Metalinguist: enhancing hate speech detection with cross-lingual meta-learning

Abstract

The rise of social media has led to an increase in hate speech. Hate speech is generally described as a deliberate act of aggression aimed at a particular group, intended to harm or marginalize them based on specific attributes of their identity. While positive interactions in diverse communities can greatly enhance confidence, it is important to acknowledge that negative remarks such as hate speech can weaken community unity and present a significant impact on people’s well-being. This highlights the need for improved monitoring and guidelines on social media platforms to protect individuals from discriminatory and harmful actions. Despite extensive research on resource-rich languages, such as English and German, the detection and analysis of hate speech in less-resourced languages, such as Norwegian, remains underexplored. Addressing this gap, our study leverages a metalinguistic approach that uses advanced meta-learning techniques to enhance the detection capabilities across bilingual texts, effectively linking technical advancements directly to the pressing social issue of hate speech. In this study, we introduce techniques that adapt models that deal with hate speech detection within the same languages (intra-lingual), across different languages (cross-lingual), and techniques that adapt models to new languages with minimal extra training, independent of the model type (cross-lingual model-agnostic meta-learning-based approaches) for bilingual text analysis in Norwegian and English. Our methodology incorporates attention mechanisms (components that help the model focus on relevant parts of the text) and adaptive learning rate schedulers (tools that adjust the learning speed based on performance). Our methodology incorporates components that help the model focus on relevant parts of the text (attention mechanisms) and tools that adjust the learning speed based on performance (adaptive learning rate schedulers). We conducted various experiments using language-specific and multilingual transformers. Among these, the combination of Nor-BERT and LSTM with zero-shot and few-shot model-agnostic meta-learning achieved remarkable F1 scores of 79% and 90%, highlighting the effectiveness of our proposed framework.