PromptASTE: Prompting a Dataset from Pre-trained Language Models for Unsupervised Aspect Sentiment Triplet Extraction

Anonymous

Keywords: unsupervised learning, aspect-based sentiment triplet extraction, prompt learning, data generation

Abstract: Aspect sentiment triplet extraction (ASTE) is a sentiment analysis task that aims to extract views' sentiment polarity, expression, and target (aspect). While the zero-shot scenario for the sentence or aspect-level sentiment has made much progress in recent years, zero-shot ASTE remains unstudied because of its far more complex data structure. This paper challenges this remaining problem and proposes the first unsupervised method for aspect sentiment triplet extraction, which even does not require any training on human-annotated data. Based on the previous discovery of the pre-trained language model (PLM)'s awareness of sentiment, we further leverage the masked language model (MLM) to prompt an ASTE dataset with automatically annotated labels. Our method, PromptASTE, fills in a series of prompts to generate a dataset for related aspects and views. The dataset is then used to train an ASTE model for prediction. Training on PromptASTE results in models with an outstanding capability in discerning sentiment polarities and targeted aspects. Our model sets the first and strong baseline on unsupervised ASTE.

Paper Type: long

Research Area: Sentiment Analysis, Stylistic Analysis, and Argument Mining