DotnetBadWordDetector 2.0.0

.NET Bad Word Detector and Filter

This is a fast and robust library that detects offensive language within text strings. It currently supports English, Portuguese, and Spanish languages.

How It Works

This library uses a logistic regression ML.NET model trained on thousands of human-labeled words. The trained model is embedded as a resource in this library and is consulted on every prediction.

Why Use This Library?

Unlike other .NET profanity detection libraries that rely solely on static hard-coded lists of bad words — such as ProfanityDetector — which uses this list stored in memory, this library uses a machine learning approach that can detect creative substitutions and misspellings (e.g., "h0us3" instead of "house"). This makes it much harder to bypass.

Performance

In benchmarks, this library was up to 618 times faster than the most downloaded .NET package for detecting profanity. For 100 successive predictions, it was approximately 24 times faster.

PC specs: Dell Inspiron 13, i7 8th gen, 16 GB RAM.

Installation

dotnet add package DotnetBadWordDetector

How to Use

Create the detector

using DotnetBadWordDetector;

// English only (default)
var detector = new ProfanityDetector();

// Or load all supported languages: English, Spanish, and Portuguese
var detectorAll = new ProfanityDetector(allLocales: true);

Check if a word is offensive

if (detector.IsProfane("example")) {
    // Word is classified as offensive
}

Check if a phrase contains any offensive words

if (detector.IsPhraseProfane("this is an example")) {
    // Phrase contains at least one offensive word
}

Get the probability that a word or phrase is offensive

float probWord = detector.GetProfanityProbability("example");
float probPhrase = detector.GetPhraseProfanityProbability("this is an example");

Mask offensive words in a phrase

string cleanText = detector.MaskProfanity("this is an example", '*');

This will replace any detected offensive words with asterisks or your chosen character.

Model Quality

Model quality metrics evaluation
--------------------------------
Accuracy: 98.43%
AUC: 99.49%
F1 Score: 97.25%

Notes

This library is not perfect: it is not 100% accurate and it is context-free — meaning it cannot detect profane phrases made of individually inoffensive words. Definitions of "profanity" can vary by culture. This library uses human-labeled data, which might not align perfectly with all contexts.

Tips

• Keep the detector instance in memory for better performance — it uses very little memory (less than 100 KB).
• Be cautious when enabling all locales together, as it may produce more false positives in multilingual texts.

Contributing

Contributions are welcome! Feel free to open an issue or submit a pull request with suggestions for new features, languages, or improvements.