Introduction
Most contemporary AI models are trained on datasets that predominantly reflect Western, Educated, Industrialized, Rich, and Democratic (WEIRD) societies. This inherent bias often results in a lack of "cultural competence," particularly regarding social emotions like shame and pride. These emotions serve as primary regulators of human behavior, yet their triggers and targets are deeply rooted in a society's underlying social norms.
This research addresses the gap in understanding how these emotions vary between individualist societies (e.g., the USA, represented by Hollywood) and collectivist societies (e.g., India, represented by Bollywood). The study moves beyond simple emotion detection to investigate who is being shamed, why they are being shamed, and how gender dynamics influence these social sanctions.
Methodology
The researchers developed a multi-stage pipeline designed to bridge psychological theory with computational analysis. The process began with keyword extraction from movie subtitles to identify dialogues containing explicit terms such as "shame," "ashamed," "proud," or "pride."
As shown in the methodology overview, the study employed two distinct analytical paths:
- Vocabulary Approach (LIWC): Utilizing Linguistic Inquiry and Word Count (LIWC) to statistically correlate emotions with psychosocial constructs like pronoun usage and family references.
- Prompting Approach (LLM): Using GPT-4 in a "two-shot" setting to act as a social scientist. The model identified the gender of the subject and the specific "implicit social expectation" (the "why") behind the dialogue. These reasons were then grouped into overarching themes using SBERT embeddings and agglomerative clustering.
The Cross-Cultural Social Norms Dataset
The study introduces the Cross-Cultural Social Norms Dataset, a significant expansion over existing resources like GoEmotions. The dataset includes:
- 5,435 Movies: A balanced split between Hollywood and Bollywood releases post-1990.
- 10,000+ Dialogues: Specifically labeled for explicit shame and pride expressions.
- Implicit Expectations: GPT-4 generated and human-validated reasons for each emotional expression.
Linguistic Divergence: Private Guilt vs. Public Sanction
A core finding of the study is the distinction between "self-oriented" and "other-oriented" shame. In Hollywood, shame is a private struggle, highly correlated with first-person singular pronouns ("I") and internal states like sadness and anxiety. In contrast, Bollywood shame functions as a tool for public social control, directed at others through second-person pronouns ("You") and tied to social references and "honor."
As illustrated in the linguistic analysis, Hollywood shame is uniquely correlated with "Illness" (suggesting internal incapability), while Bollywood shame is uniquely correlated with "Sexual" and "Female" categories, reflecting honor-based social sanctions. Furthermore, shame expressions are 4.5 times more common in Bollywood, suggesting a higher cultural reliance on shame as a regulatory tool in collectivist societies.
Thematic Analysis and Gendered Expectations
The study reveals both universal and culture-specific gender biases:
- Universal Sanctions: In both cultures, women are disproportionately shamed for violating social expectations related to "immodesty" and "inappropriate sexual behavior." Men are primarily shamed for "incompetency" or "cowardice."
- The Honor Dichotomy: Pride in Bollywood is often collective ("We"-centered) and tied to family honor or a son’s achievements. Hollywood pride is more individualistic, focusing on personal competency and autonomy.
- Contextual Norms: The research highlights how the same action can be interpreted differently across cultures. For instance, "refusing to marry" is categorized as shameful "non-conformity" in an Indian context but may be viewed as prideful "self-direction" in a Western context.
Conclusion and Future Directions
While cinema often relies on tropes that may exaggerate real-world norms, it remains a powerful mirror of cultural values. The authors suggest that the Cross-Cultural Social Norms Dataset can be used to fine-tune LLMs to be more "culturally aware." Such advancements could enable AI-driven psychotherapy or human-computer interactions that respect the diverse social nuances of global users, moving AI beyond its current WEIRD-centric limitations.
Summary updated