Exactly replicating R text preprocessing in python

Question

I would like to preprocess a corpus of documents using Python in the same way that I can in R. For example, given an initial corpus, corpus, I would like to end up

Answer 1

CountVectorizer and TfidfVectorizer can be customized as described in the docs. In particular, you'll want to write a custom tokenizer, which is a function that takes a document and returns a list of terms. Using NLTK:

import nltk.corpus.stopwords
import nltk.stem

def smart_tokenizer(doc):
    doc = doc.lower()
    doc = re.findall(r'\w+', doc, re.UNICODE)
    return [nltk.stem.PorterStemmer().stem(term)
            for term in doc
            if term not in nltk.corpus.stopwords.words('english')]

Demo:

>>> v = CountVectorizer(tokenizer=smart_tokenizer)
>>> v.fit_transform([doc]).toarray()
array([[1, 1, 1, 2, 1, 1, 1, 1, 1]])
>>> from pprint import pprint
>>> pprint(v.vocabulary_)
{u'amaz': 0,
 u'appl': 1,
 u'best': 2,
 u'ear': 3,
 u'ever': 4,
 u'headphon': 5,
 u'pod': 6,
 u'sound': 7,
 u've': 8}

(The example I linked to actually uses a class to cache the lemmatizer, but a function works too.)

Answer 2

It seems tricky to get things exactly the same between nltk and tm on the preprocessing steps, so I think the best approach is to use rpy2 to run the preprocessing in R and pull the results into python:

import rpy2.robjects as ro
preproc = [x[0] for x in ro.r('''
tweets = read.csv("tweets.csv", stringsAsFactors=FALSE)
library(tm)
library(SnowballC)
corpus = Corpus(VectorSource(tweets$Tweet))
corpus = tm_map(corpus, tolower)
corpus = tm_map(corpus, removePunctuation)
corpus = tm_map(corpus, removeWords, c("apple", stopwords("english")))
corpus = tm_map(corpus, stemDocument)''')]

Then, you can load it into scikit-learn -- the only thing you'll need to do to get things to match between the CountVectorizer and the DocumentTermMatrix is to remove terms of length less than 3:

from sklearn.feature_extraction.text import CountVectorizer
def mytokenizer(x):
    return [y for y in x.split() if len(y) > 2]

# Full document-term matrix
cv = CountVectorizer(tokenizer=mytokenizer)
X = cv.fit_transform(preproc)
X
# <1181x3289 sparse matrix of type '<type 'numpy.int64'>'
#   with 8980 stored elements in Compressed Sparse Column format>

# Sparse terms removed
cv2 = CountVectorizer(tokenizer=mytokenizer, min_df=0.005)
X2 = cv2.fit_transform(preproc)
X2
# <1181x309 sparse matrix of type '<type 'numpy.int64'>'
#   with 4669 stored elements in Compressed Sparse Column format>

Let's verify this matches with R:

tweets = read.csv("tweets.csv", stringsAsFactors=FALSE)
library(tm)
library(SnowballC)
corpus = Corpus(VectorSource(tweets$Tweet))
corpus = tm_map(corpus, tolower)
corpus = tm_map(corpus, removePunctuation)
corpus = tm_map(corpus, removeWords, c("apple", stopwords("english")))
corpus = tm_map(corpus, stemDocument)
dtm = DocumentTermMatrix(corpus)
dtm
# A document-term matrix (1181 documents, 3289 terms)
# 
# Non-/sparse entries: 8980/3875329
# Sparsity           : 100%
# Maximal term length: 115 
# Weighting          : term frequency (tf)

sparse = removeSparseTerms(dtm, 0.995)
sparse
# A document-term matrix (1181 documents, 309 terms)
# 
# Non-/sparse entries: 4669/360260
# Sparsity           : 99%
# Maximal term length: 20 
# Weighting          : term frequency (tf)

As you can see, the number of stored elements and terms exactly match between the two approaches now.