niacin

A Python library for replacing the missing variation in your text data.

Why should I use this?

Data collected for model training necessarily undersamples the likely variance in the input space. This library is a collection of tools for inserting typical kinds of perturbations to better approximate population variance; and, for creating similar-but-incorrect examples to aid in reducing the total size of the hypothesis space. These are commonly known as ENRICHMENT and NEGATIVE SAMPLING, respectively.

How do I use this?

Functions in niacin are separated into submodules for specific data types. Functions expose a similar API, with two input arguments: the data to be transformed, and the probability of applying a specific transformation.

enrichment:

from niacin.text import en
data = "This is the song that never ends and it goes on and on my friends"
print(en.add_misspelling(data, p=1.0))

This is teh song tath never ends adn it goes on anbd on my firends

negative sampling:

from niacin.text import en
data = "This is the song that never ends and it goes on and on my friends"
print(en.add_hypernyms(data, p=1.0))

This is the musical composition that never extremity and it exit on and on my person

How do I install this?

with pip:

pip install niacin

from source:

git clone git@github.com:deniederhut/niacin.git && cd niacin && python setup.py install

If you have installed niacin from source, you can run the test suite to verify that everything is working properly. We use pytest, which you will first need to install:

pip install pytest

then you can run the library's tests with

pytest -m 'not slow'

if you would like to see the coverage report, you can do so with pytest-cov like so:

pip install pytest-cov
pytest -m 'not slow' --cov=niacin && coverage html

How can I install the optional dependencies?

If you want to use the backtranslate functionality, niacin will need pytorch and some other libraries. These can be installed as extras with:

pip install niacin[backtranslate]

If you are on macos, this might fail with a warning about your version of gcc:

Your compiler (g++) is not compatible with the compiler Pytorch was
built with for this platform, which is clang++ on darwin.

You can avoid this error by executing the following:

CFLAGS='-stdlib=libc++' pip install niacin[backtranslate]