论文标题
学分：对话状态跟踪的粗到精细序列生成
CREDIT: Coarse-to-Fine Sequence Generation for Dialogue State Tracking
论文作者
论文摘要
在对话系统中，对话状态跟踪器旨在根据整个对话历史准确地找到当前对话状态的紧凑表示。尽管以前的方法通常将对话定义为单独的三元组（{\ em domain-slot-value}）的组合，但在本文中，我们采用结构化状态表示并将对话状态跟踪作为序列生成问题。基于这个新的配方，我们提出了一个{\ bf c} oa {\ bf r} s {\ bf e} - fine {\ bf di} alogue state {\ bf t} racking（{\ bf cordre}）方法。利用结构化状态表示，这是一个明显的语言序列，我们可以通过使用策略梯度方法优化自然语言指标来进一步调整预训练的模型（通过监督学习）。像所有生成状态跟踪方法一样，信用不依赖于列举所有可能的插槽值的预定义对话本体。实验证明了我们的跟踪器实现了Multiwoz 2.0和Multiwoz 2.1数据集中五个域的共同目标准确性。
In dialogue systems, a dialogue state tracker aims to accurately find a compact representation of the current dialogue status, based on the entire dialogue history. While previous approaches often define dialogue states as a combination of separate triples ({\em domain-slot-value}), in this paper, we employ a structured state representation and cast dialogue state tracking as a sequence generation problem. Based on this new formulation, we propose a {\bf C}oa{\bf R}s{\bf E}-to-fine {\bf DI}alogue state {\bf T}racking ({\bf CREDIT}) approach. Taking advantage of the structured state representation, which is a marked language sequence, we can further fine-tune the pre-trained model (by supervised learning) by optimizing natural language metrics with the policy gradient method. Like all generative state tracking methods, CREDIT does not rely on pre-defined dialogue ontology enumerating all possible slot values. Experiments demonstrate our tracker achieves encouraging joint goal accuracy for the five domains in MultiWOZ 2.0 and MultiWOZ 2.1 datasets.
