Natural Language Processing in Biomedical Literature Mining

Abstract

The biomedical literature doubles approximately every five years, making manual curation of knowledge from published text increasingly untenable, and natural language processing models that automatically extract entities, relationships, and assertions from scientific articles now offer scalable solutions, yet their adoption into biomedical knowledge workflows varies across NLP task types and application contexts. We evaluated 214 biomedical NLP programmes across centres in Austria and Germany between 2017 and 2021, spanning five task categories: named entity recognition for biomedical entities, relation extraction between genes, diseases, and drugs, text classification and triage of articles, question answering over biomedical corpora, and knowledge graph construction from literature. A Biomedical NLP Effectiveness Index (BNLPEI) was constructed from five sub-scores -- extraction accuracy on benchmark corpora, domain-specific language model quality, real-world curation acceleration, cross-corpus generalisation, and integration into downstream knowledge systems -- with weights from regression against sustained adoption into biomedical knowledge pipelines. BNLPEI correlated with adoption at r = +0.83 and discriminated adopted from non-adopted methods with an AUC of 0.880. Named entity recognition scored highest (mean BNLPEI 0.822), while knowledge graph construction trailed at 0.596. Only 34.6 percent exceeded the 0.75 threshold. Extraction accuracy carried the largest weight (beta = +0.278), followed by curation acceleration (beta = +0.230).