kalebab98/EthioMart-Amharic-NER

# 📦 **EthioMart Amharic NER System** **EthioMart** aims to become the central hub for Telegram-based e-commerce in Ethiopia by aggregating business data such as **product names**, **prices**, and **locations** from multiple independent vendor channels. This project builds an **Amharic Named Entity Recognition (NER)** system to extract and structure such data for downstream analytics and business decision-making. --- ## 🔍 **Project Overview** With the growing use of **Telegram for commerce in Ethiopia**, vendors are scattered across isolated channels. This project solves the **fragmentation problem** by: - 🔄 **Scraping** real-time messages and media from Telegram vendor channels - 🧹 **Preprocessing** Amharic text using custom tokenization and normalization - 🏷️ **Labeling** key entities in Amharic: `Product`, `Price`, `Location` - 🤖 **Fine-tuning transformer-based models**: `XLM-RoBERTa`, `mBERT`, `AfroXLMR` - 📊 **Generating vendor analytics** to support micro-lending decisions --- ## 📁 **Project Structure** ```bash ethiomart-amharic-ner/ │ ├── data/ │ ├── raw/ # Raw scraped data │ └── processed/ # Cleaned and labeled datasets │ ├── amharic_ner_data.conll │ ├── notebooks/ │ └── 01_data_ingestion.ipynb # Telegram scraping logic │ └── Task-3NERmodel.ipynb │ └── task-4a.ipynb │ └──Task-5a.ipynb │ └──Task-6a.ipynb │ ├── scripts/ │ └── 01_text_preprocessing.py # Cleaning, normalization, tokenization │ ├── README.md ├── requirements.txt └── .gitignore ``` # From Tasks 3–6 This repository contains the implementation of Tasks 3 through 6 of the **B5W4 Challenge – Amharic E-Commerce Data Extractor**, which focuses on fine-tuning transformer models for Amharic Named Entity Recognition (NER), comparing models, interpreting predictions, and building a FinTech-ready vendor analytics engine. --- ## ✅ Task 3: Fine-Tune NER Model (Amharic) ### Goal: Train a transformer-based model (AfroXLMR) on annotated Amha …