Next word prediction in Bangla using Deep Learning Techniques

Abstract

Next-word prediction is an essential feature in contemporary text input systems, significantly enhancing typing speed, improving efficiency, and reducing the likelihood of errors. This feature is particularly advantageous for users with physical disabilities and individuals seeking to enhance their productivity in digital writing. This paper delves into the development of an advanced next-word prediction system tailored for the Bangla

language, utilizing both traditional probabilistic modeling techniques and state-of-the- art deep learning approaches. The system harnesses a combination of n-gram models,

ranging from unigram to 5-gram, alongside deep learning methodologies, specifically

Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks. The n- gram models provide a probabilistic foundation for word prediction, capturing the

immediate context within the text. In contrast, the sequential models, LSTM and GRU, are adept at capturing long-term dependencies and contextual relationships within Bangla text, which is crucial for accurate next-word prediction. Our extensive experiments reveal that the LSTM model consistently outperforms the GRU model in

terms of prediction accuracy, offering a more reliable and effective approach for next- word prediction in Bangla. The LSTM model achieved an accuracy of 99.38% for the 5- gram dataset, while the GRU model achieved a peak accuracy of 80.10% for the 4-gram dataset. This research marks a significant contribution to the development of efficient Bangla text input systems, laying the groundwork for further advancements in language modeling and contextual text prediction. The implications of our findings extend beyond the scope of this study, offering potential applications in various domains requiring language processing and user interface design for Bangla-speaking populations. By bridging traditional probabilistic methods with cutting-edge deep learning techniques, our work showcases the potential of integrating diverse modeling strategies to enhance the performance and reliability of text prediction systems. This synergy between established and innovative approaches underscores the value of a comprehensive methodology in tackling complex linguistic challenges .