Prompting Generative Language Model with Guiding Augmentation for Aspect Sentiment Triplet Extraction

Abstract

Aspect Sentiment Triplet Extraction (ASTE) is a challenging task in natural language processing concerning the automatic extraction of (aspect term, opinion term, sentiment polarity) triplets from a given text. Current end-to-end generative methods achieved high results by treating it as a sequence generation task with a generative pretrained language mode (e.g., T5). However, these architectures usually suffered from the objective gaps between the pre-training tasks and fine-tuning tasks, leading to suboptimal results. Further more, they can only provide information on what is a valid triplet, but no explicit guidance on what is not a triplet, which can not fully capture the correlation between aspects and opinions. To address above issues, we propose the generative prompt to bridge the gap between pre-training and fine-tuning of generative pretrained language model via text infilling task. And we propose guiding augmentation, which drops the aspect or opinion in the sentence by depicting a tree structure to generate diverse similar sentences and new target sequences. In this way, the main differences between these augmented samples are the dropped aspect or opinion term, and the model can understand the ASTE task knowledge better through the explicit variant constraints. Experimental results confirm that our method outperforms previous state-of-the-art (SOTA) methods on four public ASTE datasets.