SSD_PLATE (Sub-Slot Dialogue dataset license plate number domain)

可以通过对话中的多轮交互逐段提供槽值，尤其是对于一些重要信息，例如电话号码和姓名。这是日常生活中普遍存在的现象，但在以往的工作中却很少关注。为了填补这一空白，本文定义了一个名为基于子槽的面向任务的对话（SSTOD）的新任务，并构建了一个中文对话数据集SSD，以促进对SSTOD的研究。该数据集包括来自四个不同领域的总共 40K 对话和 500K 话语：中文姓名、电话号码、身份证号码和车牌号码。数据用子槽值、槽值、对话状态和动作进行了很好的注释。我们在 SSTOD 中发现了一些新的语言现象和交互方式，这对为任务构建对话代理提出了严峻挑战。我们在 SSTOD 上测试了三个最先进的对话模型，发现它们无法在四个域中的任何一个域上很好地处理任务。我们还通过以插件方式涉及槽知识来研究改进的模型。应该做更多的工作来应对在现实生活中广泛存在的 SSTOD 提出的新挑战。Multi-turn interactions in conversations can be used to provide slot values segment by segment, especially for critical information such as phone numbers and names. This is a widespread phenomenon in daily life, but it has received little attention in previous research. To fill this gap, this paper defines a novel task named Sub-slot-based Task-oriented Dialogue (SSTOD), and constructs a Chinese dialogue dataset SSD to promote research on SSTOD. The dataset consists of a total of 40K conversations and 500K utterances across four distinct domains: Chinese names, phone numbers, ID card numbers, and license plate numbers. The data is thoroughly annotated with sub-slot values, slot values, dialogue states, and dialogue acts. We discover several new linguistic phenomena and interaction patterns in SSTOD, which pose significant challenges for building task-oriented dialogue agents. We test three state-of-the-art dialogue models on SSTOD and find that they fail to handle the task well across all four domains. We also investigate improved models by incorporating slot knowledge in a plugin-based manner. More work is needed to address the new challenges posed by SSTOD, which is prevalent in real-world scenarios.

搜集汇总

背景与挑战

背景概述

SSD_PLATE是一个中文对话数据集，专注于子槽面向任务的对话（SSTOD），涵盖车牌号码等四个领域，包含40K对话和500K话语，用于支持多轮交互中逐段提供槽值的研究。该数据集旨在解决以往工作中较少关注的现实语言现象，为对话代理的改进提供基础。