Skip to content

Python library for fast approximate string matching using Jaro and Jaro-Winkler similarity

License

Notifications You must be signed in to change notification settings

rapidfuzz/JaroWinkler

Repository files navigation

JaroWinkler

Continuous Integration PyPI package version Python versions
GitHub license

JaroWinkler is a library to calculate the Jaro and Jaro-Winkler similarity. It is easy to use, is far more performant than all alternatives and is designed to integrate seemingless with RapidFuzz.

⚡ Quickstart

>>> from jarowinkler import *

>>> jaro_similarity("Johnathan", "Jonathan")
0.8796296296296297

>>> jarowinkler_similarity("Johnathan", "Jonathan")
0.9037037037037037

🚀 Benchmarks

The implementation is based on a novel approach to calculate the Jaro-Winkler similarity using bitparallelism. This is significantly faster than the original approach used in other libraries. The following benchmark shows the performance difference to jellyfish and python-Levenshtein.

Benchmark JaroWinkler

⚙️ Installation

You can install this library from PyPI with pip:

pip install jarowinkler

JaroWinkler provides binary wheels for all common platforms.

Source builds

For a source build (for example from a SDist packaged) you only require a C++14 compatible compiler. You can install directly from GitHub if you would like.

pip install git+https://github.com/rapidfuzz/JaroWinkler.git@main

📖 Usage

Any algorithms in JaroWinkler can not only be used with strings, but with any arbitrary sequences of hashable objects:

from jarowinkler import jarowinkler_similarity


jarowinkler_similarity("this is an example".split(), ["this", "is", "a", "example"])
# 0.8666666666666667

So as long as two objects have the same hash they are treated as similar. You can provide a __hash__ method for your own object instances.

class MyObject:
    def __init__(self, hash):
        self.hash = hash

    def __hash__(self):
        return self.hash

jarowinkler_similarity([MyObject(1), MyObject(2)], [MyObject(1), MyObject(2), MyObject(3)])
# 0.9111111111111111

All algorithms provide a score_cutoff parameter. This parameter can be used to filter out bad matches. Internally this allows JaroWinkler to select faster implementations in some places:

jaro_similarity("Johnathan", "Jonathan", score_cutoff=0.9)
# 0.0

jaro_similarity("Johnathan", "Jonathan", score_cutoff=0.85)
# 0.8796296296296297

JaroWinkler can be used with RapidFuzz, which provides multiple methods to compute string metrics on collections of inputs. JaroWinkler implements the RapidFuzz C-API which allows RapidFuzz to call the functions without any of the usual overhead of python, which makes this even faster.

from rapidfuzz import process

process.cdist(["Johnathan", "Jonathan"], ["Johnathan", "Jonathan"], scorer=jarowinkler_similarity)
array([[1.       , 0.9037037],
       [0.9037037, 1.       ]], dtype=float32)

👍 Contributing

PRs are welcome!

  • Found a bug? Report it in form of an issue or even better fix it!
  • Can make something faster? Great! Just avoid external dependencies and remember that existing functionality should still work.
  • Something else that do you think is good? Do it! Just make sure that CI passes and everything from the README is still applicable (interface, features, and so on).
  • Have no time to code? Tell your friends and subscribers about JaroWinkler. More users, more contributions, more amazing features.

Thank you ❤️

⚠️ License

Copyright 2021 - present maxbachmann. JaroWinkler is free and open-source software licensed under the MIT License.