Entity resolution is the problem of nding co-referent instances, which at the same time describe the same topic. It is an important component of data integration systems and is indispensable in linked data publication process. Entity resolution has been a subject of extensive research; however, seeking for a perfect resolution algorithm remains a work in progress. Many approaches have been proposed for entity resolution. Among them, supervised entity resolution has been revealed as the most accurate approach [6, 2]. Meanwhile, con guration-based matching [2, 3, 5, 4] attracts most studies because of its advantages in scalability and interpretation. In order to match two instances of di erent repositories, con guration-based matching algorithms estimate the similarities between the values of the same attributes. After that, these similarities are aggregated into one matching score. This score is used to determine whether two instances are co-referent or not. The declarations of equivalent attributes, similarity measures, similarity aggregation, and acceptance threshold are speci ed by a matching con guration, which can be automatically optimized by a learning algorithm. Con guration learning using genetic algorithm has been a research topic of some studies [2, 5, 3]. The limitation of genetic algorithm is that it costs numerous iterations for reaching the convergence. We propose cLearn as a heuristic algorithm that is e ective and more e cient. cLearn can be used to enhance the performance of any con guration-based entity resolution system.
A con guration speci es the property mappings, similarity measures, similarity aggregation strategy, and matching acceptance threshold. Property mappings and similarity measures are combined together into similarity functions. Given series of initial similarity functions, similarity aggregation options, and the labeled instances pairs, the mission of cLearn is to select the optimal con guration.
cLearn begins with the consideration of each single similarity function and then checks their combinations. When checking the new combination this algorithm applies a heuristic for selecting most potentially optimal con guration. Concretely, the heuristic accepts the new combination if only its performance Training size 5% Varied by subset SOcbSjLecINtCTo+recfLearn 0.894
0.464 is better than that of the combined elements. This heuristic is reasonable as a series of similarity functions that reduces the performance has little possibility of generating a further combination with improvement. In addition to nding for similarity functions, the algorithm also optimizes the similarity aggregator and matching acceptance threshold.
cLearn is implemented as part of ScSLINT framework, and its source code is available at http://ri-www.nii.ac.jp/ScSLINT.
Table 1 reports the comparison between cLearn and other supervised systems,
including ObjectCoref [
cLearn is e cient as the average numbers of con gurations that cLearn has
to check before stopping is only 246. This number is promising because it is
much smaller than that of using genetic algorithm, which is reported in [
With the e ectiveness, potential e ciency, and small training data requirement of cLearn on a real dataset like OAEI 2010, we believe that cLearn has promising application in supervised entity resolution, including using active learning strategy to even reduce the annotation e ort.