Cross-domain Sentiment Classification with Prompt Pre-training and Tuning

Sentiment classification, as an essential task for understanding the sentiment content of texts, has garnered extensive attention. However, existing methods face numerous challenges, such as the mismatch between training and testing data in specific domains, which results in insufficient adaptability of models to new domains. Prompt learning is a parameter-efficient strategy that effectively enables pre-trained models to adapt to downstream tasks. In this paper, we propose a novel two-stage method called PT<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>, which consists of two key components: prompt pre-training and prompt fine-tuning. During the prompt pre-training phase, we employ masked language modeling and cross-domain prompt learning, aiming to uncover latent sentiments and acquire cross-domain knowledge. In the prompt fine-tuning phase, we apply the pre-trained model to specific cross-domain sentiment classification tasks, adjusting the model parameters through additional sentiment prompt templates to improve its performance on these specific tasks. We conducted experiments on publicly available datasets and established nine different cross-domain settings. The results indicate that the model, employing our proposed $\mathrm{PT}^{2}$, significantly outperforms previous state-of-the-art approaches in cross-domain sentiment classification tasks.