Fine-Grained Emotion Analysis of Arabic Tweets: A Multi-target Multi-label Approach

Abstract

Emotion Analysis (EA) is the task of determining the emotion of a given piece of text. This is an important task with many applications especially when applied to tweets. However, existing work take a rather coarse-grained approach by assuming that each tweet has a single emotion and that this emotion has no intensity. In this work, we take a fine-grained approach by considering cases where a single tweet may have several emotions (multi-label) each with possibly different intensity (multi-target). Moreover, unlike existing work, which consider languages such as English and Chinese, we focus on the Arabic language, a severely under-studied language despite its importance. We build the first dataset (to the best of our knowledge) of Arabic tweets annotated for emotion analysis as a multi-label multi-target problem. Two human experts participated in the annotation process and Cohen's Kappa measure was used to determine their concordance.