Multimodal metaphor detection based on distinguishing concreteness

Chang Su, Weijie Chen, Ze Fu, Yijiang Chen

Abstract: Metaphors are a common linguistic phenomenon, and metaphor identification plays an essential role in metaphor processing. Most existing metaphor computing techniques use only texts to gain features, but we acquire additional knowledge from other modalities. At present, the multimodal model in the metaphor field is in the exploratory stage, and the few multimodal models available are still relatively crude. We propose a multimodal metaphor detection method according to the idea that different types of words are suitable for different modality calculations. First, our proposed framework uses a fine-grained concreteness calculation method based on part of speech to distinguish abstract and concrete words. We then choose a different appropriate modal feature and a different metaphor computational method for words with different concreteness. Additionally, we also improve the use of image features in the field of metaphor detection.