Biomedical and clinical English model packages for the Stanza Python NLP library.

Area of Science:

• Computational linguistics
• Bioinformatics
• Medical informatics

Background:

• Natural Language Processing (NLP) is crucial for extracting information from unstructured biomedical and clinical text.
• Existing NLP tools often require significant adaptation or lack performance for specialized scientific domains.
• The Stanza library provides a foundation for developing domain-specific NLP capabilities.

Purpose of the Study:

• To develop and evaluate novel neural NLP packages tailored for syntactic analysis and named entity recognition (NER) in biomedical and clinical English.
• To extend the capabilities of the Stanza library for specialized scientific language processing.
• To provide researchers with accessible and high-performance NLP tools for medical text.

Main Methods:

• Implementation and training of NLP pipelines using the Stanza library, incorporating public (CRAFT treebank) and private annotated radiology reports.
• Development of neural network-based models for tokenization, part-of-speech tagging, lemmatization, dependency parsing, and NER.
• Comparative evaluation against established NLP libraries (CoreNLP, scispaCy) and state-of-the-art models (BioBERT, BioNLP CRAFT winners).

Main Results:

• Achieved superior performance in syntactic analysis compared to retrained scispaCy and CoreNLP models, matching top systems from the CRAFT shared task.
• Demonstrated substantial outperformance over scispaCy for NER and achieved performance comparable or superior to BioBERT.
• Highlighted significant computational efficiency of the developed systems.

Conclusions:

• Introduced new, user-friendly biomedical and clinical NLP packages for the Stanza library.
• The packages offer state-of-the-art performance and are optimized for ease of use.
• All models are publicly released to support further research, with an online demonstration provided.