Scalable entity resolution for Web product descriptions

Damir Vandic, Flavius Frasincar, Uzay Kaymak, Mark Riezebos

Abstract: Highlights • We propose a scalable approach for multi-source entity resolution. • We investigate hundreds of blocking schemes on three real-world datasets. • A blocking scheme has a source, a tokenizer, and a transformer. • The experiments are done for perfect and imperfect matching functions • Title model words blocking scheme gives the best efficiency/effectiveness tradeoff. Abstract Consumers are increasingly using the Web to find product information and make online purchases. This is reflected by the ongoing growth of worldwide e-commerce sales figures. Entity resolution is an important task that supports many services that have arisen from this growth, such as Web shop aggregators. In this paper, we propose a scalable framework for multi-source entity resolution. Our blocking approach employs model words to produce blocks that make our solution highly effective and efficient for the considered domains. An in-depth evaluation, performed using millions of experiments and three large datasets (on consumer electronics and software products), shows that our model words-based approach outperforms other approaches in most cases. Furthermore, we also evaluate our approach with an imperfect similarity function and find that model words-based blocking schemes provide the best blocks with respect to the F1-measure.