Detoxipy is a simple text classifier used to identify toxic comments with the aim that it can be used by developers to quickly flag potentially harmful or hateful comments in any applications which involve textual communication. The project currently contains two classifiers, the Binary Classifier and the Multilabel Classifier which solve the same task, but provide slightly different outputs.

The Binary Classifier is a simple classifier which outputs a 1 if a comment is toxic, and a 0 if not. For example, the non-toxic comment of "Have a nice day!" will output a 0. The Binary Classifier is the most efficient algorithm which can determine whether or not a comment is toxic in approximately 10 milliseconds.

The Multilabel Classifier is slightly more complex as comments can fall into several of up to seven unique categories. These categories are the following:

• Toxic - A comment which can generally be described as harmful
• Severe Toxic - A comment which is very harmful and will often also fall into one or more of the other categories
• Obscene - These comments are typically viewed as overtly sexual, and/or morally offensive
• Threat - Any comment which indicates intentions to harm, typically mentally, physically, or emotionally, another person
• Insult - These comments include generally hateful comments towards another person
• Identity Hate - These comments are typically insults towards someones gender, race, sexual orientation, or others
• Non-Toxic - Any comment that doesn't fall into the previous categories is determined to be non-toxic

Any comment can fall into one or multiple of the first six comments, such as being "toxic", "severe toxic", and "obscene". The only exception comes when a comment is determined non-toxic, where it can only fall into that single category. While the Multilabel Classifier provides greater details on what classification of toxicity a comment falls into, it takes longer to generate a response compared to the Binary Classifier while still being responsive by averaging approximately 90 milliseconds to complete.