MultiMatch: Low-Resource Generalized Entity Matching Using Task-Conditioned Hyperadapters in Multitask Learning

Generalized Entity Matching (GEM) is a variant of entity matching that identifies whether entity descriptions from diverse data sources with heterogeneous data formats refer to the same real-world entity. State-of-the-art single-task fine-tuning approaches have shown limitations in handling scenarios with entity distribution shifts, particularly in low-resource settings, and can also require significant amounts of computationally expensive fine-tuning when applied to the GEM problem. This paper addresses these challenges by deploying task-conditioned adapters for low-resource GEM. We present MultiMatch, which explores the benefits of sharing knowledge across related tasks while improving the efficiency and accuracy of models used for GEM. Furthermore, we propose a loss composition strategy that leverages the heteroscedastic uncertainty of individual tasks to adjust the loss terms for each task before computing the overall loss. Empirically, we observe regulatory effects on the model’s variance. Lastly, we analyze the carbon impact of fine-tuning different systems. Results are promising: our approach generalizes over eight GEM benchmarking tasks while reducing \(CO_2\) emissions by 85.0%.