Attention Feature Fusion Graph Convolutional Network for Target-Oriented Opinion Words Extraction.

Target-Oriented Opinion Words Extraction (TOWE) is a challenging sequence extraction task aimed at extracting opinion words corresponding to the opinion targets for a given sentence. Enhancing the performance of TOWE requires careful consideration of the semantic information within the sentence, particularly in relation to the opinion words and opinion targets. Although utilizing graph convolutional operations on the syntactic dependency tree allows for the utilization of syntactic dependency information, these operations do not effectively balance the degree of dependency on syntactic parsing results. This paper proposes an Attention Feature Fusion Graph Convolutional Network (AFFGCN) to address the issue. The proposed method enriches the feature representation of nodes through Graph Convolutional Networks (GCN) and captures the sequence features of the sentence using a Bi-LSTM. The Global Feature-aware Attention Module (GFA) guides the model to learn the global feature representation of the sentence to determine the absolute importance of a single word in the sentence. The Neighborhood-aware Attention in Feature Fusion Encoding Module (FFE) fully considers the syntactic structure of sentences to construct a high-quality syntactic perception representation. The experimental results demonstrate the effectiveness of our proposed method. The performance of AFFGCN is comparable to or even better than the state-of-the-art TOWE baseline models. [ABSTRACT FROM AUTHOR]