How the Google Toxicity Tool (Google’s Perspective API) works:

Tool available at: https://www.perspectiveapi.com/#/home

From the paper analyzing 4chan /pol/ at https://arxiv.org/pdf/2001.07487.pdf (see >>7891288 lb):

"6.2 Toxicity

Next, we set to score the content of the posts according to how toxic, inflammatory, profane, insulting, obscene, or spammy the text is. To this end, we use Google’s Perspective API [26], which offers several models for scoring text trained over crowdsourced annotations. We focus on the following 7 models:

– TOXICITY and SEVERE_TOXICITY: quantify how rude or disrespectful a comment is; note that the latter is less sensitive to messages that include positive uses of curse words compared to the former.

– INFLAMMATORY: how likely it is for a message to “inflame” discussion.

– PROFANITY: how likely a message contains swear or curse words.

– INSULT: how likely it is that a message contains insulting or negative content towards an individual or group of individuals.

– OBSCENE: how likely a message contains obscene language.

– SPAM: how likely a message is spam.

Note that INFLAMMATORY, OBSCENE, and SPAM models use data obtained from the New York Times."