Graph-boosted active learning for multi-source entity resolution

DOI:	https://doi.org/10.1007/978-3-030-88361-4_11
URL:	https://www.springerprofessional.de/graph-boosted-...
Weitere URL:	https://www.researchgate.net/publication/354932883...
Dokumenttyp:	Konferenzveröffentlichung
Erscheinungsjahr:	2021
Buchtitel:	The Semantic Web – ISWC 2021 : 20th international semantic web conference, ISWC 2021, virtual event, October 24–28, 2021, proceedings
Titel einer Zeitschrift oder einer Reihe:	Lecture Notes in Computer Science
Band/Volume:	12922
Seitenbereich:	182-199
Veranstaltungstitel:	ISWC 2021 (20, 2021)
Veranstaltungsort:	online
Veranstaltungsdatum:	24.-28.10.2021
Herausgeber:	Hotho, Andreas ; Blomqvist, Eva ; Dietze, Stefan ; Fokoue, Achille ; Ding, Ying ; Barnaghi, Payam ; Haller, Armin ; Dragoni, Mauro ; Alani, Harith
Ort der Veröffentlichung:	Berlin [u.a.]
Verlag:	Springer
ISBN:	978-3-030-88360-7 , \|978-3-030-88360-7 , 978-3-030-88362-1
ISSN:	0302-9743 , 1611-3349
Sprache der Veröffentlichung:	Englisch
Einrichtung:	Fakultät für Wirtschaftsinformatik und Wirtschaftsmathematik > Information Systems V: Web-based Systems (Bizer 2012-)
Fachgebiet:	004 Informatik
Freie Schlagwörter (Englisch):	entity resolution , link discovery , multi-source entity matching , active learning
Abstract:	Supervised entity resolution methods rely on labeled record pairs for learning matching patterns between two or more data sources. Active learning minimizes the labeling effort by selecting informative pairs for labeling. The existing active learning methods for entity resolution all target two-source matching scenarios and ignore signals that only exist in multi-source settings, such as the Web of Data. In this paper, we propose ALMSER, a graph-boosted active learning method for multi-source entity resolution. To the best of our knowledge, ALMSER is the first active learning-based entity resolution method that is especially tailored to the multi-source setting. ALMSER exploits the rich correspondence graph that exists in multi-source settings for selecting informative record pairs. In addition, the correspondence graph is used to derive complementary training data. We evaluate our method using five multi-source matching tasks having different profiling characteristics. The experimental evaluation shows that leveraging graph signals leads to improved results over active learning methods using margin-based and committee-based query strategies in terms of F1 score on all tasks.