Structured Emotion Analysis from Arabic Text

Abstract

In the field of Natural Language Processing (NLP), emotion analysis aims to map textual content with a predefined set of human emotions, typically including joy, anger, fear, surprise, disgust, and sadness. Current state-of-the-art research mainly focuses on identifying emotions in text using categories inspired by psychological theories, such as Ekman’s (1992) basic emotions. Despite the importance of emotion detection, most analyses are shallow and insufficient for tasks that require a deeper understanding of emotional meaning in context. Such applications necessitate addressing key questions, including identifying the cause that triggered the emotion (Cause), determining who experienced it (Experiencer), and more generally addressing structural questions such as who did what (Cue), to whom (Target), why (Cause), and how (Manner). This doctoral thesis aims to propose original and effective solutions to address the lack of resources and models dedicated to the structural analysis of emotions in Arabic text. To achieve this, we introduce a novel approach for analyzing the argument structure of emotions in Arabic, leveraging recent advances in Transformer-based architectures and, in particular, the capabilities of large language models (LLMs) for Arabic. The main contributions of this thesis are multifold. The first contribution consists of the construction and annotation of the first Arabic corpus dedicated to structured emotion analysis, named ‘AraERL’. The thesis also provides an in-depth examination of the impact of each semantic argument on the performance of emotion identification. In addition, it explores the use of ChatGPT for annotating Arabic texts with seman- tic roles and emotions through an interlingual annotation projection approach. The work further evaluates ChatGPT’s ability to accurately translate English semantic and emotional annotation into Arabic. Finally, it offers a comprehensive comparison of the performance of open large language models for these tasks.