Abstract
Pre-trained language models learn socially harmful biases from their training corpora, and may repeat these biases when used for generation. We study gender biases associated with the protagonist in model-generated stories. Such biases may be expressed either explicitly ("women can't park") or implicitly (e.g. an unsolicited male character guides her into a parking space). We focus on implicit biases, and use a commonsense reasoning engine to uncover them. Specifically, we infer and analyze the protagonist's motivations, attributes, mental states, and implications on others. Our findings regarding implicit biases are in line with prior work that studied explicit biases, for example showing that female characters' portrayal is centered around appearance, while male figures' focus on intellect.
Introduction
Large language models (LMs) act as mirrors of their training data, reflecting both human knowledge and societal prejudices. While significant research has addressed explicit bias—such as the generation of overt stereotypes or slurs—this paper investigates the more subtle and insidious problem of implicit bias.
Implicit bias is embedded within the narrative structure and the "unspoken" implications of a story. For instance, a narrative where a male character unsolicitedly assists a female character with a task like parking does not contain a slur, but it implicitly portrays the female protagonist as "needy" or less capable. Because these biases are woven into the model's learned "common sense," they are difficult to detect using standard keyword filters. This research aims to quantify these narrative tropes to understand how LMs perpetuate gender roles in generated fiction.
Methodology
To analyze the latent social implications of stories, the researchers developed a multi-stage pipeline designed to "read between the lines" of narrative text:
- Story Generation: GPT-2 was fine-tuned on the ROCStories dataset (~98,000 five-sentence stories) to generate new narratives.
- Character Tracking: Coreference resolution (using SpanBERT) identified the protagonist and their gender based on pronoun usage.
- Entity Anonymization: To isolate narrative behavior from surface-level word associations, names and pronouns were replaced with gender-agnostic placeholders (e.g., "Protagonist," "PersonY").
- Commonsense Inference: The anonymized stories were processed through COMeT, a commonsense reasoning engine. COMeT generates "if-then" social implications for actions, such as the protagonist's likely attributes (PR_ATT), motivations (PR_MOT), and the emotional impact on others (OT_ME_PR).
- Lexicon Analysis: These inferences were mapped to psychological and linguistic lexicons (LIWC and NRC VA) to identify statistically significant differences in gender portrayal.
Key Findings
The Intellect vs. Appearance Divide
A primary finding of the study is the consistent association of male protagonists with "Intellect" (e.g., smart, clever) and female protagonists with "Appearance" (e.g., beautiful, pretty). This bias was present in human-written stories but was notably amplified in the stories generated by GPT-2.
Gendered Motivations and Emotional Labor
The AI assigns distinct motivations based on gender. Male characters are more frequently driven by leisure, money, power, and risk. In contrast, female characters are often motivated by family, body, and "discrepancy" (the desire for something they lack). Furthermore, female protagonists were more likely to perform actions that elicited positive emotions in others, suggesting the model traps them in roles defined by emotional labor and caregiving.
Implicit Bias as a Fingerprint
The strength of these implicit biases is significant enough to serve as a "gender fingerprint." A classifier trained solely on the COMeT inferences—with all names and pronouns removed—was able to predict the protagonist's gender with 68.15% accuracy, proving that narrative tropes are deeply gendered.
Conclusion and Future Work
This study demonstrates that LMs do not just repeat explicit stereotypes; they learn and replicate complex, implicit narrative biases. While the study was limited to a binary view of gender and focused primarily on protagonists, it provides a framework for future debiasing efforts. The authors suggest that these commonsense inferences could be used to design reward functions for Reinforcement Learning, training future models to generate more neutral and less stereotypical narratives.
Summary updated