An End-to-End Re-Evaluation of Table Entity-Linkers
Martin Pekár Christensen, Katja Hose, Matteo Lissandrini
Zenodo (CERN European Organization for Nuclear Research)
Problems Identified (5)
Incomplete Entity Linker Evaluation: Existing evaluations of table entity linkers are incomplete because they focus on specific applications and aggregated output metrics rather than component effectiveness and scalability.
Overlooked Candidate Generation: Candidate generation is a crucial entity linking step that is commonly overlooked.
Entity Linker Irreproducibility: Many entity linkers are difficult to reproduce due to lack of open source code or reliance on irreproducible endpoints and datasets.
Incomplete Entity Linker Evaluation: Existing evaluations of table entity linkers are incomplete because they focus on specific applications and aggregated output metrics rather than component effectiveness and scalability.
Overlooked Candidate Generation: Candidate generation is a crucial entity linking step that is commonly overlooked.
Proposed Solutions (4)
Component-Level Linker Evaluation: The paper provides an in-depth analysis and taxonomy of state-of-the-art entity linkers by evaluating the quality and scalability of individual entity linking components.
Multi-Benchmark KG Linker Evaluation: The paper evaluates entity linkers on four existing entity linking benchmarks using DBpedia and Wikidata.
Component-Level Linker Evaluation: The paper provides an in-depth analysis and taxonomy of state-of-the-art entity linkers by evaluating the quality and scalability of individual entity linking components.
Multi-Benchmark KG Linker Evaluation: The paper evaluates entity linkers on four existing entity linking benchmarks using DBpedia and Wikidata.
Results (3)
Real-World Bottleneck Identification:
Candidate Generation Bottleneck:
Linker Reproducibility Limitation:
Research Domain
Knowledge graph entity linking for tables