rahelFM/EthioMart-Amharic-NER

# # EthioMart Amharic Named Entity Recognition (NER) This project focuses on fine-tuning a Named Entity Recognition (NER) model for the **Amharic language**, specifically targeting e-commerce-related entities such as **Product Names**, **Prices**, and **Locations** from Telegram channel messages of the fictional marketplace "EthioMart". ## Project Structure ```bash EthioMart-Amharic-NER/ ├── data/ │ ├── raw_data.csv │ └── sample_for_labeling.csv ├── labels/ │ └── labeled_data.conll ├── scripts/ │ ├── preprocessing.py │ ├── labeling_app.py │ ├── finetuning.py │ └── evaluation.py ├── amharic-ner-model/ # fine-tuned model saved here └── README.md ✅ Tasks ✅ Task 1: Data Collection & Cleaning Extracted 50 product-related messages from Telegram exports. Cleaned and normalized Amharic text. ✅ Task 2: Manual Annotation Built an interactive Streamlit labeling tool (labeling_app.py). Annotated tokens with BIO format: B-Product, I-Product, B-LOC, B-PRICE, etc. Saved in CoNLL format at labels/labeled_data.conll. ✅ Task 3: Fine-Tuning Used xlm-roberta-base pretrained on multilingual NER. Fine-tuned on labeled Amharic product data. Script: scripts/finetuning.py Model Info Base Model: Davlan/xlm-roberta-base-ner-hrl Training Epochs: 3 Labels: B-Product, I-Product B-LOC, I-LOC B-PRICE, I-PRICE O (outside entity) Dependencies pip install pandas streamlit transformers datasets seqeval Running the Project 1. Label Tokens streamlit run scripts/labeling_app.py 2. Fine-Tune Model python scripts/finetuning.py 3. Evaluate Model python scripts/evaluation.py Authors Rahel Sileshi Abdisa License MIT License