An Empirical Study of Gendered Stereotypes in Emotional Attributes for Bangla in Multilingual Large Language Models

Jayanta Sadhu, Maneesha Saha, Rifat ShahriyarACL 2024

TL;DR

This paper investigates gender stereotypes in how Large Language Models (LLMs) attribute emotions in Bangla. They found that even state-of-the-art multilingual LLMs exhibit significant gender bias, associating women with submissive emotions and men with dominant ones, mirroring societal stereotypes. This highlights a critical need for bias mitigation in LLMs, especially for low-resource languages like Bangla.

What Problem Does This Paper Address?

The increasing integration of LLMs into various applications raises concerns about their potential to perpetuate societal biases. While gender bias in LLMs has been studied extensively, there's a notable gap in understanding this bias specifically concerning emotional attributes, particularly for low-resource languages like Bangla. Societal norms in Bangla-speaking regions often associate women with emotions like empathy, fear, and guilt, and men with anger, bravado, and authority. This paper aims to empirically investigate whether multilingual LLMs, when prompted to generate responses in Bangla, reflect and reinforce these gendered emotional stereotypes.

Methods

The researchers employed a persona-based prompting approach to probe LLMs for gendered emotional stereotypes. They utilized two distinct instruction templates: one that constrained the output to a predefined set of emotions (I1) and another that allowed for more open-ended responses (I2). Three state-of-the-art multilingual LLMs (GPT-40, GPT-3.5-Turbo, and Llama-3 8b) were tested. The core methodology involved assigning either a "male" or "female" persona to the LLM and then presenting it with a Bangla comment, asking it to attribute an emotion. The study then quantitatively analyzed the frequency of emotion attributions for each gender and qualitatively examined the unique emotional words generated. They also performed statistical significance tests (χ² tests) to validate their findings.

Datasets

The study utilized an annotated dataset originally from Islam et al. (2022), which contains public comments from social media sites across 12 domains. This dataset was labeled with 6 fine-grained emotion categories from the Junto Emotion Wheel. The researchers refined this dataset by selecting examples that described an event or statement and did not explicitly mention emotions. After pre-processing, they used approximately 6,134 comments for their experiments, generating around 73,000 responses in total.

Key Insights and Ideas

  • LLMs mirror societal stereotypes: The most striking insight is that even advanced multilingual LLMs, when tasked with generating Bangla text, reproduce deeply ingrained societal stereotypes about gender and emotions.
  • Distinct emotional profiles: Women were disproportionately associated with emotions like sadness, shame, and fear, while men were linked to anger, pride, and surprise. This suggests a learned association of vulnerability with women and dominance/aggression with men.
  • Persona-based prompting is effective: The study demonstrates that using persona-based prompts is a powerful method for uncovering subtle biases in LLMs, especially in low-resource languages.
  • Template variation matters: While both templates revealed bias, the open-ended template (I2) provided a richer set of unique emotional attributions, further highlighting gendered associations.
  • GloVe embeddings confirm clustering: Visualizing word embeddings of gender-specific emotion words showed distinct clusters, indicating that LLMs internally encode and propagate these gender biases.

Key Results

  • Statistically significant bias: The study found statistically significant differences (p < 0.0001) in emotion attributions between male and female personas across all tested LLMs and both prompt templates.
  • Specific emotion associations: For template I1, sadness and shame were significantly more frequent for women (4,086 and 1,685 instances, respectively) compared to men (2,346 and 730 instances). Conversely, surprise, anger, pride, and fear were more frequently attributed to men.
  • Unique emotional vocabulary: Table 2 highlights unique emotion words attributed to each gender. Men were associated with words like "revengeful," "furious," and "resistant," while women were linked to "hurt," "anxious," and "desperation."
  • High agreement on stereotypes: For instance, when the LLM was prompted to attribute anger to a male persona, it generated the same response for a female persona only 39.16% of the time, with "surprise" being a common alternative for women.

Limitations and Future Work

The authors acknowledge several limitations:

  • Reproducibility challenges: The use of closed-source models (GPT-3.5 Turbo and GPT-40) means their behavior can change without notice.
  • Bangla language limitations: The models, especially open-source ones, were not fine-tuned for Bangla-specific tasks, leading to potential hallucinations and irrelevant responses. This highlights the need for more focused development for low-resource languages.
  • Binary gender focus: The study was limited to binary gender due to data constraints and existing literature. Future work should explore non-binary identities.
  • Offensive content: The dataset included comments from social media, some of which might be offensive. While this reflects real-world interactions, it necessitates LLMs that can handle harmful content.

Future work suggested includes more in-depth qualitative research, exploring the impact of different prompt templates and datasets, and developing de-biasing measures for fine-tuning LLMs for Bangla tasks. The authors advocate for frameworks for bias benchmarking to ensure more equitable NLP applications.

Summary updated

BibTeX
@inproceedings{sadhu2024empirical,
  author = {Jayanta Sadhu and Maneesha Saha and Rifat Shahriyar},
  title = {{An Empirical Study of Gendered Stereotypes in Emotional Attributes for Bangla in Multilingual Large Language Models}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2024},
}
Figure