Comparing DL and BERT Methods on Commits’ Frustration Detection

Frustration and negative emotions can affect the efficiency and productivity of software developers. This is especially true in the OSS (open-source software) community, where the toxicity that pervades the community diminishes the developers’ interest in contributing. While prior investigations have used surveys and biosensors to detect developers’ emotions, these approaches are inefficient and not applicable to large numbers of developers, due to cost- and ethic-related concerns.

We proposed six Deep Learning and BERT models for frustration detection. And we compared their performances among the models to select the best models; these were tested on commits mined from OSS repositories written in Python, Java and C/C++. By studying the original commits and model structures, we analysed the reasons behind the training and testing results, and the different types of errors. Finally, we summarised the experimental results with analysis and discussed the future work.

Results: I got a score of 84 (High Distinction).

You can read my report here