Skip to main content

Biomedical Named Entity Recognition and Entity Linking for Enterprise use cases

Project description

Maturity level-1

Kazu - Biomedical NLP Framework

Find our docs here

Kazu - Biomedical NLP Framework

Note: the recent 2.0 release has large elements of backwards incompatibility if you are using a custom model pack and curations.

Welcome to Kazu (Korea AstraZeneca University), a python biomedical NLP framework built in collaboration with Korea University, designed to handle production workloads.

This library aims to simplify the process of using state of the art NLP research in production systems. Some of the research contained within are our own, but most of it comes from the community, for which we are immensely grateful.

If you want to use Kazu, please cite our EMNLP 2022 publication! (citation link)

Please click here for the TinyBERN2 training and evaluation code



Python version 3.9 or higher is required (tested with Python 3.11).


pip install kazu

or download the wheel from the release page and install locally.

If you intend to use Mypy on your own codebase, consider installing Kazu using:

pip install kazu[typed]

This will pull in typing stubs for kazu's dependencies (such as types-requests for Requests) so that mypy has access to as much relevant typing information as possible when type checking your codebase. Otherwise (depending on mypy config), you may see errors when running mypy like:

.venv/lib/python3.10/site-packages/kazu/steps/linking/post_processing/ error: Library stubs not installed for "requests" [import]

Getting the model pack

For most functionality, you will also need the Kazu model pack. This is tied to each release, and can be found on the release page. Once downloaded, extract the archive and:

export KAZU_MODEL_PACK=<path to the extracted archive>

Kazu is highly configurable (using Hydra), although it comes preconfigured with defaults appropriate for most literature processing use cases. To make use of these, and process a simple document:

import hydra
from hydra.utils import instantiate

from import Document
from kazu.pipeline import Pipeline
from kazu.utils.constants import HYDRA_VERSION_BASE
from pathlib import Path
import os

# the hydra config is kept in the model pack
cdir = Path(os.environ["KAZU_MODEL_PACK"]).joinpath("conf")

    version_base=HYDRA_VERSION_BASE, config_path=str(cdir), config_name="config"
def kazu_test(cfg):
    pipeline: Pipeline = instantiate(cfg.Pipeline)
    text = "EGFR mutations are often implicated in lung cancer"
    doc = Document.create_simple_document(text)

if __name__ == "__main__":


Licensed under Apache 2.0.

Kazu includes elements under compatible licenses (full licenses are in relevant files or as indicated):

  • Some elements are a modification of code licensed under MIT by Explosion.AI - see the README here.
  • The doc build process ('s linkcode_resolve function) uses code modified from pandas, in turn modified from numpy. See PANDAS_LICENSE.txt and NUMPY_LICENSE.txt
  • Elements of the model distillation code are inspired by or modified from Huawei Noah's Ark Lab TinyBERT and DMIS-Lab's BioBERT. See the details in, and
  • PLSapbertModel is inspired by the code from sapbert, licensed under MIT. See the file for details, and see the SapBert section below regarding use of the model.
  • GildaUtils in the file is modified from Gilda. See the file for full details including the full BSD 2-Clause license.
  • The AbbreviationFinderStep uses KazuAbbreviationDetector, which is a modified version of SciSpacy's abbreviation finding algorithm, licensed under Apache 2.0 - see the files for full details.
  • The JWTAuthenticationBackend Starlette Middleware in is originally from starlette-jwt, licensed under BSD 3-Clause.
  • The AddRequestIdMiddleware Starlette Middleware in is modified from 'CustomHeaderMiddleware' in the Starlette Middleware docs. This is licensed under BSD 3-Clause along with the rest of Starlette.
  • The kazu-jvm folder includes files like gradelw and gradelw.bat distributed by gradle under Apache 2.0 - see the files for details.
  • kazu/ contains AutoNameEnum, which is AutoName from the Python Enum Docs licensed under Zero-Clause BSD.

Dataset licences

For the version of each ontology currently in use, please see the 'data_origin' field in kazu/conf/ontologies

Under Creative Commons Attribution-Share Alike 3.0 Unported Licence


ChEMBL data is from


CLO data is from


UBERON data is from

Under Creative Commons Attribution 4.0 Unported License


MONDO data is from


CELLOSAURUS data is from

Gene Ontology

Gene Ontology data is from

Other licenced datasets and models


This service/product uses the Human Phenotype Ontology (version information). Find out more at

Freely licenced under

Sebastian Köhler, Michael Gargano, Nicolas Matentzoglu, Leigh C Carmody, David Lewis-Smith, Nicole A Vasilevsky, Daniel Danis, Ganna Balagura, Gareth Baynam, Amy M Brower, Tiffany J Callahan, Christopher G Chute, Johanna L Est, Peter D Galer, Shiva Ganesan, Matthias Griese, Matthias Haimel, Julia Pazmandi, Marc Hanauer, Nomi L Harris, Michael J Hartnett, Maximilian Hastreiter, Fabian Hauck, Yongqun He, Tim Jeske, Hugh Kearney, Gerhard Kindle, Christoph Klein, Katrin Knoflach, Roland Krause, David Lagorce, Julie A McMurry, Jillian A Miller, Monica C Munoz-Torres, Rebecca L Peters, Christina K Rapp, Ana M Rath, Shahmir A Rind, Avi Z Rosenberg, Michael M Segal, Markus G Seidel, Damian Smedley, Tomer Talmy, Yarlalu Thomas, Samuel A Wiafe, Julie Xian, Zafer Yüksel, Ingo Helbig, Christopher J Mungall, Melissa A Haendel, Peter N Robinson,

The Human Phenotype Ontology in 2021,

Nucleic Acids Research, Volume 49, Issue D1, 8 January 2021, Pages D1207–D1217,


Open Targets datasets are kindly provided by, which are free for commercial use cases

Ochoa, D. et al. (2021). Open Targets Platform: supporting systematic drug–target identification and prioritisation. Nucleic Acids Research.


The Stanza framework:

Peng Qi, Yuhao Zhang, Yuhui Zhang, Jason Bolton and Christopher D. Manning. 2020. Stanza: A Python Natural Language Processing Toolkit for Many Human Languages. In Association for Computational Linguistics (ACL) System Demonstrations. 2020.

Biomedical NLP models are derived from:

Yuhao Zhang, Yuhui Zhang, Peng Qi, Christopher D. Manning, Curtis P. Langlotz.
Biomedical and Clinical English Model Packages in the Stanza Python NLP Library,
Journal of the American Medical Informatics Association. 2021.


Biomedical scispacy models are derived from

Mark Neumann, Daniel King, Iz Beltagy, Waleed Ammar
ScispaCy: Fast and Robust Models for Biomedical Natural Language Processing
Proceedings of the 18th BioNLP Workshop and Shared Task
ACL 2019


Kazu uses a distilled form of SAPBERT, from

Fangyu Liu, Ehsan Shareghi, Zaiqiao Meng, Marco Basaldella, Nigel Collier
Self-Alignment Pretraining for Biomedical Entity Representations
ACL 2021


GLiNER: Generalist Model for Named Entity Recognition using Bidirectional Transformer.
Urchade Zaratiana, Nadi Tomeh, Pierre Holat, Thierry Charnois


Kazu's SethStep uses Py4j to call the SETH mutation finder.

Thomas, P., Rocktäschel, T., Hakenberg, J., Mayer, L., and Leser, U. (2016).
SETH detects and normalizes genetic variants in text
Bioinformatics (2016)


Kazu's OpsinStep uses Py4j to call OPSIN: Open Parser for Systematic IUPAC nomenclature.

Daniel M. Lowe, Peter T. Corbett, Peter Murray-Rust, and Robert C. Glen
Chemical Name to Structure: OPSIN, an Open Source Solution
Journal of Chemical Information and Modeling 2011 51 (3), 739-753
DOI: 10.1021/ci100384d

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

kazu-2.0.0.tar.gz (197.4 kB view hashes)

Uploaded Source

Built Distribution

kazu-2.0.0-py3-none-any.whl (258.2 kB view hashes)

Uploaded Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page