TL;DR使用LLM(Large Pre-Trained Language Models)评估了在口语任务导向对话中的性能,结果表明LLMs默认情况下对口头噪音不够鲁棒,但在正确的口头TOD数据集上进行微调/训练可以获得更强的性能。
Abstract
Large pre-trained language models have demonstrated state-of-the-art performance in different downstream tasks, including dialogue state tracking and end-to-end →