Aspect Target Sentiment Classification with Instance Retrieval-Based Prompt Template Selection

Aspect Target Sentiment Classification (ATSC) aims to predict the sentiment polarity of a sentence's specific aspect terms. Fine-tuned pre-trained language models (PLMs) that achieve excellent results in ATSC typically use a large amount of training data from the same domain. However, annotation is a costly task. To address this issue, we propose a retrieval-based instance-level prompt template selection method that better solves the ATSC problem through natural language prompts. This method uses an external retrieval source to measure the quality of prompt templates and performs the prompt template selection task as a sentence correctness judgment process. To avoid deviating from the real low-resource scenario, we use a large language model to generate a pseudo-label dataset as the external retrieval source for our method. Our experiment shows that, in the zero-shot setting, our method outperforms the baseline prompt method by 8.21 % in accuracy and 9.68% in macro-Fl. In the few-shot setting, our method also significantly surpasses it. In addition, we also demonstrate that our method can solve cross-domain data issues.