Affiliation:
1. Shandong Normal University
2. Weifang Institute of Technology
Abstract
Abstract
Recently, PLMs (pre-trained language models) and syntax-based GCNs (graph convolutional neural networks) have been widely and deeply applied in aspect-based sentiment analysis, effectively promoting the development of this field. However, these methods rarely make use of external knowledge, such as knowledge of high-quality sentiment lexicons, which can be a powerful complement to these methods if the prior knowledge can be effectively used. In this paper, to use external knowledge, we propose a Knowledge Enhanced Graph Convolutional Network for aspect-based sentiment analysis (KEGCN). KEGCN automatically calculates the sentiment eigenvalue of each word and extracts strong sentiment words, then retrains the PLM representations of strong sentiment words to integrate external sentiment knowledge into our model. Experimental results on several benchmark datasets demonstrate the effectiveness of our model, which outperforms previous baseline models in terms of both accuracy and macro-F1.
Publisher
Research Square Platform LLC