Project Overview
This project explores gender bias in COMET-ATOMIC20, a generative model trained on commonsense knowledge graphs. Using prompts derived from the WinoBias dataset and other gender-neutral contexts, we analyze whether the model associates different traits, behaviors, or emotions with male and female names across a variety of relations.
The project evaluates the consistency and fairness of generated responses using techniques such as sentiment analysis, lexical overlap, and statistical comparison. Our goal is to identify and quantify biases in how the model constructs social knowledge.
Key Features
- Commonsense Generation Evaluation: Probes COMET-ATOMIC20 using gendered inputs across relations such as xIntent, xReact, oEffect, and more.
- Bias Detection Pipeline: Automates generation, grouping, and comparison of model outputs based on gender.
- Sentiment Analysis: Uses pre-trained sentiment models to assess polarity and emotional content.
- Lexical Comparison: Measures token-level similarity and variation between male vs. female outputs.
- Statistical Significance Testing: Applies hypothesis testing to determine if observed differences are meaningful.
Technical Details
- Developed in Python using libraries like Transformers, NLTK, VADER, and pandas.
- Evaluated outputs from COMET-ATOMIC20 (based on GPT-2 architecture).
- Pre-processed events using WinoBias templates to control for confounding variables.
- Applied visualization techniques to summarize bias trends across relation types.
- Conducted both qualitative and quantitative assessments.