Exploring Sensory Knowledge and Pre-training Language Models for Chinese Metaphor Detection

Processing of metaphorical languages has been one of the most challenging tasks in Natural Language Processing (NLP). Recent work utilizing neural models has received notable results on metaphor detection. However, the characteristic of metaphorical languages showing an intimate relationship with sensory experiences has not been given enough attention in NLP. This study proposes an innovative model for the task of Chinese metaphor detection by incorporating the conceptual knowledge of sensory experiences into a neural network model. Experiments show that our model outperforms the state-of-the-art baseline models significantly, hence contributing to the ongoing efforts to incorporate neuro-cognitive data in NLP tasks. In addition, the effectiveness of our model helps to deepen our understanding of metaphor that sensory experiences form the crucial part of the embodied nature of metaphorical languages.