Abstractive Text Summarization for Bengali Language Using Attention Mechanisms

Abstract

This paper offers a Bengali-to-Bengali abstractive text summarization approach based on a sequence-to-sequence Recurrent Neural Network (RNN) model with attention mechanisms, to create concise, concluding summaries of longer Bengali texts. The proposed model was established to address the significant need for summarization tools focusing on the Bengali language, especially in low-resource NLP situations. This study was organized, and the research was clearly defined in terms of obtaining a dataset from legitimate sources, methodical preprocessing including normalization and tokenization, and applying an encoder-decoder model with modifications that used attention mechanisms. The encoder captures the contextual meaning of the inputs, and the decoder produces the output summary based on the input representations and attention weights that were learned. The end-to-end process from data management and raw data through to model output was presented through architectural figures and step-by-step descriptions. and coherent summaries, while reflecting the intent of the original author. Other limitations such as resource constraints, the small size of annotated Bengali corpora, and the layers of complexity in Bengali language were noted. Despite these limitations, the model successfully achieved its goal of establishing the effectiveness of neural methods for Bengali abstractive summarization. The significance of the work provides an important stepping stone for future research efforts in Bengali, as well as, the fields of Bengali NLP and text generation. Future work in these areas can focus on increasing the datasets size, improving upon the linguistic rules, and, employing more advanced deep learning models and techniques such as transformers, to enhance the model’s performance and linguistic diversity.