Navigating tenses in Bengali sentences: A stacked ensemble model for enhanced prediction.

Tense classification in Bengali sentences is a fundamental yet unsolved problem of Bangla natural language processing (NLP) which is essential for tasks like machine translation, sentiment analysis, grammar correction, writing assistance and sentence generation. This study addresses this gap by proposing a robust stacked ensemble model designed for accurate automatic tense classification in Bengali sentences. To support this, we construct a novel Bengali corpus "BengaliTenseCorpus" comprising 13,500 manually collected and meticulously labelled sentences, categorized into three tense classes: Present (0), Past (1) and Future (2). The sentences gathered from diverse sources including news articles, songs, poems and novels, went through rigorous preprocessing techniques to preserve linguistic integrity and improve performance on data. The proposed architecture integrates predictions from five base models- Random Forest, Support Vector Machine, XGBoost classifier, Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) into the meta model- neural network to build a stacked ensemble framework. Experimental results demonstrate that this ensemble model outperforms individual models, achieving a classification accuracy of 85% on test data. This work presents the first large-scale Bengali tense classification system combining machine learning and deep learning methods in a stacked ensemble framework, establishing a strong performance benchmark for Bangla NLP with practical applications in intelligent writing tools, grammar assistance, and language learning. The findings highlight how well ensemble-based systems can capture the intricacies of Bengali verb morphology. To further boost the development of Bangla language models and applications, future extensions of this work may involve expanding the dataset, exploring transformer-based models, and incorporating tense-to-tense morphological conversion.