Adaptive Table Filling for Entity and Relation Extraction using Deformable Attention and Dynamic Convolution

• Jin Ren,
• Jin Chen,

Vol. 20, No. 1, January 31, 2026
10.3837/tiis.2026.01.003, Download Paper (Free):

• Joint relation extraction
• named entity recognition
• deformable attention mechanism
• dynamic convolution
• natural language processing.

Abstract

In joint entity and relation extraction tasks, capturing long-range semantic dependencies often requires high computational resources, limiting the practical deployment of many Transformer-based models. To address this challenge, we introduce DYTNet, a lightweight and efficient framework that reformulates entity-relation extraction as a table-filling task, treating sentence-level interactions as structured cells in a grid. Each cell represents either an entity label or a relation annotation, enabling structured modeling of token-level interactions. The architecture integrates two complementary components: a deformable attention mechanism that dynamically selects semantically relevant positions to model global context, and a dynamic convolution module that adaptively captures local dependencies with reduced computational overhead. Together, they enhance the model’s ability to extract both long-range and short-range features while significantly reducing memory usage and inference time. We further propose a weighted feature fusion strategy to balance the contributions from both modules, improving overall representational robustness. Comprehensive experiments on three widely used benchmarks—CoNLL04, ACE05, and ADE—show that DYTNet achieves superior accuracy in both entity and relation extraction tasks, while maintaining lower FLOPs and memory footprint compared to strong baselines like SpERT and GraphER. The results highlight DYTNet’s potential for real-world applications, especially in resource-constrained environments where both performance and efficiency are critical. Our work demonstrates that high-quality joint extraction can be achieved without relying on excessively large models.

Statistics

Cite this article