图卷积神经网络在中文对话情感分析中的应用

Application of Graph Convolution Neural Network for Sentiment Analysis of Chinese Dialogue Texts

针对目前对话文本情感分析中大部分模型忽略说话者情感的相互影响作用这一问题,为了有效识别对话文本中说话者本身表达的情感类别,充分考虑对话者之间的情感因素,提出一种基于图卷积神经网络的对话情感分析方法。首先,使用BiGRU将对话文本进行序列上下文编码,获得话语文本表征;然后,依据说话者对话顺序构造一个有向图,利用图卷积神经网络获取每个话语文本新的文本表征向量;最后,连接得到的两个话语表征向量,采用基于相似度的注意力机制获得最终的话语文本表示,从而进行情感分类。在dailydialog中文语料库上的实验结果表明,BiGRU结合GCN模型的方法相比于CNN和BiLSTM模型,在对话文本情感分类方面的准确率大约提高了15%,且F1值也有明显提高,取得较好的情感分类效果。

Aiming at the problem that most of the models ignore the interaction of the speakers’emotion in the current dialogue text emotion analysis,in order to effectively identify the emotional categories expressed by the speakers themselves in the dialogue texts,this paper fully considers the emotional factors among the interlocutors,and proposes a dialogue emotional analysis method based on the graph convolution neural network.Firstly,BiGRU is used to encode the dialogue text in sequence context to obtain the discourse text representation.Then,a directed graph is constructed according to the order of conversation,and the new text representation vector of each discourse text is obtained by using the graph convolution neural network.Compared with CNN and BiLSTM,the accuracy of the algorithm is improved by about 15%,and the F1 score is also improved,achieving good emotion classification effect.