NILINKER: Attention-based approach to NIL Entity Linking.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Natural Language Processing

Background:

• Unlinkable (NIL) entities pose a significant challenge for Named Entity Linking (NEL) systems.
• Current NIL entity approaches are limited to general domains and do not adequately address the complexities of specialized fields like biomedical sciences.
• The rapid growth of biomedical literature exacerbates the NIL entity problem.

Purpose of the Study:

• To develop a novel model, NILINKER, for addressing the challenge of NIL entities in the biomedical domain.
• To improve the performance of Named Entity Linking (NEL) by effectively linking biomedical NIL entities to relevant knowledge bases.
• To introduce a comprehensive evaluation dataset, EvaNIL, for training and assessing NIL entity linking models.

Main Methods:

• NILINKER employs a candidate retrieval module specifically designed for biomedical NIL entities.
• A neural network utilizing an attention mechanism identifies relevant concepts from target knowledge bases (MEDIC, CTD-Chemicals, ChEBI, HP, CTD-Anatomy, Gene Ontology-Biological Process).
• The EvaNIL dataset, comprising over 846,000 documents and 1 million annotations, was created for evaluating NIL entity linking models.

Main Results:

• NILINKER successfully identifies and links biomedical NIL entities to relevant knowledge base concepts.
• Integration of NILINKER into the REEL graph-based NEL model demonstrated a notable increase in NEL performance.
• The EvaNIL dataset provides a robust benchmark for future research in biomedical NIL entity linking.

Conclusions:

• NILINKER offers a significant advancement in handling NIL entities within the biomedical domain.
• The proposed model and dataset contribute to more accurate and effective Named Entity Linking in scientific literature.
• This work paves the way for improved information extraction and knowledge discovery in biomedical research.