HarinaChohan/PSA_Kalenjin_Kiswahili

A project of Machine Translation that used NLLB-200 to translate Kiswahili PSAs and English PSAs to Kalenjin # PSA_Kalenjin_Kiswahili A machine translation project that used NLLB-200 to translate Kiswahili PSAs and English PSAs to Kalenjin ## Introduction The Kalenjin languages are a family of a dozen Southern Nilotic languages spoken in Kenya, eastern Uganda, and northern Tanzania. Most public service announcements are in English. Language barriers reduce access to critical health and safety information. I want to help bridge the gap for people who understand Kiswahili and Kalenjin. ## Data collection The PSA data was collected via Kenyan news and government websites and parsed to keep only PSAs and not news briefings and stories. The framework for choosing PSAs is attached here ## The data The PSA dataset was 5992 different PSA sentences across 5 domains: 1. Health 2. Security 3. Agriculture 4. Education 5. Governance 6. Civic education English-Kiswahili pairs after cleaning/splitting: 6821 ## The Kalenjin dataset It was used to train the NLLB-200 model to translate from Kalenjin to Kiswahili. It was downloaded from Enabling machine translatio… It was then cleaned by removing duplicates and empty rows and removing long sentence pairs, as PSAs are not long sentences/paragraphs, resulting in a total of 26414 sentence pairs ## The model MODEL_NAME = "facebook/nllb-200-distilled-600M" This was fine-tuned over 3 epochs with a batch size of 4. With all of this, the model took 5 to 6 hours to train using the Google Colab T4 GPU. ## Demo Used Gradio ## Evaluation The translation was not very accurate. chrF and BLEU were both used; however, chrF is more informative here as Kalenjin is a rich, low-resource language, as it works at the character level and is less punishing about small spelling/morphology differences. ## Limitations 1. Needs a powerful GPU to run; on normal PCs, it took 5 to 6 hours to train and execute 2. Can still struggle with translation nuance ## Recommendations and future work 1. Collect more PSA and Kalenjin data 2. Add more East African …