Abstract:
In recent years, people have been using social media platforms to express their feelings and share their mental health struggles openly and anonymously. This surge has motivated many researchers to take advantage of social media as a valuable resource of data to detect severe depression. However, existing approaches have significant limitations as they rely on datasets suffering from a wide disparity between negative instances and positive instances, ignoring the middle ground containing depression-mimicking states. This thesis considers new mimicking states (stress, anxiety, sadness, sarcasm, and complaints) that are often misidentified as real depression due to their overlapping nature in the language and expressions used. We propose an automated system for detecting severe depression that features two main modules: content level and user level depression detection. The study encompasses the extensive evaluation of seven LLMs on the content level. We observe that finetuning significantly enhances the performance of all the evaluated LLMs. RoBERTa emerged as the top performer among the tested models, achieving an impressive AUC of 98.53% on the test set. For the user-level module, we utilize the best model, RoBERTa, along two main criteria: the user’s content severity and posting history. This approach is crucial as it allows us to capture the nuanced variations in user behavior and content, thereby enhancing the accuracy and reliability of our classification system. Overall, our work illustrates the potential of using LLMs in developing an accurate depression detection system that will contribute to a reduction in the overall prevalence of untreated depression cases.
