Bilingual AI-Powered Next-Word Prediction for Tamil AAC: A Low-Resource NLP Approach
Year: 2025
Type: article
Abstract: — this study addresses the challenges of next-word prediction for Tamil, a low-resource language, within the context of Augmentative and Alternative Communication (AAC) systems. The proposed solution is a bilingual framework that utilizes Tamil-English translation to bridge the gap between the limited resources available for Tamil and the extensive resources for English. By utilizing pre-trained translation models, the framework enhances the prediction capabilities of models such as LSTM, BiLSTM, GRU, and BERT to support AAC systems. These models are evaluated on their ability to predict the next word across different sentence lengths (less than 5 words, 6–10 words, and more than 10 words). Results show that LSTM, BiLSTM, and GRU models perform well across sentence lengths, with GRU achieving the highest precision for shorter sentences. However, challenges such as semantic loss during translation were addressed using advanced neural machine translation techniques, ensuring context preservation. The framework successfully improves next-word prediction accuracy, contextual relevance, and overall performance by utilizing the bilingual translation approach. This methodology offers a scalable solution for other low-resource languages, promoting better accessibility and inclusion in multilingual AAC applications. Additionally, the framework addresses practical challenges such as semantic preservation during translation and provides a roadmap for real-world user testing to validate its effectiveness in AAC scenarios. Future work could explore advanced models like GPT for further improvements in prediction quality. This framework achieves notable BLEU scores of 0.75 for shorter sentences, demonstrating its ability to retain semantic accuracy during translation. GRU models outperformed others, especially for short sentences, achieving the highest precision and accuracy. The framework is scalable for other low-resource languages, paving the way for inclusive AAC solutions globally.
Institution: Thiagarajar College of Engineering
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6949071491
FWCI:
Cites:
Cited by:
Related to: