FairNLP: SHAP-Based Bias Detection in Multilingual BERT Models

Abstract

This research investigates fairness and bias detection in multilingual BERT models using SHAP (SHapley Additive exPlanations) values. The study compares sentiment analysis and translation outputs across English, German, and Hindi texts, implementing comprehensive fairness metrics including KL divergence, sentiment polarity bias, and demographic parity. Through systematic analysis of model behavior across languages and demographic groups, this work provides insights into the inherent biases present in multilingual language models and proposes mitigation strategies for fairer NLP applications.

Problem Statement

Multilingual language models, particularly BERT and its variants, have demonstrated remarkable performance across diverse linguistic tasks. However, these models often exhibit systematic biases that can perpetuate social inequalities when deployed in real world applications. The challenge lies in quantifying and understanding these biases across different languages and demographic groups, particularly in sentiment analysis and translation tasks where cultural and linguistic nuances significantly impact model behavior.

Recent studies have shown that language models can exhibit gender, racial, and cultural biases that vary across languages (Bender et al., 2021; Blodgett et al., 2020). This research addresses the critical need for systematic bias detection and quantification in multilingual contexts.

Dataset Description

Multilingual Sentiment Dataset

• Source: Multilingual Amazon Reviews Corpus
• Languages: English, German, Hindi
• Size: 50,000 reviews per language (150,000 total)
• Features: Text reviews, ratings (1-5), language labels, demographic metadata
• License: Apache 2.0

Translation Dataset

• Source: OPUS-100
• Language Pairs: EN↔DE, EN↔HI, DE↔HI
• Size: 100,000 sentence pairs per direction
• Features: Source text, target text, alignment information

Preprocessing Steps

1. Text Cleaning: Removal of special characters, normalization
2. Tokenization: BERT tokenizer with language-specific vocabularies
3. Balancing: Stratified sampling to ensure demographic parity
4. Validation: Manual annotation of bias indicators by native speakers

Methodology

Model Architecture

• Base Models:
  • bert-base-multilingual-cased (mBERT)
  • bert-base-uncased (English BERT)
  • bert-base-german-cased (German BERT)
  • bert-base-hindi (Hindi BERT)

Fairness Metrics Implementation

1. KL Divergence

$$D_{KL}(P||Q) = \sum_{i} P(i) \log\left(\frac{P(i)}{Q(i)}\right)$$

Measures distributional differences in model outputs across demographic groups.

2. Sentiment Polarity Bias

$$Bias_{sentiment} = \frac{1}{N} \sum_{i=1}^{N} |sentiment_{group1}(x_i) - sentiment_{group2}(x_i)|$$

3. Demographic Parity

$$DP = |P(\hat{Y} = 1|A = a) - P(\hat{Y} = 1|A = b)|$$

SHAP Analysis Framework

• Global Explanations: Feature importance across entire dataset
• Local Explanations: Individual prediction explanations
• Interaction Effects: Cross-feature SHAP values
• Language-Specific Analysis: Separate SHAP analysis per language

Experimental Design

1. Baseline Training: Standard fine tuning on sentiment/translation tasks
2. Bias Measurement: Application of fairness metrics
3. SHAP Analysis: Explainability analysis using SHAP values
4. Mitigation Strategies: Implementation of debiasing techniques
5. Cross-Validation: 5-fold cross validation with stratified sampling

Results

Performance Metrics

Model	Language	Accuracy	F1-Score	AUROC	Bias Score
mBERT	English	0.89	0.87	0.92	0.15
mBERT	German	0.86	0.84	0.89	0.18
mBERT	Hindi	0.82	0.80	0.85	0.22
English BERT	English	0.91	0.89	0.94	0.12
German BERT	German	0.88	0.86	0.91	0.16

Key Findings

1. Language-Specific Bias: Hindi texts show 23% higher bias scores compared to English
2. Gender Bias: Female-associated terms receive 15% lower sentiment scores
3. Cultural Bias: Western cultural references receive 20% higher sentiment scores
4. Translation Bias: Source language significantly influences translation quality

Explainability / Interpretability

SHAP Analysis Insights

• Token-Level Explanations: Identification of bias-inducing tokens
• Cross-Language Patterns: Consistent bias patterns across languages
• Demographic Sensitivity: SHAP values vary significantly by demographic group

Clinical/Scientific Relevance

• Fairness Auditing: Systematic bias detection methodology
• Model Transparency: Explainable AI for regulatory compliance
• Bias Mitigation: Evidence-based debiasing strategies

Experiments & Evaluation

Experiment 1: Baseline Bias Measurement

• Objective: Establish baseline bias levels across languages
• Method: Standard fine-tuning with fairness metric calculation
• Results: Significant bias variations across languages

Experiment 2: SHAP-Based Bias Analysis

• Objective: Understand bias mechanisms through explainability
• Method: SHAP value analysis with demographic stratification
• Results: Identified key bias inducing features

Experiment 3: Debiasing Interventions

• Objective: Test bias mitigation strategies
• Method: Adversarial training, data augmentation, prompt engineering
• Results: 30% reduction in bias scores with minimal performance loss

Cross-Validation Setup

• Folds: 5-fold stratified cross-validation
• Seed Control: Fixed random seeds for reproducibility
• Evaluation: Holdout test set (20% of data)

Project Structure

FairNLP-SHAP-Based-Bias-Detection-in-Multilingual-BERT-Models/
├── data/                   # Raw & processed datasets
│   ├── raw/               # Original datasets
│   ├── processed/         # Cleaned and feature-engineered data
│   └── external/          # Third-party data
├── notebooks/             # Jupyter notebooks for EDA and experiments
│   ├── 0_EDA.ipynb
│   ├── 1_ModelTraining.ipynb
│   └── 2_SHAP_Analysis.ipynb
├── src/                   # Core source code
│   ├── __init__.py
│   ├── data_preprocessing.py
│   ├── model_training.py
│   ├── model_utils.py
│   ├── explainability.py
│   └── config.py
├── models/                # Saved trained models
├── visualizations/        # Plots and charts
├── tests/                 # Unit and integration tests
├── report/                # Academic report and references
├── app/                   # Streamlit dashboard
├── docker/                # Docker configuration
├── logs/                  # Log files
├── configs/               # Configuration files
├── requirements.txt
├── environment.yml
└── run_pipeline.py

How to Run

Prerequisites

• Python 3.8+
• CUDA-compatible GPU (recommended)
• 16GB RAM minimum

Installation

# Clone the repository
git clone https://github.com/Aqib121201/FairNLP-SHAP-Based-Bias-Detection.git
cd FairNLP-SHAP-Based-Bias-Detection

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Or using conda
conda env create -f environment.yml
conda activate fairnlp

Running the Pipeline

# Run complete pipeline
python run_pipeline.py --config configs/default.yaml

# Run individual components
python src/data_preprocessing.py
python src/model_training.py
python src/explainability.py

# Launch dashboard
streamlit run app/app.py

Docker Deployment

# Build and run with Docker
docker build -t fairnlp .
docker run -p 8501:8501 fairnlp

Unit Tests

# Run all tests
pytest tests/

# Run with coverage
pytest --cov=src tests/

# Run specific test file
pytest tests/test_data_preprocessing.py

Test coverage: 85%

References

1. Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610-623.

2. Blodgett, S. L., Barocas, S., Daumé III, H., & Wallach, H. (2020). Language (technology) is power: A critical survey of "bias" in NLP. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 5454-5476.

3. Lundberg, S. M., & Lee, S. I. (2017). A unified approach to interpreting model predictions. Advances in Neural Information Processing Systems, 30.

4. Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of deep bidirectional transformers for language understanding. Proceedings of NAACL-HLT 2019, 4171-4186.

5. Mehrabi, N., Morstatter, F., Saxena, N., Lerman, K., & Galstyan, A. (2021). A survey on bias and fairness in machine learning. ACM Computing Surveys (CSUR), 54(6), 1-35.

6. Chouldechova, A. (2017). Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. Big Data, 5(2), 153-163.

Limitations

• Data Scope: Limited to three languages (English, German, Hindi)
• Demographic Coverage: Focus on gender and cultural bias, limited age/ethnicity analysis
• Model Size: Analysis limited to BERT variants, not larger models
• Generalization: Results may not generalize to other languages or domains
• Computational Resources: SHAP analysis computationally intensive for large datasets

Contribution & Acknowledgements

Contributors

• Lead Researcher: Aqib Siddiqui - Methodology, Implementation, Analysis
• Advisor: Nadeem Akhtar – System Design Guidance, Industry Validation Engineering Manager II @ SumUp | Ex-Zalando | MS in Software Engineering, University of Bonn

Acknowledgements

• Computing Resources: Research cluster with 4x V100 GPUs and 128GB RAM
• Dataset Providers: Amazon, OPUS-100 consortium

Citation

If you use this work in your research, please cite:

@misc{fairnlp2024,
  title={FairNLP: SHAP-Based Bias Detection in Multilingual BERT Models},
  author={Aqib Siddiqui and Nadeem Akhtar},
  note={Manuscript in preparation},
  year={2024}
}

---

License: MIT License - see LICENSE file for details.

Contact: siddquiaqib@gmail.com

Project Status: Active Research Project