Improving Coverage and Novelty of Abstractive Text Summarization Using Transfer Learning and Divide and Conquer Approaches

Abstract

Automatic Text Summarization (ATS) models yield outcomes with insufficient coverage of crucial details and poor degrees of novelty. The first issue resulted from the lengthy input, while the second problem resulted from the characteristics of the training dataset itself. This research employs the divide-and-conquer approach to address the first issue by breaking the lengthy input into smaller pieces to be summarized, followed by the conquest of the results in order to cover more significant details. For the second challenge, these chunks are summarized by models trained on datasets with higher novelty levels in order to produce more human-like and concise summaries with more novel words that do not appear in the input article. The results demonstrate an improvement in both coverage and novelty levels. Moreover, we defined a new metric to measure the novelty of the summary. Finally, we investigated the findings to discover whether the novelty is influenced more by the dataset itself, as in CNN/DM, or by the training model and its training objective, as in Pegasus.