Dual-enhanced generative model with graph attention network and contrastive learning for aspect sentiment triplet extraction

Currently, generative models are showing exceptional abilities to identify and generate triplets expressed within sentences within the field of Aspect Sentiment Triplet Extraction (ASTE). Although these models are capable of recognizing terms and sentiment representations, they are not fully capable of generating multi-word aspects and opinion terms. In response to these challenges, this paper presents a dual-enhanced generative model with graph attention network and contrastive learning for aspect sentiment triplet extraction (GAC). In the GAC model, we construct a graph triplet loss module, which integrates dependency syntactic information to deepen the understanding of complex sentence structures, and utilizes graph attention network to explicitly define the dependencies between words, which makes the model better at recognizing aspects and opinions within complex structures. Furthermore, we designed the triplet representation contrastive learning module, which significantly enhances the model's ability to identify complex sentiment types and differentiate aspect and opinion terms composed of single words and sentences by capturing the internal connections between sentiment types and term lengths. In the experimental section, the paper tests two public datasets. According to the results, the GAC model outperforms existing methods in generating triplets, confirming the efficiency and advancement of our approach in tackling the ASTE challenges. Specifically, on different subsets (14lap, 14res, 15res, 16res) of the ASTE-Data-v2 and ASTE-Data-v1 datasets, the F1 scores of our method were 66.47%, 76.01%, 69.04%, 76.25% and 64.14%, 76.44%, 68.94%, 76.37%, respectively.