Genial...: Automatic Irony Detection in Spanish Tweets

Abstract

Irony is a form of non-literal speech that can alter the meaning of an utterance. Understanding irony may greatly impact Natural Language Processing (NLP) tasks such as sentiment analysis or stance detection. While a growing body of NLP research has started to focus on irony detection, little work has been conducted for other languages such as Spanish. This thesis aims to contribute to research in Spanish irony detection by, taking as a basis an existing dataset (IroSvA) for irony detection in Spanish, revising it and enriching it with annotations for irony types. The improved dataset constitutes the first corpus including labels for types of irony in Spanish. Furthermore, we undertake crosslingual experimentation on irony detection in three different evaluation settings: monolingual, multilingual, and crosslingual. For these experiments, Italian and English datasets were employed in addition to the Spanish ones. Results show that irony does not transfer easily across languages except in the case of Italian to Spanish, for which the results are surprisingly good. Furthermore, training on multiple languages does not help to improve results for irony detection. Results also demonstrate that monolingual language models perform better than multilingual ones. Finally, the thesis offers a detailed and comprehensive analysis and discussion on the difficulties in annotating and learning to detect irony.