LRER: A Low-Resource Entity Resolution Framework with Hybrid Information

Entity resolution (ER) is a key task in data integration and data science applications. The state-of-the-art approaches in ER have been achieved through designing various deep learning-based methods. However, these methods typically require a large amount of high-quality labeled resources, which are difficult to obtain in the real world. In this paper, we propose a novel ER framework called LRER, which efficiently maintains the low-resource settings while using hybrid information from different perspectives of entities. In the first stage of LRER, we construct a self-supervised contrastive pre-training module, which contains reasonable contrastive instances and a novel training objective to pre-train the representation model without using any labeled data. In the second stage of LRER, we propose a hybrid information fine-tuning module, which contains a structure-aware processing for relational entities to capture the structural information and fine-tune the model with hybrid information. Experimental results on real-world datasets show that LRER outperforms the existing methods under low-resource settings.