scispacy

Why we included this project

scispaCy is built for anyone doing text mining over research papers, clinical notes, or other biomedical text. General-purpose spaCy models tend to stumble on dense medical vocabulary, so this project adds a custom tokenizer plus a POS tagger, syntactic parser, and entity span detector trained on biomedical corpora. Models load the same way as standard spaCy ones, so existing pipelines take it in with little refactoring. Separate NER models trained on CRAFT, JNLPBA, BC5CDR, and BIONLP13CG recognize genes, chemicals, and disease mentions, and add-on pipes like the abbreviation detector and UMLS entity linker handle the messier parts of scientific writing. With small, medium, large, and SciBERT-backed variants, you can trade speed against accuracy depending on whether you are processing millions of abstracts in batch or running a live annotation service.