Prashant Kapil and Asif Ekbal, IIT Patna, India

In recent years, the increasing propagation of hate speech on social media has encouraged researchers to address the problem of hateful content identification. To build an efficient hate speech detection model, a large number of annotated data is needed to train the model. To solve this approach we utilized eleven datasets from the hate speech domain and compared different transformer encoder-based approaches such as BERT, and ALBERT in single-task learning and multi-task learning (MTL) framework. We also leveraged the eight sentiment and emotion analysis datasets in the training to enrich the features in the MTL setting. The stacking based ensemble of BERT-MTL and ALBERT-MTL is utilized to combine the features from best two models. The experiments demonstrate the efficacy of the approach by attaining state-of-the-art results in all the datasets. The qualitative and quantitative error analysis was done to figure out the misclassified tweets and the effect of models on the different data sets.

BERT, Multi-task learning, Hate speech, Transformer, Ensemble.