Prompt Learning for Subjective and Objective Recognition of Text

Subjective expressions generally contain the user's emotional opinions and thus can better represent the sentiment information of the originating text reviews, while objective descriptions usually appear in declarative sentences that do not reflect the sentiment polarity of the comment text. Therefore, subjective and objective recognition techniques for text are necessary for natural language processing tasks (e.g. opinion mining, sentiment analysis). However, most existing deep learning-based methods need large human-annotated data, but labeling large amounts of high-quality emotional texts is labor-intensive. The lack of labeled data has become a bottleneck restricting the deep algorithm. To cope with the shortage of labeled data, we employ Prompt Learning strategy in Subjective and Objective recognition task of text, thus proposing a simple PLSO method. PLSO can effectively inherit the abundant prior knowledge from the pre-trained model and learn high-level semantic features of text by reducing the differences between upstream and downstream tasks. Meanwhile, we introduce triplet loss as a regular term in the training objective, which guides the model to capture the category distribution of data in joint training. Benefiting from the above advantages, the PLSO approach significantly outperforms the other baseline methods in the data sparsity scenario, demonstrating its superiority and effectiveness.