DIAL-RE: LLM-Assisted False-Negative Recovery for Biomedical Relation Extraction

Incomplete relation annotations create a difficult acquisition problem: uncertain entity pairs can refine a classifier's boundary, while high-confidence unannotated pairs may reveal missed relations. We present ours, an iterative framework that allocates queries to both regions and uses an LLM to check candidate relations against textual evidence. Accepted supervision augments a biomedical relation extractor across training rounds. We evaluate the framework on BioRED and CirrhoRED, a liver-cirrhosis corpus containing 800 documents with relation-annotation agreement of kappa=0.9325. On BioRED, the Full configuration achieves 69.85 gold-entity (GE) micro-F1, improving over the shared encoder baseline by 3.03 percentage points. On CirrhoRED, it achieves 71.46 GE micro-F1, a gain of 2.62 points. The corresponding end-to-end scores are 59.47 on BioRED and 63.73 on CirrhoRED. Performance differences among configurations vary across datasets, and the iteration trajectory peaks before the final round. Evidence-level analyses illustrate the role of relation polarity and span extraction in supervision admission.