아주대학교

목차

Ⅰ Introduction 1
1.1. Motivation and Background 1
1.2. Problem Definition 2
1.3. Contributions 5
Ⅱ Related Works 7
2.1. Text Difficulty and CEFR-based Classification 7
2.2. Human-centered Explainable AI (XAI) 8
2.3. Counterfactual Examples as Interactive Units 10
2.4. Reciprocal Learning and Human-AI Co-conceptualization 12
Ⅲ Proposed Method 14
3.1. Overview of Framework 14
3.2. Adaptive Classifier 15
3.3. Counterfactual Generator 16
3.4. Expert Feedback Interface 17
3.5. Contrastive Learning 18
3.6. Conceptualization 19
Ⅳ Experimental Design 22
4.1. Dataset and Setup 22
4.2. Expert Participants and Procedure 23
4.3. Contrastive Learning Implementation 27
V Results and Analysis 29
5.1. RQ1: Model Alignment and Performance Change 30
5.2. RQ2: Expert Judgement Evolution and Code Refinement. 35
5.3. RQ3: Conceptualization through Counterfactual Interaction 44
VI In-depth Analysis of Explanation-Feedback-Intervention 50
6.1. Evaluation Framework for Human-AI Collaboration 50
6.2. Comparative Analysis of Interaction Loops 55
6.3. Structural Efficacy of Counterfactual Examples 65
Ⅶ Discussion 68
7.1. Limitations 68
7.2. Future work 69
7.3. Refining CEFR Guidelines through Quantitative Control Codes 71
Ⅷ Conclusion 73
8.1. Summary of Findings 73
8.2. Contributions to Human-AI Conceptualization Methodology 74
References 76
Appendix 84
Appendix 1: Examples of Counterfactual Examples 84
Appendix 2: Prompt Used to Generate Counterfactuals 85
Appendix 3: Loop-wise Control Code 86
Appendix 4: Per-code Explanations 87
Appendix 5: Detailed Statistical Analysis 89