Deciphering the Reviewer's Aspectual Perspective: A Joint Multitask Framework for Aspect and Sentiment Extraction from Scholarly Peer Reviews

Peer reviews are one of the most important artifacts in scholarly communications. Peer reviews can serve as a rich source of knowledge discovery from texts that are human-generated and also opinionated on the paper under scrutiny. Reviewers comment on several implicit aspects of the paper (Originality, Soundness, Clarity, Appropriateness, etc.) where they sometimes appreciate, sometimes discuss, or sometimes question or criticize the work. Hence, correctly understanding the reviewer's aspectual perspective on the paper is crucial for chairs/editors to take a stand and also for the authors to respond or revise accordingly. In this paper, we introduce MASEPR, a novel multitask deep neural architecture to jointly discover the aspects and associated sentiments from the peer review texts. Our proposed approach leverages the knowledge sharing between aspect and sentiment lexicons to generate predictions. We outperform the standard baselines by a significant margin. We also make our codes available at https://github.com/cruxieu17/MASEPR.