Decoding Dehumanization: Leveraging NLP to Identify Dehumanizing Language and Its Targets

With the rise of social media, hate speech has become increasingly widespread and potent.A notable type of hate speech is dehumanization, which has been hypothesized to enable harmful behavior towards members of society.The large-scale automatic identification of dehumanizing language and its targets can potentially aid online content moderation efforts.We compiled a list of dehumanizing terms (DHT), which refer to words that strips individuals or groups of their humanity, and target identities (TI), the individuals or groups that are dehumanized.Using this DHT list, we selected Reddit conversations for annotation of TI.As a result, we created the first dataset of English conversation threads where both DHT and their TI are annotated.We show that large pre-trained language models can be used effectively to automatically verify DHTs and identify their TIs.