Multimodal and Multilingual Hate Speech Detection

Workshop topic and description

As the number and availability of social media platforms grow, the spread of hate speech among online communities (such as Twitter, Facebook, Reddit, Youtube, and so on) is also dramatically increasing. While there is no internationally agreed-upon definition, hate speech is broadly defined as a speech targeted to a given community or group with the potential of inciting violence towards them (Jacobs et al. 2000, Walker 1994). Davidson et al. (2017) discriminate between hate speech (languages used to express hatred towards a target group and incites violence) and offensive speech (usage of rude, hurtful, derogatory, obscene, or insulting language to upset or embarrass people). In this workshop, we broadly define hate speech as “inappropriate language” that is used in online communities, which can be expressed via text, image, or video and could be ultimately handled using an automatic approach.

Furthermore, we also invite researchers who work on the multilingual aspects of this topic. The current developments on multilingual transformer models attract NLP researchers in multiple domains. The recent work by Ghosh Roy et al. (2020) showcases how to build hate speech detection systems with a pre-trained multilingual Transformer-based text encoder. In this workshop, we would like to discuss the challenges, approaches, frameworks, and technologies that could facilitate hate speech detection in multilingual environments and the consequences and implications of hate speech detection approaches for multilingual setup.

Moreover, multimodal aspects, day by day, become an integral part of those above-mentioned communication mediums (Twitter, Reddit, Facebook, etc.). If the message producer provides two sources of information together to deliver his/her message, then it suffices to assume that the meaning is distributed into both modalities to some degree. Particularly, in determining whether a multimodal tweet or memes accompanied with an image or video carries hateful content, uni-modal approaches can easily fail in case of cynicism: e.g. a tweet/meme with a very innocent-looking text accompanied by a very targeted and offensive image content or the other way round, see Figure 1. To mitigate this problem, all the existing modalities of input should be taken into account instead of relying on text-only content. As a result of this need, the number of public multimodal datasets is becoming more available by providing nice testbeds for NLP researchers.