ALL-IN-ONE: Multi-Task Learning BERT models for Evaluating Peer Assessments

• DOI: 10.48550/arxiv.2110.03895
• Publication Date: 2021-01-01
• Authors: Qinjin Jia; Jialin Cui; Yunkai Xiao; Chengyuan Liu; Parvez Rashid; Edward F. Gehringer
• Keywords: Feature Learning; Feature (linguistics); Representation

Peer assessment has been widely applied across diverse academic fields over the last few decades and demonstrated its effectiveness. However, advantages of peer can only be achieved with high-quality reviews. Previous studies have found that review comments usually comprise several features (e.g., contain suggestions, mention problems, use a positive tone). Thus, researchers attempted to evaluate peer-review by detecting different using various machine learning deep models. there is no single study investigates multi-task (MTL) model detect multiple simultaneously. This paper presents two MTL models for evaluating leveraging state-of-the-art pre-trained language representation BERT DistilBERT. Our results demonstrate BERT-based significantly outperform previous GloVe-based methods around 6% in F1-score on tasks feature, further improves performance while reducing size.