Aspect-sited hybrid graph neural networks for sentiment classification
Qi Wu · 2022
Aspect-based sentiment classification is a fine-grained task aimed at inferring the sentiment polarities over specific aspect words in a sentence. Unlike previous approaches that mostly model directly on a given sentence, syntactic dependency trees constructed by graph neural networks have been used in recent years to describe the inner connections between aspect words and related sentiment polarities in sentences with good results. However, we found that the way to use graph neural networks is to model directly on a single kind of multilayer graph neural network, and this approach may not fully explore the inner connections between aspect words and context words in a sentence. Also, the direct modeling of the sentences does not utilize the sequence position information in the sentences and mistakenly identifies irrelevant context words as clues for judging aspect sentiment. In this paper, we propose a new model, ASH-GNNs, which integrates various kinds of graph neural network implementations of semantic dependency trees describing the intrinsic connections between aspects words and related sentiment words, and adds relative position to the input sentences to guarantee the sequence position information in the sentences. The experimental results show that ASH-GNNs outperform all the baseline models.