Affect Recognition in Code Review: An In-situ Biometric Study of Reviewer's Affect

Abstract Code reviews are an important practice in software development that increases team productivity and improves product quality. They are also examples of remote, computer-mediated asynchronous communications which are prone to the loss of affective information. Prior research has focused on sentiment analysis in source codes, as positive affect has been linked to developer productivity. Although methods of sentiment analysis have advanced, challenges remain due to numerous domain-specific expressions, subtle nuance, and indications of sentiment. In this paper, we uncover the potential for 1) nonverbal behavioral signals such as conventional typing, and 2) indirect physiological measures (eye gaze, GSR, touch pressure) to reveal genuine affective states in in situ code review in a large software company. Nonverbal behavioral signals of 33 professional software developers were recorded unobtrusively while they worked on their daily code reviews. After analyzing these signals using Linear Mixed Effect Models, we observe that affect presented in the written comments is associated with prolonged typing duration. Using physiological features, a trained Random Forest classifier can predict post-task valence with 90.0% accuracy (F1-score = 0.937) and arousal with 83.9% accuracy (F1-score = 0.856). The results show promise for the creation of intelligent affect-aware interfaces for code review.