188宝金博页面版

Bootstrapping Active Name Disambiguation withCrowdsourcingYu Cheng, Zhengzhang Chen, Jiang Wang, Ankit Agrawal, Alok ChoudharyEECS Department, Northwestern University{ych133,zzc472,jwa368,ankit,choudhar}@eecs.northwestern.eduABSTRACTName disambiguation is a challenging and important prob-lem in many domains, such as digital libraries, social mediamanagement and people search systems. Traditional meth-ods, based on direct assignment using supervised machinelearning techniques, seem to be the most ef f ective, but theirperformances are highly dependent on the amount of train-ing data, while large data annotation can be expensive andtime-consuming requiring hours of manual inspection by adomain expert. To ef f i ciently acquire labeled data, we pro-pose a bootstrapping algorithm for the name disambiguationtask based on active learning and crowdsourced labeling. Weshow that the proposed method can leverage the advantagesof exploration and exploitation by combining two strategies,thereby improving the overall quality of the training data atminimal expense. The experimental results on two datasetsDBLP and ArnetMiner demonstrate the superiority of ourframework over existing methods.Categories and Subject DescriptorsH.3.3 [Information Systems]: Information Storage andRetrieval—Information RetrievalGeneral TermsAlgorithms, ExperimentationKeywordsName Disambiguation, Active Learning, Crowdsourcing, Boot-strapping1. INTRODUCTIONName disambiguation has been viewed as a very impor-tant problem in many domains, such as digital libraries,social media management [2] and people search systems.Given a large set of entity names, the task is to deter-mine which names are referring to the same underlying en-Permission to make digital or hard copies of all or part of this work forpersonalorclassroomuseisgrantedwithoutfeeprovidedthatcopiesarenotmade or distributed for prof t or commercial advantage and that copies bearthis notice and the full citation on the f rst page. Copyrights for componentsof this work owned by others than ACM must be honored. Abstracting withcredit is permitted. To copy otherwise, or republish, to post on servers or toredistribute to lists, requires prior specif c permission and/or a fee.CIKM’13, Oct. 27–Nov. 1, 2013, San Francisco, CA, USA.Copyright 2013 ACM 978-1-4503-2263-8/13/10 ...$15.00.http://dx.doi.org/10.1145/2505515.2507858 .tity. To solve this problem, there are many approaches pro-posed in the recent years [5, 8, 7, 6]. Generally existingmethods mainly fall into three categories: supervised-based,unsupervised-based and constraint-based [9]. Among all themethods, the supervised-based approaches are considered tobe the most ef f ective ones [3]. However, their ef f ectivenessare highly dependent on the amount of training data avail-able.For a large amount of training examples, the annota-tion work is expensive and time-consuming requiring hoursof manual inspection by domain experts. In response, re-searchers have exploited active learning techniques to helpwith the labeling ef f ort problems. In [3], Ferrera et al. pro-posed an active sampling strategy based on association rulesto discriminate the author names. Wang et al. [9] introducedthe active name disambiguation problem and presented theADANA method to select the data for human labeling. Bycarefully selecting the informative samples, these approachescan largely reduce the amount of the data needed for con-structing the training set. But the annotators still may gettired when asked to go through hundreds of hard-to-labelsamples in the labeling process, which is very tedious anderror-prone. Moreover, previous active strategies for namedisambiguation either select the data based on a single un-certainty measure [9] or just select the most potentially erro-neous ones [3]. These criteria perform exploitation focuseson regions that are dif f i cult to learn and would overlookthose highly representative samples (exploration). This sam-pling bias may signif i cantly limit the active learning perfor-mance.On the other hand, social computing through services suchas Amazon Mechanical Turk (MTurk) have made it possi-ble for researchers to acquire suf f i cient quantities of crowd-sourced labels. Crowdsourcing distributes problem solvingto a broader community for requesting annotation and en-able acquiring labeled data at less expensive cost.In this paper, we propose a novel active learning algo-rithm, Active Data Augmentation, to exploit crowdsourcingtechniques for name disambiguation. The proposed methodcombines discriminative features and crowdsourced label-ing in a bootstrapping framework. Our hypothesis is thatby combining the two strategies, we could balance the ex-ploration and exploitation sampling. We f i rst formulatethe name disambiguation as a graph partitioning problem,and then we propose an algorithm to actively acquire la-beled data by combining discriminative feature labeling andcrowdsourcing with bootstrapping. Our approach is able toachieve both exploration and exploitation advantages, and1213