An in-depth analysis of pre-trained embeddings for entity resolution: An in-depth analysis of pre-trained embeddings for entity resolution: A. Zeakis et al.

Recent works on entity resolution (ER) leverage deep learning techniques that rely on language models to improve effectiveness. These techniques are used both for blocking and matching, the two main steps of ER. Several language models have been tested in the literature, with fastText and BERT variants being most popular. However, there is no detailed analysis of their strengths and weaknesses. We cover this gap through a thorough experimental analysis of 12 popular pre-trained language models over 17 established benchmark datasets. First, we examine their relative effectiveness in blocking, unsupervised matching and supervised matching. We enhance our analysis by also investigating the complementarity and transferability of the language models and we further justify their relative performance by looking into the similarity scores and ranking positions each model yields. In each task, we compare them with several state-of-the-art techniques in the literature. Then, we investigate their relative time efficiency with respect to vectorization overhead, blocking scalability and matching run-time. The experiments are carried out both in schema-agnostic and schema-aware settings. In the former, all attribute values per entity are concatenated into a representative sentence, whereas in the latter the values of individual attributes are considered. Our results provide novel insights into the pros and cons of the main language models, facilitating their use in ER applications. © The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2024.