C3D+P: A summarization method for interactive entity resolution
C3D+P: A summarization method for interactive entity resolution
复制标题
C3D P：一种交互式实体解析的概括方法
DOI：
10.1016/j.websem.2015.05.004
复制
发表时间：
2015-12-01
影响因子：
2.5
通讯作者：
Qu, Yuzhong
中科院分区：
文献类型：
--
作者：
Cheng, Gong;Xu, Danyun;Qu, Yuzhong
实体解析是数据集成中的一项基本任务。最近对这个问题的研究，包括主动学习，众包和按需付费的方法，已经开始让人类用户参与到循环中来执行交互式实体解析任务，即邀请人类用户判断两个实体描述是否指的是同一个真实世界的实体。这个判断过程需要工具支持，特别是当实体描述包含大量特征（即属性-值对）时。为了便于判断，在这篇文章中，我们建议选择，从实体描述，一个子集的关键功能作为一个摘要，显示和判断的人类用户。倾向于选择的特征是反映两个实体之间最多的共同点和最多的冲突的特征，以及携带关于它们的最大量的特征和多样性信息的特征。然后对选定的特征进行分组和排序，以提高可读性并进一步加快判断速度。实验结果表明，我们的方法生成的摘要帮助用户更有效地判断（3.57-3.78倍的速度）比整个实体描述，而不会显着损害判断的准确性。通过我们的方法实现的准确性也高于现有的总结方法。(C)2015 Elsevier B. V.版权所有。
Entity resolution is a fundamental task in data integration. Recent studies of this problem, including active learning, crowdsourcing, and pay-as-you-go approaches, have started to involve human users in the loop to carry out interactive entity resolution tasks, namely to invite human users to judge whether two entity descriptions refer to the same real-world entity. This process of judgment requires tool support, particularly when entity descriptions contain a large number of features (i.e. property-value pairs). To facilitate judgment, in this article, we propose to select, from entity descriptions, a subset of critical features as a summary to be shown and judged by human users. Features preferred to be selected are those that reflect the most commonalities shared by and the most conflicts between the two entities, and that carry the largest amount of characteristic and diverse information about them. Selected features are then grouped and ordered to improve readability and further speed up judgment. Experimental results demonstrate that summaries generated by our method help users judge more efficiently (3.57-3.78 times faster) than entire entity descriptions, without significantly hurting the accuracy of judgment. The accuracy achieved by our method is also higher than those achieved by existing summarization methods. (C) 2015 Elsevier B.V. All rights reserved.