Biomedical ontology alignment: an approach based on representation learning.

Area of Science:

• Natural Language Processing
• Machine Learning
• Bioinformatics

Background:

• Representation learning shows promise in NLP but has limited impact on ontology matching.
• Traditional methods rely on feature engineering, which is less effective for complex semantic relationships.

Purpose of the Study:

• To develop a novel representation learning approach specifically for the ontology matching task.
• To improve the accuracy and efficiency of aligning disparate ontologies.

Main Methods:

• Embedding ontological terms into a high-dimensional Euclidean space.
• Utilizing a novel phrase retrofitting strategy to inscribe semantic similarity onto pre-trained word vectors.
• Incorporating a denoising autoencoder for outlier detection to enhance performance.

Main Results:

• Achieved an F-score of 94% in aligning the Adult Mouse Anatomical Dictionary with the Foundational Model of Anatomy ontology (FMA).
• Demonstrated state-of-the-art performance by achieving F-scores of 93.2% and 89.2% when aligning FMA to NCI Thesaurus and SNOMED CT, respectively.
• The novel framework significantly improved upon existing ontology matching systems.

Conclusions:

• The proposed representation learning approach effectively captures semantic similarity using terminological embeddings.
• The method is well-suited for ontology matching, offering a new pathway for addressing this challenge.
• Results indicate the potential for broader applications in knowledge representation and semantic integration.