A Benchmark Study of the Contemporary Toxicity Detectors on Software Engineering Interactions

TL;DR

This study evaluates the effectiveness of existing toxicity detection tools on large-scale software engineering communication datasets, revealing significant performance degradation and proposing improvements for SE-specific toxicity detection.

Contribution

It provides a comprehensive empirical evaluation of toxicity detectors on SE data, introduces a manual labeling rubric, and offers recommendations for SE-specific toxicity detection improvements.

Findings

All tools' performance degraded significantly on SE datasets.

Performance was worse on formal communication like code reviews.

Retraining improved some models' accuracy.

Abstract

Automated filtering of toxic conversations may help an Open-source software (OSS) community to maintain healthy interactions among the project participants. Although, several general purpose tools exist to identify toxic contents, those may incorrectly flag some words commonly used in the Software Engineering (SE) context as toxic (e.g., 'junk', 'kill', and 'dump') and vice versa. To encounter this challenge, an SE specific tool has been proposed by the CMU Strudel Lab (referred as the `STRUDEL' hereinafter) by combining the output of the Perspective API with the output from a customized version of the Stanford's Politeness detector tool. However, since STRUDEL's evaluation was very limited with only 654 SE text, its practical applicability is unclear. Therefore, this study aims to empirically evaluate the Strudel tool as well as four state-of-the-art general purpose toxicity detectors on a large scale SE dataset. On this goal, we empirically developed a rubric to manually label toxic SE interactions. Using this rubric, we manually labeled a dataset of 6,533 code review comments and 4,140 Gitter messages. The results of our analyses suggest significant degradation of all tools' performances on our datasets. Those degradations were significantly higher on our dataset of formal SE communication such as code review than on our dataset of informal communication such as Gitter messages. Two of the models from our study showed significant performance improvements during 10-fold cross validations after we retrained those on our SE datasets. Based on our manual investigations of the incorrectly classified text, we have identified several recommendations for developing an SE specific toxicity detector.