Title Page
Contents
Abstract 10
초록 11
Chapter 1. Introduction 12
1.1. Overview 12
1.2. Contribution 14
Chapter 2. Background 15
2.1. Task-Oriented Dialogue 15
2.2. ChatGPT Capabilities 17
Chapter 3. Methodology 20
3.1. Methodology 20
3.1.1. Adapter for each task (NLU, DST, NLG) 20
3.1.2. Metric-Aware Reinforcement Learning for DST & NLG module 22
3.1.3. ChatGPT Refinement Process for DST 24
Chapter 4. Experiments 27
4.1. Experimental Setup 27
4.1.1. Datasets 27
4.1.2. Baselines & Settings 27
4.2. Experimental Results 29
4.2.1. Dialogue State Tracking 29
4.2.2. End-to-End Response Generation 31
4.2.3. Further Analysis of Reinforcement Learning 32
4.2.4. ChatGPT Refinement Process Qualitative Analysis 35
Chapter 5. Conclusion 38
5.1. Summary 38
5.2. Limitations 39
Appendices 40
A. Units of Adapters 40
B. w/o Reinforcement Learning of TOATODsmall[이미지참조] 41
C. Implementation Details 41
Bibliography 42
TABLE 3.1. Comparing each parameter size of pre-trained and trainable. 20
TABLE 4.1. Joint Goal Accuracy for DST results. 30
TABLE 4.2. Inform, Success, BLEU, Combined Score for NLG. 31
TABLE 4.3. Task performance of TOATODbase before and after applying REINFORCE.[이미지참조] 32
TABLE 4.4. Hyperparameter experiment with a and b on the NLG task. 33
TABLE 4.5. Hyperparameter experiment with a on the DST task. 34
TABLE 1. Adapter units experiment results. 40
TABLE 2. Task performance of TOATODsmall before and after apply- ing REINFORCE.[이미지참조] 41
FIGURE 1.1. Overview of the Task-Optimized Adapters for an End-to-End Task-Oriented Dialogue System 12
FIGURE 3.1. Architecture of the sub-modules in TOATOD. 21
FIGURE 3.2. Overview of Dialogue State Refinement Process 24
FIGURE 3.3. Prompt examples of the restaurant domain. 25
FIGURE 4.1. Effect of hyperparameter a 34
FIGURE 4.2. Example of the cases of wrong refinement. 35