先学习可能性，再选择最佳方案：通过文本游戏分离语言中的一对多关系（英文版）

英文标题：Learn What Is Possible, Then Choose What Is Best: Disentangling One-To-Many Relations in Language Through Text-based Games中文摘要：利用离散潜在变量和知识蒸馏的方法，从多样的预训练数据集中捕捉了一系列不同的行为，并将这种多样性应用于学生模型中，显著提高了对话系统生成回复的准确性。英文摘要：Language models pre-trained on large self-supervised corpora, followed bytask-specific fine-tuning has become the dominant paradigm in NLP. Thesepre-training datasets often have a one-to-many structure--e.g. in dialoguethere are many valid responses for a given context. However, only som