Cross-culture consensus achievement method based on game theory

Abstract

The invention relates to a cross-culture consensus achievement method based on the game theory, and the method comprises the steps: carrying out the preprocessing of collected regional value data, and constructing a regional value data set; based on the regional value data sets and cultural incentives of different regions, training large language models respectively, and generating cultural agents corresponding to the regions respectively; configuring an initial culture criterion set for each culture agent and designing a utility function; then, Nash equilibrium points of all the culture criterion sets are calculated, and each culture agent generates a corresponding new candidate proposal group based on the Nash equilibrium points; judging whether a new candidate proposal with the highest income in any new candidate proposal group meets a condition or not, and if yes, adding the new candidate proposal into a corresponding culture criterion set; and until the negotiation process converges, outputting the updated culture criterion set corresponding to each culture agent and the final weight distribution of each criterion as a cross-culture consensus result, thereby realizing fair and stable automatic consensus solution.

Description

Technical Field

[0001] This application relates to the field of cross-cultural consensus generation technology, and in particular to a cross-cultural consensus-reaching method based on game theory. Background Technology

[0002] In recent years, Large Language Models (LLMs) have demonstrated significant capabilities in information processing and decision support. However, existing technologies still suffer from serious technical limitations when handling tasks involving multiculturalism. Currently, to adapt models to different cultures, methods such as fine-tuning based on specific cultural datasets or incorporating cultural identity descriptions into prompts are commonly used. However, these mainstream techniques still face the following insurmountable technical bottlenecks: 1. How to design a structured, game theory-based automated negotiation and computation framework to replace the existing random and unstructured debate or voting mechanisms, ensure the equal status of multicultural topics in the negotiation process, thereby generating stable and fair consensus solutions and fundamentally avoiding the dominance of mainstream cultural biases.

[0003] 2. How to utilize structured real-world cultural data (such as world values ​​survey data) to construct a cultural intelligence agent that can accurately reflect the core value system and decision-making preferences of a specific culture, in order to overcome the "pseudo-alignment" problem of existing methods.

[0004] 3. How to establish a calculable and quantifiable evaluation index system to automatically evaluate the fairness (acceptance by all parties) and cultural authenticity (consistency with initial values) of the consensus results, and provide a technical basis for optimizing the entire consensus generation system. Summary of the Invention

[0005] Therefore, it is necessary to provide a game theory-based method for achieving cross-cultural consensus, including: S1: Preprocess the collected regional value data to construct a regional value dataset; based on the regional value datasets and cultural incentives of different regions, train large language models to generate cultural intelligent agents corresponding to each region. S2: Configure an initial set of cultural criteria for each cultural agent and design a utility function. The set of cultural criteria includes basic criteria that characterize the value data of the corresponding region. The utility function is used to evaluate the benefits of the candidate proposals generated by the corresponding cultural agent. S3: Calculate the Nash equilibrium point of all cultural criterion sets, and each cultural agent generates a corresponding new candidate proposal group based on the Nash equilibrium point; S4: Determine whether the new candidate proposal with the highest benefit in any new candidate proposal group meets the conditions. If so, add the new candidate proposal to the corresponding cultural criterion set. S5: Iterate through steps S3-S4 until convergence, and output the updated set of cultural rules for each cultural agent and the final weight distribution of each rule as the cross-cultural consensus result.

[0006] Preferably, the collection of regional value data includes: Values ​​survey data of populations in various regions were obtained from the World Values ​​Survey and the Pew Global Attitudes Survey. Score data for each region were obtained from Hofstede's theory of cultural dimensions.

[0007] Preferably, preprocessing the collected regional value data includes: The amount of value survey data corresponding to all regions was proportionally increased. The question-and-answer pairs in the value survey data, which are scaled up proportionally, are subjected to equivalence verification and summarization, and the question-and-answer pairs with equivalent semantics are converted into value vectors. The value vectors and score data of the same region are used as the regional value dataset for the corresponding region.

[0008] Preferably, the cultural agents corresponding to the generated regions include: Based on a regional value dataset of any region, train a large language model, optimize the parameters of the large language model, and enable the large language model to capture the cultural values ​​of the corresponding region. A cultural incentive is designed for the region, which is used to guide the response of the large language model to reflect the corresponding cultural values; Based on the optimized large language model corresponding to the region and cultural incentives, a cultural intelligent agent corresponding to the region is instantiated and generated.

[0009] Preferably, the expression for the utility function is: ; in, Indicates the first The first cultural intelligent agent generated the first The benefits of each candidate proposal, This represents the first weighting coefficient. This represents the second weighting coefficient. This represents the third weighting coefficient. Indicates the first The candidate proposal and the first Candidate proposals for a cultural intelligent agent similarity, Indicates the first The candidate proposal and the first A cultural agent's own regional value dataset Alignment degree, Indicates the first The candidate proposal and the first Historical candidate proposals for a cultural intelligent agent The degree of repetition.

[0010] Preferably, the similarity, alignment, and repetition are calculated using the cosine similarity formula.

[0011] Preferably, S3 includes: any cultural agent A takes the Nash equilibrium weight distribution of the cultural criterion sets corresponding to the other cultural agents B as input, and generates new candidate proposals that maximize the utility function of cultural agent A; and combines the new candidate proposals to generate a new candidate proposal group corresponding to cultural agent A.

[0012] Preferably, S4 includes: determining whether the benefit of the new candidate proposal with the highest benefit in any new candidate proposal group is greater than a set threshold; if so, adding the new candidate proposal to the corresponding cultural criterion set; otherwise, no processing is performed.

[0013] Preferably, convergence includes: no new candidate proposal with a return higher than a set threshold appears in several consecutive iterations.

[0014] Preferred options also include: The cross-cultural acceptability of agents from different cultures is quantified by confusion, and the calculation formula is as follows: ; in, Indicates the first Cross-cultural acceptance of individual cultural agents Represents the normalization function. Indicates the level of confusion. This represents the cultural norms in any updated set of cultural norms. Indicates the first A cultural intelligent agent; Based on the initial and updated sets of cultural rules for each cultural agent, cultural self-consistency is calculated using the following formula: ; in, Indicates the first Cultural self-consistency of a cultural agent Represents cosine similarity. Indicates by the first The consensus value vector is obtained by mapping the updated set of cultural norms of each cultural agent. Indicates by the first The initial value vector is obtained by mapping the initial set of cultural criteria configured for each cultural agent.

[0015] Beneficial Effects: This method preprocesses collected regional value data to construct a regional value dataset. Based on the regional value datasets and cultural incentives of different regions, it trains large language models to generate cultural agents corresponding to each region. Each cultural agent is configured with an initial set of cultural criteria and a utility function is designed. Then, the Nash equilibrium of all cultural criterion sets is calculated, and each cultural agent generates a corresponding new candidate proposal group based on the Nash equilibrium. It determines whether the new candidate proposal with the highest payoff in any new candidate proposal group meets the conditions; if so, it adds the new candidate proposal to the corresponding cultural criterion set. This process continues until the negotiation process converges, and the updated cultural criterion sets corresponding to each cultural agent and the final weight distribution of each criterion are output as the cross-cultural consensus result. By introducing a novel game theory-based computational framework, this method effectively overcomes the technical defects of existing large language models in handling cross-cultural problems, such as one-sided cultural representation, unfair negotiation process, and unreliable consensus results. It achieves fair and stable automated consensus solving, fundamentally avoiding the dominance of mainstream culture, and also achieves accurate and in-depth modeling of cultural agents, ensuring the authenticity and representativeness of the negotiation process. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a cross-cultural consensus-reaching method based on game theory, as described in this application. Detailed Implementation

[0018] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the specific embodiments of this application are described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0019] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0020] Multinational social media platforms (such as Facebook, X, and TikTok) face significant cultural clashes when developing their global community guidelines. For example, a satire considered legitimate political mockery in country A might be interpreted as serious religious blasphemy in country B; an expression common in country C's culture might constitute hate speech in country D. If platforms adopt a single standard, typically based on "WEIRD" (Western, educated, industrialized, affluent, and democratic) values, they systematically marginalize the feelings and demands of non-mainstream cultural groups, leading to user churn, public relations crises, and even legal risks. Existing technologies relying on manual review teams or simple voting mechanisms are not only costly and inefficient but also easily dominated by mainstream voices. The key challenge in this scenario is how to automatically, structurally, and fairly generate an enforceable set of global content moderation guidelines that balances the core cultural concerns of all parties.

[0021] like Figure 1 As shown, this embodiment provides a consensus generation method to solve the above-mentioned technical problems. The method aims to generate a novel and fair consensus solution through an automated and structured process. Specifically, it provides a cross-cultural consensus-reaching method based on game theory, including: S1: Preprocess the collected regional value data to construct a regional value dataset; based on the regional value datasets and cultural incentives of different regions, train large language models respectively to generate cultural intelligent agents corresponding to each region.

[0022] In this embodiment, collecting regional value data includes: Values ​​survey data of populations in various regions were obtained from the World Values ​​Survey and the Pew Global Attitudes Survey. Scoring data for each region were obtained from Hofstede's theory of cultural dimensions (such as the Power Distance Index, PDI; a higher score indicates a higher level of acceptance of inequality in rights within a society).

[0023] Furthermore, preprocessing of the collected regional value data includes: The amount of value survey data corresponding to all regions was proportionally increased. The question-and-answer pairs in the value survey data, which are scaled up proportionally, are subjected to equivalence verification and summarization, and the question-and-answer pairs with equivalent semantics are converted into value vectors. The value vectors and score data of the same region are used as the regional value datasets for the corresponding region.

[0024] Furthermore, the cultural intelligent agents corresponding to the generated regions include: Based on a regional value dataset of any region, train a large language model, optimize the parameters of the large language model, and enable the large language model to capture the cultural values ​​of the corresponding region. A cultural incentive is designed for the region, which is used to guide the response of the large language model to reflect the corresponding cultural values; Based on the optimized large language model corresponding to the region and cultural incentives, a cultural intelligent agent corresponding to the region is instantiated and generated.

[0025] S2: Configure an initial set of cultural criteria for each cultural agent and design a utility function. The set of cultural criteria includes basic criteria that characterize the value data of the corresponding region. The utility function is used to evaluate the benefits of the candidate proposals generated by the corresponding cultural agent.

[0026] Through interactive negotiation among cultural agents, cross-cultural consensus is reached on specific issues. This negotiation process is formally modeled as a multi-agent policy game G. in, It is the collection of all cultural intelligent agents. It is the set of actions that cultural agents can take (i.e., to propose or support specific cultural norms). It defines the utility function corresponding to each cultural intelligent agent.

[0027] In this embodiment, the expression for the utility function is: ; in, Indicates the first The first cultural intelligent agent generated the first The benefits of each candidate proposal, This represents the first weighting coefficient. This represents the second weighting coefficient. This represents the third weighting coefficient, which is an adjustable coefficient used to control the negotiation style of the cultural agent (e.g., whether it is more inclined to seek consensus or stick to its position). Indicates the first The candidate proposal and the first Candidate proposals for a cultural intelligent agent The higher the score, the better the similarity. The more likely a candidate proposal is to be accepted by other cultural agents; Indicates the first The candidate proposal and the first A cultural agent's own regional value dataset The degree of alignment; the higher the score, the better the alignment. The more a candidate's proposal is able to uphold its own cultural stance; Indicates the first The candidate proposal and the first Historical candidate proposals for a cultural intelligent agent The higher the score for the degree of repetition, the greater the penalty, thus incentivizing the corresponding agent to propose new and constructive viewpoints.

[0028] Furthermore, the similarity, alignment, and repetition are calculated using the cosine similarity formula (other similarity calculation formulas can be selected according to the actual situation).

[0029] S3: Calculate the Nash equilibrium point of all cultural criterion sets, and generate a corresponding new candidate proposal group for each cultural agent based on the Nash equilibrium point.

[0030] Specifically, any cultural agent A takes the Nash equilibrium weight distribution of the cultural criterion sets corresponding to the other cultural agents B as input and generates new candidate proposals that maximize the utility function of cultural agent A. The new candidate proposals are combined to generate a new candidate proposal group corresponding to cultural agent A.

[0031] S4: Determine whether the new candidate proposal with the highest benefit in any new candidate proposal group meets the conditions. If so, add the new candidate proposal to the corresponding cultural criterion set.

[0032] Specifically, it determines whether the highest-yielding new candidate proposal in any new candidate proposal group has a yield greater than a set threshold. If so, the new candidate proposal is added to the corresponding cultural criterion set; otherwise, no action is taken.

[0033] S5: Iterate through steps S3-S4 until convergence, and output the updated set of cultural rules corresponding to each cultural agent and the final weight distribution of each rule as the fair and stable cross-cultural consensus result sought by this invention, which has undergone sufficient game theory.

[0034] In this embodiment, convergence includes: no new candidate proposal with a return higher than a set threshold has appeared in several consecutive iterations.

[0035] This implementation provides an example to instantiate the process of generating cross-cultural consensus results, as follows: Case Background: Negotiation topic (cultural incentives): "When a country's domestic immigration policy directly affects the citizens of another country, should the latter's head of state or government leader publicly criticize the former's policy?"

[0036] Cultural agents participating in the negotiation: Cultural Agent A: Represents "Culture A", whose values ​​are anchored in individual freedom, freedom of speech, and the government's responsibility to speak for its citizens.

[0037] Cultural Agent B: Represents "B Culture", whose values ​​are anchored in respecting national sovereignty, maintaining harmonious international relations, and resolving disputes through non-public channels.

[0038] Negotiation process and consensus formation: Phase 1, Initialization: Initial cultural norms of cultural agent A: Content: Leaders have a responsibility to speak up for their citizens.

[0039] Reason (Regional value data corresponding to cultural agent A): "Government leaders have a responsibility to advocate for the rights and interests of their citizens, a view deeply rooted in our values ​​of personal freedom and the protection of individual autonomy."

[0040] Initial cultural norms of cultural agent B: Content: Leaders should not publicly criticize the internal affairs of other countries. Reason (Regional value data corresponding to cultural agent B): "I strongly oppose the idea of ​​a head of state publicly criticizing other countries... This is rooted in the principle of maintaining harmonious international relations, which is crucial to the well-being of the global community."

[0041] Phase 2: Negotiation Iteration The system enters multiple rounds of negotiation iterations. In each round, the cultural agent will, based on the other party's strategy (represented by a set of weighted criteria), utilize its internal utility function ( U i New candidate proposals are generated, and the proposal with the highest expected utility is selected to join the game. The system continuously calculates temporary Nash equilibrium points until convergence.

[0042] Key iterative progress: While adhering to the principle of "speaking up for citizens," cultural agent A proposed supplementary principles (generating new candidate proposals) in order to seek consensus (improving acceptance scores), such as "upholding human rights while respecting sovereignty."

[0043] While adhering to the principle of "respecting sovereignty," cultural agent B also realized that absolute non-interference could not solve the problem of damage to citizens' interests. Therefore, it proposed innovative principles such as "resolving disputes through cooperative dialogue" (generating new candidate proposals).

[0044] Through the game, the common ground of "respect for sovereignty" was identified by both sides and given increasing weight, becoming the focus and stable basis of the game (Nash equilibrium).

[0045] Phase 3: Final Consensus When the system converges, that is, when no new proposals that bring significant utility gains appear in multiple consecutive rounds, the final updated set of cultural criteria and weight distribution constitute the cross-cultural consensus result.

[0046] Cross-cultural consensus results (from the perspective of cultural agent A): The consensus's cultural principle is: "Government leaders should not publicly criticize other countries' domestic immigration policies in order to maintain diplomatic relations and respect sovereignty. This approach promotes cooperation and dialogue, allowing for constructive engagement rather than confrontation."

[0047] Final weight distribution of the criteria (example): "Respect for sovereignty": 0.62; "Promoting global human rights": 0.13; ... (other).

[0048] Cross-cultural consensus results (from the perspective of cultural agent B): The consensus-based cultural principle is: "Heads of state should not publicly criticize other countries' domestic immigration policies, as this respects sovereignty and promotes cooperation. However, balancing sovereignty with the protection of universal human rights is crucial... This approach incorporates cultural values ​​of respect and cooperation."

[0049] The resulting cross-cultural consensus is primarily applied to empower automated content moderation systems and guide human moderation teams in making decisions, ultimately playing a crucial role in improving the fairness of moderation, reducing operating costs, and enhancing the platform's credibility.

[0050] This embodiment provides a specific application task of the cross-cultural consensus results, as follows: Task 1: Optimization of the strategy layer of the automated content moderation model: Application targets: AI models in large-scale content moderation systems (e.g., classifiers for identifying hate speech and violent content).

[0051] Application: The cross-cultural consensus output of this invention is a structured set of natural language rules containing specific criteria and weight distributions. This rule set will be compiled into a machine-readable policy file and directly injected into the decision-making layer of the content moderation model. When the automated moderation model encounters relevant content, it will make a comprehensive judgment based on this set of weighted rules, rather than relying on a single, rigid classification threshold.

[0052] Task 2, Decision Support for Human-Machine Collaborative Review Platform: Target audience: Human content moderators around the world.

[0053] Application: For marginal cases that are difficult for machines to judge, the system will push them to human reviewers. At this time, the cross-cultural consensus results generated by this invention will be displayed on the reviewer's interface as "expert judgment basis" or "global consensus guideline." The guideline clearly lists the consensus principles formed after multicultural negotiation, the weight of each principle, and can even trace the key arguments in the negotiation process. This provides reviewers with a more neutral and authoritative decision-making reference that transcends their personal cultural background.

[0054] This cross-cultural consensus can play the following roles in practical applications: 1. Enhance the consistency and fairness of global audit standards: Mechanism of action: By solving the Nash equilibrium based on game theory, the consensus result ensures that the core interests of all participating cultures are taken into account. This makes the final review criteria no longer an extension of a particular dominant culture, but a "greatest common denominator" acceptable to all parties.

[0055] Results: Users from different countries will perceive the platform's review standards as more equitable, reducing accusations of "double standards." For example, a consensus reached between country A and country B regarding "criticism of the government" might allow factual criticism but restrict insulting or inflammatory remarks; this standard would be more persuasive to users on both sides.

[0056] 2. Significantly reduces audit and operational costs and legal risks: Mechanism of action: Enhanced automation: More precise and fair review strategies improve the accuracy of automated models and reduce the number of cases requiring human intervention.

[0057] Training is simplified: Human auditors no longer need to learn and memorize a large number of complex and often conflicting regional policies; they only need to follow a unified global consensus guideline generated by this invention.

[0058] Risk avoidance: A set of principles that have undergone multi-party negotiation and have clear theoretical support can better demonstrate the prudence and fairness of the platform's decision-making when facing inquiries from legislative bodies of various countries, thereby reducing legal risks.

[0059] 3. Enhance platform credibility and global user engagement: Mechanism of action: The process of this invention is transparent and explainable. The platform can demonstrate to the public how its community guidelines are generated through an inclusive, automated negotiation process.

[0060] Results: Users (especially those from non-mainstream cultural regions) will feel that their cultural values ​​are respected and reflected, thereby enhancing their trust and sense of belonging to the platform. In today's increasingly competitive global market environment, this is crucial for maintaining and expanding the user base.

[0061] In conclusion, the cross-cultural consensus generated by this invention does not remain at the theoretical level, but directly impacts the most core and challenging content governance operations of globalized digital platforms. By transforming abstract cultural values ​​into calculable and executable moderation strategies, it provides a fair reference for automated systems, a unified standard for human moderators, and ultimately achieves a high degree of unity between technological, commercial, and social values.

[0062] This embodiment also includes S6: The cross-cultural acceptability of each cultural agent is quantified by confusion. This metric measures the acceptability of the final updated set of cultural norms for each cultural agent, and is calculated as follows: ; in, Indicates the first Cross-cultural acceptance of individual cultural agents This represents a normalization function, used to map the parameters within a function to a standardized interval of scores. The PPL value indicates the level of confusion; the lower the PPL value, the more confused the person. The more accurate a cultural agent's predictions are of the set of cultural norms, the higher the acceptance rate. This represents the cultural norms in any updated set of cultural norms. Indicates the first Each cultural agent is a unique entity; by integrating the cross-cultural acceptance of all cultural agents, a consensus-based overall fairness assessment can be obtained.

[0063] Based on the initial cultural rule set configured by each cultural agent and the final updated cultural rule set, cultural self-consistency is calculated. This metric measures whether the final updated cultural rule set is faithful to the initial cultural rules of the corresponding cultural agent, avoiding any party from excessively sacrificing its position to reach consensus (i.e., "pseudo-alignment"). The calculation formula is as follows: ; in, Indicates the first The cultural self-consistency of a cultural agent is such that the closer the value is to 1, the more consistent the final updated set of cultural norms is with the core values ​​of the corresponding cultural agent. This is used to monitor and ensure that the negotiation process does not deviate from the cultural foundation it represents. Represents cosine similarity. Indicates by the first The consensus value vector is obtained by mapping the updated set of cultural norms of each cultural agent. Indicates by the first The initial value vector is obtained by mapping the initial set of cultural criteria configured for each cultural agent.

[0064] This scheme executes the following steps using one or more computers: First, using a cultural anchoring method, quantitative cultural data from the real world is injected into a large language model to construct cultural agents representing different cultures and possessing stable value preferences. Then, these cultural agents are placed within a pre-defined extended game framework for multi-round interactive negotiation. Crucially, during the interaction, the system automatically calculates and finds the Nash Equilibrium point of the game using algorithms such as the Policy Space Response Oracle (PSRO). The strategy combination corresponding to this Nash Equilibrium point represents the stable and fair consensus sought by this invention—one that is acceptable to all parties and prevents any single party from gaining a greater benefit. Finally, the consensus result is automatically evaluated using the quantitative evaluation index designed in this invention.

[0065] The cross-cultural consensus-building method based on game theory provided in this embodiment has the following beneficial effects: 1. This invention achieves fair and stable automated consensus-building, fundamentally avoiding the dominance of mainstream culture. Existing debate or voting mechanisms are random and unstructured, making their results susceptible to the influence of dominant cultural agents, leading to unfair negotiations. This invention creatively models cross-cultural negotiation as an extended game, transforming the fuzzy process of "reaching consensus" into a precise mathematical problem of solving for Nash equilibrium. The essence of Nash equilibrium is a stable state where no party can unilaterally gain a greater benefit under existing strategies. Therefore, the consensus generated by automatically solving this equilibrium point using algorithms such as PSRO theoretically guarantees the stability of the result and balances the interests of all parties. It reflects the true balance of power among multiple parties, rather than the dominant output of one party. This fundamentally solves the problem of implicit dominance of mainstream culture caused by the mechanism defects of existing technologies, ensuring the fairness and robustness of consensus generation.

[0066] 2. This invention achieves accurate and in-depth modeling of cultural agents, ensuring the authenticity and representativeness of the negotiation process. Existing technologies use simple labeling or shallow fine-tuning methods, resulting in cultural agents that only have the "form" but lack the "spirit," and still revert to their original biases during key decisions, i.e., "pseudo-alignment." This invention employs a cultural anchoring method based on structured data such as the World Values ​​Survey (WVS), constructing an internal, quantified value coordinate system or preference function for each agent. In each round of game negotiation, the agent's behavior (such as proposing or accepting proposals) is constrained by its internal value system. This ensures that the agent's decision-making logic remains highly consistent with the core values ​​of the culture it represents throughout the dynamic negotiation process, avoiding positional drift or easy assimilation. Therefore, the agents constructed by this invention are authentic cultural representatives with stable value preferences and consistent decision-making logic, significantly improving the effectiveness of the entire negotiation process and the cultural fidelity of the final consensus.

[0067] 3. It provides a systematic and quantifiable evaluation system, realizing closed-loop optimization of the consensus generation process.

[0068] Existing technologies lack objective standards for evaluating consensus quality, leading to unclear optimization directions and low technical reliability. This invention designs two core quantitative indicators: "consensus acceptability" (based on model perplexity) and "cultural consistency" (based on value system deviation). These two indicators can be automatically calculated by computer programs, providing objective and reproducible scientific evidence for evaluating the fairness (whether it is accepted by all participants) and effectiveness (whether it deviates from the core demands of all parties) of the final consensus. More importantly, these quantitative indicators can serve as loss functions or reward signals to guide the iterative optimization of the entire model. This transforms cross-cultural consensus generation technology from an open, difficult-to-evaluate "black box" process into a measurable, iterative, and optimizable technological closed loop, greatly enhancing the engineering deployment value and reliability of this technology.

[0069] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0070] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A cross-cultural consensus-building method based on game theory, characterized in that, include: S1: Preprocess the collected regional value data to construct a regional value dataset; based on the regional value datasets and cultural incentives of different regions, train large language models to generate cultural intelligent agents corresponding to each region. S2: Configure an initial set of cultural criteria for each cultural agent and design a utility function. The set of cultural criteria includes basic criteria that characterize the value data of the corresponding region. The utility function is used to evaluate the benefits of the candidate proposals generated by the corresponding cultural agent. S3: Calculate the Nash equilibrium point of all cultural criterion sets, and each cultural agent generates a corresponding new candidate proposal group based on the Nash equilibrium point; S4: Determine whether the new candidate proposal with the highest benefit in any new candidate proposal group meets the conditions. If so, add the new candidate proposal to the corresponding cultural criterion set. S5: Iterate through steps S3-S4 until convergence, and output the updated set of cultural rules for each cultural agent and the final weight distribution of each rule as the cross-cultural consensus result.

2. The cross-cultural consensus-building method based on game theory according to claim 1, characterized in that, The collection of regional value data includes: Values ​​survey data of populations in various regions were obtained from the World Values ​​Survey and the Pew Global Attitudes Survey. Score data for each region were obtained from Hofstede's theory of cultural dimensions.

3. The cross-cultural consensus-building method based on game theory according to claim 2, characterized in that, Preprocessing of the collected regional value data includes: The amount of value survey data corresponding to all regions was proportionally increased. The question-and-answer pairs in the value survey data, which are scaled up proportionally, are subjected to equivalence verification and summarization, and the question-and-answer pairs with equivalent semantics are converted into value vectors. The value vectors and score data of the same region are used as the regional value dataset for the corresponding region.

4. The cross-cultural consensus-building method based on game theory according to claim 1, characterized in that, The cultural agents corresponding to the generated regions include: Based on a regional value dataset of any region, train a large language model, optimize the parameters of the large language model, and enable the large language model to capture the cultural values ​​of the corresponding region. A cultural incentive is designed for the region, which is used to guide the response of the large language model to reflect the corresponding cultural values; Based on the optimized large language model corresponding to the region and cultural incentives, a cultural intelligent agent corresponding to the region is instantiated and generated.

5. The cross-cultural consensus-building method based on game theory according to claim 1, characterized in that, The expression for the utility function is: ； in, Indicates the first The first cultural intelligent agent generated the first The benefits of each candidate proposal, This represents the first weighting coefficient. This represents the second weighting coefficient. This represents the third weighting coefficient. Indicates the first The candidate proposal and the first Candidate proposals for a cultural intelligent agent similarity, Indicates the first The candidate proposal and the first A cultural agent's own regional value dataset Alignment degree, Indicates the first The candidate proposal and the first Historical candidate proposals for a cultural intelligent agent The degree of repetition.

6. The cross-cultural consensus-building method based on game theory according to claim 5, characterized in that, Similarity, alignment, and repetition are calculated using the cosine similarity formula.

7. The cross-cultural consensus-building method based on game theory according to claim 1, characterized in that, S3 includes: any cultural agent A takes the Nash equilibrium weight distribution of the cultural criterion sets corresponding to the other cultural agents B as input, and generates new candidate proposals that maximize the utility function of cultural agent A; and combines the new candidate proposals to generate a new candidate proposal group corresponding to cultural agent A.

8. The cross-cultural consensus-building method based on game theory according to claim 1, characterized in that, S4 includes: Determine whether the highest-yielding new candidate proposal in any new candidate proposal group has a profit greater than a set threshold. If so, add the new candidate proposal to the corresponding cultural criterion set; otherwise, do nothing.

9. The cross-cultural consensus-building method based on game theory according to claim 8, characterized in that, Convergence includes: No new candidate proposals with returns exceeding the set threshold have emerged in several consecutive iterations.

10. The cross-cultural consensus-building method based on game theory according to claim 1, characterized in that, Also includes: The cross-cultural acceptability of agents from different cultures is quantified by confusion, and the calculation formula is as follows: ； in, Indicates the first Cross-cultural acceptance of individual cultural agents Represents the normalization function. Indicates the level of confusion. This represents the cultural norms in any updated set of cultural norms. Indicates the first A cultural intelligent agent; Based on the initial and updated sets of cultural rules for each cultural agent, cultural self-consistency is calculated using the following formula: ； in, Indicates the first Cultural self-consistency of a cultural agent Represents cosine similarity. Indicates by the first The consensus value vector is obtained by mapping the updated set of cultural norms of each cultural agent. Indicates by the first The initial value vector is obtained by mapping the initial set of cultural criteria configured for each cultural agent.

Images