Unified Medical Language System resources improve sieve-based generation and Bidirectional Encoder Representations...

Area of Science:

• Natural Language Processing
• Medical Informatics

Background:

• Concept normalization links text phrases to ontology concepts, crucial for tasks like relation extraction and information retrieval.
• Accurate concept normalization requires robust systems to handle the complexity of clinical text.

Purpose of the Study:

• To present a generate-and-rank concept normalization system for clinical text.
• To evaluate the system's performance in the 2019 National NLP Clinical Challenges Shared Task Track 3.

Main Methods:

• A sieve-based candidate generation system using Lucene indices, Unified Medical Language System (UMLS) preferred terms, and synonyms.
• A BERT-based neural network classifier for ranking candidate concepts, incorporating UMLS semantic types via a regularizer.

Main Results:

• The generate-and-rank system achieved third place out of 33 participants.
• The system's accuracy improved from 79.44% (candidate generator alone) to 81.66% and later to 83.56% post-evaluation.
• Performance surpassed the previous state-of-the-art accuracy of 76.35%.

Conclusions:

• The generate-and-rank framework effectively integrates UMLS features with neural networks for accurate concept normalization.
• Prioritizing UMLS preferred terms and utilizing the semantic type regularizer enhances prediction quality.
• The model demonstrates robust performance, even on concepts not encountered during training.