Data Augmentation for Entity Resolution: A Comparative Evaluation

Data augmentation is an important technique to enhance the performance of models in different fields, such as computer vision and natural language processing. In entity resolution, few papers have applied general or task-specific data augmentation techniques, and systematic knowledge is sparse. This study evaluates eight rule-based, model-based, and sample interpolation by comparing them systematically across 10 datasets and two transformer-based models or entity resolution matching systems: BERT and Ditto, leveraging DistiIBERT. The results show the potential of data augmentation to increase per-formance, although the influence of the technique used depends heavily on the model and dataset. Here, MixDA and random word insertion performed best, so they are most suitable for recommendation to practitioners dealing with entity resolution tasks.