Pronoun Bias in Emotion Classification

This repository contains code to evaluate pronoun-induced emotion prediction drift in pretrained NLP models. It shows how changing the subject pronoun in a sentence ("he", "she", or "they")—while keeping the sentence otherwise identical—can lead to different predicted emotions by widely used classifiers.

---

📊 What This Code Does

• Uses a CSV file of 1,000 sentence triplets, each differing only in subject pronoun.
• Runs predictions using a Hugging Face model (e.g., distilbert-base-uncased-go-emotions-student).
• Measures label mismatches between:
  • he vs she
  • he vs they
  • she vs they
• Visualizes emotion drift using confusion matrix heatmaps.
• Prints example mismatches for qualitative understanding.

---

📁 Dataset Format

The dataset (gender_emotion_sentences_1000.csv) should have the following structure:

he	she	they
He stayed silent.	She stayed silent.	They stayed silent.
He helped a friend.	She helped a friend.	They helped a friend.
...	...	...

---

🚀 How to Run

1. Install dependencies:

pip install transformers pandas matplotlib seaborn