Highly Generalizable Models for Multilingual Hate Speech Detection
This work addresses the need for scalable hate speech detection across multiple languages for content moderation platforms, though it is incremental in applying existing embeddings to a new dataset.
The paper tackled multilingual hate speech detection by compiling a dataset of 11 languages and using language-agnostic embeddings to develop generalizable models, achieving performance improvements in cross-lingual training scenarios.
Hate speech detection has become an important research topic within the past decade. More private corporations are needing to regulate user generated content on different platforms across the globe. In this paper, we introduce a study of multilingual hate speech classification. We compile a dataset of 11 languages and resolve different taxonomies by analyzing the combined data with binary labels: hate speech or not hate speech. Defining hate speech in a single way across different languages and datasets may erase cultural nuances to the definition, therefore, we utilize language agnostic embeddings provided by LASER and MUSE in order to develop models that can use a generalized definition of hate speech across datasets. Furthermore, we evaluate prior state of the art methodologies for hate speech detection under our expanded dataset. We conduct three types of experiments for a binary hate speech classification task: Multilingual-Train Monolingual-Test, MonolingualTrain Monolingual-Test and Language-Family-Train Monolingual Test scenarios to see if performance increases for each language due to learning more from other language data.