Autonext: Optimised Recurrent and Context-Aware Statistical Models for Predictive Text Input in Low-Resource Indic Languages
ACM Transactions on Asian and Low-Resource Language Information Processing · 31.08.2026 · 25 (8) · 1-25
DOI: 10.1145/3822405 ↗Kurz gefasst
Fachartikel in der Zeitschrift ACM Transactions on Asian and Low-Resource Language Information Processing, online erschienen am 23.07.2026. Verfasst von Tanmay Somkuwar und 1 weiteren Autor:innen. Die Arbeit nennt 73 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Indian Institute of Technology Gandhinagar, Science and Engineering Research Board.
Abstract (Originalsprache)
Text entry systems allow users to input information and interact with digital environments, with predictive input technologies further enhancing efficiency by anticipating user intent and offering real-time suggestions for words or letters. To date, the use of predictive approaches for letters and words in low-resource languages is still limited. This work proposes a set of deep learning-based optimised recurrent language models (ORLMs), along with statistical context-aware smoothed entropy-based prediction by partial matching methods (CASE-based PPM) for next-letter and next-word prediction. We evaluated these RNN-based deep learning models and PPM-based statistical methods in real time using various input sentences in six languages: English, Hindi, Marathi, Gujarati, Punjabi, and Kannada. We conducted four experiments to assess these approaches across multiple standard and practical performance metrics, with a comparative analysis aimed at real-world deployment in predictive applications. The total percentage of correct characters or letters achieved for CASE-based PPM-5 and PPM-6 was \(90.63\pm 3.33\) %, \(92.85\pm 4.93\) %, respectively, across all languages. The optimised deep learning models, using long short-term memory and gated recurrent unit techniques, showed better suitability for word prediction applications, achieving accuracies of \(93.63\pm 2.15\) % and \(94.01\pm 2.08\) %, keystroke savings of \(51.59\pm 16.69\) % and \(50.25\pm 14.95\) %, and hit rates of \(62.75\pm 27.41\) % and \(60.575\pm 25.38\) %, respectively, across all six languages. These results demonstrate the effectiveness of the proposed approaches, which consistently outperform state-of-the-art predictive methods across various performance metrics and hold significant potential for developing text entry applications in low-resource languages.
Förderung
- Indian Institute of Technology Gandhinagar
- Science and Engineering Research Board
Zitieren
APA
Somkuwar, T., & Meena, Y. (2026). Autonext: Optimised Recurrent and Context-Aware Statistical Models for Predictive Text Input in Low-Resource Indic Languages. ACM Transactions on Asian and Low-Resource Language Information Processing, 25(8), 1-25. https://doi.org/10.1145/3822405
Vancouver
Somkuwar T, Meena Y. Autonext: Optimised Recurrent and Context-Aware Statistical Models for Predictive Text Input in Low-Resource Indic Languages. ACM Transactions on Asian and Low-Resource Language Information Processing. 2026;25(8):1-25. doi:10.1145/3822405
BibTeX
@article{Somkuwar2026,
author = {Somkuwar, Tanmay and Meena, Yogesh},
title = {Autonext: Optimised Recurrent and Context-Aware Statistical Models for Predictive Text Input in Low-Resource Indic Languages},
journal = {ACM Transactions on Asian and Low-Resource Language Information Processing},
year = {2026},
volume = {25},
number = {8},
pages = {1--25},
doi = {10.1145/3822405}
}
EUVIMED