A noise tolerant and schema-agnostic blocking technique for entity resolution

Abstract

The increasing use of Web systems has become a valuable source of semi-structured data. In this context, the Entity Resolution (ER) task emerges as a fundamental step to integrate multiple knowledge bases or identify similarities between the data items (i.e., entities). Usually, blocking techniques are widely applied as an initial step of ER approaches in order to avoid computing similarities between all pairs of entities (quadratic cost). In practice, heterogeneous and noisy data increase the difficulties faced by blocking techniques, since these issues directly interfere the block generation. To address these challenges, we propose the NA-BLOCKER technique, which is capable of tolerating noisy data to extract information regarding the data schema and generate high-quality blocks. NA-BLOCKER applies Locality Sensitive Hashing (LSH) to hash the attribute values of entities and enable the generation of high-quality blocks, even with the presence of noise in the attribute values. In our experimental evaluation, we use five real-world datasets, and highlight that NA-BLOCKER presents better results regarding effectiveness compared to the state-of-the-art technique. In terms of efficiency, NA-BLOCKER produces, on average, 34% less comparisons. However, due to the cost introduced by LSH, it results in an increase of the execution time at around 30%, on average.