Towards Better Graph-based Cross-document Relation Extraction via Non-bridge Entity Enhancement and Prediction Debiasing targets relation extraction where evidence is distributed across multiple documents.

The paper argues that prior cross-document systems over-focus on bridge entities while ignoring non-bridge entities that still provide semantic association. It also addresses prediction bias caused by many NA instances in CodRED.

Why we're excited

The proposed system builds a unified entity graph over target, bridge, and non-bridge entities, then uses a graph recurrent network plus debiasing. That combination is interesting because it improves both evidence representation and decision calibration.

The paper reports state-of-the-art results on CodRED closed and open settings and releases code.

Our community read

This is one of the most directly relevant patterns for multi-document intelligence, literature discovery, and dossier-style extraction. Those systems live or die by whether evidence routing works.

The limitation is that operational cross-document IE requires more than relation modeling. Retrieval, document bag construction, deduplication, provenance, and false-positive control all become part of the product surface.

Source

• arXiv: 2406.16529