Commonsense knowledge graph-based adapter for aspect-level sentiment classification

Attention-and graph-based models are widely applied to existing aspect-level sentiment classification (ALSC) tasks. In spite of the effectiveness, most of these studies ignore commonsense knowledge of aspects. When sentences have the same syntactic structures and opinion words, the sentiment polarities toward aspects can be different. Moreover, no easy and flexible method has been proposed to infuse knowledge into existing ALSC models. For these reasons, we propose a novel commonsense knowledge graph-based adapter (CKGA) for ALSC. Firstly, we link aspects to a knowledge graph end extract an aspect-related sub-graph. Then, a pre-trained language model and the knowledge graph embedding are utilized to encode the commonsense knowledge of entities based on which the corresponding knowl-edge is extracted with a graph convolutional networks. Specifically, CKGA is an adapter-based model which can be added to existing models in a simple way without modifying the original models. Experimental results on three benchmark datasets illustrate that state-of-the-art ALSC models can be sig-nificantly improved with CKGA. Thus, CKGA can leverage external knowledge to enhance the sentiment delivery on the task of ALSC.(c) 2023 Elsevier B.V. All rights reserved.

Automated summary

Attention-and graph-based models are commonly used in aspect-level sentiment classification tasks. However, most studies neglect the commonsense knowledge of aspects. This paper proposes a novel commonsense knowledge graph-based adapter (CKGA) that can improve the performance of existing models by incorporating external knowledge. Experimental results show that CKGA significantly enhances state-of-the-art aspect-level sentiment classification models.