yeezyyoba/amharic-sentiment-analysis

Amharic sentiment analysis using multilingual transformers (Afro-XLM-R) fine-tuned on AfriSenti — low-resource NLP for Ethiopian social media text # Amharic Sentiment Analysis Platform ### Low-Resource NLP with Multilingual Transformers | End-to-End NLP Project --- ## Overview An end-to-end NLP project that fine-tunes a multilingual transformer (Afro-XLM-R) on Amharic social media text to classify sentiment as **positive**, **negative**, or **neutral**. Amharic is a low-resource language spoken by 50+ million people in Ethiopia — making this a meaningful contribution to African NLP research. The project compares a classical TF-IDF + SVM baseline against a state-of-the-art transformer, with a live deployed web application for real-time inference. --- ## Problem Statement Sentiment analysis tools exist for English, French, Arabic — but almost nothing exists for Amharic. Ethiopian businesses, researchers, and policymakers have no automated way to understand public opinion expressed in the national language. This project builds that tool. --- ## Dataset **AfriSenti-SemEval 2023 — Amharic Subset** - Source: HuggingFace — shmuhammad/AfriSenti-twitter-sentiment - Language: Amharic (am) - Labels: Positive, Negative, Neutral - Domain: Twitter/social media text --- ## Project Structure ``` amharic-sentiment-analysis/ │ ├── data/ │ ├── raw/ # Original AfriSenti dataset files │ ├── processed/ # Cleaned and tokenized data │ └── external/ # Reference data │ ├── notebooks/ │ ├── 01_EDA.ipynb # Dataset exploration │ ├── 02_preprocessing.ipynb # Text cleaning pipeline │ ├── 03_baseline_model.ipynb # TF-IDF + SVM baseline │ ├── 04_transformer_finetuning.ipynb # Afro-XLM-R fine-tuning │ ├── 05_evaluation.ipynb # Model comparison │ └── 06_error_analysis.ipynb # Error & attention analysis │ ├── src/ │ ├── data/ # Data loading and preprocessing │ ├── models/ # Model training utilities │ └── visualization/ # Plotting functions │ ├── app/ │ └── app.py # Streamlit web appli …