Aspect-level sentiment analysis: A survey of graph convolutional network methods

Aspect-level sentiment analysis (ALSA) is the process of collecting, processing, analyzing, inferring, and synthesizing subjective sentiments of entities contained in texts at the aspect level. The development of social networks has been driven by the on-going appearance of vast numbers of short documents, such as those in which opinions are expressed and comments are made. The text in these documents reflects users' emotions related to entities. The ALSA of these short texts plays an important role in solving various problems in life. Particularly in e-commerce, manufacturers can use sentiment analysis to determine users' orientations, adapt their products to perfection, identify potential users, and pinpoint users that influence other users. Therefore, improving the performance of ALSA methods has recently attracted the interest of researchers. Currently, four main types of ALSA methods are available: knowledge-based, machine learning-based, hybrid-based, and most recently, graph convolutional network (GCN)-based. This study is the first survey to focus on reviewing the proposed methods for ALSA using GCN methods. In this paper, we propose a novel taxonomy to divide GCN-based ALSA models into three categories based on the types of knowledge extraction. We present and compare GCN-based ALSA methods following our taxonomy comprehensively. Common benchmark datasets and text representations that are often used in GCN-based methods are also discussed. In addition, we discuss five challenges and suggest seven future research directions for GCN-based ALSA methods. The findings of our survey are expected to provide the necessary guidelines for beginners, practitioners, and new researchers to improve the performance of ALSA methods.

Automated summary

Aspect-level sentiment analysis (ALSA) is crucial in social networks, especially in e-commerce. This study provides a comprehensive survey on GCN-based ALSA methods, proposing a novel taxonomy and discussing benchmark datasets and text representations commonly used. The study also highlights future research directions and challenges in this field.