dlite-tools/NLPiper

NLPiper is a package that agglomerates different NLP tools and applies their transformations in the target document.

NLPiper is a package that agglomerates different NLP tools and applies their transformations in the target document.

Goal

Lightweight package integrated with a universe of frameworks to pre-process documents.

---

Installation

You can install NLPiper from PyPi with pip or your favorite package manager:

pip install nlpiper

---

Optional Dependencies

Some transformations require the installation of additional packages. The following table explains the optional dependencies that can be installed:

Package	Description
bs4	Used in CleanMarkup to remove HTML and XML from the document.
gensim	Used in GensimEmbeddings for document embedding extraction.
hunspell	Used in Stemmer and SpellCheck to normalize the document.
nltk	Used in RemoveStopWords to remove stop words from the document.
numpy	Used in some document's transformations.
sacremoses	Used in MosesTokenizer to tokenize the document.
spacy	Used in SpacyTokenizer to tokenize the document and could also be used for extracting entities, tags, etc..
stanza	Used in StanzaTokenizer to tokenize the document and could also be used for extracting entities, tags, etc.
torchtext	Used in TorchTextEmbeddings for document embedding extraction.

To install the optional dependency needed for your purpose you can run:

pip install nlpiper[<package>]

You can install all of these dependencies at once with:

pip install nlpiper[all]

The package can be installed using pip:

pip install nlpiper

For all transforms be available: pip install 'nlpiper[all]', otherwise, just install the packages needed.

Usage

Define a Pipeline:

>>> from nlpiper.core import Compose
>>> from nlpiper.transformers import cleaners, normalizers, tokenizers
>>> pipeline = Compose([
...                    cleaners.CleanNumber(),
...                    tokenizers.BasicTokenizer(),
...                    normalizers.CaseTokens()
... ])
>>> pipeline
Compose([CleanNumber(), BasicTokenizer(), CaseTokens(mode='lower')])

Generate a Document and Document structure:

>>> from nlpiper.core import Document
>>> doc = Document("The following character is a number: 1 and the next one is not a.")
>>> doc
Document(
    original='The following character is a number: 1 and the next one is not a.',
    cleaned='The following character is a number: 1 and the next one is not a.',
    tokens=None,
    embedded=None,
    steps=[]
)

Apply Pipeline to a Document:

>>> doc = pipeline(doc)
>>> doc
Document(
    original='The following character is a number: 1 and the next one is not a.',
    cleaned='The following character is a number:  and the next one is not a.',
    tokens=[
        Token(original='The', cleaned='the', lemma=None, stem=None, embedded=None),
        Token(original='following', cleaned='following', lemma=None, stem=None, embedded=None),
        Token(original='character', cleaned='character', lemma=None, stem=None, embedded=None),
        Token(original='is', cleaned='is', lemma=None, stem=None, embedded=None),
        Token(original='a', cleaned='a', lemma=None, stem=None, embedded=None),
        Token(original='number:', cleaned='number:', lemma=None, stem=None, embedded=None),
        Token(original='and', cleaned='and', lemma=None, stem=None, embedded=None),
        Token(original='the', cleaned='the', lemma=None, stem=None, embedded=None),
        Token(original='next', cleaned='next', lemma=None, stem=None, embedded=None),
        Token(original='one', cleaned='one', lemma=None, stem=None, embedded=None),
        Token(original='is', cleaned='is', lemma=None, stem=None, embedded=None),
        Token(original='not', cleaned='not', lemma=None, stem=None, embedded=None),
        Token(original='a.', cleaned='a.', lemma=None, stem=None, embedded=None)
    ],
    embedded=None,
    steps=['CleanNumber()', 'BasicTokenizer()', "CaseTokens(mode='lower')"]
)

Available Transformers

Cleaners

Clean document as a whole, e.g. remove HTML, remove accents, remove emails, etc.

• CleanURL: remove URL from the text.
• CleanEmail: remove email from the text.
• CleanNumber: remove numbers from text.
• CleanPunctuation: remove punctuation from text.
• CleanEOF: remove end of file from text.
• CleanMarkup: remove HTML or XML from text.
• CleanAccents: remove accents from the text.

Tokenizers

Tokenize a document after cleaning is done (Split document into tokens)

• BasicTokenizer: Split tokens by spaces in the text.
• MosesTokenizer: Split tokens using Moses tokenizer (https://github.com/alvations/sacremoses)
• StanzaTokenizer: Split tokens using Stanza tokenizer (https://github.com/stanfordnlp/stanza)

Normalizer

Applies on the token level, e.g. remove stop-words, spell-check, etc.

• CaseTokens: lower or upper case all tokens.
• RemovePunctuation: Remove punctuation from resulting tokens.
• RemoveStopWords: Remove stop-words as tokens.
• VocabularyFilter: Only allow tokens from a pre-defined vocabulary.
• Stemmer: Get the stem from the tokens.
• SpellCheck: Spell check the token, if given max distance will calculate the Levenshtein distance from the token with the suggested word and if lower the token is replaced by the suggestion else will keep the token. If no maximum distance is given if the word is not correctly spelt then will be replaced by an empty string.

Embeddings

Applies on the token level, converting words by embeddings

• GensimEmbeddings: Use Gensim word embeddings.
• TorchTextEmbeddings: Applies word embeddings using torchtext models Glove, CharNGram and FastText.

Document

Document is a dataclass that contains all the information used during text preprocessing.

Document attributes:

• original: original text to be processed.
• cleaned: original text to be processed when document is initiated and then attribute which Cleaners and Tokenizers work.
• tokens: list of tokens that where obtained using a Tokenizer.
• steps: list of transforms applied on the document.
• embedded: document embedding.

token:

• original: original token.
• cleaned: original token at initiation, then modified according with Normalizers.
• lemma: token lemma (need to use a normalizer or tokenizer to obtain).
• stem: token stem (need to use a normalizer to obtain).
• ner: token entity (need to use a normalizer or tokenizer to obtain).
• embedded: token embedding.

Compose

Compose applies the chosen transformers into a given document. It restricts the order that the transformers can be applied, first are the Cleaners, then the Tokenizers and lastly the Normalizers and Embeddings.

It is possible to create a compose using the steps from a processed document:

>>> doc.steps
['CleanNumber()', 'BasicTokenizer()', "CaseTokens(mode='lower')"]
>>> new_pipeline = Compose.create_from_steps(doc.steps)
>>> new_pipeline
Compose([CleanNumber(), BasicTokenizer(), CaseTokens(mode='lower')])

It is also possible to rollback the steps applied to a document:

>>> new_doc = Compose.rollback_document(doc, 2)
>>> new_doc
Document(
    original='The following character is a number: 1 and the next one is not a.',
    cleaned='The following character is a number:  and the next one is not a.',
    tokens=None,
    embedded=None,
    steps=['CleanNumber()']
)
>>> doc
Document(
    original='The following character is a number: 1 and the next one is not a.',
    cleaned='The following character is a number:  and the next one is not a.',
    tokens=[
        Token(original='The', cleaned='the', lemma=None, stem=None, embedded=None),
        Token(original='following', cleaned='following', lemma=None, stem=None, embedded=None),
        Token(original='character', cleaned='character', lemma=None, stem=None, embedded=None),
        Token(original='is', cleaned='is', lemma=None, stem=None, embedded=None),
        Token(original='a', cleaned='a', lemma=None, stem=None, embedded=None),
        Token(original='number:', cleaned='number:', lemma=None, stem=None, embedded=None),
        Token(original='and', cleaned='and', lemma=None, stem=None, embedded=None),
        Token(original='the', cleaned='the', lemma=None, stem=None, embedded=None),
        Token(original='next', cleaned='next', lemma=None, stem=None, embedded=None),
        Token(original='one', cleaned='one', lemma=None, stem=None, embedded=None),
        Token(original='is', cleaned='is', lemma=None, stem=None, embedded=None),
        Token(original='not', cleaned='not', lemma=None, stem=None, embedded=None),
        Token(original='a.', cleaned='a.', lemma=None, stem=None, embedded=None)
    ],
    embedded=None,
    steps=['CleanNumber()', 'BasicTokenizer()', "CaseTokens(mode='lower')"]
)

---

Development Installation

git clone https://github.com/dlite-tools/NLPiper.git
cd NLPiper
poetry install

To install an optional dependency you can run:

poetry install --extras <package>

To install all the optional dependencies run:

poetry install --extras all

---

Contributions

All contributions, bug reports, bug fixes, documentation improvements, enhancements and ideas are welcome.

A detailed overview on how to contribute can be found in the contributing guide on GitHub.

---

Issues

Go here to submit feature requests or bugfixes.

---

License and Credits

NLPiper is licensed under the MIT license and is written and maintained by Tomás Osório (@tomassosorio), Daniel Ferrari (@FerrariDG), Carlos Alves (@cmalves, João Cunha (@jfecunha)