Nahom32/FAI-Amharic

Explaining Amharic dataset with XAI. # Debiasing Hate Speech Detection Model ## Overview This project focuses on developing and evaluating machine learning models to detect hate speech while addressing bias and fairness across demographic groups. The goal is to build models that not only perform well in accuracy but also minimize disparities in predictions among different ethnic or social groups, ensuring equitable treatment. --- ## Features * Replicate hate speech detection models (e.g., LSTM, BiLSTM) with debiasing techniques such as adversarial training. * Measures standard classification metrics: accuracy, precision, recall, and F1-score. * Evaluates fairness using group-wise metrics like true positive rate (TPR), false positive rate (FPR), false negative rate (FNR), and selection rates. * Quantifies disparities across groups to assess fairness gaps. * Supports analysis of bias mitigation effectiveness. --- ## Motivation Hate speech detection models can unintentionally propagate or amplify biases against certain demographic groups, leading to unfair treatment and misclassification. This project aims to identify and mitigate such biases, promoting fair and responsible AI deployment. --- ## Getting Started ### Requirements * Python 3.7+ * TensorFlow * gensim * Libraries: scikit-learn, pandas, numpy, ### Installation ```bash git clone github.com cd FAI-Amharic pip install -r requirements.txt ``` ### Usage 1. **Data Preparation:** Load and preprocess datasets with group labels indicating demographic categories. 2. **Training:** Train baseline and adversarially debiased models using provided scripts. 3. **Evaluation:** Run evaluation scripts to calculate performance and fairness metrics. 4. **Analysis:** Review metrics to understand trade-offs between accuracy and fairness. --- ## Results * Debiased models generally reduce disparities in TPR and FPR across groups. * Adversarial training shows improved fairness in true positive rates, while some trade-of …