lavu1/BembaTranslationOffline

Offline Bemba–English neural machine translation and conversational AI for low-resource NLP, built with mBART, MarianMT and mT5. # Bemba–English Offline Translation and Conversational AI A resource-conscious neural machine translation and conversational AI system for Bemba and English, developed as part of a Master of Information Systems research project in Zambia. The project is designed around a practical question: **can useful African language technology run locally on modest hardware, without sending a user's text to a cloud service?** Try the verified public demonstration · View the researcher's profile · Hugging Face profile ## Research contribution - A curated Bemba–English corpus of approximately **350,000 parallel sentences**. - Parameter-efficient adaptation of multilingual sequence-to-sequence models. - A local inference application supporting translation and conversation. - Quantisation experiments intended for constrained computing environments. - Automatic evaluation complemented by ratings from **50 native Bemba speakers**. | Evaluation signal | Reported result | | --- | ---: | | BLEU | 35.9 | | METEOR | 52.4 | | Perplexity | 6.76 | | Native-speaker cultural-fit rating | 4.1 / 5 | These are thesis-stage results from the project's documented evaluation setup. They should not be treated as universal performance guarantees or compared directly with scores produced on different test sets. ## What the application does - Translates Bemba to English with a fine-tuned mBART model. - Translates English to Bemba with a locally stored MarianMT model. - Supports Bemba conversation with a fine-tuned encoder–decoder model. - Supports English conversation by translating through Bemba locally. - Provides a lightweight Flask interface and keeps conversation history on the device. - Selects Apple Silicon, CUDA or CPU execution automatically. ## System architecture ```mermaid flowchart LR UI["Local Flask interface"] --> MODE{"Selected task"} MODE -->|"Bemba → English"| MBART["Fine-tuned mBART"] MODE -->|"English → Bemba"| MARIAN["MarianMT"] MODE -->|"Conversation"| CHAT["Bemba dial …