nahomWM/nlp-based-amharic-hate-speech-detection

This project implements a hate speech detection system specifically designed for Amharic language text. # Amharic Hate Speech Detection A machine learning project for detecting hate speech in Amharic text using Natural Language Processing (NLP) techniques. This classifier can identify three categories: **hate**, **normal**, and **offensive** speech. ## Author **Nahom_WM** ## Table of Contents - Overview - Dataset - Features - Project Structure - Installation - Usage - Model Performance - Visualization - Technologies Used - License ## Overview This project implements a hate speech detection system specifically designed for Amharic language text. The system uses Logistic Regression with TF-IDF (Term Frequency-Inverse Document Frequency) vectorization to classify Amharic messages into three categories: - **Hate**: Messages containing hate speech - **Normal**: Regular, non-offensive messages - **Offensive**: Messages that are offensive but not necessarily hate speech ## Dataset ### Data Source The dataset used in this project is sourced from **Hugging Face Datasets**: - **Dataset Name**: `uhhlt/amharichatespeechranlp` - **Source**: Hugging Face - Amharic Hate Speech RANLP - **Description**: A curated dataset of Amharic text messages labeled for hate speech detection - **Labels**: hate, normal, offensive ### Dataset Statistics The dataset contains thousands of Amharic text messages with the following distribution: - **Hate**: ~45% of messages - **Normal**: ~37% of messages - **Offensive**: ~18% of messages ### Data Preprocessing The preprocessing pipeline includes: 1. **Normalization of Amharic characters**: Merging different forms of similar characters (e.g., ሐ, ሑ, ሒ → ሀ) 2. **Removal of @USER handles**: Anonymizing user mentions 3. **Punctuation removal**: Cleaning special characters 4. **Number removal**: Removing numeric characters 5. **Whitespace normalization**: Removing extra spaces ## Features - ✅ **Automated data download** from Hugging Face - ✅ **Text preprocessing** with Amharic character normalization - ✅ **TF-IDF feature extraction** (5000 ma …