Intermediate training of BERT for product matching

Peeters, Ralph ; Bizer, Christian ; Glavaš, Goran

URL:	http://ceur-ws.org/Vol-2726/paper1.pdf
Weitere URL:	http://ceur-ws.org/Vol-2726/
Dokumenttyp:	Konferenzveröffentlichung
Erscheinungsjahr:	2020
Buchtitel:	DI2KG 2020 : Proceedings of the 2nd International Workshop on Challenges and Experiences from Data Integration to Knowledge Graphs co-located with 46th International Conference on Very Large Data Bases (VLDB 2020), Tokyo, Japan, August 31, 2020
Titel einer Zeitschrift oder einer Reihe:	CEUR Workshop Proceedings
Band/Volume:	2726
Seitenbereich:	1-2
Veranstaltungstitel:	DI2KG 2020
Veranstaltungsort:	Online
Veranstaltungsdatum:	31.08.2020
Herausgeber:	Piai, Federico
Ort der Veröffentlichung:	Aachen, Germany
Verlag:	RWTH Aachen
ISSN:	1613-0073
Sprache der Veröffentlichung:	Englisch
Einrichtung:	Fakultät für Wirtschaftsinformatik und Wirtschaftsmathematik > Information Systems V: Web-based Systems (Bizer 2012-) Fakultät für Wirtschaftsinformatik und Wirtschaftsmathematik > Text Analytics for Interdisciplinary Research (Juniorprofessur) (Glavaš 2017-2021)
Fachgebiet:	004 Informatik
Freie Schlagwörter (Englisch):	e-commerce , product matching , deep learning
Abstract:	Transformer-based models like BERT have pushed the state-of the-art for a wide range of tasks in natural language processing. General-purpose pre-training on large corpora allows Transformers to yield good performance even with small amounts of training data for task-specific fine-tuning. In this work, we apply BERT to the task of product matching in e-commerce and show that BERT is much more training data efficient than other state-of-the-art methods. Moreover, we show that we can further boost its effectiveness through an intermediate training step, exploiting large collections of product offers. Our intermediate training leads to strong performance (>90% F1) on new, unseen products without any product-specific fine-tuning. Further fine-tuning yields additional gains, resulting in improvements of up to 12% F1 for small training sets. Adding the masked language modeling objective in the intermediate training step in order to further adapt the language model to the application domain leads to an additional increase of up to 3% F1.