Different lexicon-based approaches to emotion identification in Portuguese tweets

Abstract

This paper presents the existing literature on the identification of emotions and describes various lexica-based approaches and translation strategies to identify emotions in Portuguese tweets. A dataset of tweets was manually annotated to evaluate our classifier and also to assess the difficulty of the task. A lexicon-based approach was used in order to classify the presence or absence of eight different emotions in a tweet. Different strategies have been applied to refine and improve an existing and widely used lexicon, by means of automatic machine translation and aligned word embeddings. We tested six different classification approaches, exploring different ways of directly applying resources available for English by means of different translation strategies. The achieved results suggest that a better performance can be obtained both by improving a lexicon and by directly translating tweets into English and then applying an existing English lexicon.