The impact of learning Unified Medical Language System knowledge embeddings in relation extraction from biomedical...

Area of Science:

• Biomedical Natural Language Processing
• Bioinformatics
• Computational Linguistics

Background:

• Accurate relation extraction from biomedical texts is crucial for knowledge discovery.
• Existing methods often struggle to fully leverage the rich information within biomedical terminologies.
• The Unified Medical Language System (UMLS) Metathesaurus offers a valuable resource for enhancing such systems.

Purpose of the Study:

• To investigate the impact of knowledge embeddings (KEs) derived from the UMLS Metathesaurus on biomedical relation extraction quality.
• To develop and evaluate a novel system, REKE, that incorporates KEs for improved relation extraction.

Main Methods:

• Two forms of KEs were learned from the UMLS Metathesaurus: lexicalized (LKEs) and unlexicalized.
• A knowledge embedding encoder (KEE) was developed to learn these embeddings.
• The REKE system was designed to integrate LKEs or unlexicalized KEs for relation types and their arguments.

Main Results:

• The REKE system incorporating KEs advanced the state of the art in relation extraction on two diverse datasets (i2b2/VA and Drug-Drug Interaction).
• Lexicalized knowledge embeddings (LKEs) demonstrated superior performance, achieving F1 scores of 78.2 and 82.0.
• The system successfully incorporated knowledge from the UMLS Metathesaurus through novel KE representations.

Conclusions:

• Incorporating LKEs informed by the UMLS Metathesaurus shows significant promise for biomedical relation extraction.
• The REKE system establishes new state-of-the-art results, highlighting the effectiveness of this approach.
• The study underscores the importance and subtleties of integrating KEs into relation extraction systems.