Chinese Implicit Sentiment Analysis Combining Sequence and Dependent Structure Features

Abstract:
In the process of sentiment analysis on social platform texts, we found a large number of implicit sentiment texts. For this type of text, traditional sequence models cannot identify them well. Therefore, this paper proposes a method of fusing the sequence information of the text with the dependency syntax information in order to obtain more valuable information in the short text. We build a graph based on the syntax tree and use Graph Convolutional Neural Network (GCN) to learn the overall information. At the same time, the text sequence information is introduced for node features to realize the fusion of two kinds of information. Experiments show that our model has strong superiority in the recognition of implicit emotions compared to traditional sequence models and pre-training models such as ERNIE.