Getting Started

Augmenty is an augmentation library based on spaCy for augmenting texts. Augmenty differs from other augmentation libraries in that it corrects (as far as possible) the token, sentence and document labels under the augmentation.

Installation

Before we get ahead of ourselves let us just install the required packages:

pip install augmenty
# install the spacy pipeline
python -m spacy download en_core_web_md

Introduction

Augmenty is an augmentation library for spaCy, consisting of many different augmenters. To get an idea of all the available augmenters you can always try out the use:

import augmenty

augmenters = augmenty.augmenters()

for augmenter in augmenters:
    print(augmenter)

spacy.combined_augmenter.v1
spacy.orth_variants.v1
spacy.lower_case.v1
random_casing_v1
char_replace_random_v1
char_replace_v1
keystroke_error_v1
remove_spacing_v1
char_swap_v1
upper_case_v1
spongebob_v1
paragraph_subset_augmenter_v1
random_starting_case_v1
conditional_token_casing_v1
token_insert_v1
token_insert_random_v1
duplicate_token_v1
random_synonym_insertion_v1
token_dict_replace_v1
wordnet_synonym_v1
token_replace_v1
word_embedding_v1
letter_spacing_augmenter_v1
spacing_insertion_v1
token_swap_v1
da_æøå_replace_v1
da_historical_noun_casing_v1
ents_replace_v1
per_replace_v1
ents_format_v1

To get get more information about an individual augmenter you can always simply use help for instance let is say you want to know more about the upper case augmenter you could run: help(augmenters["upper_case_v1"]).

After you have an idea about the augmenter you wish to use loading in augmenters in augmenty is made easy using the load command given and given the arguments:

upper_case_augmenter = augmenty.load("upper_case_v1", level=1.00)  # 100% uppercase

Applying the augmentation

Augmenters in augmenty always take in a spaCy Language pipeline and an spaCy Example so that it can be easily used while training workflows, however, augmenty also allows for easy application of augmenters to raw text and spaCy Docs.

Why examples and not just raw text?

A spaCy example consist of two documents, the labelled document, containing all the correct labels including document classification such as whether a tweet is positive or negative and token classiification such as Part-of-speech-tags and named entities. When augmenting the Example augmenty seeks to correct these tags in accordance with the augmentation. As the raw text does not include these labels it as naturally not possible. For instance if I was to swap two tokens I would want to swap their corresponding labels as well. When swapping tokens augmenty even respect entities and sentences as to not split an entity or swap tokens across sentence borders. You can naturally turn this of if you wish to.

Applying augmentations on Docs

import spacy

nlp = spacy.load("en_core_web_lg")
docs = nlp.pipe(
    [
        "Augmentation is a wonderful tool for obtaining higher performance on limited data.",
        "You can also use it to see how robust your model is to changes.",
    ]
)

augmented_docs = augmenty.docs(docs, augmenter=upper_case_augmenter, nlp=nlp)

for doc in augmented_docs:
    print(doc)

AUGMENTATION IS A WONDERFUL TOOL FOR OBTAINING HIGHER PERFORMANCE ON LIMITED DATA.
YOU CAN ALSO USE IT TO SEE HOW ROBUST YOUR MODEL IS TO CHANGES.

Applying augmentations on text

We can also try it out on text. Let us also try out a new augmenter for replacing entities. Remember you can always use help(augmenters["ents_replace.v1"]) to figure out which inputs the augmenter takes and see and example.

texts = [
    "Augmenty is a wonderful tool for augmentation which is made using SpaCy developed by Explosion."
]

ent_augmenter = augmenty.load(
    "ents_replace_v1", level=1.00, ent_dict={"ORG": [["The SpaCy Universe"]]}
)

augmented_texts = augmenty.texts(texts, augmenter=ent_augmenter, nlp=nlp)

for text in augmented_texts:
    print(text)

The SpaCy Universe is a wonderful tool for augmentation which is made using The SpaCy Universe developed by Explosion.

Customizing augmenters

Augmenty is more than a list of augmenters and also contains utilities for dealing with augmenters such as combining and moderating augmenters.

Combining augmenters

We can start of by combing the entity augmenter with an augmenter which replaces words with their synonym based on wordnet.

synonym_augmenter = augmenty.load("wordnet_synonym_v1", level=1, lang="en")

combined_aug = augmenty.combine([ent_augmenter, synonym_augmenter])

augmented_texts = augmenty.texts(texts, augmenter=combined_aug, nlp=nlp)

for text in augmented_texts:
    print(text)

The SpaCy Universe is a grand putz for augmentation which is ca-ca expend The SpaCy Universe educate by Detonation.

Moderating Augmenters

Certain augmenters apply augmentation at different levels. For instance the augmenter keystroke_error.v1 augments examples based on keyboard distances, where each character has a chance to be replaced with a neightbouring character. However, we might wish to apply this augmentation to 5% of characters, but only apply it 50% of the training samples. Using augmenty.set_doc_level we can add this last part to any augmenter, thus allowing for more flexibility when using the model.

keystroke_augmenter = augmenty.load(
    "keystroke_error_v1", keyboard="en_qwerty_v1", level=0.05
)  # 5% if characters

keystroke_augmenter = augmenty.set_doc_level(
    keystroke_augmenter, level=0.5
)  # 50% of texts

texts = [
    "Augmenty is a wonderful tool for augmentation.",
    "Augmentation is a wonderful tool for obtaining higher performance on limited data.",
    "You can also use it to see how robust your model is to changes.",
]

augmented_texts = augmenty.texts(texts, augmenter=keystroke_augmenter, nlp=nlp)

for text in augmented_texts:
    print(text)

Auvnenty is a wondereul tool for augmentation.
Augmentation is a wonderful tool for obtaining higher performance on limited data.
You can also use it to see how robust your model 6s to changex.

Similarly one might wish the augment to instead of simply yielding the augmented example also yield the original, such that the trained model always see the actual data.

token_swap_augmenter = augmenty.load("token_swap_v1", level=0.20)
token_swap_augmenter = augmenty.yield_original(
    token_swap_augmenter
)  # yield both the augmented and original example

augmented_texts = augmenty.texts(texts, augmenter=token_swap_augmenter, nlp=nlp)

for text in augmented_texts:
    print(text)

Augmenty a is wonderful tool for augmentation.
Augmenty is a wonderful tool for augmentation.
Augmentation is a wonderful tool for obtaining performance higher on limited data.
Augmentation is a wonderful tool for obtaining higher performance on limited data.
You can use also it to see how robust your is to model changes.
You can also use it to see how robust your model is to changes.

Applying augmentation to Examples or a Corpus

Examples consists of two docs, one containing the predictions of the model, the other containing the gold labelled document. For this example we will load the DaNE dataset. DaNE includes the Danish dependency treebank additionally tagged for named entities. Here we will use synonym replacement to augment a corpus.

To load the corpus we will use DaCy which we will install using:

pip install dacy

And then we can apply the methods:

from dacy import datasets

train, dev, test = datasets.dane(splits=["train", "dev", "test"])

from spacy.lang.da import Danish

nlp_da = Danish()

synonym_augmenter = augmenty.load("wordnet_synonym.v1", level=0.2, lang="da")
augmented_corpus = [
    e for example in test(nlp_da) for e in synonym_augmenter(nlp_da, example)
]

Using Faker for entity replacing

Augmenty allow for easy integrations of existing libraries such as Faker. Faker is an open-source python library which can generate various entities such as names. It is compatibile with multiple languages and has a lot of language-specific entity generators.

To install Faker:

pip install faker

import spacy
from faker import Faker

nlp_en = spacy.load("en_core_web_lg")
texts = "my name is Kenneth Enevoldsen"
doc = nlp_en(texts)

# entity replace augmenter takes a generator returning list of strings
fake = Faker()


def faker_name():
    while True:
        yield fake.name().split()


ent_dict = {"PERSON": faker_name()}

ent_augmenter = augmenty.load("ents_replace_v1", level=1.00, ent_dict=ent_dict)

docs = list(augmenty.docs([doc], augmenter=ent_augmenter, nlp=nlp_en))
print(docs[0])

my name is Dalton Wright

Creating and Contributing Augmenters

After using augmenty you might want to create and contribute an augmenter. Most augmenters can be created based on already existing augmenters. For instance the augmenter per_replace_v1, which replaces names in a text is a special case of the augmenter ents_replace_v1 with better handling of first and last names. If you want to create an augmenter from scratch following spaCy’s guide on creating custom augmenters is a good start. You can always use augmenters from augmenty as inspiration as well. If you find yourself in troubles feel free to ask in the augmenty forums.

When you are satisfied with your augmenter feel free submit a pull request to add the augmenter to augmenty.