Word-level and phrase-level strategies for figurative text identification

Metaphors are a common language tool used in communication that widely exist in natural language. Metaphor detection is beneficial to improving the performance of natural language processing tasks such as machine translation. The current metaphor detection method captures the semantic incoherence of the metaphorical word and its surrounding text. However, it ignores the different contributions of phrases and words. In addition, word information and its contextual information should be deeply integrated. We propose a learning framework that combines word representations, contextual representations and combined representations. Specifically, we establish word-level and phrase-level attention mechanisms to learn enhanced feature representations. For the word-level attention, we extract word embedding, part-of-speech (POS) and distance features as multilevel representations and use multi-attention to obtain weight information. For the phrase-level attention, the IndRNN and self-attention are employed to obtain the deep semantic representation of the sentence. By using this strategy, our model has the ability to mine richer semantic representations. Experiments on the VUA metaphor corpus show that our method achieves an F-score of 69.7% on the test dataset, thereby surpassing the scores of all the existing models by a considerable margin.