Abstract: A large number of harmful texts on the Internet adopt implicit and euphemistic expressions to evade detection by censorship systems. Most of the current work focuses on explicit harmful speech and cannot effectively detect implicit harmful text. This paper investigates the detection of implicit euphemistic harmful text in Chinese using a multitask learning approach, where euphemistic sentence recognition is used to assist harmful text detection. Firstly, methods for integrating euphemistic language vocabulary features are explored to enhance the model’s representation of implicit meanings. Subsequently, contrastive learning is applied to enhance latent semantic representations and extract common features from implicitly harmful discourse. Finally, a multitask learning framework is constructed by combining euphemistic sentence recognition tasks with harmful text detection tasks, aiming to improve the detection performance through shared multitask parameters and multifeature fusion loss functions. The experimental results demonstrate the effectiveness of the model in detecting implicit harmful text.

Key words: implicit harmful text, euphemism, multitask learning,  , learning, contrastive learning

摘要： 互联网中大量有害文本采用了隐晦的委婉表达形式，以躲避审查系统.目前大多研究都集中在明确或显性的有害言论上，无法有效地检测伪装的隐式有害文本形式.因此，开展基于多任务学习的中文隐式委婉表达有害文本检测研究，提出了一个隐式有害文本检测模型(IHTDKML)，将委婉句子识别任务用于辅助有害文本检测任务.首先，研究委婉语词汇特征融合方法，提高模型对隐含含义的表征能力；随后，研究了基于对比学习的潜在含义知识增强，学习到共享含义的隐性有害言论的共同特征；最后，联合委婉句识别任务和有害文本检测任务构建多任务学习框架，通过多任务参数共享和多特征融合损失函数提高模型的检测性能.实验结果全面展示了模型在检测隐性有害文本任务上的有效性.

关键词: 隐性有害文本, 委婉语, 多任务学习, 提示学习, 对比学习