Enhancing Grid Tagging Schema with Hybrid Contrastive Learning for Aspect-level Sentiment Triplet Extraction

Abstract

Aspect-based sentiment triplet extraction (ASTE) is a fine-grained opinion mining to jointly extract a triplet consisting of aspect terms, opinion terms, and sentiment polarity. Previous works tried incorporating extra dependency information by stacking complex graph convolution networks (GCN) on transformers to learn the relationship between the aspects and their associated opinion. While the number of existing datasets is still insufficient, the GCN-accompanied redundant parameters in each method require more samples to provide supervised information, leading to the model being trapped by overfitting. Instead of stacking GCNs, this study proposes an enhanced grid tagging scheme with hybrid contrastive learning, exploiting a self-supervised training approach from the negative samples. Specifically, sentiment-aware contrastive learning (SCL) is applied by introducing word-level continuous sentiment information in the valence-arousal space. Syntactic-aware graph contrastive learning (SGCL) is used to guide the language model to gain the ability to capture dependencies. Supervised tagging contrastive learning (STCL) is further adapted to guide the language model to distinguish both positive and negative samples. Hybrid contrastive learning can provide extra supervision information but only introduce tolerable parameters. Extensive experimental results on the ASTE-Data-V2 datasets show that the proposed model significantly outperforms previous baselines.