vtext

NLP in Rust with Python bindings

This package aims to provide a high performance toolkit for ingesting textual data for machine learning applications.

Features

• Tokenization: Regexp tokenizer, Unicode segmentation + language specific rules
• Stemming: Snowball (in Python 15-20x faster than NLTK)
• Token counting: converting token counts to sparse matrices for use in machine learning libraries. Similar to CountVectorizer and HashingVectorizer in scikit-learn but will less broad functionality.
• Levenshtein edit distance; Sørensen-Dice, Jaro, Jaro Winkler string similarities

Usage

Usage in Python

vtext requires Python 3.6+ and can be installed with,

pip install vtext

Below is a simple tokenization example,

>>> from vtext.tokenize import VTextTokenizer
>>> VTextTokenizer("en").tokenize("Flights can't depart after 2:00 pm.")
["Flights", "ca", "n't", "depart" "after", "2:00", "pm", "."]

For more details see the project documentation: vtext.io/doc/latest/index.html

Usage in Rust

Add the following to Cargo.toml,

[dependencies]
vtext = "0.2.0"

For more details see rust documentation: docs.rs/vtext

Benchmarks

Tokenization

Following benchmarks illustrate the tokenization accuracy (F1 score) on UD treebanks ,

and the English tokenization speed,

Text vectorization

Below are benchmarks for converting textual data to a sparse document-term matrix using the 20 newsgroups dataset, run on Intel(R) Xeon(R) CPU E3-1270 v6 @ 3.80GHz,

Note however that these two estimators in vtext currently support only a fraction of scikit-learn's functionality. See benchmarks/README.md for more details.

License

vtext is released under the Apache License, Version 2.0.