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.
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.
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 is anywhere from 300 - 4000 times faster than
profanity-filter in this benchmark!
This table speaks for itself:
|Package||Test Accuracy||Balanced Test Accuracy||Precision||Recall||F1 Score|
See the How section below for more details on the dataset used for these results.
$ pip install profanity-check
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.']) #  predict(['fuck you']) #  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
More on How/Why It Works
Special thanks to the authors of the datasets used in this project.
profanity-check was 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
CalibratedClassifierCV. It uses a Bag-of-words model to vectorize input strings before feeding them to a linear classifier.
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!
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.