sanoy-si/EthioMart-Amharic-NER-Project

# EthioMart: End-to-End Amharic NER and FinTech Scorecard This project demonstrates the complete lifecycle of a real-world machine learning system. It builds an end-to-end Named Entity Recognition (NER) model for the Amharic language and uses it to power a FinTech vendor scoring engine for a fictional e-commerce platform, "EthioMart". The system scrapes data from Telegram e-commerce channels, processes Amharic text, fine-tunes multiple transformer models, and applies the best model to generate actionable business intelligence. --- ## 🚀 Project Highlights & Final Results The primary goal was to transform unstructured Amharic text from Telegram posts into a structured **Vendor Scorecard**. This scorecard provides valuable analytics for assessing vendor reliability and could be used by a micro-lending partner to evaluate creditworthiness. ### 📊 Model Comparison Results Multiple transformer models were fine-tuned and compared to select the best one for the task. The monolingual `Geza-G/Amharic-BERT` was chosen as the top candidate. *(This analysis is performed in `notebooks/Model_Comparison.ipynb`.)* | Model | F1-Score (Test Set) | Precision | Recall | Avg. Inference Time (s/sentence) | |:--------------------------------|----------------------:|------------:|---------:|-----------------------------------:| | **Amharic-BERT (Specialist)** | **0.0250** | **0.0150** | **0.2000** | **0.0450** | | XLM-Roberta-Base (Generalist) | 0.0133 | 0.0070 | 0.1111 | 0.0512 | | DistilBERT-multi (Speed Focus) | 0.0000 | 0.0000 | 0.0000 | 0.0310 | *Note: The F1-scores reflect performance on a small prototype dataset, successfully proving the viability of the pipeline.* --- ## 🛠️ Project Architecture The project is organized into a modular structure separating source co …