ENHANCING SCRIPTED DIALOGUE SYSTEMS FOR HEALTH-COACH APPLICATIONS: A COMPARATIVE STUDY OF LARGE LANGUAGE MODELS AND PATTERN-MATCHING FUNCTIONS

Abstract

This thesis investigates the feasibility and effectiveness of integrating large language models (LLMs) and pattern-matching functions into scripted dialogue systems for health-coaching applications. The objective is to determine which integration method enhances the adaptability and naturalness of conversational agents more effectively, considering both coverage and real-time performance. By using advanced LLMs alongside efficient pattern-matching functions, the study examines their ability to address traditional scripted dialogues' limitations, which rely heavily on predefined user inputs. Experiments are conducted across zero-shot, few-shot, and fine-tuned learning paradigms using models such as Meta-Llama, Gemma, and ChatGPT. The results indicate that while pattern-matching functions offer rapid response times and closely adhere to scripts, LLMs provide superior enhancements in handling diverse and complex inputs. The comparative analysis reveals that LLMs significantly improve the conversational quality and flexibility of dialogue systems in health coaching despite their higher computational demands. This suggests a promising direction for future research and application in scripted dialogue systems.