A novel span-based Knowledge-enhanced framework for aspect sentiment triplet extraction☆

Aspect Sentiment Triplet Extraction (ASTE) is a subtask of Aspect-based Sentiment Analysis, which aims to identify all aspect sentiment triplets in given sentences. ASTE is an important research task to discover sentimental opinions in online social media. Existing ASTE models are usually based on pipeline or joint strategy. The pipeline-based methods may suffer from error propagation. The joint-based methods could avoid error propagation in an end-to-end manner, which have become the more popular choice. Among these joint-based ASTE models, the syntactic information from the dependency tree is widely used. However, the dependency tree may fail to connect the aspect and opinion terms in some cases, which causes the lack of interaction in triplets. Inspired by the work of Liang et al. on the Aspect Sentiment Classification (ASC) task, we integrate common-sense knowledge into the ASTE task, which can bring more important information to the modeling of context representation. To address these limitations, this paper first involves common-sense sentiment knowledge along with syntactic dependency knowledge to get better representations of contexts. The syntactic and common-sense knowledge can contribute to the interpretability when extracting the sentiment elements. We also introduce a self-attention mechanism based on the orthogonal loss function to better capture the interactions between words. We evaluated our model on four public datasets with recently proposed baselines in F1-score metric, especially 3.62%, 4.14%, 3.58% and 2.22% higher on 14Res, 14Lap, 15Res, and 16Res, compared to the SSJE approach, respectively. Experimental results show that our method outperforms SOTA models on three datasets.