LLM-based selection boosts entity disambiguation performance

Researchers have developed a modular approach to entity disambiguation (ED) that separates candidate retrieval from entity selection. By using a large language model (LLM) for selection and a training-free BM25 retriever for candidate generation, they achieved a new state-of-the-art performance on the ZELDA benchmark, improving inKB micro-F1 from 82.3 to 86.3. This decoupled system also allows for abstention when the correct entity is not found among candidates, leading to a 90.7 F1 score in an evaluation that rewards correct abstentions. AI

IMPACT This research could improve knowledge graph construction and retrieval by decoupling entity selection from training, potentially leading to more accurate and efficient information extraction.

RANK_REASON The item is an academic paper detailing a new methodology for entity disambiguation. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• alphaXiv
• arXiv
• BM25
• CatalyzeX
• DagsHub
• entity disambiguation
• Gotit.pub
• Hugging Face
• ScienceCast
• ZELDA

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →