Tools
Tools
Atusiuba, Amalachukwu Adaeze (2025) Detecting Depression Over Time: Fusing Emoji and Text Representations with Transformer and CLIP Architectures. Masters thesis, Dublin, National College of Ireland.
Preview |
PDF (Master of Science)
Download (2MB) | Preview |
Preview |
PDF (Configuration Manual)
Download (445kB) | Preview |
Abstract
Depression is a mental disorder that can severely impact an individual's life and potentially lead to self-harm. Social media has become an avenue for bringing people together to share their opinions, and this is particularly found among people who are isolated or suffer from different forms of depression. Previous research has used different Natural Language Processing and Artificial Intelligence techniques to detect depressive disorders in people on various social media platforms, including predicting whether a text is depressed or not depressed. Current detection approaches focus on isolated textual posts, overlooking the importance of temporal context and non-verbal cues, such as emojis. To address this gap, this research introduces a longitudinal multimodal framework that analyses sequences of user posts over time, combining CLIP-based emoji embeddings with Transformer-based textual models. The framework also incorporates emoji-to-emotion sentiment mapping to enhance the emotional context of each post. Our experiments compare baseline models with progressively complex architectures, showing that the Transformer + CLIP Fusion Model outperforms others, achieving 91.3% accuracy and strong recall. This system has potential real-world applications as an early intervention tool for social media monitoring, and mental health support platforms, making it a step forward in AI-driven public health solutions.
Actions (login required)
 |
View Item |