Transfer Learning for Spanish Emotion Detection
Luna De Bruyne and Véronique Hoste

In NLP, emotions and the possibilities to automatically identify them have been studied primarily on English data, mainly using supervised machine learning approaches. Indeed, supervised machine learning has proven powerful in solving many NLP tasks, including emotion detection, but its dependence on high-quality annotated data and their lack of generalizability (across domains, genres and emotion frameworks) is problematic. We wish to investigate how transfer learning (Pan and Yang, 2010) can tackle this data acquisition problem for low-resource languages.

Chen et al. (to appear) propose a method to transfer knowledge from English to Chinese and Arabic for sentiment classification, using adversarial training. Here, sentiment and language classification are learned simultaneously, aiming to optimize the performance on sentiment classification. However, by hampering the language classification task, language-independent features will be learned.

In this exploratory study, we want to test the proposed methodology of Chen et al. (to appear) to transfer knowledge from an English emotion detection system to other languages, focusing on Spanish. We will use the English and Spanish data from SemEval-2018 (Mohammad et al., 2018). The datasets have the same domain (topic-independent tweets) and label spaces, but differ in language (English versus Spanish). We will assess the transfer learning system’s performance by comparing it with a translation-based approach and a traditional system that is trained from scratch.