Efficiently Fine-Tuned Flan T5 LLM for Happiness Oriented Dialogue Summarization

In conversational AI applications, helping people to get emotional independence and a good mental health is becoming a popular theme. Summarizing such kind of dialogues is a task which has not been attempted in previous research. In this research paper, we fine-tune the FLAN-T5 large language model using the Low Rank Adaptation (LoRA) technique specifically for the task of dialogue summarization on a custom dataset derived from the book Happiness Unlimited. We evaluate the model’s performance against zero-shot, few-shot, and fully fine-tuned FLAN-T5 approaches through human evaluation and ROUGE scores to demonstrate its efficiency and effectiveness in generating concise, informative summaries with reduced computational requirements. The dataset which have been used is based on a conversation between two persons in the book. The output of this research work is an LLM which can help users to gain a deeper understanding of concepts such as emotional independence, being stable in chaos etc, which are discussed in the book. The performance of fine tuned LLM is compared with other state of the art techniques such as Zero shot, Few shot in context learning and the full fine tuned version of FLAN-T5 based on Human Evaluation and ROUGE scores for informativeness with respect to human generated summaries. Both type of evaluation shows that our LORA based efficiently fine tuned model is far better than Zero shot and Few shot learning. The ROUGE scores also indicate that our fine tuned model achieves similar values as the full fine tuned model but uses far less compute and storage resources.