A multi-layer representation mixing method based on BERT for entity resolution

The Entity Resolution (ER) task aims to link entities from various sources to a single real-world entity. Despite significant progress made by deep learning-based ER models, their performance still heavily relies on large-scale, high-quality annotated data. To overcome these limitations, this paper introduces a novel solution called Multi-Layer Semantic Integration, implemented as the Mixer model. The model integrates raw and augmented data representations through linear interpolation of BERT’s multi-layer embeddings, forming a hybrid feature representation that combines both original and augmented information. To further evaluate the model’s efficacy, experiments were conducted to determine which layer produces optimal mixing performance in ER tasks. Additionally, a dual-channel module is incorporated after the mixing module. This module utilizes the complementary feature extraction capabilities of CNN and BiLSTM networks. Experimental results demonstrate that the proposed model outperforms existing approaches across datasets from various domains.