Abstract
Biomedical entity linking (EL) consists of named entity recognition (NER) andnamed entity disambiguation (NED). EL models are trained on corpora labeled bya predefined KB. However, it is a common scenario that only entities within asubset of the KB are precious to stakeholders. We name this scenario partialknowledge base inference: training an EL model with one KB and inferring on thepart of it without further training. In this work, we give a detaileddefinition and evaluation procedures for this practically valuable butsignificantly understudied scenario and evaluate methods from threerepresentative EL paradigms. We construct partial KB inference benchmarks andwitness a catastrophic degradation in EL performance due to dramaticallyprecision drop. Our findings reveal these EL paradigms can not correctly handleunlinkable mentions (NIL), so they are not robust to partial KB inference. Wealso propose two simple-and-effective redemption methods to combat the NILissue with little computational overhead.