Validating Pseudo-label Dataset for Japanese Hate Speech Detection

Currently, labeled datasets for hate speech detection in Japanese are insufficient. Therefore, we show that it is possible to construct a high-performance model using a pseudo-labeled dataset. This method can reduce the cost and time required to create a labeled dataset for Japanese hate speech detection. In the validation, we compared the performance when trained on a human-labeled dataset and a pseudo-labeled dataset. As a result, the F1 score of the human-labeled dataset was 0.5443, while that of the pseudo-labeled dataset with LLM was 0.5073. Thus, this method can be used to construct a high-performance hate speech detection model without human labeling.