A pipeline for NLP projects using SkLearn
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.
nlp_pipeline_manager make it python installable Aug 17, 2018
prototyping add base code Aug 12, 2018
.gitignore make it python installable Aug 17, 2018
README.md Update README.md Aug 17, 2018
setup.py add install requirements Aug 17, 2018


NLP Pipeline Manager

The most frustrating part of doing NLP for me is keeping track of all the different combinations of cleaning functions, stemmers, lemmatizers, vectorizers, models, etc. I almost always resort to writing some awful function that hacks those bits together and then prints out some scoring piece. To help manage all of this better, I've developed a pipelining system that allows the user to load all of the pieces into a class and then let the class do the management for them.


Clone this repo. Go to the directory where it is cloned and run:

python setup.py install

nlp_pipeline_manager will then install to your machine and be available. This project assumes python 3 and requires NLTK and SkLearn.

Examples of using the pipeline:

from nlp_pipeline_manager import nlp_preprocessor

corpus = ['BOB the builder', 'is a strange', 'caRtoon type thing']
nlp = nlp_preprocessor()
array([[1, 1, 0, 0, 0, 1, 0, 0],
       [0, 0, 0, 1, 1, 0, 0, 0],
       [0, 0, 1, 0, 0, 0, 1, 1]])

Loading a stemmer into the pipeline (we actually pass in the stemming method):

from nltk.stem import PorterStemmer

nlp = nlp_preprocessor(stemmer=PorterStemmer().stem)

The pipeline allows users to set:

  • Vectorizer (using SkLearn classes)
  • Cleaning Function (user can just provide a function name without the parens at the end)
  • Tokenizer (Either an NLTK tokenizer or a function that takes a string and returns a list of tokens)
  • Stemmer (Can be any function that takes in a word and returns a root form - be it a stemmer or a lemmatizer)

If the user wants to provide a cleaning function, it must accept 3 arguments. The text, the tokenizer, and the stemmer. Here's an example extra cleaning function:

def clean_text(text, tokenizer, stemmer):
        A naive function to lowercase all words and clean them quickly
        with a stemmer if it exists.
        cleaned_text = []
        for post in text:
            cleaned_words = []
            for word in tokenizer(post):
                low_word = word.lower()
                if stemmer:
                    low_word = stemmer(low_word)
            cleaned_text.append(' '.join(cleaned_words))
        return cleaned_text

ML with the pipeline

It's quick and easy to create modeling pipelines that wrap around the preprocessor. Two example pipes are shown, one for classification and one for topic modeling. Here's an example of using the classification pipe:

from nlp_pipeline_manager import supervised_nlp
from nlp_pipeline_manager import nlp_preprocessor
from sklearn.naive_bayes import MultinomialNB

nlp = nlp_preprocessor()

nlp_pipe = supervised_nlp(MultinomialNB(), nlp)
nlp_pipe.fit(ng_train_data, ng_train_targets)
nlp_pipe.score(ng_test_data, ng_test_targets)