Neural Locality Sensitive Hashing for Blocking in Entity Blocking

Anonymous

Abstract: Locality-sensitive hashing (LSH) is an algorithmic technique that hashes similar input items into the same “buckets” with high probability. It is a basic primitive in several large-scale data processing applications, including nearest-neighbor search, entity resolution, clustering, etc. In this work, we focus on the blocking phase in the entity resolution task. The goal of blocking is to avoid comparing all entity pairs by filtering out unmatched pairs. For this purpose, existing LSH functions that are based on generic similarity metric like Jaccard similarity, can only capture the occurrence of words while the semantics of the texts are ignored. On the other hand, several work have proposed to use language models to vectorize the data items and use the similarity of embeddings to find candidate pairs. However, it is still a challenge to fine-tune the language models so that the obtained embeddings can precisely capture the similarity of item pairs for ranking purpose. To this end, we propose NLSHBlock (Neural-LSH Block), a blocking approach that is based on pre-trained language models and fine-tuned with a novel LSH-inspired loss function. We evaluate the performance of Neural-LSH on the blocking stage of entity resolution, and show that it out-performs existing methods by a large margin on a wide range of datasets.

Paper Type: long

Research Area: Information Retrieval and Text Mining