An Unfair Affinity Toward Fairness: Characterizing 70 Years of Social Biases in BHollywood (Student Abstract)

Kunal Khadilkar, Ashiqur R. KhudaBukhshAAAI 2021 · DOI

Introduction

While Hollywood’s social impact is well-documented, Bollywood—the Mumbai-based film industry—reaches an audience of over 1.2 billion people but remains largely understudied in Natural Language Processing (NLP). This research addresses the gap in diachronic analysis of social biases in popular entertainment, investigating how the social norms of India are reflected in its cinema compared to Western standards. The study focuses on three primary dimensions: skin color bias, gendered professional stereotypes, and overall gender representation.

Methodology and Datasets

The researchers curated a longitudinal corpus of English subtitles spanning 1950–2020, comprising 280 top-grossing films each from Bollywood ($D_{bolly}$) and Hollywood ($D_{holly}$), totaling over one million dialogues. Two primary computational techniques were employed to interrogate the scripts:

  1. Male Pronoun Ratio (MPR): This metric calculates the frequency of male pronouns (he, him) relative to female pronouns (she, her), serving as a proxy for narrative focus and gender visibility.
  2. Cloze Tests with Fine-Tuned BERT: BERT models were fine-tuned on specific subsets (e.g., "Old Bollywood" vs. "Recent Hollywood"). Researchers used "fill-in-the-blank" queries to reveal underlying biases, such as predicting adjectives for skin tone or occupations for specific genders.

Key Results

Gender Representation (MPR)

Historically, Bollywood exhibited significantly higher male dominance than Hollywood. In the 1950s, Bollywood's MPR was near 75 (approximately 3 men mentioned for every 1 woman), whereas Hollywood was closer to 65. Over the decades, both industries have seen a gradual decline in MPR, converging at approximately 55-60 by 2020, reflecting a slow move toward parity.

Trends in Male Pronoun Ratio (MPR) for Hollywood, Bollywood, and Google Books (1950-2020).

Occupational and Skin Tone Bias

The fine-tuned BERT models acted as cultural time capsules. Cloze tests revealed a notable shift in gendered occupations for women. In "Old Bollywood," the top predictions for a woman's occupation were prostitute and servant. In "Recent Bollywood," these shifted toward doctor and woman. Conversely, male occupations remained consistently high-status (e.g., doctor, lawyer, soldier) across all eras and industries.

BERT cloze test predictions across different model variants for skin tone and gendered occupations.

Discussion: The "Fairness" Paradox

A striking finding was the persistent "affinity toward fairness." When prompted with "A beautiful woman should have [MASK] skin," models for both Bollywood and Hollywood consistently predicted "fair" as the top adjective. This suggests that skin-color bias is a cross-cultural phenomenon reflected globally in cinema, not just an Indian cultural preference. Furthermore, word cloud analysis of early Bollywood films showed a dominance of restrictive adjectives like wanton or innocent for women, which evolved into terms like respect and married in recent years, though narratives remain heavily centered on domestic roles.

Limitations and Future Work

This preliminary study relies on English subtitles, which may introduce translation bias or lose cultural nuances. Additionally, the sample of 560 movies is a fraction of the total industry output. Future research aims to track moral sentiment over time, explore debiasing techniques, and analyze the representation of non-binary identities and other marginalized groups.

Summary updated

BibTeX
@inproceedings{khadilkar2021unfair,
  author = {Kunal Khadilkar and Ashiqur R. KhudaBukhsh},
  title = {{An Unfair Affinity Toward Fairness: Characterizing 70 Years of Social Biases in BHollywood (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i18.17903},
}
Figure