Abstract
Online hate speech has become a major social problem owing to the rapid growth of Internet communities. Relying on anonymity, people use hateful or abusive language for groups who are different from them. As these terms vary by region and are reflected in local languages, it is important to build robust hate speech detection models for each local language. We propose an ensemble of several Bidirectional Encoder Representations from transformers (BERT)-based models to enhance English and Korean hate speech detection. Parallel Model Fusion (PMF) requires the results of BERT-based models and a final estimator called meta-learner. During each cross-validation, validation and testing results were used to train and test the PMF data. PMF test data are calculated using Majority Voting Integration or Weighted Probabilistic Averaging. Popular machine learning algorithms such as Random Forest, Logistic Regression, Gaussian Naïve Bayes, and Support Vector Machine are employed as meta-learners for PMF. The proposed model outperformed previous studies and the single-model approach in English and Korean, with accuracies of 85% and 89%, respectively, for each dataset. This study demonstrates improved automatic hate speech detection and encourage not only studies on English hate speech detection but also further work on non-English hate speech detection.