rs-natural

rs-natural (the natural crate) is a natural language processing library for Rust that bundles a set of classic text-analysis building blocks into one dependency. It provides string-similarity metrics, tokenization, n-gram generation, phonetic matching, and simple statistical classifiers so you can add language features to a Rust program without wiring together several separate crates.

The library covers Jaro-Winkler and Levenshtein distance, word and sentence tokenizing, padded and unpadded n-grams, Soundex phonetics, naive Bayes classification, and TF-IDF scoring, with optional Serde support for serializing trained classifiers. It is a pragmatic, approachable toolkit aimed at everyday text-processing tasks rather than deep-learning-scale NLP.

What You Get

• String-distance functions for Jaro-Winkler and Levenshtein similarity
• Tokenizers for splitting text into words and sentences
• N-gram generation with optional padding
• Soundex phonetic matching plus naive Bayes and TF-IDF classifiers with Serde serialization

Common Use Cases

• Fuzzy string matching and typo-tolerant lookups
• Categorizing short text with a naive Bayes classifier
• Ranking documents or terms by TF-IDF relevance

Under The Hood

Architecture - The crate is organized one concern per module (src/distance.rs, tokenize.rs, ngram.rs, phonetics.rs, classifier.rs, tf_idf.rs) re-exported through lib.rs, so each feature is an independent set of free functions or small structs (for example SoundexWord, the naive Bayes classifier, and the TF-IDF model) that operate directly on &str and Vec inputs. There is no central engine or shared state; callers pull in only the modules they need.

Tech Stack - Pure Rust with a minimal dependency footprint; the main optional dependency is Serde (behind a serde_support feature) to serialize trained naive Bayes and TF-IDF models. It builds as a standard Cargo crate with no build scripts or native code.

Code Quality - Modules include unit tests for the distance, tokenizer, and classifier logic, and the code is small and readable. The authors openly flag it as experimental and unoptimized (notably the naive Bayes path), and it does not promise backward compatibility, so it favors clarity over performance tuning.

API Design - The surface is beginner-friendly: import a function such as levenshtein_distance or soundex and call it, or construct a classifier struct and train it. Naming maps directly to well-known algorithms, and the README shows copy-paste examples for each module, keeping the ramp-up short.