TextAugment: Improving Short Text Classification through Global Augmentation Methods

You have just found TextAugment.

TextAugment is a Python 3 library for augmenting text for natural language processing applications. TextAugment stands on the giant shoulders of NLTK, Gensim v3.x, and TextBlob and plays nicely with them.

Acknowledgements

Cite this paper when using this library. Arxiv Version

@inproceedings{marivate2020improving,
  title={Improving short text classification through global augmentation methods},
  author={Marivate, Vukosi and Sefara, Tshephisho},
  booktitle={International Cross-Domain Conference for Machine Learning and Knowledge Extraction},
  pages={385--399},
  year={2020},
  organization={Springer}
}

Table of Contents

• Features
• Citation Paper
  • Requirements
  • Installation
  • How to use
    • Word2vec-based augmentation
    • WordNet-based augmentation
    • RTT-based augmentation
• Easy data augmentation (EDA)
• An easier data augmentation (AEDA)
• Mixup augmentation
  • Implementation
• Acknowledgements

Features

• Generate synthetic data for improving model performance without manual effort
• Simple, lightweight, easy-to-use library.
• Plug and play to any machine learning frameworks (e.g. PyTorch, TensorFlow, Scikit-learn)
• Support textual data

Citation Paper

Improving short text classification through global augmentation methods.

Requirements

• Python 3

The following software packages are dependencies and will be installed automatically.

$ pip install numpy nltk gensim==3.8.3 textblob googletrans 

The following code downloads NLTK corpus for wordnet.

nltk.download('wordnet')

The following code downloads NLTK tokenizer. This tokenizer divides a text into a list of sentences by using an unsupervised algorithm to build a model for abbreviation words, collocations, and words that start sentences.

nltk.download('punkt')

The following code downloads default NLTK part-of-speech tagger model. A part-of-speech tagger processes a sequence of words, and attaches a part of speech tag to each word.

nltk.download('averaged_perceptron_tagger')

Use gensim to load a pre-trained word2vec model. Like Google News from Google drive.

import gensim
model = gensim.models.KeyedVectors.load_word2vec_format('./GoogleNews-vectors-negative300.bin', binary=True)

You can also use gensim to load Facebook's Fasttext English and Multilingual models

import gensim
model = gensim.models.fasttext.load_facebook_model('./cc.en.300.bin.gz')

Or training one from scratch using your data or the following public dataset:

• Text8 Wiki

• Dataset from "One Billion Word Language Modeling Benchmark"

Installation

Install from pip [Recommended]

$ pip install textaugment
or install latest release
$ pip install git+git@github.com:dsfsi/textaugment.git

Install from source

$ git clone git@github.com:dsfsi/textaugment.git
$ cd textaugment
$ python setup.py install

How to use

There are three types of augmentations which can be used:

• word2vec

from textaugment import Word2vec

• fasttext

from textaugment import Fasttext

• wordnet

from textaugment import Wordnet

• translate (This will require internet access)

from textaugment import Translate

Fasttext/Word2vec-based augmentation

See this notebook for an example

Basic example

>>> from textaugment import Word2vec, Fasttext
>>> t = Word2vec(model='path/to/gensim/model'or 'gensim model itself')
>>> t.augment('The stories are good')
The films are good
>>> t = Fasttext(model='path/to/gensim/model'or 'gensim model itself')
>>> t.augment('The stories are good')
The films are good

Advanced example

>>> runs = 1 # By default.
>>> v = False # verbose mode to replace all the words. If enabled runs is not effective. Used in this paper (https://www.cs.cmu.edu/~diyiy/docs/emnlp_wang_2015.pdf)
>>> p = 0.5 # The probability of success of an individual trial. (0.1<p<1.0), default is 0.5. Used by Geometric distribution to selects words from a sentence.

>>> word = Word2vec(model='path/to/gensim/model'or'gensim model itself', runs=5, v=False, p=0.5)
>>> word.augment('The stories are good', top_n=10)
The movies are excellent
>>> fast = Fasttext(model='path/to/gensim/model'or'gensim model itself', runs=5, v=False, p=0.5)
>>> fast.augment('The stories are good', top_n=10)
The movies are excellent

WordNet-based augmentation

Basic example

>>> import nltk
>>> nltk.download('punkt')
>>> nltk.download('wordnet')
>>> from textaugment import Wordnet
>>> t = Wordnet()
>>> t.augment('In the afternoon, John is going to town')
In the afternoon, John is walking to town

Advanced example

>>> v = True # enable verbs augmentation. By default is True.
>>> n = False # enable nouns augmentation. By default is False.
>>> runs = 1 # number of times to augment a sentence. By default is 1.
>>> p = 0.5 # The probability of success of an individual trial. (0.1<p<1.0), default is 0.5. Used by Geometric distribution to selects words from a sentence.

>>> t = Wordnet(v=False ,n=True, p=0.5)
>>> t.augment('In the afternoon, John is going to town', top_n=10)
In the afternoon, Joseph is going to town.

RTT-based augmentation

Example

>>> src = "en" # source language of the sentence
>>> to = "fr" # target language
>>> from textaugment import Translate
>>> t = Translate(src="en", to="fr")
>>> t.augment('In the afternoon, John is going to town')
In the afternoon John goes to town

EDA: Easy data augmentation techniques for boosting performance on text classification tasks

This is the implementation of EDA by Jason Wei and Kai Zou.

https://www.aclweb.org/anthology/D19-1670.pdf

See this notebook for an example

Synonym Replacement

Randomly choose n words from the sentence that are not stop words. Replace each of these words with one of its synonyms chosen at random.

Basic example

>>> from textaugment import EDA
>>> t = EDA()
>>> t.synonym_replacement("John is going to town", top_n=10)
John is give out to town

Random Deletion

Randomly remove each word in the sentence with probability p.

Basic example

>>> from textaugment import EDA
>>> t = EDA()
>>> t.random_deletion("John is going to town", p=0.2)
is going to town

Random Swap

Randomly choose two words in the sentence and swap their positions. Do this n times.

Basic example

>>> from textaugment import EDA
>>> t = EDA()
>>> t.random_swap("John is going to town")
John town going to is

Random Insertion

Find a random synonym of a random word in the sentence that is not a stop word. Insert that synonym into a random position in the sentence. Do this n times

Basic example

>>> from textaugment import EDA
>>> t = EDA()
>>> t.random_insertion("John is going to town")
John is going to make up town

AEDA: An easier data augmentation technique for text classification

This is the implementation of AEDA by Karimi et al, a variant of EDA. It is based on the random insertion of punctuation marks.

https://aclanthology.org/2021.findings-emnlp.234.pdf

Implementation

See this notebook for an example

Random Insertion of Punctuation Marks

Basic example

>>> from textaugment import AEDA
>>> t = AEDA()
>>> t.punct_insertion("John is going to town")
! John is going to town

Mixup augmentation

This is the implementation of mixup augmentation by Hongyi Zhang, Moustapha Cisse, Yann Dauphin, David Lopez-Paz adapted to NLP.

Used in Augmenting Data with Mixup for Sentence Classification: An Empirical Study.

Mixup is a generic and straightforward data augmentation principle. In essence, mixup trains a neural network on convex combinations of pairs of examples and their labels. By doing so, mixup regularises the neural network to favour simple linear behaviour in-between training examples.

Implementation

See this notebook for an example

Built with ❤ on

• Python

Authors

• Joseph Sefara (http://www.speechtech.co.za)
• Vukosi Marivate (http://www.vima.co.za)

Acknowledgements

Cite this paper when using this library. Arxiv Version

@inproceedings{marivate2020improving,
  title={Improving short text classification through global augmentation methods},
  author={Marivate, Vukosi and Sefara, Tshephisho},
  booktitle={International Cross-Domain Conference for Machine Learning and Knowledge Extraction},
  pages={385--399},
  year={2020},
  organization={Springer}
}

Licence

MIT licensed. See the bundled LICENCE file for more details.