Span-Based Pair-Wise Aspect and Opinion Term Joint Extraction with Contrastive Learning

Abstract

Pair-wise aspect and opinion term extraction (PAOTE) is a subtask of aspect-based sentiment analysis (ABSA), aiming at extracting aspect terms (AT) and corresponding opinion terms (OT) in the review sentences in the form of aspect-opinion pairs (AO), which provides global profiles about products and services for users. Existing methods cannot identify the boundary information of AT and OT precisely. In addition, pairing errors are easily caused if there are multiple AT and OT in a sentence. In our work, to address above limitations, we design a span-based joint extraction framework with contrastive learning (SJCL) to enhance both term extraction and pairing in PAOTE. For term extraction, we utilize span-based convolutional neural network (CNN) to merge contextual syntactic and semantic features to identify the boundaries of aspect and opinion terms precisely and combine contrastive learning (CL) to enhance the distinctiveness of different types of terms. For term pairing, we prune the original dependency trees according to part-of-speech (POS) information to remove insignificant noise information and leverage graph convolutional network (GCN) to learn pairing features, then negative sampling and contrastive learning are used to avoid mismatched aspect-opinion pairs. Experimental results on four public datasets certify the state-of-the-art performance of our model.