Comparison of Extractive and Abstractive Approaches in Automatic Text Summarization: An Evaluation on BBC-News and PubMed Datasets

This study focuses on the effectiveness of advanced and up-to-date text summarization techniques in the field of automatic text summarization. Among the extractive summarization systems examined for their performance are TextRank, LexRank, KL-Summ, and LSA, while the abstractive summarization systems include Pegasus, BART, T5, and LED. The capabilities of these state-of-the-art models have been evaluated on the BBC-News and PubMed datasets. Several evaluation metrics such as SacreBLEU, METEOR, BERTScore, and ROUGE were employed. Summaries of 50,100, and 150 words were generated for the BBC-News and PubMed datasets. The findings of the study provide a comprehensive evaluation of the performance of different summarization techniques across various summary lengths on the BBC-News and PubMed datasets. It was reported that the TextRank approach achieved notably successful results in text summarization. Among the abstractive methods investigated, LED demonstrated a strong ability to generate contextually accurate summaries. This study is considered to make significant contributions to the literature.