Autonext: Optimised Recurrent and Context-Aware Statistical Models for Predictive Text Input in Low-Resource Indic Languages

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 ± 3.33%, 92.85 ± 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 ± 2.15% and 94.01 ± 2.08%, keystroke savings of 51.59 ± 16.69% and 50.25 ± 14.95%, and hit rates of 62.75 ± 27.41% and 60.575 ± 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.