ClassiVax

Classify vaccine-related news headlines into topic categories using a Recognizing Textual Entailment (RTE) approach, powered by the LuizNeves/BERT-NLI-vaccine-v2 model.

Built on the method described in:

Neves, L. F. F., Camargo, C. Q., & Massarani, L. (2025). Boosting Transformers: Recognizing Textual Entailment for Classification of Vaccine News Coverage. Computational Communication Research, 7(1), 1–32. https://doi.org/10.5117/CCR2025.1.1.NEVE

Installation

pip install classivax

Usage

Default classes (10 vaccine-related topics from the paper)

from classivax import ClassiVaxClassifier

clf = ClassiVaxClassifier()

result = clf.classify("Twitter will begin removing vaccine misinformation")
print(result["class"])   # "Vaccine Hesitancy and Mis-/Disinformation"
print(result["score"])   # e.g. 0.999

Batch classification

headlines = [
    "Twitter will begin removing vaccine misinformation",
    "Britain begins coronavirus vaccine rollout",
    "COVID-19 vaccination key to economic recovery",
]

df = clf.classify_batch(headlines)
print(df)

Custom classes

The underlying RTE approach also supports classifying text against any set of natural-language class descriptions you define — not just the original 10 classes.

custom_classes = {
    "Vaccine Safety Debate": "The headline discusses concerns, controversies, or debates about vaccine safety",
    "Vaccine Logistics": "The headline discusses shipping, storage, cold chain, or distribution logistics for vaccines",
}

clf = ClassiVaxClassifier(classes=custom_classes)
result = clf.classify("Cold storage shortages delay vaccine shipments to rural areas")
print(result["class"])  # "Vaccine Logistics"

Note: the underlying model was fine-tuned specifically on English-language, vaccine-related headlines from 2020–2021. Using custom classes outside this domain is supported, but accuracy has not been formally evaluated beyond the original study's scope — treat results on very different domains with appropriate caution.

The 10 default classes

How it works

For each headline, the classifier tests it against every class description as a text-hypothesis pair, following the entailment framework: the headline is the text (T), and each class description is a candidate hypothesis (H). The class whose description achieves the highest entailment score is selected as the prediction.

Citation

If you use this library or the underlying model in your research, please cite:

@article{neves2025boosting,
  title={Boosting Transformers: Recognizing Textual Entailment for Classification of Vaccine News Coverage},
  author={Neves, Luiz F. F. and Camargo, Chico Q. and Massarani, Luisa},
  journal={Computational Communication Research},
  volume={7},
  number={1},
  pages={1--32},
  year={2025},
  doi={10.5117/CCR2025.1.1.NEVE}
}

License

MIT