Abstractive Summarization with Efficient Transformer Based Approach

Abstract

One of the most significant research areas is how to make a document smaller while keeping its essential information because of the rapid proliferation of online data. This information must be summarized in order to recover meaningful knowledge in an acceptable time. Text summarization is what it's called. Extractive and abstractive text summarization are the two types of summarization. In current years, the arena of abstractive text summarization has become increasingly popular. Abstractive Text Summarization (ATS) aims to extract the most vital content from a text corpus and condense it into a shorter text while maintaining its meaning and semantic and grammatical accuracy. Deep learning architectures have entered a new phase in natural language processing (NLP). Many studies have demonstrated the competitive performance of innovative architectures including recurrent neural network (RNN), Attention Mechanism and LSTM among others. Transformer, a recently presented model, relies on the attention process. In this paper, abstractive text summarization is accomplished using a basic Transformer model, a Transformer with a pointer generation network (PGN) and coverage mechanism, a Fastformer architecture and Fastformer with pointer generation network (PGN) and coverage mechanism. We compare these architectures after careful and thorough hyperparameter adjustment. In the experiment the standard CNN/DM dataset is used to test these architectures on the job of abstractive summarization.