MetaPro: A computational metaphor processing model for text pre-processing

Abstract

Metaphor is a special linguistic phenomenon, challenging diverse natural language processing tasks. Previous works focused on either metaphor identification or domain-specific metaphor interpretation, e.g., interpreting metaphors with a specific part-of-speech, metaphors in a specific application scenario or metaphors with specific concepts. These methods cannot be used directly in everyday texts. In this paper, we propose a metaphor processing model, termed MetaPro, which integrates metaphor identification and interpretation modules for text pre-processing. To the best of our knowledge, this is the first end-to-end metaphor processing approach in the present field. MetaPro can identify metaphors in a sentence on token-level, paraphrasing the identified metaphors into their literal counterparts, and explaining metaphoric multi-word expressions. It achieves state-of-the-art performance in the evaluation of sub-tasks. Besides, the model can be used as a text pre-processing method to support downstream tasks. We examine the utility of MetaPro text pre-processing on a news headline sentiment analysis task. The experimental results show that the performance of sentiment analysis classifiers can be improved with the pre-processed texts.