The WDC training dataset and gold standard for large-scale product matching

Primpeli, Anna ; Peeters, Ralph ; Bizer, Christian

DOI:	https://doi.org/10.1145/3308560.3316609
URL:	https://dl.acm.org/citation.cfm?doid=3308560.33166...
Weitere URL:	https://www2019.thewebconf.org/media/TheWebConf201...
Dokumenttyp:	Konferenzveröffentlichung
Erscheinungsjahr:	2019
Buchtitel:	Companion Proceedings of The 2019 World Wide Web Conference
Seitenbereich:	381-386
Veranstaltungstitel:	Workshop on e-Commerce and NLP at WWW 2019
Veranstaltungsort:	San Francisco, CA
Veranstaltungsdatum:	May 13-17, 2019
Herausgeber:	Liu, Ling
Ort der Veröffentlichung:	New York, NY, USA
Verlag:	ACM
ISBN:	978-1-4503-6675-5
Sprache der Veröffentlichung:	Englisch
Einrichtung:	Fakultät für Wirtschaftsinformatik und Wirtschaftsmathematik > Information Systems V: Web-based Systems (Bizer 2012-)
Fachgebiet:	004 Informatik
Freie Schlagwörter (Englisch):	deep matching , embeddings , entity resolution , evaluation data , product matching , schema.org annotations
Abstract:	A current research question in the area of entity resolution (also called link discovery or duplicate detection) is whether and in which cases embeddings and deep neural network based matching methods outperform traditional symbolic matching methods. The problem with answering this question is that deep learning based matchers need large amounts of training data. The entity resolution benchmark datasets that are currently available to the public are too small to properly evaluate this new family of matching methods. The WDC Training Dataset for Large-Scale Product Matching fills this gap. The English language subset of the training dataset consists of 20 million pairs of offers referring to the same products. The offers were extracted from 43 thousand e-shops which provide schema.org annotations including some form of product ID such as a GTIN or MPN. We also created a gold standard by manually verifying 2200 pairs of offers belonging to four product categories. Using a subset of our training dataset together with this gold standard, we are able to publicly replicate the recent result of Mudgal et al. that embeddings and deep neural network based matching methods outperform traditional symbolic matching methods on less structured data.