Bisrath1/amharic-ecommerce-ner-extractor

Fine-tuning LLMs for Amharic NER to extract product information from Telegram e-commerce channels # Amharic E-commerce Data Extractor Named Entity Recognition (NER) System A transformer-based NER system to extract products, prices, and locations from Ethiopian Telegram e-commerce channels for EthioMart's centralized platform. ## Project Overview This project, developed as part of the **10 Academy Artificial Intelligence Mastery** (18 June - 24 June 2025), builds an **Amharic E-commerce Data Extractor** to transform unstructured Telegram posts into a structured FinTech engine for EthioMart. The system extracts key entities (Product, Price, Location) from Amharic text in Ethiopian e-commerce Telegram channels, enabling a centralized platform for product discovery and vendor analysis for micro-lending. ## Business Need EthioMart aims to consolidate decentralized Telegram-based e-commerce activities in Ethiopia into a unified platform. By extracting structured data from Telegram posts, the system enables: - Seamless customer interaction with multiple vendors. - A centralized database of products, prices, and locations. - A FinTech engine to identify promising vendors for micro-lending based on engagement metrics. ## Objectives 1. Develop a repeatable workflow for data ingestion, preprocessing, and entity extraction. 2. Fine-tune a transformer-based model for Amharic NER with high F1-score accuracy. 3. Compare multiple NER models and recommend the best for EthioMart’s use case. 4. Use SHAP and LIME for model interpretability to ensure transparency. 5. Create a vendor scorecard to rank vendors for micro-lending based on activity and engagement. ## Repository Structure ``` Amharic-Ecommerce-Extractor/ ├── data/ │ ├── raw/ # Raw Telegram data (CSV/JSON) │ ├── processed/ # Preprocessed data and CoNLL files │ └── vendor_scorecard.csv # Final vendor scorecard ├── notebooks/ │ ├── task1_data_ingestion.ipynb # Data scraping and preprocessing │ ├── task2_data_labeling.ipynb # CoNLL format labeling │ ├── t …