TEXTUAL MENTAL ILLNESS DETECTION FROM SOCIAL MEDIA USING DEEP LEARNING

Abstract

Mental illnesses are one of the most prevalent public health issues, and social media is a primary source through which young people express their mental states and feelings. Therefore, textual mental illness detection from social media has become an essential research topic in natural language processing (NLP), due to its high impact on mental health analysis. Mental illness detection includes a variety of tasks, such as disorder classification, mental illness severity analysis and symptoms identification. This thesis aims to design novel deep learning-based models to address different challenges in mental illness detection. Firstly, we explore the application of the Transformer-based model, which aims to effectively extract contextual and long-term dependency information from texts. More specifically, we also conduct a preliminary study about how Transformer-based pre-trained language models (PLMs) and large language models (LLMs) perform on mental health analysis. Experiments performed on our collected dataset and public suicide-related dataset illustrate the effectiveness of our model. Moreover, the results of PLMs and LLMs show that PLMs have competitive performance, while the LLMs have room for further improvement in mental health analysis. Secondly, existing depression severity detection models mainly model the semantic information of posts while ignoring the relevant sentiment information, which can reflect the user's state of mind and is useful for severity detection. Additionally, these methods often treat all severity levels equally, making the model difficult to distinguish between closely-labelled categories. Therefore, in order to solve these shortcomings, we propose a sentiment-guided Transformer model, which efficiently fuses social media posts' sentiment information with semantic information. Furthermore, we introduce a supervised severity-aware contrastive learning framework that enables the model to better distinguish between different severity levels. Experimental results performed on two depression severity corpora illustrate our model's effectiveness. Thirdly, depression scales such as the Patient Health Questionnaire (PHQ-9) can be beneficial for understanding the symptoms of depression and improving the interpretability of detection models. Therefore, we consider utilising PHQ descriptive information related to symptoms and present a Span-based PHQ-aware and similarity contrastive network (SpanPHQ), which enhances the model's ability to capture both semantic contextual information and PHQ-9 information. Besides, a similarity contrastive learning is designed to effectively utilise the label information in identifying class-specific features. Experiments performed on two depressive symptoms identification datasets show good performance in terms of improving multiple evaluation metrics and increasing the interpretability of the model.

Date of Award	1 Aug 2024
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Sophia Ananiadou (Supervisor) & Junichi Tsujii (Supervisor)