Concept placement using BERT trained by transforming and summarizing biomedical ontology structure.

Area of Science:

• Ontology Engineering
• Natural Language Processing
• Bioinformatics

Background:

• Hierarchical positioning of concepts in ontologies relies on IS-A relationships.
• Manual identification of these relationships is time-consuming and requires specialized expertise.
• Automating this process is crucial for efficient ontology development.

Purpose of the Study:

• To develop an automated method for predicting IS-A relationships between new and existing concepts.
• To leverage language representation models for ontology concept prediction.
• To improve the efficiency and accuracy of ontology curation.

Main Methods:

• Utilized the BERT language representation model for predicting IS-A relationships.
• Converted concept neighborhood networks into "sentences" for BERT's Next Sentence Prediction (NSP).
• Employed an ontology summarization technique to refine training data and enhance model performance.

Main Results:

• The proposed method achieved an average F1 score of 0.88.
• The ontology summarization technique improved the average Recall score from 0.94 to 0.96.
• The model was trained on SNOMED CT hierarchies and applied to predict parents of new concepts.

Conclusions:

• Automated prediction of IS-A relationships using BERT is feasible and effective.
• Ontology summarization significantly boosts the performance of concept relationship prediction.
• This method offers a valuable tool for accelerating ontology development and maintenance.