🧠 Biased Neurons, Fairer Models

This repository contains the replication package for our study
“Tracing Stereotypes in Pre-trained Transformers: From Biased Neurons to Fairer Models”

The project extends the Knowledge Neurons framework (Dai et al., ACL 2022) to investigate how social biases are internally represented within pretrained transformers and how they can be localized and mitigated at the neuron level.
Our work introduces the concept of biased neurons—neurons that encode stereotypical associations—and proposes a reproducible methodology to trace and suppress them, evaluating the trade-off between bias mitigation and performance on fairness-sensitive software engineering tasks.

---

🧩 Repository Structure

├── data/
│   ├── biased_relations/
│   ├── dataset_extraction/
│   └── README.md
│
├── results/
│   ├── bert-base-cased/
│   ├── bert-large-cased/
│   ├── ....
│   ├── analysis_rq1/
│   └── README.md
│
├── src/
│   ├── 1_analyze_mlm.py
│   ├── 2_get_kn.py
│   ├── ...
│   └── README.md
│
└── README.md

---

📁 data/

Contains the datasets of biased relations and bias-activating prompts used to trace and evaluate biased neurons.
It includes:

• The nine bias categories (Age, Gender, Race/Color, etc.);
• Scripts for dataset extraction and processing;
• Documentation of the full generation pipeline.

See data/README.md for details.

---

📊 results/

Stores all experimental outputs, summaries, and statistical analyses for the three research questions (RQ1–RQ3).
Includes:

• Attribution and suppression results for each BERT model;
• Per-task evaluation data for SE tasks (e.g., sentiment, tone bearing, requirements classification);
• Scripts for data analysis and visualization.

See results/README.md for a detailed description of each file.

---

⚙️ src/

Contains the complete source code implementing our adaptation of the Knowledge Neurons framework for bias tracing and suppression.
Includes:

• Scripts for neuron attribution, refinement, and suppression;
• Plotting utilities for visual analysis;
• Evaluation scripts for SE downstream tasks.

See src/README.md for usage instructions.

---