Adequate Prompting Improves Performance of Regression Models of Emotional Content

Regression models of natural language content are widely popular in many applications, such as sentiment analysis, stance detection and emotion detection. The multidimensional Russell and Mehrabian’s Valence-Arousal-Dominance (VAD) model has been universally recognized as a valuable tool for describing emotional content. Standard methods for VAD prediction learn the relation between natural language text and VAD values by fine-tuning a pre-trained machine learning model based on Transformers (e.g. BERT) on a corpus of natural language sentences manually annotated with VAD. We investigate the potential of employing effective prompting, a technique previously proven to be advantageous in classification and other natural language processing (NLP) tasks, to enhance the VAD prediction process. Our findings reveal that with appropriate prompting, we can leverage the knowledge acquired during pre-training to improve regression performance, showcasing the benefits of this approach for VAD prediction.