Social Bias Benchmark for Generation: A Comparison of Generation and QA-Based Evaluations

2025-03-10

AI Summary

Existing social bias evaluation methods are ill-suited for long-text generation tasks. Method: We introduce BBG, the first bilingual (English/Korean) benchmark for bias evaluation in long-text generation. Inspired by QA-style bias benchmarks (e.g., BBQ), we extend the paradigm to story continuation by designing bias-sensitive bilingual story prompts via prompt engineering; we then quantify model probabilities of generating neutral versus biased continuations and systematically compare generative versus discriminative (multiple-choice) evaluation outcomes. Contribution/Results: Experiments across 10 mainstream LLMs reveal a 37–62% inconsistency rate between generative and discriminative evaluations, exposing critical limitations of prevailing discriminative assessment approaches. BBG establishes a more realistic, task-aligned paradigm for LLM bias evaluation and provides the first empirical benchmark specifically designed for long-text generation scenarios.

Technology Category

• Natural Language Processing: Generation
• Machine Learning: Ethics, Bias, and Fairness
• Philosophy and Ethics of AI: Bias, Fairness & Equity

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Social Networks and Social Media: Fairness and bias in social network and social media analysis
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Measuring social bias in large language models (LLMs) is crucial, but existing bias evaluation methods struggle to assess bias in long-form generation. We propose a Bias Benchmark for Generation (BBG), an adaptation of the Bias Benchmark for QA (BBQ), designed to evaluate social bias in long-form generation by having LLMs generate continuations of story prompts. Building our benchmark in English and Korean, we measure the probability of neutral and biased generations across ten LLMs. We also compare our long-form story generation evaluation results with multiple-choice BBQ evaluation, showing that the two approaches produce inconsistent results.

Problem

• Assess social bias in long-form generation by LLMs
• Compare generation-based and QA-based bias evaluations
• Measure bias in English and Korean LLM outputs

Innovation

• Adapted BBQ for long-form generation bias evaluation
• Used story prompts to measure LLM social bias
• Compared generation and QA-based bias evaluation results