Modular Entity Disambiguation via LLM Selection

This study evaluates decoupling entity disambiguation into a retrieval stage and an LLM-based selection stage rather than using single-task dual-encoders. Results show that combining various retrieval strategies like BM25 or dense retrievers with a centralized LLM selector can be more effective than end-to-end training.