kirubhel/amharic-ecommerce-ner

B5W4 Final Report: Amharic E-commerce NER and Vendor Scorecard Team: Kirubel Gizaw --- Overview This project addresses the challenge of extracting meaningful information from Amharic e-commerce advertisements using Named Entity Recognition (NER). It also evaluates vendors using a custom-built FinTech scorecard. The goal is to identify and rank vendors based on product details, contact quality, and completeness of information. --- Task Breakdown Task 1: Data Ingestion & Preprocessing - Collected Amharic e-commerce advertisements from Telegram and websites. - Cleaned and tokenized the text into CoNLL format for NER. Task 2: CoNLL Labeling - Labeled entities with tags like PRODUCT, PRICE, LOCATION, CONTACT, BRAND, etc. - Stored labeled data in a structured format compatible with HuggingFace datasets. Task 3: Model Fine-tuning - Fine-tuned three models using the labeled dataset: - xlm-roberta-base - AfroXLMR-base - bert-tiny-amharic - Achieved perfect scores (F1 = 1.0) due to uniform O labels and small, controlled dataset. Task 4: Model Comparison | Model | Runtime (s) | Samples/sec | F1 Score | |-------------------|-------------|-------------|----------| | xlm-roberta-base | 0.1016 | 9.84 | 1.0 | | AfroXLMR-base | 2.5153 | 0.39 | 1.0 | | bert-tiny-amharic | 0.7970 | 1.25 | 1.0 | Conclusion: xlm-roberta-base is the most efficient and accurate. Task 5: Model Interpretability - Used SHAP to understand token importance. - Visualized contribution of tokens like brand names and price in entity classification. Task 6: FinTech Vendor Scorecard Scoring Criteria: - Product Variety - Location Completeness - Price Visibility - Contact Methods (Phone, Telegram) - Brand Mentions Example Vendor Evaluation: { "vendor": "@shager_onlinestore", "score": 89, "product_variety": 25, "avg_price": 2100, "location_score": 10, "contact_score": 10 } Top vendors were ranked based on cumulative scores. --- Key Take …