Large Language Model-Based Representation Learning for Entity Resolution using Contrastive Learning

In this paper, we introduce TriBERTa, a supervised entity resolution (ER) system, that utilizes a pre-trained large language model and a triplet loss function. TriBERTa is engineered to learn representations that inherently group similar entities together while separating dissimilar ones, forming a foundational asset that can be leveraged across all steps of the ER process, including entity matching, data blocking, and data resolution. Our approach employs a two-step process: first, name entity records are fed into a Sentence Bidirectional Encoder Representations from Transformers (SBERT) model, a pre-trained language model that utilizes the BERT architecture. The language model then generates vector representations, which are subsequently fine-tuned using contrastive learning based on a triplet loss function. In the second step, the fine-tuned representations are used as input for Entity Resolution tasks in our case entity matching. Our results show that our proposed approach outperforms state-of-the-art representations, including SBERT without fine-tuning and conventional Term Frequency-Inverse Document Frequency (TF-IDF), by a margin of 3% to 19%. Additionally, in comparison to specialized end-to-end entity matching cross-encoder models, the representations generated by TriBERTa demonstrate increased robustness, maintaining consistently higher performance across a range of datasets.