How do I use sklearn CountVectorizer with both 'word' and 'char' analyzer? - python

Question

How do I use sklearn CountVectorizer with both 'word' and 'char' analyzer? http://scikit-learn.org/stable/modules/generated/sklearn.feature_extraction.text.CountVectorizer.html

I could extract the text features by word or char separately but how do i create a charword_vectorizer? Is there a way to combine the vectorizers? or use more than one analyzer?

>>> from sklearn.feature_extraction.text import CountVectorizer
>>> word_vectorizer = CountVectorizer(analyzer='word', ngram_range=(1, 2), min_df=1)
>>> char_vectorizer = CountVectorizer(analyzer='char', ngram_range=(1, 2), min_df=1)
>>> x = ['this is a foo bar', 'you are a foo bar black sheep']
>>> word_vectorizer.fit_transform(x)
<2x15 sparse matrix of type '<type 'numpy.int64'>'
    with 18 stored elements in Compressed Sparse Column format>
>>> char_vectorizer.fit_transform(x)
<2x47 sparse matrix of type '<type 'numpy.int64'>'
    with 64 stored elements in Compressed Sparse Column format>
>>> char_vectorizer.get_feature_names()
[u' ', u' a', u' b', u' f', u' i', u' s', u'a', u'a ', u'ac', u'ar', u'b', u'ba', u'bl', u'c', u'ck', u'e', u'e ', u'ee', u'ep', u'f', u'fo', u'h', u'he', u'hi', u'i', u'is', u'k', u'k ', u'l', u'la', u'o', u'o ', u'oo', u'ou', u'p', u'r', u'r ', u're', u's', u's ', u'sh', u't', u'th', u'u', u'u ', u'y', u'yo']
>>> word_vectorizer.get_feature_names()
[u'are', u'are foo', u'bar', u'bar black', u'black', u'black sheep', u'foo', u'foo bar', u'is', u'is foo', u'sheep', u'this', u'this is', u'you', u'you are']

Solution 1

You can pass a callable as the analyzer argument to get full control over the tokenization, e.g.

>>> from pprint import pprint
>>> import re
>>> x = ['this is a foo bar', 'you are a foo bar black sheep']
>>> def words_and_char_bigrams(text):
...     words = re.findall(r'\w{3,}', text)
...     for w in words:
...         yield w
...         for i in range(len(w) - 2):
...             yield w[i:i+2]
...             
>>> v = CountVectorizer(analyzer=words_and_char_bigrams)
>>> pprint(v.fit(x).vocabulary_)
{'ac': 0,
 'ar': 1,
 'are': 2,
 'ba': 3,
 'bar': 4,
 'bl': 5,
 'black': 6,
 'ee': 7,
 'fo': 8,
 'foo': 9,
 'he': 10,
 'hi': 11,
 'la': 12,
 'sh': 13,
 'sheep': 14,
 'th': 15,
 'this': 16,
 'yo': 17,
 'you': 18}

Solution 2

You can combine arbitrary feature extraction steps with the FeatureUnion estimator: http://scikit-learn.org/dev/modules/pipeline.html#featureunion-combining-feature-extractors

In this case this is probably less efficient than larsmans solution, but might be easier to use.

Comments

[alvas]

pardon the noobness, what does r'\w{3,} mean? is it the same as [a-zA-Z0-9_]?

[Fred Foo]

@alvas: \w is [a-zA-Z0-9_]; {3,} means "three or more" so it skips short tokens. You might want to experiment with different REs to capture tokens; this is one of my baseline ones.