Prompt-Based Generative Multi-label Emotion Prediction with Label Contrastive Learning

Multi-label emotion prediction, which aims to predict emotion labels from text, attracts increasing attention recently. It is ubiquitous that emotion labels are highly correlated in this task. Existing state-of-the-art models solve multi-label emotion prediction in sequenceto-sequence (Seq2Seq) manner, while such label correlations are merely leveraged in decoding side. In this work, we propose an emotion prediction framework to jointly generate emotion labels and template sentences via Seq2Seq language model. On the one hand, our template-based natural language generation method makes better use of generative language model compared with generating label sequences in the prior Seq2Seqbased generative classification model. On the other hand, we introduce the Correlation-based Label Prompts (CLP) through soft prompt learning and contrastive learning, which enables our model to further consider emotion label correlations in encoding side. To demonstrate the effectiveness of our prompt-based generative multi-label emotion prediction model, we perform experiments on the GoEmotions and SemEval 2018 datasets, achieving competitive results, outperforming 7 baselines w.r.t. 3 evaluation metrics. In-depth analyses show the generation manner is much more impressive compared with generating label sequences and our model is particularly effective in label correlation modeling.