RiSAWOZ

为了缓解多领域数据的短缺并为面向任务的对话建模捕获话语现象，我们提出了 RiSAWOZ，这是一个具有丰富语义注释的大型多领域中文绿野仙踪数据集。 RiSAWOZ 包含 11.2K 人对人 (H2H) 多轮语义注释对话，超过 150K 话语跨越 12 个域，比以前所有带注释的 H2H 对话数据集都要大。单域对话和多域对话都构建，分别占65%和35%。每个对话都带有全面的对话注释，包括自然语言描述形式的对话目标、领域、对话状态以及用户和系统方面的行为。除了传统的对话注释外，我们还特别提供了对话中话语现象的语言注释，例如省略号和共指，这对于对话共指和省略号解析任务很有用。除了完全注释的数据集外，我们还详细描述了数据集的数据收集过程、统计和分析。报告了一系列基准模型和结果，包括自然语言理解（意图检测和槽填充）、对话状态跟踪和对话上下文到文本生成，以及共指和省略号解析，有助于未来研究的基线比较在这个语料库上。

To alleviate the shortage of multi-domain dialogue data and capture discourse phenomena for task-oriented dialogue modeling, we propose RiSAWOZ, a large multi-domain Chinese Wizard-of-Oz dataset with rich semantic annotations. RiSAWOZ contains 11.2K human-to-human (H2H) multi-turn semantically annotated dialogues, with over 150K utterances spanning 12 domains, which is larger than all previously released annotated H2H dialogue datasets. Both single-domain and multi-domain dialogues are constructed, accounting for 65% and 35% of the total dataset respectively. Each dialogue is paired with comprehensive annotations, including dialogue goals, domains, dialogue states, and user and system behaviors formulated as natural language descriptions. In addition to standard dialogue annotations, we also provide linguistic annotations for discourse phenomena occurring in the dialogues, such as ellipsis and coreference, which are critical for dialogue coreference resolution and ellipsis resolution tasks. Apart from the fully annotated dataset, we also elaborate on the data collection process, statistics and analysis of the dataset. We present a suite of benchmark models and corresponding experimental results, covering natural language understanding (intent detection and slot filling), dialogue state tracking, dialogue context-to-text generation, as well as coreference resolution and ellipsis resolution, which can serve as baseline comparisons for future research on this corpus.

搜集汇总

背景与挑战

背景概述

RiSAWOZ是一个大规模多领域中文绿野仙踪数据集，包含11.2K人对人对话和超过150K话语，覆盖12个域，并提供丰富的语义注释，如对话目标、状态和行为。该数据集旨在缓解多领域数据短缺问题，支持任务导向对话建模，包括自然语言理解、对话状态跟踪、文本生成以及共指和省略号解析任务。