EXPLORING ABSTRACTIVE SUMMARIZATION OF PRE-TRAINED MODELS: A STUDY ON GPT-2, T5, PEGASUS AND BART

Abstract

Text summarization, a significant challenge within Natural Language Processing (NLP), purposes to distill huge volumes of content into brief, coherent summaries. As the amount of text data continues to expand, the need for efficient and accurate summarization methods has grown, making it a critical task across various sectors. Despite advancements in summarization, especially with the emergence of transfer learning that leverage the capabilities of pre-trained models, questions remain about which models perform best for summarization on specific datasets. This study evaluates how well the pre-trained models—GPT-2, T5, Pegasus, and BART perform for the abstractive text summarization task. We employ the datasets MultiNews, WikiSum, DUC & CNN/Daily Mail, consisting of news related articles accompanied by their human-generated summaries. In the proposed work, we fine-tuned each model on these datasets through transfer learning, by carefully adjusting the parameters. The metrics employed to assess the performance of each model are ROUGE and METEOR. After rigorous experiments, the results indicate that the T5 model beats the others in abstractive summarization on these datasets, achieving superior ROUGE scores of R1, R2 and RL equal to 63.04%, 42.22%, 47.82% respectively and an average precision of 90.10%. Lastly, the further exploration of future research directions is provided.