Task-Oriented Dialogue Understanding and Evaluation using Minimal Data
COM3 Level 2
MR24, COM3 02-64
Abstract:
In the rapidly evolving user interface landscape, Task-oriented Dialogue (TOD) Systems are set to revolutionize our interaction with technology by shifting from static website browsing to dynamic conversational engagements. The success of these systems hinges on Dialogue State Tracking (DST), a crucial component for understanding user preferences and enabling precise database queries, thus ensuring response relevance and accuracy.
This thesis defense introduces a holistic approach to improving TOD Systems, focusing on the challenges of DST and domain adaptation in low-resource scenarios. It starts by presenting three innovative strategies: human-augmented data generation, TOD-specific automated data augmentation, and a Parameter-Efficient Fine-Tuning (PEFT) method, aimed at enhancing system adaptability to new domains. It also proposes a novel evaluation metric, Granular Change Accuracy (GCA), offering a more precise assessment of DST model performance and addressing existing metric limitations. Furthermore, it explores the integration of Instruction-tuned Large Language Models (LLMs) via a new framework named CESAR to enhance compositional task management in dialogue modeling, facilitating seamless downstream task integration. This defense endeavors to advance TOD systems, contributing significantly to the fields of natural language processing and AI-driven dialogue interfaces.