Logo Lanfrica

SnitTeshome/PSA-MT-Project

Domain:

natural language processing

Record type:

project
Creator:
Sni
Host:
Machine translation of Kenyan Public Service Announcements (English/Kiswahili ↔ indigenous languages) using few-shot cross-lingual transfer learning. # DSA4020_PSA_Project_G1 ## Kenyan Multilingual PSA Machine Translation **DSA 4020A — Natural Language Processing | Semester Project** - Prepared under the supervision of Dr. Edward Ombui A proof-of-concept multilingual machine translation (MT) system for Kenyan Public Service Announcements (PSAs) — translating between English/Kiswahili and selected under-resourced indigenous languages, and deploying the result as a working demo. > **The group tried multiple modeling approaches in parallel and tracks all of them in full, not just the "winning" one** — pretrained-model fine-tuning (mT5/NLLB) and dictionary-prompted LLM translation. See §8 and `docs/results_summary.md` for every metric reported, on a shared BLEU basis wherever the raw outputs allowed recomputing it. ### Quick navigation | Want to... | Go to | |---|---| | See every result, every metric, both approaches | `docs/results_summary.md` | | Read the full experiment log (24 sections, warts and all) | `docs/ekegusii_transfer_learning.md` | | Browse all documentation by week | `docs/README.md` | | Understand the dataset / reproduce it | `data/README.md` | | Run something yourself | §10 below | | See what's honestly still incomplete | §11 below | | See the project poster | `docs/PSA_Group1_Poster_Portrait.pdf` | Every folder in this repo (`data/`, `docs/`, `scripts/`, `notebooks/`, `experiments/`, `src/`) has its own `README.md` — open any of them on GitHub and you'll land on an explanation of what's inside and how it connects to the rest, not a bare file list. ## Team Members — Group 1 | # | Name | ID Number | |---|---|---| | 1 | Snit Teshome | 670552 | | 2 | Bradley Azegele | 668341 | | 3 | Kyeremateng Martin | 669217 | | 4 | Kemo Dibassy | 669111 | | 5 | Samantha Nyatichi Masaki | 670455 | ## 1. Project Overview **Goal:** build a deployable digital public good that translates PSAs between English/Kiswahili and under-resourced Kenyan languages — Ekegusii (Bantu), Dholuo (Nilotic), and Somali (Cus …