The paper “Enhancing cross-encoders using knowledge graph hierarchy for medical entity linking in zero- and few-shot scenarios”, co-authored by LASIGE members Pedro Ruas and Francisco M. Couto, along with Fernando Gallego and Francisco J. Veredas, has been published in Knowledge-Based Systems, a prestigious journal in artificial intelligence and data science.

The paper introduces an innovative approach to improve Medical Entity Linking (MEL) by leveraging relationships from knowledge graphs, such as hierarchies and synonyms, to enhance BERT-based cross-encoders. For clinical texts containing medical mentions, the method refines training data with enriched triplets, enabling stronger connections to large knowledge bases like the Unified Medical Language System (UMLS).

This strategy can be integrated into natural language processing pipelines, streamlining the analysis of electronic health records and reducing the effort required for semi-automatic annotation. Potential benefits include more effective medical text processing and improved access to clinical information, particularly for under-resourced languages like Spanish or Portuguese. The paper is available here.