A study of multilingual toxic text detection approaches under imbalanced sample distribution

Multilingual characteristics, lack of annotated data, and imbalanced sample distribution are the three main challenges for toxic comment analysis in a multilingual setting. This paper proposes a multilingual toxic text classifier which adopts a novel fusion strategy that combines different loss functions and multiple pre-training models. Specifically, the proposed learning pipeline starts with a series of pre-processing steps, including translation, word segmentation, purification, text digitization, and vectorization, to convert word tokens to a vectorized form suitable for the downstream tasks. Two models, multilingual bidirectional encoder representation from transformers (MBERT) and XLM-RoBERTa (XLM-R), are employed for pre-training through Masking Language Modeling (MLM) and Translation Language Modeling (TLM), which incorporate semantic and contextual information into the models. We train six base models and fuse them to obtain three fusion models using the F1 scores as the weights. The models are evaluated on the Jigsaw Multilingual Toxic Comment dataset. Experimental results show that the best fusion model outperforms the two state-of-the-art models, MBERT and XLM-R, in F1 score by 5.05% and 0.76%, respectively, verifying the effectiveness and robustness of the proposed fusion strategy.