Agnus: Robust Entity Disambiguation using LLMs

ACL ARR 2025 May Submission7944 Authors

Abstract:

Entity disambiguation (ED) is the process of disambiguating entities relating to a knowledge base and a necessary step of the entity linking workflow. With the advent of pretrained generative large language models (LLM), the field of natural language processing has been revolutionised, yet related techniques for ED are scarce.
In this paper, we introduce Agnus an approach leveraging pretrained LLM contextual knowledge to disambiguate entities.
We mitigate challenges posed by modern LLMs: order-dependant bias for candidate options, hallucinations and evaluation data contamination. We reach state-of-the-art results in 4 datasets, beating prior work by 3.7% on average for zero-shot configurations, provide code and a novel synthetic dataset for entity disambiguation.

Paper Type: Long

Research Area: Information Extraction

Research Area Keywords: entity disambiguation, llm, synthetic data

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Publicly available software and/or pre-trained models, Data resources

Languages Studied: English

Submission Number: 7944