Cross-Cultural Survey Simulation Based on Calibrated Value-Based Personality: Reducing Prediction Errors for Underrepresented Groups

This paper proposes a value-based personality construction method. By extracting core cultural dimensions from survey responses and calibrating response diversity, it significantly reduces prediction errors in cross-cultural survey simulations, especially improving simulation performance for underrepresented groups. This method has important application value in scenarios such as market research, policy evaluation, survey design optimization, and social science research, and helps build more fair and inclusive AI systems.

Large Language Models (LLMs) are widely used in survey simulations, including scenarios like market research (simulating crowd reactions before product launch), policy evaluation (predicting policy acceptance), survey design optimization (testing questionnaires), and social science research (exploring hypotheses). However, cross-cultural simulations with LLMs have limitations: they reflect the dominant cultural perspective in training data, perform poorly in simulating underrepresented groups, and may lead to issues such as cultural blind spots in global products, biases in policy-making, and distorted research conclusions.

Existing personification prompting methods rely on indirect proxy variables such as sociodemographic characteristics or the Big Five personality traits, which have the following problems: lack of values (values, not demographic characteristics, truly shape opinions), simplified cultural dimensions (unable to capture deep cultural dimensions like individualism/collectivism), and distortion of underrepresented groups (amplifying biases in training data).

Core ideas: Value priority (as the core dimension of personality), data-driven (learning relationships from actual survey responses), cultural dimension mapping (mapping to frameworks like Hofstede). Specific steps: 1. Value extraction (select survey questions reflecting deep values, analyze response patterns and map to cultural dimensions); 2. Text description generation (natural language expression, contextualized examples, comparative explanations); 3. Personality sampling and aggregation (sample from the value distribution of the target group, aggregate group-level predictions after multi-personality simulation).

LLM simulation responses have insufficient diversity issues: excessive consensus (underestimating marginal opinions), underestimated variance (distribution variance smaller than reality), and missing extreme values. Calibration strategies: diversity enhancement (adjusting sampling and temperature parameters), distribution matching (matching real data distribution characteristics), and opinion preservation (not distorting the average opinion of the group). Calibration effects: response distribution is closer to reality, capturing extreme values and long tails, increasing authenticity while maintaining accuracy.

Evaluation setup: Using representative survey data from multiple countries, predicting various issues such as policy attitudes and social values, and comparing with existing demographic personality methods. Core results: Overall prediction errors are reduced, with the largest improvement for underrepresented groups, and the performance gap between mainstream and marginal groups is narrowed. Specific findings: Mild improvements in high-representation countries (US, UK), over 50% reduction in errors in low-representation countries (some African and Asian countries), and significant improvements in cultural dimensions like power distance.

1. Value shift: Treat values as the core of personality construction, collect value data, understand cultural dimensions, and cross-culturally validate the effectiveness of the method; 2. Balance diversity and accuracy: Pay attention to opinion distribution, quantify uncertainty, and capture extreme views; 3. Fairness and inclusiveness: Emphasize underrepresented groups, use technology to mitigate data biases, and continuously monitor differences in simulation performance.

Limitations: Complexity of value measurement, cultural dimension selection based on classic frameworks, dynamic changes of values over time, and causal relationships needing in-depth analysis. Future directions: Explore other cultural theory frameworks, develop automated value extraction methods, study longitudinal survey applications, and explore multilingual cross-cultural simulations.