Task-Constrained Optimization for Entity-Relation Extraction

ICLR 2026 Conference Submission17619 Authors

Keywords: Entity-relation extraction, Multi-Task Learning, Priority-Aware Optimization

Abstract: Multi-task learning for entity-relation extraction often suffers from implicit task interference and lacks explicit mechanisms to enforce structural task prioritization. We propose Hyperbolic Barrier-based Adaptive Hierarchical Optimization, a constraint-driven optimization framework that formulates entity recognition as a dynamic hard constraint via a numerically stable hyperbolic barrier function, while adaptively reweighting relation classification through a curriculum-based thresholding strategy. This principled framework enforces strict task prioritization throughout training, yielding absolute improvements of up to $6.4\%$ in triplet F1 across five entity-relation extraction benchmarks. Furthermore, the proposed method generalizes effectively to structurally divergent domains such as recommender systems. Our findings underscore that explicitly modeling task hierarchies through constrained optimization represents a critical yet underexplored paradigm for achieving stable and effective multi-task learning.

Primary Area: optimization

Submission Number: 17619