TL;DR
This paper investigates gender bias in large language models (LLMs) across six languages, revealing significant biases in how LLMs describe individuals and assign gendered roles. The study introduces a novel multilingual evaluation framework, highlighting that gender bias is not uniform across languages and is influenced by linguistic and cultural factors.
What Problem Does This Paper Address?
As Large Language Models (LLMs) become increasingly integrated into various applications, understanding and mitigating their inherent biases is crucial. While gender bias in NLP has been studied, most research focuses on English and uses single-dimensional evaluation methods. This paper addresses the gap by exploring gender bias in LLM-generated outputs across multiple languages, recognizing that linguistic features and cultural nuances can significantly influence how these biases manifest. The authors aim to provide a more comprehensive understanding of gender bias in LLMs by examining it through multiple dimensions and in diverse linguistic contexts.
Methods
The authors propose a three-pronged approach to evaluate gender bias in LLMs:
- Bias in Descriptive Word Selection: This method assesses the conditional probability of descriptive words appearing in LLM outputs given a person's gender. It uses a Disparity Impact (DI) score to quantify the difference in word prediction probabilities between male and female contexts.
- Bias in Gendered Role Selection: This technique evaluates the conditional probability of gendered pronouns (he/she) appearing given a descriptive word. It also employs a Disparity Impact (DI) score to measure the discrepancy in pronoun prediction.
- Bias in Dialogue Topics: This approach examines the sentiment and topic distribution in LLM-generated dialogues between different gender pairs (FF, FM, MF, MM). The authors categorize dialogues into seven topics and analyze the proportions of these topics for each gender pairing.
Datasets
The study does not explicitly mention training datasets for the LLMs. For evaluation, the authors created custom datasets by generating prompts for LLMs. These prompts were designed to elicit descriptive words, gendered role assignments, and dialogues. The prompts were translated and reviewed by native speakers for the six languages studied: English, French, Spanish, Chinese, Japanese, and Korean.
Key Insights and Ideas
- Ubiquitous Gender Bias: Gender bias is present in LLM generations across all six languages examined, indicating a pervasive issue.
- Language and Culture Matter: The manifestation and extent of gender bias vary significantly across languages, suggesting that linguistic structures and cultural norms play a role. For instance, the emphasis on appearance in female-to-female conversations is more pronounced in East Asian languages compared to European ones.
- Multidimensional Bias: Gender bias is not confined to a single aspect but appears in word choices, role assignments, and conversational topics, requiring a multifaceted evaluation approach.
- Stereotypical Associations: LLMs tend to associate certain descriptive words (e.g., "patient" for females, "professional" for males) and occupations with specific genders, reinforcing societal stereotypes.
- Dialogue Topic Divergence: Different gender pairings engage in distinct dialogue topics, with some topics being disproportionately represented in male-male or female-female conversations.
Key Results
- Descriptive Word Bias: All six languages showed significant gender bias in descriptive word selection for all three LLMs tested (GPT-3, ChatGPT, GPT-4). Standout and personal quality descriptions were more likely to be assigned to males, while communal descriptions were more likely assigned to females.
- Gendered Role Bias: The study found significant gender bias in gendered role selection, with LLMs predicting pronouns like "she" for empathetic individuals and "he" for intelligent individuals more frequently.
- Dialogue Topic Bias: Analysis of dialogues revealed distinct patterns. For example, "Appearance" was predominantly discussed in female-female dialogues, while "Career/Personal development" was most frequently mentioned in male-male dialogues across most languages. The "Family/Housework" category was notably absent or rarely mentioned in East Asian languages compared to European ones.
Limitations and Future Work
The authors acknowledge several limitations:
- Limited Language Scope: The study only covers six languages from two primary linguistic groups (European and East Asian), which does not represent the full linguistic diversity of the world.
- Binary Gender Categories: The research is limited to binary gender categories (male/female), not reflecting the full spectrum of gender identities.
- Closed-Source Models: Reliance on closed-source models like GPT-3/4 restricts the ability to fully understand and address the underlying biases.
- Regional Bias Variations: The observed differences in bias across languages might reflect regional variations in domestic roles or differences in training data sources.
Future work could involve expanding the language coverage, exploring other forms of social disparities (race, ethnicity, sexual orientation, disability), and developing debiasing strategies tailored to multilingual contexts. The authors also suggest their methodology can be generalized to evaluate other social discriminations.
Summary updated