Tools
Tools
Srivastava, Diya (2024) Utilizing Counselor-Client Dialogues to Develop a Memory-Efficient Mental Health Question-Answering System with Large Language Models. Masters thesis, Dublin, National College of Ireland.
Preview |
PDF (Master of Science)
Download (4MB) | Preview |
Preview |
PDF (Configuration Manual)
Download (2MB) | Preview |
Abstract
Healthy living for an individual most certainly encompasses Mental well-being enabling humans to endure emotions. A healthy state of mind is premiere to a healthy life, yet nearly millions of individuals globally suffer from mental disorders such as anxiety, depression, and PTSD, with the COVID-19 pandemic further exacerbating this crisis. Traditional therapy is often challenging due to scarcity of psychologists, expensive sessions or apprehensions associated with people belonging to different demographics. Hence, in response this research explores the advancements in Natural Language Generation(NLG) domain of Artificial Intelligence(AI) to conduct Virtual Therapy. The study proposes the use of sequence-to-sequence large Language Models built on decoder-transformer architecture, leveraging Parameter efficient Fine-tuning technique like LoRa and Memory Efficient Quantization strategy to develop a mental health domain specific question-answering system in a resource constraint environment. The study experiments with Flan T-5-Base, Tiny Llama-1.1B, Llama-2 7B, Gemmma-2 2B and GPT-Neo 2.7B to prospect their performance after being fine-tuned on ’MentalChat16k’ dataset of question-answer pair from a therapist, client conversation. The study evaluates model generated outputs qualitatively, while Conducting quantitative analysis on diverse LLMs by computing BLEU, BERT and ROUGE Score. Concluding with Gemma-2 achieving 0.5 ROUGE-1 score outperforming other models, while Llama-2 prevails in delivering more empathetic and coherent responses.
Actions (login required)
 |
View Item |