augtxt 0.3.0

Text augmentation.

augtxt – Text Augmentation

Yet another text augmentation python package.

Table of Contents

• Usage

  • `augtxt.augmenters - Pipelines <#pipelines>`__

    • `wordaug - Word Augmentations <#word-augmentations>`__

    • `sentaug - Sentence Augmentations <#sentence-augmentations>`__

  • `augtxt.typo - Typographical Errors <#typographical-errors-tippfehler>`__

  • `augtxt.punct - Interpunctation Errors <#interpunctation-errors-zeichensetzungsfehler>`__

  • `augtxt.wordsubs - Word substitutions <#word-substitutions>`__

• Appendix

  • Installation

  • Commands

  • Support

  • Contributing

Usage

import augtxt
import numpy as np

Pipelines

Word Augmentations

The function augtxt.augmenters.wordaug applies randomly different augmentations to one word. The result is a simulated distribution of possible word augmentations, e.g. how are possible typological errors distributed for a specific original word. The procedure does not guarantee that the original word will be augmented.

from augtxt.augmenters import wordaug
import augtxt.keyboard_layouts as kbl
import numpy as np
from collections import Counter

settings = [
    {
        'p': 0.04,
        'fn': 'typo.drop_n_next_twice',
        'args': {'loc': ['m', 'e'], 'keep_case': True}
    },
    {
        'p': 0.04,
        'fn': 'typo.swap_consecutive',
        'args': {'loc': ['m', 'e'], 'keep_case': True}
    },
    {
        'p': 0.02,
        'fn': 'typo.pressed_twice',
        'args': {'loc': 'u', 'keep_case': True}
    },
    {
        'p': 0.02,
        'fn': 'typo.drop_char',
        'args': {'loc': ['m', 'e'], 'keep_case': True}
    },
    {
        'p': 0.02,
        'fn': 'typo.pressed_shiftalt',
        'args': {'loc': ['b', 'm'], 'keymap': kbl.macbook_us, 'trans': kbl.keyboard_transprob}
    },
]

np.random.seed(seed=42)
word = "Blume"
newwords = []
for i in range(1000):
    newwords.append( wordaug(word, settings) )

Counter(newwords)

Sentence Augmentations

The function augtxt.augmenters.sentaug applies randomly different augmentations to a) at least one word in a sentence, or b) not more than a certain percentage of words in a sentence. The procedure guarantees that the sentence is augmented.

The functions also allows to exclude specific strings from augmentation (e.g. exclude=("[MASK]", "[UNK]")). However, these strings cannot include the special characters .,;:!? (incl. whitespace).

from augtxt.augmenters import wordaug
import augtxt.keyboard_layouts as kbl
import numpy as np

settings = [
    {
        'weight': 2, 'fn': 'typo.drop_n_next_twice',
        'args': {'loc': 'u', 'keep_case': True}
    },
    {
        'weight': 2, 'fn': 'typo.swap_consecutive',
        'args': {'loc': 'u', 'keep_case': True}},
    {
        'weight': 1, 'fn': 'typo.pressed_twice',
        'args': {'loc': 'u', 'keep_case': True}
    },
    {
        'weight': 1, 'fn': 'typo.drop_char',
        'args': {'loc': 'u', 'keep_case': True}
    },
    {
        'weight': 1, 'fn': 'typo.pressed_shiftalt',
        'args': {'loc': ['b', 'm']}
    },
]

np.random.seed(seed=42)
exclude = ["[MASK]", "[UNK]"]
sentence = 'Die Lehrerin [MASK] einen Roman.'
augmentations = sentaug(sentence, settings=settings, exclude=exclude, num_augmentations=10, pmax=0.1)
assert len(augmentations) == 10

Typographical Errors (Tippfehler)

The augtxt.typo module is about augmenting characters to mimic human errors while using a keyboard device.

Swap two consecutive characters (Vertauscher)

A user mix two consecutive characters up.

• Swap 1st and 2nd characters: augtxt.typo.swap_consecutive("Kinder", loc=0) (Result: iKnder)

• Swap 1st and 2nd characters, and enforce letter cases: augtxt.typo.swap_consecutive("Kinder", loc=0, keep_case=True) (Result: Iknder)

• Swap random i-th and i+1-th characters that are more likely at the end of the word: np.random.seed(seed=123); augtxt.typo.swap_consecutive("Kinder", loc='end')

Add double letter (Einfüger)

User presses a key twice accidentaly

• Make 5th letter a double letter: `augtxt.typo.pressed_twice("Eltern", loc=4) (Result: Elterrn)

Drop character (Auslasser)

User presses the key not enough (Lisbach, 2011, p.72), the key is broken, finger motion fails.

• Drop the 3rd letter: augtxt.typo.drop_char("Straße", loc=2) (Result: Staße)

Drop character followed by double letter (Vertipper)

Letter is left out, but the following letter is typed twice. It’s a combination of augtxt.typo.pressed_twice and augtxt.typo.drop_char.

from augtxt.typo import drop_n_next_twice
augm = drop_n_next_twice("Tante", loc=2)
# Tatte

Pressed SHIFT, ALT, or SHIFT+ALT

Usually SHFIT is used to type a capital letter, and ALT or ALT+SHIFT for less common characters. A typo might occur because these special keys are nor are not pressed in combination with a normal key. The function augtxt.typo.pressed_shiftalt such errors randomly.

from augtxt.typo import pressed_shiftalt
augm = pressed_shiftalt("Onkel", loc=2)
# OnKel, On˚el, Onel

The keymap can differ depending on the language and the keyboard layout.

from augtxt.typo import pressed_shiftalt
import augtxt.keyboard_layouts as kbl
augm = pressed_shiftalt("Onkel", loc=2, keymap=kbl.macbook_us)
# OnKel, On˚el, Onel

Further, transition probabilities in case of a typo can be specified

from augtxt.typo import pressed_shiftalt
import augtxt.keyboard_layouts as kbl

keyboard_transprob = {
    "keys": [.0, .75, .2, .05],
    "shift": [.9, 0, .05, .05],
    "alt": [.9, .05, .0, .05],
    "shift+alt": [.3, .35, .35, .0]
}

augm = pressed_shiftalt("Onkel", loc=2, keymap=kbl.macbook_us, trans=keyboard_transprob)

Interpunctation Errors (Zeichensetzungsfehler)

Remove PUNCT and COMMA tokens

The PUNCT (.?!;:) and COMMA (,) tokens carry syntatic information. An use case

import augtxt.punct
text = ("Die Lehrerin [MASK] einen Roman. "
        "Die Schülerin [MASK] ein Aufsatz, der sehr [MASK] war.")
augmented = augtxt.punct.remove_syntaxinfo(text)
# 'Die Lehrerin [MASK] einen Roman Die Schülerin [MASK] ein Aufsatz der sehr [MASK] war'

Merge two consequitive words

The function augtxt.punct.merge_words removes randomly whitespace or hyphens between words, and transform the second word to lower case.

import augtxt.punct

text = "Die Bindestrich-Wörter sind da."

np.random.seed(seed=23)
augmented = augtxt.punct.merge_words(text, num_aug=1)
assert augmented == 'Die Bindestrich-Wörter sindda.'

np.random.seed(seed=1)
augmented = augtxt.punct.merge_words(text, num_aug=1)
assert augmented == 'Die Bindestrichwörter sind da.'

Word substitutions

The augtxt.wordsubs module is about replacing specific strings, e.g. words, morphemes, named entities, abbreviations, etc.

Pseudo-synonyms from pretrained word embedding

The semantic similarity between two words can be measured with a similarity metric (e.g. jaccard score, cosine similarity) between the corresponding word vectors from pretrained word embeddings (e.g. word2vec, GloVe, and fastText).

Furthermore, we compute the character-level (non-semantically) k-shingle based jaccard similarity to exclude near duplicates, or resp. to favor semantic similar words with a different spelling.

fastText

1. Download a language-specifc pretrained fastText embedding, e.g.

augtxt_downloader.py --fasttext --lang=de

2. Tokenize the whole corpus, and create a list of unique words, e.g.

import itertools
token_seqs = [["Das", "ist", "ein", "Satz", "."], ["Dies", "ist", "ein", "anderer", "Satz", "."]]
vocab = set(itertools.chain(*token_seqs))
# {'anderer', 'Satz', '.', 'Das', 'ein', 'Dies', 'ist'}

3. Lookup up synonyms. Make sure that the lang parameter corresponds to the --lang parameter in step (1).

import augtxt.wordsubs

synonyms = augtxt.wordsubs.pseudo_synonyms_fasttext(
    vocab, lang='de',
    max_neighbors=25,
    min_vector_score=0.65,  # Jaccard similarity btw. fastText vectors
    max_shingle_score=0.35,  # Jaccard similarity btw. k-shingles
    kmax=8,  # the k in k-shingle
    n_max_wildcards=1  # number of wildcards in each shingle
)

We prefer the term pseudo-synonyms because the results can be considered to be inaccurate. For example, #einleitungssatz was identified as similar to the token satz. It contains a hashtag # what can be considered as inaccurate.

{
    'anderer': ['verschiedener', 'einiger', 'vieler', 'diverser', 'sonstiger',
                'etlicher', 'einzelner', 'bestimmter', 'ähnlicher'],
    'satz': ['sätze', 'anfangssatz', 'schlussatz', 'eingangssatz', 'einleitungssatzes',
             'einleitungsssatz', 'einleitungssatz', 'behauptungssatz', 'beispielsatz',
             'schlusssatz', 'anfangssatzes', 'einzelsatz', '#einleitungssatz',
             'minimalsatz', 'inhaltssatz', 'aufforderungssatz', 'ausgangssatz'],
    '.': [',', '🎅'],
    'das': ['welches', 'solches'],
    'ein': ['weiteres'],
    'dies': ['was', 'umstand', 'dass']
}

The function augtxt.wordsubs.pseudo_synonyms_fasttext stores all pseudo-synonyms inside a buffer file on your HDD in the background. In order to call this buffer file directly, and to avoid preprocessing with fastText again, you can use the following function:

import itertools
import augtxt.wordsubs
token_seqs = [["Das", "ist", "ein", "Satz", "."], ["Dies", "ist", "ein", "anderer", "Satz", "."]]
vocab = set(itertools.chain(*token_seqs))
synonyms = augtxt.wordsubs.lookup_buffer_fasttext(vocab, lang='de')

Please note: When using the function augtxt.wordsubs.pseudo_synonyms_fasttext with fastText pretrained models, then (1) you have to cite Bojanowski et. al. (2017), and (2) the subsequently derived data, e.g. the augmented examples, fall under the CC BY-SA 3.0 license.

Using pseudo-synonym dictionaries to augment tokenized sequences

It is recommend to filter vocab further. For example, PoS tag the sequences and only augment VERB and NOUN tokens.

import itertools
import augtxt.wordsubs
import numpy as np

original_seqs = [["Das", "ist", "ein", "Satz", "."], ["Dies", "ist", "ein", "anderer", "Satz", "."]]
vocab = set([s.lower() for s in itertools.chain(*original_seqs) if len(s) > 1])

synonyms = augtxt.wordsubs.lookup_buffer_fasttext(
    vocab, lang='de')

np.random.seed(42)
augmented_seqs = augtxt.wordsubs.synonym_replacement(
    original_seqs, synonyms, num_augm=10, keep_case=True)

# check results for 1st sentence
for s in augmented_seqs[0]:
    print(s)

References

• Lisbach, B., 2011. Linguistisches Identity Matching. Vieweg+Teubner, Wiesbaden. https://doi.org/10.1007/978-3-8348-9791-6

• Bojanowski, P., Grave, E., Joulin, A., Mikolov, T., 2017. Enriching Word Vectors with Subword Information. arXiv:1607.04606 [cs].

Appendix

Installation

The augtxt git repo is available as PyPi package

pip install augtxt>=0.2.4
pip install git+ssh://[e-mail]/ulf1/augtxt.git

Commands

Install a virtual environment

python3 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt
pip install -r requirements-dev.txt
pip install -r requirements-demo.txt

(If your git repo is stored in a folder with whitespaces, then don’t use the subfolder .venv. Use an absolute path without whitespaces.)

Python commands

• Check syntax: flake8 --ignore=F401 --exclude=$(grep -v '^#' .gitignore | xargs | sed -e 's/ /,/g')

• Run Unit Tests: pytest

Publish

pandoc README.md --from markdown --to rst -s -o README.rst
python setup.py sdist
twine upload -r pypi dist/*

Clean up

find . -type f -name "*.pyc" | xargs rm
find . -type d -name "__pycache__" | xargs rm -r
rm -r .pytest_cache
rm -r .venv

Support

Please open an issue for support.

Contributing

Please contribute using Github Flow. Create a branch, add commits, and open a pull request.