A Study on the Feasibility to Detect Hate Speech in Swedish
Johan Fernquist, Oskar Lindholm, Lisa Kaati, Nazar Akrami · 2019
Hate speech in digital environments is becoming a societal challenge. To deal with the problem, techniques that automatically detect hate speech have been developed by social media companies as well as researchers. Hate can be expressed in many different ways, which makes it difficult to detect automatically using algorithms. Also, how hate is expressed depends heavily on the language. The effectiveness of automatic detection techniques is still to be improved in many languages. In this paper, we attempt to detect hate speech in Swedish using machine learning. We compare different pre-trained language models that are fine-tuned on a corpus of hateful comments. To examine how well our models would work in a real scenario, we used a set of randomly selected comments from a Swedish discussion forum. The results showed that using pre-trained language models provides a better result than using a baseline SVM model, but it also reveals that detecting hate speech in the wild is challenge that need more research.