CRI:CRD Creation of a Goal-Oriented, Human-Machine Spoken Dialog Corpus and an Infrastructure for Dialog System Evaluation

正在对一套小型人机对话语料库(“微型语料库”)进行转录和注释，并分发给对话领域的研究人员。这些语料库的原始数据已经存在，并是使用两个现有的系统(来自AT&amp；T Labs的DiSCoH和来自卡内基梅隆大学的Conquest)生成的。产生的语料库正在分发给实地的研究人员，目的是就语料库的组成和注释征求反馈意见，以最大限度地支持人机对话交互领域的研究。现正整理这项工作的反馈意见，并将其分发回社区。在与2007年HLT/NAACL会议同时举行的讲习班上讨论这项工作的结果，为大规模收集这类数据制定一套准则提供了机会。工作间的配置方便了对话社区的许多研究人员出席。此外，NSF的支持使研讨会能够确保来自北美和国际中心的研究人员的广泛参与。反馈和征集过程中产生的讨论和文件为编写社区支持的NSF CRI：CRD计划提案提供了基础。提供系统收集和注解的数据有助于在人机对话研究方面取得进展，这反过来又有助于开发更复杂和可广泛获取的信息获取技术。

A set of small-scale human-computer dialogue corpora ("micro-corpora") is being transcribed and annotated, and distributed to researchers in the area of dialogue. The raw data for these corpora already exists and has been generated using two existing system (DiSCoH from AT&T labs and ConQuest from Carnegie Mellon University). The resulting corpora are being distributed to researchers in the field with the goal of soliciting feedback on the corpus composition and annotation that can best support research in the field of human-machine dialogue interaction. Feedback from this exercise is being collated and disseminated back to the community. Discussion on the outcome of this exercise at a workshop collocated with the 2007 HLT/NAACL conference provides the opportunity to develop a set of guidelines for the large scale collection of such data. The workshop collocation makes attendance convenient for many researchers in the dialogue community. In addition, support from NSF allows the workshop to ensure broad participation by researchers from both North America and international centers. Discussions and documents generated in the feedback and solicitation process provide a basis for the preparation of a community-supported proposal to the NSF CRI:CRD program. The availability of systematically collected and annotated data supports progress in human-computer dialogue research which in turn enables the development of more sophisticated and broadly-accessible technologies for information access.