Reproduce “End-to-end Biomedical Entity Linking with Span-based Dictionary Matching” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shogo Ujiie, Hayate Iso, Shuntaro Yada, Shoko Wakamiya, Eiji Aramaki
A team from Nara Institute of Science and Technology developed an end-to-end model for biomedical entity linking that integrates span-based neural representations with dictionary-matching features. The model addresses the "zero-shot problem" by normalizing concepts not present in training data, achieving NEN F1-scores of 0.704 on NCBID and 0.597 on BC5CDR for unseen terms, while also securing top performance on standard tasks with F1-scores of 0.891 for NER and 0.854 for NEN on the NCBID corpus.
Key insights
- Explicit dictionary-matching features are effective for solving the "zero-shot problem" in biomedical entity linking, enabling the model to normalize concepts unseen during training.
- The hybrid architecture, combining data-driven neural representations with knowledge-based dictionary features in an end-to-end span-based framework, improves overall performance for both NER and NEN tasks.
- Dictionary information offers a valuable, complementary signal that refines predictions even for concepts present in the training data, enhancing the neural model's learned contextual understanding.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.