Alt-profanity-check

Alt profanity check is a drop-in replacement of the profanity-check library for the not so well maintained https://github.com/vzhou842/profanity-check:

A fast, robust Python library to check for profanity or offensive language in strings. Read more about how and why profanity-check was built in this blog post.

Our aim is to follow scikit-learn's (main dependency) versions and post models trained with the same version number, example alt-profanity-check version 1.2.3.4 should be trained with the 1.2.3.4 version of the scikit-learn library.

For joblib which is the next major dependency we will be using the latest one which was available when we trained the models.

Last but not least we aim to clean up the codebase a bit and maybe introduce some features or datasets.

Logo from https://www.istockphoto.com/vector/comic-text-phrase-pop-art-wtf-gm846106188-138614723

Changelog

See CHANGELOG.md

How It Works

profanity-check uses a linear SVM model trained on 200k human-labeled samples of clean and profane text strings. Its model is simple but surprisingly effective, meaning profanity-check is both robust and extremely performant.

Why Use profanity-check?

No Explicit Blacklist

Many profanity detection libraries use a hard-coded list of bad words to detect and filter profanity. For example, profanity uses this wordlist, and even better-profanity still uses a wordlist. There are obviously glaring issues with this approach, and, while they might be performant, these libraries are not accurate at all.

A simple example for which profanity-check is better is the phrase

"You cocksucker"* - profanity thinks this is clean because it doesn't have
"cocksucker"* in its wordlist.

Performance

Other libraries like profanity-filter use more sophisticated methods that are much more accurate but at the cost of performance. A benchmark (performed December 2018 on a new 2018 Macbook Pro) using a Kaggle dataset of Wikipedia comments yielded roughly the following results:

Package	1 Prediction (ms)	10 Predictions (ms)	100 Predictions (ms)
profanity-check	0.2	0.5	3.5
profanity-filter	60	1200	13000
profanity	0.3	1.2	24

profanity-check is anywhere from 300 - 4000 times faster than profanity-filter in this benchmark!

Accuracy

This table speaks for itself:

Package	Test Accuracy	Balanced Test Accuracy	Precision	Recall	F1 Score
profanity-check	95.0%	93.0%	86.1%	89.6%	0.88
profanity-filter	91.8%	83.6%	85.4%	70.2%	0.77
profanity	85.6%	65.1%	91.7%	30.8%	0.46

See the How section below for more details on the dataset used for these results.

Installation

$ pip install alt-profanity-check

For older Python versions

Following Scikit-learn, Python3.6 is not supported after its 1.0 version if you are using 3.6 pin alt-profanity-check to 0.24.2.

Usage

You can test from the command line:

profanity_check "Check something" "Check something else"

from profanity_check import predict, predict_prob

predict(['predict() takes an array and returns a 1 for each string if it is offensive, else 0.'])
# [0]

predict(['fuck you'])
# [1]

predict_prob(['predict_prob() takes an array and returns the probability each string is offensive'])
# [0.08686173]

predict_prob(['go to hell, you scum'])
# [0.7618861]

Note that both predict() and predict_prob return numpy arrays.

More on How/Why It Works

How

Special thanks to the authors of the datasets used in this project. profanity-check hence also alt-profanity-check is trained on a combined dataset from 2 sources:

t-davidson/hate-speech-and-offensive-language, used in their paper Automated Hate Speech Detection and the Problem of Offensive Language
the Toxic Comment Classification Challenge on Kaggle.

profanity-check relies heavily on the excellent scikit-learn library. It's mostly powered by scikit-learn classes CountVectorizer, LinearSVC, and CalibratedClassifierCV. It uses a Bag-of-words model to vectorize input strings before feeding them to a linear classifier.

Why

One simplified way you could think about why profanity-check works is this: during the training process, the model learns which words are "bad" and how "bad" they are because those words will appear more often in offensive texts. Thus, it's as if the training process is picking out the "bad" words out of all possible words and using those to make future predictions. This is better than just relying on arbitrary word blacklists chosen by humans!

Caveats

This library is far from perfect. For example, it has a hard time picking up on less common variants of swear words like "f4ck you" or "you b1tch" because they don't appear often enough in the training corpus. Never treat any prediction from this library as unquestionable truth, because it does and will make mistakes. Instead, use this library as a heuristic.

Developer Notes

Create a virtual environment from the project
pip install -r development_requirements.txt

Retraining data

With the above in place:

cd profanity_check/data
python train_model.py

Uploading to PyPi

Currently trying to automate it using Github Actions; see: .github/workflows/package_release_dry_run.yml.

Setup:

Set up your "~/.pypirc" with the appropriate token
pip install -r requirements_for_uploading.txt which installs twine

New Version:

With x.y.z as the version to be uploaded:

First tag:

git tag -a vx.y.z -m "Version x.y.z"
git push --tags

Then upload:

python setup.py sdist
twine upload dist/alt-profanity-check-x.y.z.tar.gz

barbuza/alt-profanity-check