Logo Lanfrica

Naty-12/Amharic-NER-Data-Collector_week4

Domaine:

natural language processing

Type de record:

project
Créateur:
Nat
Hôte:
# Amharic-NER-Data-Collector # 🛍️ EthioMart Amharic Named Entity Recognition (NER) System EthioMart aims to become the central hub for all Telegram-based e-commerce activities in Ethiopia. With the rise of decentralized vendor channels on Telegram, the need for a unified and intelligent platform has become critical. This project focuses on building a fine-tuned Amharic Named Entity Recognition (NER) system that extracts key business-related entities (e.g., Product, Price, Location) from text and images shared across Ethiopian Telegram e-commerce channels. --- ## 🚀 Project Objectives - **Data Collection**: Ingest real-time messages and media from 5+ Telegram e-commerce channels. - **Preprocessing**: Clean and structure Amharic text data for downstream NLP tasks. - **Labeling**: Create a high-quality, custom NER dataset following CoNLL format. - **Model Training**: Fine-tune large pre-trained multilingual models (e.g., XLM-R, mBERT) for Amharic NER. - **Model Comparison**: Evaluate and compare models using precision, recall, F1-score. - **Interpretability**: Use SHAP and LIME to explain model predictions and ensure transparency. - **Business Insight**: Populate EthioMart’s centralized product database with extracted structured data. --- ## 📦 Extracted Entity Types ### Primary Entities: - **PRODUCT**: Product names or types - **PRICE**: Monetary values in birr (e.g., "25,000 ብር") - **LOCATION**: City/region references (e.g., "አዲስ አበባ") ### Optional Entities: - **DELIVERY_FEE**: Additional shipping costs (e.g., "በነፃ መመዝገቢያ") - **CONTACT_INFO**: Phone numbers or Telegram usernames --- ## 🗂️ Project Structure ``` ├── data/ │ ├── raw/ # Unprocessed raw Telegram messages │ ├── cleaned/ # Cleaned Amharic texts │ ├── images/ # Downloaded product/media images │ └── labeled/ # CoNLL-style labeled datasets ├── notebooks/ # Jupyter notebooks for EDA and model training ├── scripts/ # Python scripts for scraping, preprocessing, training ├── models/ # Saved models and chec …