The impact of fine-tuning on entity resolution: An experimental evaluation

Fine-tuning pre-trained language models has become the state-of-the-art approach for Entity Resolution (ER), but this has created a divide between two dominant architectures: fast-but-less-accurate bi-encoders and accurate-but-slow cross-encoders. However, a concrete gap in prior ER benchmarking remains unresolved: existing studies often evaluate architectures in isolation or on limited datasets. It remains unclear which base models and architectures are best suited for the diverse range of real-world ER datasets, each with unique characteristics and performance bottlenecks. This paper bridges this gap through an extensive empirical evaluation. We systematically compare three popular pre-trained models (MiniLM, MPNet, and BGE) across three distinct architectural paradigms: a pre-trained bi-encoder, a fine-tuned bi-encoder, and a fine-tuned cross-encoder. We tested these combinations on eight diverse real-world and semi-synthetic datasets, analyzing their performance, training costs, and final resolution times. Our results reveal a clear accuracy-vs-efficiency trade-off, identifying the fine-tuned bi-encoder as the optimal balance between performance and practical resolution speed. More importantly, we demonstrate that fine-tuning is not a universal solution. Its effectiveness is highly contingent on the dataset: it provides substantial gains on specialized domains by fixing pre-existing performance gaps but is detrimental to performance on datasets where pre-trained models are already well-aligned. These findings provide a practical guide for practitioners on selecting the optimal model and architecture based on their specific data and application requirements.