Prompt and Target Template guided Diversified Paraphrasing for Aspect Sentiment Tuple Prediction

Aspect-based sentiment analysis aims to extract and predict sentiment elements from a comment sentence, including aspect term, aspect category, opinion term and sentiment polarity. Existing generation-based methods either generate sentiment elements in a fixed order, neglecting their interdependencies, or solely employ sentiment elements arranged in multiple orders as generation targets, which fail to leverage the label’s semantic information. We propose a prompt and target sequence template guided diversified paraphrasing method to fully utilize the knowledge acquired by the model during the pre-training stage, the generation order of sentiment elements and label’s semantic information. Specifically, we construct corresponding target sequence templates containing natural language expressions and marker symbols for different orders, design order-based prompts to guide the model in generating sentiment elements, and further obtain the final prediction results by aggregation and filtering. The extensive experiments on benchmark datasets for multiple tasks demonstrate that the proposed method improves performance significantly across all settings, particularly in lowresource scenarios, compared to previous state-of-the-art methods.