"What Has Been Said Cannot Be Taken Back": A Toxic Speech Detection Framework for TikTok using Whisper and Perspective API

Abstract

["""On social media platforms, such as TikTok, toxic speech is a common problem. With a focus on videos from the 2020 US presidential election, this study suggests a framework for spotting toxic speech in TikTok videos. For the purpose of transcribing and analyzing spoken content in TikTok videos, the framework combines a speech-to-text algorithm and a toxicity detection API. The findings show that TikTok videos have varying amounts of toxic speech, with the majority of texts scoring low for toxicity. With the help of BERTopic, semantic characteristics extraction, dominant topics like Joe Biden's actions and discussions of race and politics are identified. Sentiment analysis shows different emotional tones across topics. It is also shown that there may be a correlation between some sentiments and higher levels of toxicity by looking at the relationship between toxicity and sentiment. These findings provide insights into the characteristics of toxic speech in TikTok videos. The results contribute to the development of strategies for content moderation and the promotion of healthier online communities. Future research should address limitations and further explore toxic speech on video-based social media platforms.""]