An Interactive Graph Convolutional Network of Reconstructing Dependence Matrix with External Knowledge
Aspect-based sentiment analysis aims to predict the emotional polarity of a particular aspect in a given sentence.However,most of the existing researches ignore the effective use of external knowledge in the process of improving the model,which leads to the failure of the model to accurately identify the affective polarity of aspect words.In addition,considering only limited contextual information may cause the model to ignore the importance and influence of aspect words in the sentence.Based on this,we propose a graph convolutional network method combining external knowledge to construct dependency enhancement matrix.The word embeddings and position embeddings are fused to obtain sentence feature vectors containing position information firstly.At the same time,emotional common sense information is introduced to extend the original dependency matrix and form a new dependency enhancement matrix(DEM),which highlights the weight information of emotion words to a certain extent,and makes up for the shortage of the original de-pendency tree to capture edge labels.In addition,we construct an information interactive network(IIN)to extract global sentence information through the interaction between aspect words and context.Finally,experimental results on five benchmark datasets show a 1 to 2 percentage points improvement in accuracy and Macro-F1,respectively,compared to the baseline model.
国家科技期刊平台
NSTL
万方数据
