; and selecting the candidate strategy with the smallest maximum regret value as the optimal strategy for this round.

[0033] The three outcomes of reaching a deal, negotiation breakdown, and negotiation stalemate are assigned a preset payoff score. Combined with the derived probability distribution, the expected payoff of each strategy against different types of adversaries is calculated, forming a payoff matrix. Regret is calculated by subtracting the expected payoff of the current strategy from the highest possible payoff in that scenario for each adversary type. The resulting value represents the missed opportunity by choosing that strategy over the optimal one. The maximum possible regret value for each strategy under all adversary types is determined, and the strategy with the lowest maximum regret value is selected as the optimal and lowest-risk option for this round of negotiations.

[0034] The present invention also provides a procurement negotiation data processing system based on a large model of a digital human, comprising the following modules: The thinking expansion module is used to obtain the negotiation data of the opponent in the current negotiation round and the multi-dimensional negotiation goals preset by this party; based on the negotiation data of the opponent and the multi-dimensional negotiation goals, multiple thinking flows are initialized in parallel in the pre-built negotiation knowledge graph, and each thinking flow generates a candidate thinking tree through hierarchical expansion; during the expansion process, when the confidence of the branch to be expanded in any thinking tree is lower than the preset threshold, cross-flow grafting is triggered, and the nodes in other thinking flows with confidence higher than the preset threshold are merged into the current thinking tree as new branches.

[0035] The deduction module is used to calculate the multidimensional negotiation state vector containing the bid advantage, long-term cooperation value and performance reliability for the leaf nodes of each candidate thinking tree, and screen out non-dominated leaf nodes as the set of candidate strategies; call multiple preset opponent portrait game models, conduct N rounds of adversarial deduction for each candidate strategy in the future, and generate a probability distribution matrix containing multiple deduction outcomes.

[0036] The negotiation data generation module is used to determine the optimal strategy of this round in the set of candidate strategies based on the probability distribution matrix and the minimum maximum regret criterion, and generate structured negotiation response data based on the thinking path corresponding to the optimal strategy.

[0037] In a preferred embodiment, based on the counterparty negotiation data and multi-dimensional negotiation goals, multiple thought flows are initialized in parallel in a pre-constructed negotiation knowledge graph, and each thought flow generates a candidate thought tree through hierarchical expansion, including: taking the three dimensions of optimal price, shortest delivery cycle, and most stable cooperative relationship as the main goals, initializing three independent thought flows respectively, and setting each main goal as the root node of the corresponding thought flow; for any node in each thought flow, with the negotiation point represented by the node as the center, the related arguments, potential risks and response plans are retrieved in the negotiation knowledge graph as sub-nodes, and breadth-first expansion is performed layer by layer until the expansion depth reaches the preset number of layers to form a candidate thought tree.

[0038] In a preferred embodiment, when the confidence of the branch to be expanded of any thinking tree is lower than a preset threshold, cross-flow grafting is triggered, and the nodes in other thinking flows with confidence higher than the preset threshold are merged into the current thinking tree as new branches, including: calculating the confidence of the branch to be expanded as the cumulative product of the confidences of all nodes on the path from the root node to the current node to be expanded; when the confidence of the branch to be expanded of thinking tree A is lower than the preset threshold, pausing the expansion of the branch; retrieving all nodes in other thinking trees whose path confidence from the root node to the node is higher than the preset threshold, and selecting the node with the highest path confidence as the grafting node; copying and connecting the grafting node and the complete subtree of the grafting node to the branch to be expanded that triggered the pause in thinking tree A to form a new expanded branch.

[0039] In a preferred embodiment, a multi-dimensional negotiation state vector including the bid advantage, long-term cooperation value and performance reliability is calculated for each leaf node of the candidate thought tree, including: calculating the bid advantage , the calculation formula is: ,in, is the bid advantage, is the counterparty’s historical average quote, Quote for the current leaf node strategy; calculate the long-term cooperation value , the historical cooperation years, historical average order amount and supplier rating score are normalized and calculated by weighted summation. The calculation formula is: , where is the long-term cooperation value, 、 、 are the normalized values ​​of historical cooperation years, historical average order amount and supplier rating score, 、 、 is the corresponding preset weight coefficient, and Calculate performance reliability , the calculation formula is: , where is the performance reliability, The number of orders delayed in the past year. The total number of orders in the past year.

[0040] In a preferred embodiment, multiple preset opponent portrait game models are called, and N rounds of adversarial deductions are carried out for each candidate strategy to generate a probability distribution matrix containing multiple deduction outcomes, including: calling an aggressive price suppression model, a robust value-oriented model, and a compromising relationship maintenance model as the multiple opponent portrait game models; taking any candidate strategy as our first-round input, simulating the opponent's response and our subsequent two-round response in the three models respectively until three rounds of deduction are completed; counting the frequencies of the three outcomes of reaching a deal, negotiation breakdown, and negotiation stalemate at the end of the deduction under each model, and converting the frequencies into probabilities to obtain a probability distribution matrix of the deduction outcomes.

[0041] In a preferred embodiment, according to the probability distribution matrix, the minimum-maximum regret criterion is used to determine the optimal strategy for this round in the set of candidate strategies, including: presetting benefit values ​​for the three deduction outcomes of reaching a deal, negotiation breakdown, and negotiation stalemate, and combining the probability distribution matrix to calculate the expected benefit value of each candidate strategy under different opponent portrait game models to form a benefit matrix; finding the highest expected benefit value that can be achieved by all candidate strategies under each opponent portrait game model; calculating the regret matrix, in which the calculation method of each element is: the highest expected benefit value under the opponent portrait game model minus the expected benefit value of a specific strategy under the opponent portrait game model; determining the maximum regret value of each candidate strategy in all opponent portrait game models; and selecting the candidate strategy with the smallest maximum regret value as the optimal strategy for this round.

[0042] In the description of this specification, “a plurality of” means at least two, for example, two, three or more, etc., unless otherwise clearly defined.

[0043] Although this specification has shown and described several embodiments of the present invention, it is obvious to those skilled in the art that such embodiments are provided by way of example only. Those skilled in the art will think of many changes, modifications and substitutions without departing from the idea and spirit of the present invention.

Claims

1. A procurement negotiation data processing method based on a large model of digital human, characterized by: The following steps are involved: Obtain the negotiation data of the counterparty in the current negotiation round and the multi-dimensional negotiation goals preset by the party; Based on the negotiation data of the counterparty and the multi-dimensional negotiation goals, multiple thought flows are initialized in parallel in the pre-built negotiation knowledge graph, and each thought flow is expanded hierarchically to generate a candidate thought tree; During the expansion process, when the confidence of the to-be-expanded branch of any thinking tree is lower than a preset threshold, cross-flow grafting is triggered, and nodes in other thinking flows with confidence higher than the preset threshold are merged into the current thinking tree as new branches; Calculate a multi-dimensional negotiation state vector for each leaf node of the candidate mindset tree, including bid dominance, long-term cooperation value, and contract performance reliability, and select non-dominated leaf nodes as the candidate strategy set. Invoke multiple preset counterparty profiling game models to conduct N rounds of adversarial deductions for each candidate strategy, generating a probability distribution matrix containing multiple deduction outcomes. According to the probability distribution matrix, the minimum-maximum-regret criterion is adopted to determine the optimal strategy of this round in the candidate strategy set, and structured negotiation response data is generated based on the thinking path corresponding to the optimal strategy.

2. The procurement negotiation data processing method based on a large model of digital human according to claim 1 is characterized in that: Based on the negotiation data of the counterparty and the multi-dimensional negotiation goals, multiple thought flows are initialized in parallel in the pre-built negotiation knowledge graph, and each thought flow generates a candidate thought tree through hierarchical expansion, including: With the best price, shortest delivery cycle, and strongest cooperative relationship as the main goals, initialize three independent thought flows respectively, and set each main goal as the root node of the corresponding thought flow; For any node in each thought flow, centering on the negotiation points represented by the node, the related arguments, potential risks and response plans are retrieved in the negotiation knowledge graph as sub-nodes, and breadth-first expansion is performed layer by layer until the expansion depth reaches the preset number of layers to form a candidate thought tree.

3. The procurement negotiation data processing method based on a large model of digital human according to claim 1 is characterized in that: When the confidence of the to-be-expanded branch of any thought tree is lower than a preset threshold, cross-flow grafting is triggered, and nodes in other thought flows with confidence higher than the preset threshold are merged into the current thought tree as new branches, including: The confidence of the branch to be expanded is calculated as the cumulative product of the confidences of all nodes on the path from the root node to the current node to be expanded; When the confidence level of the branch to be expanded in thinking tree A is lower than the preset threshold, the expansion of the branch is suspended; Retrieve all nodes in other thinking trees whose path confidence from the root node to the node is higher than the preset threshold, and select the node with the highest path confidence as the grafting node; The grafted node and the complete subtree of the grafted node are copied and connected to the branch to be expanded that triggers the pause in the thinking tree A to form a new expansion branch.

4. The procurement negotiation data processing method based on a large model of digital human according to claim 1 is characterized in that: The multi-dimensional negotiation state vector including the bid advantage, long-term cooperation value and performance reliability is calculated for each leaf node of the candidate thinking tree, including: Calculate the bid advantage , the calculation formula is: ,in, is the bid advantage, is the counterparty’s historical average quote, Quote for the current leaf node strategy; Calculating the long-term value of cooperation , the historical cooperation years, historical average order amount and supplier rating score are normalized and calculated by weighted summation. The calculation formula is: ,in, For the long-term cooperation value, 、 、 are the normalized values ​​of historical cooperation years, historical average order amount and supplier rating score, 、 、 is the corresponding preset weight coefficient, and ; Calculating performance reliability , the calculation formula is: ,in, For contract performance reliability, The number of orders delayed in the past year. The total number of orders in the past year.

5. The procurement negotiation data processing method based on a large model of digital human according to claim 1 is characterized in that: The method calls multiple preset opponent portrait game models, conducts N rounds of adversarial deductions for each candidate strategy, and generates a probability distribution matrix containing multiple deduction outcomes, including: Invoke the aggressive price suppression model, the robust value-oriented model, and the compromising relationship maintenance model as multiple counterparty profiling game models; Using any candidate strategy as our first-round input, we simulate the opponent's response and our subsequent two-round response in each of the three models until we complete three rounds of deductions. Under each model, the frequency of occurrence of the three outcomes of reaching a deal, negotiation breakdown, and negotiation stalemate at the end of the deduction is counted, and the occurrence frequency is converted into probabilities to obtain the probability distribution matrix of the deduction outcomes.

6. The procurement negotiation data processing method based on a large model of digital human according to claim 1 is characterized in that: Determining the optimal strategy of this round in the set of candidate strategies using the minimum-maximum-regret criterion according to the probability distribution matrix includes: Preset payoff values ​​for the three deduction outcomes of deal completion, negotiation breakdown, and negotiation stalemate, and calculate the expected payoff value of each candidate strategy under different counterparty profile game models based on the probability distribution matrix to form a payoff matrix; Find the highest expected return that can be achieved by all candidate strategies under each opponent profile game model; Calculate a regret matrix, where each element is calculated as follows: the highest expected return under the opponent profile game model minus the expected return of a specific strategy under the opponent profile game model; Determine the maximum regret value of each candidate strategy in all opponent profile game models; The candidate strategy with the smallest maximum regret value is selected as the optimal strategy for this round.

7. A procurement negotiation data processing system based on a large model of digital human, characterized by: Includes the following modules: The thought expansion module is used to obtain the negotiation data of the other party in the current negotiation round and the multi-dimensional negotiation goals preset by the party; based on the negotiation data of the other party and the multi-dimensional negotiation goals, multiple thought flows are initialized in parallel in the pre-built negotiation knowledge graph, and each thought flow is hierarchically expanded to generate a candidate thought tree; During the expansion process, when the confidence of the to-be-expanded branch of any thinking tree is lower than a preset threshold, cross-flow grafting is triggered, and nodes in other thinking flows with confidence higher than the preset threshold are merged into the current thinking tree as new branches; The deduction module is used to calculate a multi-dimensional negotiation state vector for each leaf node of the candidate thinking tree, including bid advantage, long-term cooperation value, and performance reliability, and screen out non-dominated leaf nodes as the candidate strategy set. It calls multiple preset opponent portrait game models to conduct N rounds of adversarial deduction for each candidate strategy, generating a probability distribution matrix containing multiple deduction outcomes. The negotiation data generation module is used to determine the optimal strategy of this round in the set of candidate strategies based on the probability distribution matrix and the minimum maximum regret criterion, and generate structured negotiation response data based on the thinking path corresponding to the optimal strategy.

8. The procurement negotiation data processing system based on a large model of digital human according to claim 7 is characterized in that: Based on the negotiation data of the counterparty and the multi-dimensional negotiation goals, multiple thought flows are initialized in parallel in the pre-built negotiation knowledge graph, and each thought flow generates a candidate thought tree through hierarchical expansion, including: With the best price, shortest delivery cycle, and strongest cooperative relationship as the main goals, initialize three independent thought flows respectively, and set each main goal as the root node of the corresponding thought flow; For any node in each thought flow, centering on the negotiation points represented by the node, the related arguments, potential risks and response plans are retrieved in the negotiation knowledge graph as sub-nodes, and breadth-first expansion is performed layer by layer until the expansion depth reaches the preset number of layers to form a candidate thought tree.

9. The procurement negotiation data processing system based on a large model of digital human according to claim 7 is characterized in that: When the confidence of the to-be-expanded branch of any thought tree is lower than a preset threshold, cross-flow grafting is triggered, and nodes in other thought flows with confidence higher than the preset threshold are merged into the current thought tree as new branches, including: The confidence of the branch to be expanded is calculated as the cumulative product of the confidences of all nodes on the path from the root node to the current node to be expanded; When the confidence level of the branch to be expanded in thinking tree A is lower than the preset threshold, the expansion of the branch is suspended; Searching for all nodes in other thinking trees whose path confidence from the root node to the node is higher than the preset threshold, and selecting the node with the highest path confidence as the grafting node; The grafted node and the complete subtree of the grafted node are copied and connected to the branch to be expanded that triggers the pause in the thinking tree A to form a new expansion branch.

10. The procurement negotiation data processing system based on a large model of digital human according to claim 7 is characterized in that: The multi-dimensional negotiation state vector including the bid advantage, long-term cooperation value and performance reliability is calculated for each leaf node of the candidate thinking tree, including: Calculate the bid advantage , the calculation formula is: ,in, is the bid advantage, is the counterparty’s historical average quote, Quote for the current leaf node strategy; Calculating the long-term value of cooperation , the historical cooperation years, historical average order amount and supplier rating score are normalized and calculated by weighted summation. The calculation formula is: , where is the long-term cooperation value, 、 、 are the normalized values ​​of historical cooperation years, historical average order amount and supplier rating score, 、 、 is the corresponding preset weight coefficient, and ; Calculating performance reliability , the calculation formula is: , where is the performance reliability, The number of orders delayed in the past year. The total number of orders in the past year.

Images