Comparative Analysis of BanglaT5 and Pointer Generator Network for Bengali Abstractive Story Summarization

Sequence-to-sequence models, especially transformer-based ones, are the recent baseline for generating a summary of text documents and various natural language processing (NLP) tasks. Transformer-based language models can perform well for text summarization. However, for text summarization, these models often need help to capture the actual context of a summary. This paper compares the BanglaT5 seq-to-seq model with a pointer-generator network for generating novel words with a feature-rich encoder. We aim to build an automatic text summarizer that can summarize literary works. We present a new dataset, namely the Bangla short story collection of Rabindranath Tagore. Technically, this will be a step towards Bangla literature revealing something new. Evaluations from ROUGE-1, ROUGE-2, and ROUGE-L scores show that the pointer-generator network with linguistic features significantly enhances text summarization. We got better ROUGE scores for a pointer-generator network that are explained in the result analysis section.