Advances in Biomedical Entity Linking and Continual Adaptation

ADAPT-BioEL: A Zero-Shot Biomedical Entity Matching and Linking – Work in Progress by Katerina Gkirtzou

Biomedical entity matching and linking are foundational tasks for transforming unstructured clinical text into structured, interoperable knowledge. Despite substantial progress, current state-of-the-art knowledge extraction systems rely heavily on large, ontology-specific annotated corpora—resources that exist for only a small fraction of biomedical ontologies, limiting scalability across the many biomedical ontologies that lack such resources. 

In this talk, I will introduce ADAPT-BioEL, an ongoing project that aims to advance biomedical Named Entity Recognition and Entity Linking through a novel two-stage framework that combines robust span representation learning with ontology-adaptive linking. I will present the design principles behind the approach, discuss the current stage of development, along with early experimental results and key technical challenges encountered during development. 

MoSE: Mixture-of-Specialized-Experts for Efficient Continual Adaptation by Van-Tuan Tran 

Continual learning and adaptive AI systems must efficiently incorporate new knowledge without degrading performance on previously learned tasks, while also operating under computational and memory constraints. This talk presents MoSE (Mixture-of-Specialized-Experts), a framework designed to enable efficient continual adaptation through modular expert architectures. 

The approach leverages a mixture-of-experts paradigm in which specialized subnetworks are dynamically selected or updated to handle new tasks or domains, reducing catastrophic forgetting and improving parameter efficiency compared to monolithic continual learning models. The presentation will discuss the architecture design, routing strategies, and adaptation mechanisms, as well as potential applications in evolving data environments and adaptive AI systems. Preliminary results and ongoing research challenges related to scalability, expert specialization, and training stability will also be discussed.