📄 Learning Dynamic Context Augmentation for Global Entity Linking

Summary

Global (collective) entity linking improves disambiguation by enforcing coherence among all entities in a document. However, it relies on strong coherence assumptions and costly approximate inference, such as loopy belief propagation, which scales poorly. The authors propose Dynamic Context Augmentation (DCA), which links mentions in a single sequential pass. Each decision is enriched with information from entities already linked, plus their Wikipedia neighbours (entities with inlinks to them). Soft attention and hard top-K pruning reduce noise from irrelevant or wrongly linked entities. DCA works as a plug-and-enhance module on top of existing local models, demonstrated with ETHZ-Attn and Berkeley-CNN. The authors train it two ways: supervised ranking with a max-margin loss, using gold history, and reinforcement learning, which accounts for errors carried forward to later decisions. Experiments show accuracy comparable to global models at much lower cost. Running time grows linearly with the number of candidates, and memory stays low. This makes collective entity linking practical for large-scale production.

Key points

• DCA processes each document once, using previously linked entities and their neighbours as accumulated context. Global methods such as Ment-Norm and Deep-ED instead repeatedly refine all assignments.
• Adding neighbour entities (2-hop) over linked entities alone (1-hop) gives statistically significant average relative gains (p < 0.005):
  • 0.72% over 1-hop DCA-(SL/RL);
  • 3.56% over a baseline-SL model without DCA.
• Combining soft and hard attention outperforms simple averaging of previous entities, and hard pruning adds a small further gain. DCA-RL proved robust to noise on the CWEB dataset.
• Compared with Ment-Norm and Deep-ED, DCA's running time scales linearly with candidate count, and its memory use stays nearly flat. It used up to 80% less energy than Ment-Norm.
• An auxiliary entity-typing system (PER, GPE, ORG, UNK) trained on AIDA-train reached 95% accuracy on AIDA-A. Code and data are released at github.com/YoungXiyuan/DCA.