Toxicité et sentiment : comment l'étude des sentiments peut aider la détection de toxicité
|Abstract:||Automatic toxicity detection of online content is a major research field nowadays. Moderators cannot filter manually all the messages that are posted everyday and users constantly find new ways to circumvent classic filters. In this master’s thesis, I explore the benefits of sentiment detection for three majors challenges of automatic toxicity detection: standard toxicity detection, making filters harder to circumvent, and predicting conversations at high risk of becoming toxic. The two first challenges are studied in the first article. Our main intuition is that it is harder for a malicious user to hide the toxic sentiment of their message than to change a few toxic keywords. To test this hypothesis, a sentiment detection tool is built and used to measure the correlation between sentiment and toxicity. Next, the sentiment is used as features to train a toxicity detection model, and the model is tested in both a classic and a subversive context. The conclusion of those tests is that sentiment information helps toxicity detection, especially when using subversion. The third challenge is the subject of our second paper. The objective of that paper is to validate if the sentiments of the first messages of a conversation can help predict if it will derail into toxicity. The same sentiment detection tool is used, in addition to other features developed in previous related works. Our results show that sentiment does help improve that task as well.|
|Document Type:||Mémoire de maîtrise|
|Open Access Date:||14 December 2019|
|Collection:||Thèses et mémoires|
All documents in CorpusUL are protected by Copyright Act of Canada.