基于混合架构神经网络的攻击性言论识别与分类研究

Research on the Recognition and Classification of Offensive Comments Based on Hybrid Architecture Neural Networks

为提高中文攻击性言论识别能力,文中设计了一种基于混合架构的神经网络模型。该模型首先采用BERT对输入的文本序列进行编码,得到文本中每个词语的动态词向量表示;然后应用BiLSTM进一步增强对文本语义的理解,并通过CNN来捕捉局部特定短语或词汇的关键语义特征。实验结果表明,相较于单一架构的神经网络模型,该模型能更好地应用于中文攻击性言论识别任务,具备更高的识别准确性。

In order to improve the recognition ability of Chinese offensive speech,a neural networks model based on hybrid architecture is designed in this paper.The model first uses BERT to encode the input text sequence to obtain the dynamic word vector representation of each word in the text;then uses BiLSTM to further enhance the understanding of text semantics,and uses CNN to capture the key semantic features of local specific phrases or words.Experimental results show that compared with the neural networks model with a single architecture,the model can be better applied to the task of Chinese offensive speech recognition and has higher recognition accuracy.