Transformer based contextualization of pre-trained word embeddings for irony detection in Twitter

Human communication using natural language, specially in social media, is influenced by the use of figurative language like irony. Recently, several workshops are intended to explore the task of irony detection in Twitter by using computational approaches. This paper describes a model for irony detection based on the contextualization of pre-trained Twitter word embeddings by means of the Transformer architecture. This approach is based on the same powerful architecture as BERT but, differently to it, our approach allows us to use in-domain embeddings. We performed an extensive evaluation on two corpora, one for the English language and another for the Spanish language. Our system was the first ranked system in the Spanish corpus and, to our knowledge, it has achieved the second-best result on the English corpus. These results support the correctness and adequacy of our proposal. We also studied and interpreted how the multi-head self-attention mechanisms are specialized on detecting irony by means of considering the polarity and relevance of individual words and even the relationships among words. This analysis is a first step towards understanding how the multi-head self-attention mechanisms of the Transformer architecture address the irony detection problem.