多尺度上下文与动态词典增强的中文毒性言论检测OA

Current Chinese toxicity detection models are generally limited in their ability to handle the"semantic vari-ants"that are expressed in an obscure and diverse manner due to their isolated text analysis approach.To break this limita-tion,this paper proposes a multi-scale context and dynamic dictionary enhanced Chinese toxicity detection method(GLL-Tox).The core idea of this method is to construct a systematic analysis framework,placing the text to be tested in a three-dimensional information field of"macro-micro-prior"for comprehensive judgment.Specifically,the method captures the macro semantic distribution of the corpus through macro semantic benchmark modeling to provide a stable discrimination benchmark;it retrieves similar samples through the local environment feature module to construct a micro context to enhance the robustness of variant recognition;and it dynamically injects domain prior knowledge into the semantic focus through the dictionary feature enhancement module.Experiments on the TOXICN and COLDataset public datasets show that the F1 scores of GLL-Tox reach 83.98%and 83.65%respectively,significantly improving by 3.68 and 3.35 percentage points compared to the RoBERTa baseline model,and outperforming existing comparison methods in all aspects.This study proves that systematically integrating macro corpus distribution,micro similar context and domain prior knowledge is an effective paradigm for improving the accuracy and robustness of Chinese toxicity detection.