Using schema.org Annotations for Training and Maintaining Product Matchers

• DOI: 10.1145/3405962.3405964
• Publication Date: 2020-08-25T04:24:02Z
• Authors: Ralph Peeters; Anna Primpeli; Benedikt Wichtlhuber; Christian Bizer
• Keywords: Schema Matching; Schema (genetic algorithms)

Product matching is a central task within e-commerce applications such as price comparison portals and online market places. State-of-the-art product methods achieve F1 scores above 0.90 using deep learning techniques combined with huge amounts of training data (e.g > 100K pairs offers). Gathering maintaining large corpora costly, it implies labeling offers matches or non-matches. Acquiring the ability to be good at thus means major investment for an company. This paper shows that manual can replaced by relying exclusively on schema.org annotations gathered from public Web. We show only training, we are able between 0.92 0.95 depending category. As new products appear everyday, important models maintained justifiable effort. In order give practical advice how maintain models, compare performance traditional unseen experiment different fine-tuning re-training strategies model maintenance, again data. Finally, Web distant supervision carries inherent noise, evaluate regards their label-noise resistance deal identifier-noise found in annotations.