Chinese Online Violent Speech Detection Based on EBLA

INTRODUCTION: The Internet's features of transcending time and space and anonymity have fostered more rampant and covert online violent speech. Thus, accurate and effective management of online public opinion is of great significance. In recent years, scholars both domestically and internationally have conducted extensive research on online violent speech detection, but current challenges include extracting semantics from diverse and implicit expressions in Chinese online violent short texts. OBJECTIVES: This paper aims to propose the EBLA model for online violent speech detection, based on the ERNIE knowledge-enhanced semantic understanding pre-training model and the BiLSTM-Attention network, to precisely identify relevant textual semantic information and provide an effective method for online content moderators. METHODS: The model is trained using publicly available Chinese datasets related to online violence. It enhances deep, sentence-level feature extraction by integrating an attention mechanism into the BiLSTM layer on top of the ERNIE pre-training model. The model consists of vector transformation, deep text feature extraction, and text classification prediction phases. RESULTS: Results show that the precision of this model in identifying Chinese online violence tasks surpasses the BERT pre-training model by 3.7% and outperforms the BiLSTM combined with the attention mechanism by 13.84%. Empirical studies on additional datasets confirm the model's robustness and transferability. CONCLUSION: The EBLA model provides a strong basis for online violent speech detection, though it has limitations such as not accounting for identity bias or dynamic speech nature. Future improvements will focus on multimodal analysis and dynamic monitoring capabilities.