# Amharic-E-commerce-Data-Extractor
## Project Overview
A pipeline for extracting, labeling, and analyzing Amharic e-commerce data from Telegram channels, with NER model training, interpretability, and vendor analytics.
## Completed Tasks
1. **Environment Setup & Library Installation**
- Installed all required Python libraries for NLP, ML, and interpretability.
2. **Data Ingestion & Preprocessing**
- Loaded Telegram data from Excel files.
- Cleaned and normalized Amharic text.
- Extracted potential entities (products, prices, locations) using regex.
3. **NER Labeling**
- Sampled messages for manual labeling.
- Provided CoNLL labeling template and exported for annotation.
- Converted labeled CSVs to CoNLL format.
4. **Model Fine-tuning**
- Loaded labeled data from CoNLL format.
- Created label mappings and tokenized data for Hugging Face models.
- Split data into train/validation sets.
- Trained transformer-based NER models (XLM-Roberta, mBERT).
- Evaluated and saved both models.
- Added inference code for new Amharic text.
5. **Model Comparison & Evaluation**
- Compared XLM-Roberta and mBERT on F1, precision, recall, accuracy, model size, and inference speed.
- Displayed entity-level metrics and recommendations.
6. **Model Interpretability**
- Added LIME analysis for NER interpretability.
- Added attention visualization for token-level interpretability.
7. **Vendor Analytics & Scorecards**
- Extracted NER entities for all messages.
- Aggregated by vendor/channel and computed business metrics.
- Visualized top vendors by product diversity, price range, and location coverage.
## Next Steps
- Export scorecard results for business use.
- Integrate with FinTech decision engines.
- Continue improving NER and analytics with more data and feedback.
---
For details, see the main notebook: `Amharic_Ecommerce_Data_Extractor.ipynb`.