Automated Labeling of Toxic Conversations on a Large Scale Software Engineering Dataset
A Benchmark Study of the Contemporary Toxicity Detectors on Software Engineering Interactions
An open-source software community may benefit from automated filtering of toxic conversations to maintain healthy interactions among the project participants.However, several general purpose tools exist to identify toxiccontents, those may incorrectly flag some words commonly used in the softwareengineering (se) context as toxic (e.g.,'junk','kill', and'dump') and viceversa.To address this challenge, an se specific tool has been proposed by the cmu strudel lab (`strudel' hereinafter) by combining the output of the perspective api with the output from a customized version of the politeness detector tool.However, 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 sedataset.On this goal, we empirically developed a rubric to manually label toxic se interactions.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 communication such as code review than on our dataset of informal communication such as gitter messages.Based on our manual investigations of the incorrectly classified text, we have identified several recommendations for developing an se specific toxicity detector.Two of the models from our study showed significant performance improvements during 10-fold cross validations after we retrained those on our se datasets.
