Multi-granularity enhanced graph convolutional network for aspect sentiment triplet extraction

Abstract

Aspect Sentiment Triple Extraction (ASTE) is an emerging sentiment analysis task, which describes both aspect terms and their sentiment polarity, as well as opinion terms that represent sentiment polarity. Some models have been presented to analyze sentence sentiment more accurately. Nonetheless, previous models have had problems, like inconsistent sentiment predictions for one-to-many, many-to-one, and sequence annotation. In addition, part-of-speech and contextual semantic information are ignored, resulting in the inability to identify complete multi-word aspect terms and opinion terms. To address these problems, we propose a Multi-granularity Enhanced Graph Convolutional Network (MGEGCN) to solve the problem of inaccurate multi-word term recognition. First, we propose a dual-channel enhanced graph convolutional network, which simultaneously analyzes syntactic structure and part-of-speech information and uses the combined effect of the two to enhance the deep semantic information of aspect terms and opinion terms. Second, we also design a multi-scale attention, which combines self-attention with deep separable convolution to enhance attention to aspect terms and opinion terms. In addition, a convolutional decoding strategy is used in the decoding stage to extract triples by directly detecting and classifying the relational regions in the table. In the experimental part, we conduct analysis on two public datasets (ASTE-DATA-v1 and ASTE-DATA-v2) to prove that the model improves the performance of ASTE tasks. In four subsets (14res, 14lap, 15res, and 16res), the F1 scores of the MGEGCN method are 75.65%, 61.62%, 67.62%, 74.12% and 74.69%, 62.10%, 68.18%, 74.00%, respectively.