A Low-Resource NLP Framework For Bilingual Next-Word Prediction In Tamil In Assistive Technology

Betsybha Evangeline¹, Anitha Dhakshina Moorthy¹

---

ABSTRACT

This study focuses on improving next-word prediction for Tamil, a low-resource language, within Assistive Technology. Limited availability of annotated Tamil language data poses a major challenge for building accurate predictive models. To address this, a bilingual prediction framework is proposed that uses Tamil–English translation to take advantage of the richer linguistic resources available in English. Pre-trained translation models are integrated with language models such as LSTM, BiLSTM, GRU, and BERT to enhance prediction performance in AAC applications. The models are evaluated across sentences of varying lengths, including short, medium, and long utterances commonly used by AAC users. Experimental results show that recurrent neural models perform consistently across all sentence lengths, with GRU achieving the highest precision for short sentences. To reduce semantic loss during translation, neural machine translation techniques designed to preserve contextual meaning are employed, leading to improved prediction accuracy and contextual relevance. The proposed framework demonstrates effective semantic retention, achieving a BLEU score of 0.75 for short sentences. Overall, the bilingual approach improves next-word prediction quality and supports more natural communication in AAC systems. The framework is scalable to other low-resource languages and provides a foundation for future real-world user evaluation and the integration of advanced generative language models to further enhance predictive performance.

Keywords: Next-word prediction, Augmentative and Alternative Communication (AAC), Bilingual framework, Tamil-English translation, Low-resource language, Multilingual AAC applications

How to cite this article: Evangeline B, Moorthy AD. A Low-Resource NLP Framework for Bilingual Next-Word Prediction in Tamil in Assistive Technology. Int J Drug Deliv Technol. 2026;16(10s): 953-968. DOI: 10.25258/ijddt.16.10s.111

Source of support: Nil.

Conflict of interest: None