biosenter - Sentence splitting for biomedical text

biosenter is a small spaCy sentence-boundary model plus markup-aware pre/post-processing, trained and evaluated specifically against PMC full text and PubMed abstracts with LLM-annotated sentence boundaries.

Why?

• Generic sentence splitters break constantly on biomedical text
• Full stops appear everywhere in biomedical text (Fig. 1, et al.) and can easily fool a parser into splitting sentence in the wrong place
• Other tools don't deal nicely with citations that hang at the ends of sentences

Install

pip install biosenter

A trained model ships with the package, so split_into_sentences works out of the box with no extra configuration. It integrates directly with bioconverters for extracting text from PMC JATS XML and PubMed baseline/update XML.

Usage

1. Direct spaCy usage

For plain, markup-free text (e.g. PubMed abstracts):

import spacy
nlp = spacy.load("biosenter")
doc = nlp("As shown in Fig. 1, expression increased. A second effect was seen in Fig. 2.")
for sent in doc.sents:
    print(sent.text)

2. Formatting & citation-aware with bioconverters

Real PMC full text carries inline markup and citation markers that spaCy won't treat nicely. Combine functionality from the bioconverters package with the split_into_sentences function to keep formatting and citations in the right sentences:

from bioconverters import pmcxml2tagged
from biosenter import split_into_sentences

# pmcxml2tagged() yields (metadata, text) pairs with inline markup and
# citations preserved, e.g.:
#   'Recurrent <italic>C. difficile</italic> infection is common
#   (<citation pmid="23079555">1</citation>).'
for meta, text in pmcxml2tagged('PMC1234567.xml'):
    for start, end, sentence in split_into_sentences(text):
        print(sentence)

Benchmarks

Sentence-boundary F1 against corpus/test/ -- 100 randomly-sampled PMC articles that the model was neither trained nor checkpoint-selected on -- excluding the formatting/citation handling. The articles used to train and evaluate with were annotated by Claude Sonnet and Opus.

Reproduce with scripts/evaluate_senter.py, which loads any of these by name (they're all just installed spaCy packages):

pip install en_core_web_sm  # or: python -m spacy download en_core_web_sm
python scripts/evaluate_senter.py --models rule en_core_web_sm biosenter/model \
  --pmc_eval_dir corpus/test

scispacy's released models pin to an older spaCy than biosenter requires (en_core_sci_sm 0.5.4 needs spacy<3.8) -- their own README recommends an isolated environment for exactly this reason, so run it separately:

pip install scispacy
pip install https://s3-us-west-2.amazonaws.com/ai2-s2-scispacy/releases/v0.5.4/en_core_sci_sm-0.5.4.tar.gz
pip install --no-deps -e .  # biosenter itself, for split_into_sentences -- skip its spacy>=3.8 pin here
python scripts/evaluate_senter.py --models en_core_sci_sm --pmc_eval_dir corpus/test

Retraining / evaluating

The bundled model, the corpus it was trained on, and the tooling to rebuild or extend either are all included:

• corpus/train/ (300), corpus/validation/ (100) and corpus/test/ (100) -- LLM-annotated PMC articles used to train, tune and score the bundled model. See corpus/README.md for the dataset schema, the annotation conventions, and how it was built.
• scripts/fetch_pmc.py -- fetch more PMC Open Access articles (random or by PMCID), filtered to commercially-redistributable licenses by default.
• scripts/build_senter_corpus.py -- turn fetched XML into corpus records with boundaries bootstrapped from the current model, ready to correct.
• scripts/flag_boundaries.py -- screen a split for likely boundary mistakes (see corpus/README.md) before/after correcting it.
• scripts/prepare_pmc_senter_corpus.py -- convert corpus/{train,validation,test} into spaCy training data.
• scripts/train_senter.py -- train a new model.
• scripts/evaluate_senter.py -- score a model against the hand-labelled eval set.

python scripts/prepare_pmc_senter_corpus.py --corpus_dir corpus/train --out_path data/senter/pmc_train.spacy
python scripts/prepare_pmc_senter_corpus.py --corpus_dir corpus/validation --out_path data/senter/pmc_val.spacy
python scripts/train_senter.py --data_dir data/senter --run_name my_run --seed 0
python scripts/evaluate_senter.py --models rule runs/my_run/model-best --pmc_eval_dir corpus/validation

License

MIT. See LICENSE.