LELA: an LLM-based Entity Linking Approach with Zero-Shot Domain Adaptation

TL;DR

This work tackles true zero-shot entity linking by proposing LELA, a modular coarse-to-fine pipeline that leverages large language models without any fine-tuning. It components include retriever-agnostic candidate generation, an LLM-based pointwise reranker, and a reasoning-driven final selection with self-consistency, all integrated into a Python/ spaCy pipeline. Across multiple benchmarks (ZESHEL, ESCO, GLADIS, ZELDA) and domains, LELA consistently outperforms prior true zero-shot methods and often rivals fine-tuned approaches, while remaining robust to the choice of underlying LLM and retrieval method. The findings suggest that inference-time reasoning can substitute for domain-specific supervision, reducing the data burden for deploying EL in new domains, and point to broader applicability of LLMs as EL baselines in future work.

Abstract

Entity linking (mapping ambiguous mentions in text to entities in a knowledge base) is a foundational step in tasks such as knowledge graph construction, question-answering, and information extraction. Our method, LELA, is a modular coarse-to-fine approach that leverages the capabilities of large language models (LLMs), and works with different target domains, knowledge bases and LLMs, without any fine-tuning phase. Our experiments across various entity linking settings show that LELA is highly competitive with fine-tuned approaches, and substantially outperforms the non-fine-tuned ones.

Figures (16)

• Figure 1: Classification of entity linking settings ("training" refers to fine-tuning after initial pre-training). LELA is in the true zero-shot setting.
• Figure 2: Overview of the LELA approach.
• Figure 3: Reranker, retriever and LLM selection prompts.
• Figure 4: Comparison of (micro-averaged) accuracy, selection accuracy and accuracy@k on ZESHEL.
• Figure 5: Output diversity histogram, using 10 samples for self consistency, on ZESHEL.