AI for social good

• Main Author: Wong, Kelly Jie Yin
• Other Authors: Erik Cambria
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science
• Online Access: https://hdl.handle.net/10356/175028
• Institution: Nanyang Technological University

This study explores the effectiveness of diverse machine learning architectures in detecting depression, with an emphasis on postpartum depression identification and advanced sentiment analysis. We evaluated various Transformer-based models and identified RoBERTa as particularly adept at identifying depression, outperforming others across accuracy, precision, recall, and F1 metrics. Employing multi-task learning strategies, including hard and soft parameter sharing, we improved our models' diagnostic accuracy. Further, we utilized Large Language Models (LLMs) with zero-shot learning, few-shot prompting, and fine-tuning with LoRA adapters on targeted datasets, achieving notable performance enhancements. A practical application of our research is the fine-tuned chatbot we developed, which has shown potential as a mental health support tool. Our experiments also extended to multimodal approaches that combined textual, visual, and temporal data using Transformer Encoders. These multimodal models surpassed the performance of text-only models and text-image hybrid models. Our results demonstrate that integrating and optimizing various computational techniques significantly boosts the detection of depression indicators. Additionally, we conducted a basic analysis utilizing explainable AI methods such as the LIME explainer to contrast the performance between the postpartum depression (PPD) model and the general depression detection model. This research confirms the promise of multimodal data in improving AI's role in mental health diagnostics and support, offering a pathway to more comprehensive solutions.