Evaluating Emotional Detection & Classification Capabilities of GPT-2 & GPT-Neo Using Textual Data

Mental health has emerged as a pressing global health public concern due to the rising prevalence of mental illnesses and the notable gap in the detection and diagnosis of these conditions. With the widespread use of text as a means of communication, analysing mental health conditions from text has become essential. Artificial intelligence (AI) for healthcare has seen rapid growth and substantial research. In recent decades, the field of emotion recognition through AI has witnessed a notable shift from statistically shallow models to deep neural networks-based models, but they often grapple with challenges related to data scarcity, interpretability, and context-awareness, which has prompted the use of Large Language Models (LLMs) as a viable alternative. However, the introduction of LLMs like Generative Pretrained Transformer (GPT)-2 & GPT-Neo has astounded the research community with its unprecedented capabilities, which were previously unseen in traditional deep models. This work evaluates the effectiveness of LLMs-based GPT-Neo-125M and GPT-2 in 3 binary textbased mental health 2-class classification tasks: stress, depression, and suicidality which is, to our knowledge, the first attempt of the kind. This paper demonstrated that GPT-2 delivered higher overall accuracy, notably achieving an accuracy score of over 0.98 across all datasets.