[1K3-ES-2] Machine learning: Event detection

[1K3-ES-2-04] A Preliminary Analysis of Offensive Language Transferability from Social Media to Video Live Streaming

〇Zhiwei Gao¹, Shuntaro Yada¹, Shoko Wakamiya¹, Eiji Aramaki¹ (1. Nara Institute of Science and Technology)

キーワード：Offensive language, Transfer learning, Deep learning, Social media, Twitch

Given the growing popularity of online games and eSports, the young generation is increasingly enjoying its video live streaming service. Streaming channels are usually combined with chat rooms, where offensive conversations often appear against the streamer or audience. Such offensive languages may cause many serious impacts on a victim’s life and even lead to teen suicide. One method of detecting offensive language is to use deep learning techniques. This research aims to detect offensive language appearing in video live streaming chats. Focusing on Twitch, the most popular live streaming platform, we created a dataset for the task of detecting offensive language. We collected 2000 chat posts across four popular game titles with genre diversity (i.e., competitive, violent, peaceful). Making use of the similarity in offensive languages among social media, we adopt the state-of-the-art models trained over the offensive language on Twitter to our Twitch data (i.e., transfer learning). Our results show that transfer from social media to live streaming is possible. However, the similarity of chat posts or target contents does not help to predict the tranferability with limited data.