Abstract: Offensive speech spread on the Internet seriously disrupts the normal network order and destroys the network environment for healthy communication. Existing detection technologies are more focused on the distinctive features in the text, and it is difficult to discover more subtle attack methods. In response to the above problems, an offensive speech detection model BSWD (BERT-based Sarcasm and Word Detection) incorporating irony mechanism is proposed. First, a model Sarcasm-BERT based on the irony mechanism is proposed to detect semantic conflicts in speech; secondly, a fine-grained vocabulary offensive feature extraction model WordsDetect is proposed to detect offensive words in speech; finally, the BSWD model is obtained by fusing the two models to improve the detection accuracy.The experimental results show that the accuracy, precision, recall, and F1 score indicators of the proposed model are generally improved by more than 2%, compared with the BERT(Bidirectional Encoder Representation from Transformers) and HateBERT methods, which significantly improves the detection performance and can better detect implicit offensive speech. Compared with the SKS (Sentiment Knowledge Sharing) and BiCHAT (Bidirectional Long Short-Term Memory with deep Convolution neural network and Hierarchical ATtention) methods, it has stronger generalization ability and robustness. The results of the ablation experiment show that the proposed model has an improvement of 0.5%~1.5%, compared with WordsDetect and Sarcasm-BERT in terms of precision, recall, and F1 score indicators, successfully integrating the advantages of each part of the work.

Key words: irony detection, offensive speech detection, fine-grained feature, obscure attack, attention mechanism 
 

摘要： 互联网上散布的攻击性言论严重扰乱了正常网络秩序，破坏了健康交流的网络环境。现有的检测技术更专注于文本中的鲜明特征，难以发现更加隐晦的攻击方式。针对上述问题，提出了融合反讽机制的攻击性言论检测模型BSWD (BERT-based Sarcasm and Word Detection)。首先，提出基于反讽机制的模型Sarcasm-BERT，以检测言论中的语义冲突；其次，提出细粒度词汇攻击性特征提取模型WordsDetect，检测言论中的攻击性词汇；最后，融合两种模型得到BSWD模型，提高检测精度。实验结果表明，与BERT(Bidirectional Encoder Representation from Transformers)、HateBERT模型相比，所提模型的准确率、精确率、召回率和F1分数指标普遍提升2%以上，显著提高了检测性能，更能发现隐含的攻击性言论；同时，与SKS(Sentiment Knowledge Sharing)、BiCHAT(Bidirectional Long Short-Term Memory BiLSTM with deep Convolution neural networkCNN and Hierarchical ATtention)模型相比，具有更强的泛化能力和鲁棒性。消融实验结果表明，与WordsDetect、Sarcasm-BERT相比，所提模型的精确率、召回率和F1分数指标都有0.5%~1.5%的提升，成功集成了每部分工作的优势能力。

关键词: 反讽检测, 攻击性言论检测, 细粒度特征, 隐晦攻击, 注意力机制