Impact of Annotation Difficulty on Automatically Detecting Problem Localization of Peer-Review Feedback

We believe that providing assessment on students’ reviewing performance will enable students to improve the quality of their peer reviews. We focus on assessing one particular aspect of the textual feedback contained in a peer review – the presence or absence of problem localization; feedback containing problem localization has been shown to be associated with increased understanding and implementation of the feedback. While in prior work we demonstrated the feasibility of learning to predict problem localization using linguistic features automatically extracted from textual feedback, we hypothesize that inter-annotator disagreement on labeling problem localization might impact both the accuracy and the content of the predictive models. To test this hypothesis, we compare the use of feedback examples where problem localization is labeled with differing levels of annotator agreement, for both training and testing our models. Our results show that when models are trained and tested using only feedback where annotators agree on problem localization, the models both perform with high accuracy, and contain rules involving just two simple linguistic features. In contrast, when training and testing using feedback examples where annotators both agree and disagree, the model performance slightly drops, but the learned rules capture more subtle patterns of problem localization.