2023 VOLUME 3 ISSUE 1
RESEARCH ARTICLE
Aspect-based sentiment analysis of online peer reviews and prediction of paper acceptance results
AUTHOR
Minghui Meng, Ruxue Han, Jiangtao Zhong, Haomin Zhou, Chengzhi Zhang
ABSTRACT
Peer reviews of academic articles contain reviewers' overall impressions and specific comments on the contributed articles, which have a lot of sentimental information. By exploring the fine-grained sentiments in peer reviews, we can discover critical aspects of interest to the reviewers. The results can also assist editors and chairmen in making final decisions. However, current research on the aspects of peer reviews is coarse-grained, and mostly focuses on the overall evaluation of the review objects. Therefore, this paper constructs a multi-level fine-grained aspect set of peer reviews for further study. First, this paper uses the multi-level aspect extraction method to extract the aspects from peer reviews of ICLR conference papers. Comparative experiments confirm the validity of the method. Secondly, various Deep Learning models are used to classify aspects' sentiments automatically, with LCFS-BERT performing best. By calculating the correlation between sentimental scores of the review aspects and the acceptance result of papers, we can find the important aspects affecting acceptance. Finally, this paper predicts acceptance results of papers (accepted/rejected) according to the peer reviews. The optimal acceptance prediction model is XGboost, achieving a Macro_F1 score of 87.43%.
KEYWORDS
Peer reviews; Aspect extraction; Sentiment analysis; Prediction of paper acceptance results