marakiteferi/NLP-Amharic-to-Afan-Oromoo-translation-assignment-

# Amharic to Afaan Oromoo Machine Translation Using T5 **Adama Science and Technology University (ASTU)** **College of Electrical Engineering and Computing** **Department of Computer Science and Engineering** | | | |---|---| | **Course** | Introduction to Natural Language Processing | | **Project title** | Amharic to Afaan Oromoo Machine Translation Using T5 | | **Submitted to** | Dr. Sintayehu Hirpassa | ### Group members | No. | Name | ID | |-----|------|-----| | 1 | Firomsa Assefa | UGR/25280/14 | | 2 | Feleke Kinfe | UGR/25571/14 | | 3 | Naol Fikadu | UGR/25588/14 | | 4 | Kaleb Seifu | UGR/26483/14 | | 5 | Maraki Teferi | UGR/25729/14 | | 6 | Elsabet Mulugeta | UGR/25592/14 | | 7 | Biniyam Megersa | UGR/25541/14 | --- Fine-tuned **mT5-small** (T5-family) seq2seq model for translating **Amharic** (Ethiopic script) into **Afaan Oromoo** / Afan Oromo (Latin script), built from a filtered NLLB-style parallel corpus and evaluated with SacreBLEU, ROUGE, and token-level metrics. **Repository:** github.com --- ## Problem statement Low-resource language pairs such as Amharic–Afan Oromo lack widely available commercial MT systems. This project trains a compact multilingual T5 variant on parallel scripture/web text, with quality filtering and script-aware preprocessing, to produce a reproducible baseline for academic study and classroom demonstration. --- ## Dataset | Path | Description | |------|-------------| | `data/raw/NLLB.am-om.am` | Amharic source lines | | `data/raw/NLLB.am-om.om` | Afan Oromo target lines (Latin) | | `data/raw/NLLB.am-om.scores` | Alignment confidence scores | | `data/processed/train.csv` | Training split | | `data/processed/val.csv` | Validation split (3,000 rows) | | `data/processed/test.csv` | Test split (3,000 rows) | | `data/predictions/results.csv` | 200 test predictions | **Preprocessing** (see `scripts/train_local.py`): - Amharic homophone normalization - M …