Detection of everyday metaphor in Spanish: annotation and evaluation

Abstract

Metaphors are pervasive in our daily utterances, which is why the automatic processing of metaphorical expressions has gained popularity in the field of Natural Language Processing, with a view to achieve a more fluid and natural interaction between humans and machines. The development of automatic tools that identify metaphors in English is several steps ahead than in other languages. However, it is important for other linguistic communities to be able to count on these resources as well. With this aim in mind, in this work we focus on the task of Metaphor Detection in Spanish both from corpus-based and computational approaches. On the one hand, we collect and manually label CoMeta: the largest publicly available dataset with metaphorical annotations in texts of general domain for the Spanish language. We address in detail the main questions derived from the application of the MIPVU guidelines used to develop the most popular metaphor corpus for English, namely the VUA corpus, to the Spanish language. On the other hand, we leverage CoMeta and multilingual pre-trained language models based on the Transformer architecture to empirically evaluate the quality of the annotations. The close performance achieved in comparison to the results obtained with the larger English VUA dataset are quite promising and encouraging for future researchers interested in using CoMeta or in developing their own corpora for their languages of interest.