ScispaCy: Fast and Robust Models for Biomedical Natural Language Processing (1902.07669v3)

Abstract: Despite recent advances in natural language processing, many statistical models for processing text perform extremely poorly under domain shift. Processing biomedical and clinical text is a critically important application area of natural language processing, for which there are few robust, practical, publicly available models. This paper describes scispaCy, a new tool for practical biomedical/scientific text processing, which heavily leverages the spaCy library. We detail the performance of two packages of models released in scispaCy and demonstrate their robustness on several tasks and datasets. Models and code are available at https://allenai.github.io/scispacy/

• The paper introduces ScispaCy, a library that enhances biomedical NLP with fast, domain-optimized models for tagging, parsing, and NER.
• It presents two model packages, en_core_sci_sm and en_core_sci_md, which achieve high accuracy in POS tagging and dependency parsing on the GENIA dataset.
• The study highlights efficient integration into Python workflows, enabling practical biomedical text analysis for research and clinical applications.

ScispaCy: Fast and Robust Models for Biomedical NLP

The paper presents ScispaCy, a Python library designed to enhance biomedical text processing by leveraging the robust SpaCy NLP framework. Given the exponential growth of publications in the biomedical field, ScispaCy aims to provide effective tools for extracting structured information from scientific documents.

Core Contributions

ScispaCy introduces two model packages: en_core_sci_sm and en_core_sci_md. These models are developed with specific considerations for the biomedical domain, offering advancements over existing tools in terms of speed, robustness, and ease of use within Python-based workflows. The models focus on traditional NLP tasks like POS tagging, dependency parsing, and NER, specifically optimized for biomedical text.

Technical Evaluation

• Processing Speed: The models exhibit competitive processing speeds, comparable to tools written in C++ and Java, such as the GENIA Tagger.
• POS Tagging and Dependency Parsing: The ScispaCy models show high accuracy on the GENIA dataset. While slightly outperformed by the Biaffine parser in terms of parsing accuracy, ScispaCy's efficiency offers a substantial advantage.
• Named Entity Recognition: ScispaCy is benchmarked on multiple datasets, demonstrating competitive baseline performance for various biomedical NER tasks. Models trained specifically on datasets like BC5CDR demonstrate effectively balanced precision and recall.

Data and Methodology

The paper releases a reformatted GENIA 1.0 corpus, adapting it to Universal Dependencies, providing an essential resource for further research in biomedical NLP. Additionally, training incorporates data from OntoNotes to enhance robustness across a broad range of text types.

Practical Implications

ScispaCy's integration into the Python ecosystem offers a gateway for applying advanced NLP in biomedical and clinical settings. Its efficient design allows straightforward incorporation into applications requiring detailed text processing without significant computational overhead.

Theoretical Implications

The framework advances domain adaptation in NLP by retraining models, enhancing them with biomedical-specific optimizations while maintaining computational efficiency. This adaptation lays the groundwork for further exploration of domain-focused NLP tools in scientific domains.

Future Directions

The paper suggests potential expansions including a more comprehensive entity linker and other pipeline features like negation detection. These additions would broaden ScispaCy's utility in clinical and biomedical NLP applications, enhancing its role in extracting and linking biomedical information.

In conclusion, ScispaCy contributes significant advancements to NLP in the biomedical domain, balancing performance with efficiency. It furnishes the research community with an adaptable and robust toolkit, supporting ongoing advancements in biomedical information extraction.