MuP-SciDocSum: Leveraging Multi-perspective Peer Review Summaries for Scientific Document Summarization

Scientific article summarization poses a challenge because the interpretability of the article depends on the objective, experience of the reader. Editors/Chairs assign experts in the domain as peer reviewers. These experts often write a summary of the article at the beginning of their reviews which offers a summarized view of their understanding (perspectives) on the given paper. Multiperspective summaries can provide multiple related but distinct perspectives of the reviewers rather than being influenced by a single summary. Here in this work, we propose a method to produce abstractive multiperspective summaries of scientific articles leveraging peer reviews. Our proposed method includes performing extractive summarization to identify the essential parts of the paper by extracting contributing sentences. In the subsequent step, we utilize the extracted pertinent information to condition a transformer-based language model comprising of a single encoder followed by multiple decoders that share weights. Our goal is to train the decoder to not only learn from a single reference summary but also to take into account multiple perspectives when generating the summary during the inference stage. Experimental results show that our approach achieves the best average ROUGE F1 Score, ROUGE-2 F1 Score, and ROUGE-L F1 Score with respect to the comparing systems. We make our code public (https://github.com/sandeep82945/Muti-percepective-summarization ) for further research.