CRF-GCN: An effective syntactic dependency model for aspect-level sentiment analysis

Attention-based models have been widely used in aspect-level sentiment analysis owing to their contextual semantic-alignment capabilities. However, these models lack a mechanism that can in-corporate syntactic tendencies. Thus, attention-based neural networks are less accurate when dealing with complex syntactic relationships and long-span syntactically dependent utterances. To solve this problem, a neural network model combining a conditional random field and a graph convolutional network is proposed. The model uses a conditional random field chain to extract the opinion span of an aspect-specific word, integrates the contextual information within the opinion span to the global nodes through a multilayer graph convolutional network using the improved position decay function, and predicts the aspect-specific sentiment polarity labels by computing the vector expressions of the global nodes. This study addresses the problem of fluctuations in model accuracy when multiple aspect words exist in an utterance by introducing global nodes into a graph convolutional network. The proposed model performed well in the experiments on all four datasets tested.(c) 2022 Elsevier B.V. All rights reserved.

Automated summary

Attention-based models are widely used in aspect-level sentiment analysis due to their contextual semantic-alignment capabilities. However, these models lack the ability to incorporate syntactic tendencies, resulting in lower accuracy when dealing with complex syntactic relationships and long-span syntactically dependent utterances. To address this problem, a neural network model combining a conditional random field and a graph convolutional network is proposed. The proposed model integrates contextual information within the opinion span to global nodes and predicts aspect-specific sentiment polarity labels by computing vector expressions of the global nodes.