Learning unsupervised contextual representations for medical synonym discovery | JoVE Visualize

Area of Science:

Natural Language Processing
Medical Informatics
Machine Learning

Background:

Accurate identification of synonymous terms in medical texts is essential for improving patient data analysis and concept linking to medical ontologies.
Medical terminology exhibits significant variation, often not captured in structured resources, necessitating methods for discovering synonyms from diverse textual data.
Discovering synonyms without relying on annotated data enables the use of large, unannotated medical corpora, reducing costs and potentially capturing broader linguistic nuances.

Purpose of the Study:

To develop and evaluate a method for unsupervised discovery of medical synonyms using contextualized deep learning models.
To leverage advanced language representations to identify synonymous medical terms, regardless of lexical similarity or dissimilarity.
To reduce the dependency on expensive, manually annotated datasets for training synonym discovery models.

Main Methods:

Utilized contextualized deep learning models (e.g., ELMo, BERT) to generate word representations that incorporate surrounding sentence context.
Integrated character-level models to effectively handle out-of-vocabulary medical terms.
Performed unsupervised synonym discovery by analyzing the learned representations to identify likely synonym pairs.

Main Results:

Contextualized deep learning representations demonstrated consistent performance improvements over non-contextualized and non-neural methods across all evaluation metrics.
The proposed method successfully identified synonymous medical terms, showcasing the effectiveness of deep learning in this task.
Evaluation was performed using the ShARe/CLEF eHealth Evaluation Lab 2013 Task 1b dataset.

Conclusions:

Contextualized models provide highly effective representations for the task of medical synonym discovery.
The developed approach offers a robust method for identifying synonyms from unannotated medical text, facilitating better information extraction and knowledge representation.
The gains in performance are expected to extend to other natural language processing tasks within the medical domain.