FAIR-SEA: Fairness Assessment in Representation for Southeast Asia

Live Demo: fairsea.streamlit.app

FAIR-SEA is a comprehensive framework and interactive dashboard designed to evaluate socio-cultural bias in Large Language Models (LLMs) specifically within the Southeast Asian context, with an initial focus on Singapore's multicultural society.

---

Table of Contents

• Background & Purpose
• What FAIR-SEA Provides
• Framework Overview
• Validation Approach
• Using the Dashboard
• Running Locally

---

Background & Purpose

Large Language Models (LLMs), predominantly trained on Western-centric datasets, can perpetuate biases that misrepresent or stereotype Southeast Asian cultures. This misalignment reduces both trust and utility when these models are deployed in diverse societies like Singapore.

The Challenge:

• Existing global bias benchmarks (BBQ, HolisticBias) lack sufficient SEA representation
• Manual red-teaming efforts (e.g., IMDA's AI Safety Exercise) are difficult to scale
• Regional biases remain largely unmeasured and unaddressed

FAIR-SEA's Solution: Transform manual bias assessments into a systematic, data-driven, and reproducible evaluation process for LLMs deployed in Southeast Asia.

---

What FAIR-SEA Provides

1. Structured Framework

• Curated dataset of prompts reflecting local contexts
• Identity attributes: race (Chinese, Malay, Indian), gender, nationality
• Multiple testing techniques: Substitution, Generation, BBQ, Classification

2. Automated Pipeline

• Python-based evaluation workflow
• Systematic prompt generation and response collection
• Reproducible analysis with fixed parameters

3. Comprehensive Reporting

• Interactive Streamlit dashboard for visualization
• Quantitative metrics (Chi-square, ANOVA, sentiment analysis)
• Qualitative insights through manual inspection
• Intersectional bias detection

---

Framework Overview

Stage 1: Scoping & Dataset Curation

Bias Categories:

• Aptitude Stereotypes: Assumptions about innate abilities based on demographics
• Behavioral Stereotypes: Expected behaviors tied to identity
• Association Stereotypes: Problematic correlations between groups and attributes

Identity Tokens:

• Names culturally representative of Singapore's major ethnic groups
• Gender markers (Male, Female)
• Nationality indicators (Singaporean, Malaysian, Indonesian, etc.)

Prompt Design: Prompts are crafted across multiple domains:

• Employment & hiring decisions
• Educational aptitude assessments
• Customer service interactions
• Cultural expectations & family dynamics

Stage 2: Model Probing & Response Collection

• Automated API-based prompting
• Systematic response logging
• Version control for reproducibility
• Multiple model comparison capability

Stage 3: Bias Evaluation & Scoring

Quantitative Metrics:

• Demographic Bias Index (DBI): Standardized measure of bias intensity
• Fairness Deviation Index (FDI): Deviation from expected fair distribution
• Jensen-Shannon Divergence (JSD): Statistical distance between distributions
• Chi-square tests: Independence testing for categorical outcomes
• ANOVA: Variance analysis for continuous outcomes
• Sentiment Analysis: Polarity and emotional tone classification

Qualitative Analysis:

• Manual review of nuanced stereotypes
• Cultural sensitivity assessment
• Context-specific harm evaluation

---

Validation Approach

Validation Performed

This project implements and validates the FAIR-SEA framework through:

1. Dataset Curation:

   • 8 distinct analysis types (D1-D4, I1-I4)
   • 500+ curated prompts with identity token variations
   • Ground truth data for comparative analysis (I1, I2)

2. Model Testing:

   • Experiments conducted on multiple LLMs (GPT-4, GPT-3.5, etc.)
   • Systematic variation of identity attributes
   • Response collection with consistent parameters

3. Statistical Analysis:

   • Implementation of all quantitative metrics
   • Intersectional bias detection across Race × Gender × Nationality
   • Visualization of bias patterns through heatmaps and charts

4. Dashboard Development:

   • Interactive interface for exploring results
   • Real-time analysis capability
   • Export functionality for further research

Key Findings Demonstrated

The validation successfully demonstrates the framework's ability to:

• Detect statistically significant bias patterns
• Identify intersectional amplification effects
• Compare model outputs against ground truth
• Provide actionable insights for model improvement

---

Using the Dashboard

Accessing the Live Demo

Visit fairsea.streamlit.app to explore the dashboard immediately—no installation required!

Dashboard Features

1. Home Page

• Overview of the FAIR-SEA framework
• Quick navigation to analysis tools
• Background on methodology and purpose

2. Analysis Dashboard

Step 1: Choose Your Data Source

• Demo Data: Pre-loaded dataset with sample LLM outputs

  • Ready to explore immediately
  • Covers all 8 analysis types
  • Includes multiple demographic combinations

• Upload Custom CSV: Analyze your own LLM outputs

  • Required columns:
    • Gender: Male, Female, etc.
    • Race: Chinese, Malay, Indian, etc.
    • Nationality: Singaporean, Malaysian, etc.
    • prompt_text: The prompt sent to the LLM
    • prompt_id_full: Format [AnalysisType]-[Nationality]-[Race]-[Gender]-[Name]-[Number]
    • llm_output: The LLM's response
    • model: Model name (e.g., gpt-4o-mini)

Step 2: Select Analysis Type Choose from 8 specialized bias analyses:

Direct Bias Tests (D1-D4):

• D1: Leadership Style Perception
• D2: Aptitude & Suitability Assessment
• D3: Tone Interpretation Bias
• D4: Cultural & Career Choice Perception

Indirect Bias Tests (I1-I4):

• I1: Occupation Prediction Bias
• I2: Industry Classification Bias
• I3: Scholarship Decision Bias
• I4: Workplace Professionalism Perception

Step 3: Run Analysis Click "Run Analysis" to generate:

• Summary statistics across demographic groups
• Statistical significance tests (Chi-square, ANOVA)
• Bias indices (DBI, FDI, JSD)
• Interactive visualizations
• Intersectional bias analysis

Step 4: Interpret Results Navigate through tabs to explore:

• By Gender: Gender-specific bias patterns
• By Race: Racial bias distributions
• By Nationality: National origin bias
• Intersectional: Combined effects across multiple attributes

3. About Page

• Detailed framework explanation
• Methodology documentation
• Validation approach
• Research background

Tips for Best Results

1. Start with Demo Data: Familiarize yourself with the interface and metrics
2. Review Analysis Descriptions: Each analysis type has specific use cases
3. Check Statistical Significance: p-values < 0.05 indicate significant bias
4. Compare Across Demographics: Look for disproportionate impacts
5. Examine Intersectional Results: Bias often amplifies at intersections

---

Running Locally

For development or offline analysis, you can run FAIR-SEA on your local machine.

Prerequisites

• Python 3.8 or higher
• pip package manager
• Virtual environment (recommended)

Installation Steps

1. Clone the Repository:

git clone https://github.com/HartantoLs/fairsea.git
cd fairsea

2. Create Virtual Environment:

# Create environment
python -m venv venv

# Activate environment
# On macOS/Linux:
source venv/bin/activate
# On Windows (PowerShell):
.\venv\Scripts\activate
# On Windows (CMD):
.\venv\Scripts\activate.bat

3. Install Dependencies:

pip install -r requirements.txt

4. Run the Automated Scripting Script:

cd prompts

• Upload Prompts.xlsx to prompts folder.
• Input OPENAI_API_KEY in consolidated_prompts.ipynb and run the Jupyter notebook.
• The resulting data file will be saved in the folder data.

5. Run the Dashboard:

cd streamlit
streamlit run app.py

The dashboard will open automatically in your browser at http://localhost:8501