Achieving Hate Speech Detection in a Low Resource Setting

Online social networks provide people with convenient platforms to communicate and share life moments. However, because of the anonymous property of these social media platforms, the cases of online hate speeches are increasing. Hate speech is defined by the Cambridge Dictionary as “public speech that expresses hate or encourages violence towards a person or group based on something such as race, religion, sex, or sexual orientation”. Online hate speech has caused serious negative effects to legitimate users, including mental or emotional stress, reputational damage, and fear for one’s safety. To protect legitimate online users, automatically hate speech detection techniques are deployed on various social media. However, most of the existing hate speech detection models require a large amount of labeled data for training. In the thesis, we focus on achieving hate speech detection without using many labeled samples. In particular, we focus on three scenarios of hate speech detection and propose three corresponding approaches. (i) When we only have limited labeled data for one social media platform, we fine-tune a per-trained language model to conduct hate speech detection on the specific platform. (ii) When we have data from several social media platforms, each of which only has a small size of labeled data, we develop a multitask learning model to detect hate speech on several platforms in parallel. (iii) When we aim to conduct hate speech on a new social media platform, where we do not have any labeled data for this platform, we propose to use domain adaptation to transfer knowledge from some other related social media platforms to conduct hate speech detection on the new platform. Empirical studies show that our proposed approaches can achieve good performance on hate speech detection in a low resource setting