sensitive-words-filter

A trie-based sensitive-word filter for Node.js. It scans text against a keyword dictionary and masks matches, while also detecting common obfuscation tricks such as inserting spaces, symbols, digits, or letters between the characters of a keyword.

Features

• Trie-based engine — keywords are compiled into a prefix tree; scanning starts from every character with per-match filler skipping
• Obfuscation detection — matches keywords even when filler characters are inserted between keyword characters
• Longest-match — when keywords overlap (e.g. AB and ABC), the longer match wins
• Case-insensitive — keywords and input text are matched case-insensitively
• Flexible keyword loading — pass an array, a single file, or a directory of keyword files
• Singleton or instances — WordFilter.instance() for a shared instance, or new WordFilter() for independent dictionaries
• Word-boundary mode for pure-English text — optional word-segmentation mode for English-only environments; restricts ASCII keywords to word-boundary matches (opt-in, off by default)
• Per-keyword lenient mode — tag a keyword as { word: 'drug', lenient: true } to let that word tolerate letter/digit fillers (d1rug matches drug); every other keyword stays strict

Install

npm install sensitive-words-dfa-filter --save

TypeScript users get full type hints out of the box — the package ships with type declarations.

Quick start

const WordFilter = require('sensitive-words-dfa-filter');

const filter = WordFilter.instance();

// Load keywords from an array (or a file / directory, see below)
filter.init(['AB', 'ABC', '治国']);

// Basic matching
console.log(filter.replace('This is "AB｜A B｜AAB｜A1B｜A@B" filter word!'));
// This is "**｜**｜A**｜A1B｜**" filter word!

console.log(filter.replace('This is "ABC｜A B C｜A1B1C｜A@B@C" filter word!'));
// This is "***｜***｜A1B1C｜***" filter word!

// Custom replacement character
console.log(filter.replace('This is "治国｜治 国｜治A国｜治1国｜治@国" filter word!', '?'));
// This is "??｜??｜??｜??｜??" filter word!

// Matching modes
// Word-boundary mode: for pure-English environments — ASCII keywords match
// only at word boundaries (admin no longer matches administrator).
filter.init(['admin'], { wordBoundary: true });
console.log(filter.replace('administrator admin admin123'));
// administrator ***** admin123

// Both knobs compose in one call: wordBoundary is global, lenient is per
// keyword — d1rug matches drug (up to 3 letter/digit fillers), while
// drugabc does not match drug anymore because of the word boundary.
filter.init(['admin', { word: 'drug', lenient: true }], { wordBoundary: true });
console.log(filter.replace('d1rug admin drugabc'));
// **** ***** drugabc

Loading keywords

init() accepts either an array of keywords or a path:

Input	Example	Notes
Array	filter.init(['赌博', '诈骗'])	Simple in-code keyword list
File	filter.init('/path/to/keywords.txt')	Reads the file; keywords separated by 、 or newlines
Directory	filter.init('/path/to/keywords')	Reads every file inside the directory, including subdirectories

Example keyword file:

赌博、诈骗、禁言、敏感词
spam
cheat

Empty entries and surrounding whitespace are ignored.

API

WordFilter.instance()

Returns the shared singleton instance.

new WordFilter()

Creates an independent filter with its own dictionary — useful when different parts of an application need different keyword sets.

init(keywords: (string | { word: string; lenient?: boolean })[], options?: { wordBoundary?: boolean }): void

Synchronously builds the matching dictionary.

• keywords — an array of keyword strings or entries ({ word, lenient }), or a path to a keyword file/directory.
• lenient (per entry) — when true, this keyword tolerates up to 3 letters/digits as fillers even in pure-ASCII matches (d1rug matches drug, d1r2u3g too; d11r22u33g does not). The cap prevents the keyword's first letter from matching across unrelated words (e.g. admin drug does not match drug from the d of admin). Off by default; note this also means A1B matches AB if AB is tagged lenient.
• options.wordBoundary — a word-segmentation mode intended for pure-English environments. When true, pure-ASCII keywords only match at word boundaries: the character immediately before and after the keyword must not be a letter, digit, or underscore (same definition as \b), so admin no longer matches inside administrator. Off by default; CJK keywords are not affected. Composes with per-keyword lenient entries in the same call.
• Calling init() again replaces the existing dictionary.

replace(searchValue: string, replaceValue?: string): string

Scans searchValue and replaces every matched keyword with replaceValue repeated once per matched keyword character (default *).

• Returns searchValue unchanged if the filter has not been initialized.

getCacheStats(): { size: number; entries: string[] }

Returns the size of the preloaded "skippable-character" cache and the cached characters (useful for debugging and performance monitoring).

clearCache(): void

Fully resets the filter: clears the preloaded character cache, the keyword dictionary, and the initialization state. Call init() again before using the filter.

Matching rules

• Case-insensitive: Text matches keyword text.
• Longest match wins: with keywords AB and ABC, input ABC is masked as ***.
• Filler characters can be skipped between keyword characters:

Filler type	Skipped by default?	Example
Symbols (@, #, ., ~, ...)	Always	A@B matches AB
Spaces / whitespace	Always	A B matches AB
CJK radicals	Always	中灬国 matches 中国
Full-width spaces / punctuation (　, 、, …, ...)	Always	治　国 matches 治国
Full-width letters / digits (Ａ-Ｚ, ａ-ｚ, ０-９)	Always, as fillers (not normalized)	治Ａ国 matches 治国; ＡＢ does not match AB
Digits (0-9)	Only when the matched part already contains non-ASCII characters	治1国 matches 治国; A1B does not match AB
Letters (a-z, A-Z)	Only when the matched part already contains non-ASCII characters	治A国 matches 治国; AAB becomes A**

The non-ASCII rule prevents false positives in pure English text: with keyword text, the input This is text only masks text itself — the leading words are left untouched.

• English keywords match as plain substrings by default — admin matches inside administrator, and sex inside sexy. For pure-English environments, pass { wordBoundary: true } to init() to switch to word-segmentation mode: ASCII keywords only match at word boundaries (e.g. init(['admin'], { wordBoundary: true }) masks the admin but leaves administrator and admin123 untouched). CJK keywords are not affected; ASCII keywords are boundary-checked regardless of the surrounding text language.
• Per-keyword lenient mode — tag a keyword as { word: 'drug', lenient: true } to allow up to 3 letter/digit fillers for that keyword even when the match is pure ASCII (d1rug matches drug). The cap stops cross-word matches (admin drug does not match drug via the d of admin). This is opt-in because it is a precision tradeoff: if AB is tagged lenient, A1B matches AB too.
• Single-character ASCII keywords are ignored (e.g. a) to avoid over-matching; single-character CJK keywords (e.g. 赌) are supported.
• Overlapping matches are merged into a single masked region: with keywords ABC and BCD, the input A B C D is masked as ****.

Performance

Measured locally (Node.js 24, v2.1.0) with an 11-keyword dictionary and ~1MB inputs:

Input	Time	Throughput
Chinese text	~540 ms	~1.9 MB/s
English text	~190 ms	~5.2 MB/s
Mixed Chinese/English	~370 ms	~2.7 MB/s

Numbers vary by machine, dictionary size, and hit rate — run npm run bench on your own hardware. The engine is a trie with per-position scanning (not an Aho-Corasick automaton), which is plenty for interactive filtering but not for multi-GB/s pipelines.

A pathological case — a single 500-character keyword over 20KB of repetitive text — takes ~0.4s, since nearly every position walks the whole keyword path.

Limitations

• Substring matching by default — English keywords match inside longer words (admin matches administrator). In pure-English environments, enable { wordBoundary: true } (see Matching rules) to switch to word-segmentation mode.
• No full-width normalization — full-width ＡＢ does not match keyword AB; full-width letters/digits are only treated as skippable fillers between keyword characters.
• Curated obfuscation coverage — the skippable-filler set is a fixed list (symbols, whitespace, CJK radicals, full-width characters). Exotic tricks such as emoji, zero-width characters, or homoglyphs are not covered.
• Memory — each replace() call allocates a small amount of memory proportional to the input length (coverage/star arrays); fine for typical messages, worth noting for very large documents.
• Pathological inputs — a very long keyword over highly repetitive text can approach O(n × keyword length).

Development

npm install
npm test        # compile and run all test suites
npm run build   # compile src/index.ts to build/index.js
npm run bench   # local performance benchmark

Test suites:

• tests/WordsFilter.test.ts — 100 fixture-based integration cases (fails the run on any failure)
• tests/WordsFilter-New.test.ts — 60 additional fixture-based cases (fails the run on any failure)
• tests/unit.test.ts — 33 focused unit tests with assertions

License

MIT