Statistical Learning and Text Classification with

NLTK and scikit-learn

Olivier Grisel
http://twitter.com/ogrisel

PyCON FR – 2010
 Applications of Text Classification
Task Predicted outcome

Spam filtering Spam, Ham

Language guessing English, Spanish, French, ...

Sentiment Analysis for Product

Positive, Neutral, Negative
Reviews
News Feed Topic Politics, Business, Technology,
Categorization Sports, ...
Pay-per-click optimal ads
Will yield money, Won't
placement
Personal twitter filter Will interest me, Won't
Malware detection in log files Normal, Malware
 Supervised Learning Overview
● Convert training data to a set of vectors of features
(input) & label (output)
● Build a model based on the statistical properties of
features in the training set, e.g.
● Naïve Bayesian Classifier
● Logistic Regression / Maxent Classifier
● Support Vector Machines
● For each new text document to classify
● Extract features
● Asked model to predict the most likely outcome
 Supervised Learning Summary
features
Training vectors
Text
Documents

Machine
Learning
Algorithm
Labels

New features
Text vector Predictive Expected
Document Model Label
 Typical features for text documents
● Tokenize document into list of words: uni-grams
['the', 'quick', 'brown', 'fox', 'jumps', 'over',   
'the', 'lazy', 'dog']
● Then chose one of:
● Binary occurrences of uni-grams:
{'the': True, 'quick': True, ...}
● Frequencies of uni-grams: nb times word_i / nb
words in document:
    {'the': 0.22, 'quick': 0.11, ...}
● TF-IDF of uni-grams (see next slides)
 Better than freqs: TF-IDF
● Term Frequency

● Inverse Document Frequency

=> No real need for stop words any more, non informative
words such as “the” are scaled done by IDF term
 More advanced features
● Instead of uni-grams use
● bi-grams of words: “New York”, “very bad”, “not
good”
● n-grams of chars: “the”, “ed ”, “ a ” (useful for
language guessing)
● And the combine with:
● Binary occurrences
● Frequencies
● TF-IDF
 NLTK
● Code: ASL 2.0 & Book: CC-BY-NC-ND
● Tokenizers, Stemmers, Parsers, Classifiers,
Clusterers, Corpus Readers
 NLTK Corpus Downloader
>>> import nltk
>>> nltk.download()
 Using a NLTK corpus
>>> from nltk.corpus import movie_reviews
>>> pos_ids = movie_reviews.fileids('pos')
>>> neg_ids = movie_reviews.fileids('neg')
>>> len(pos_ids), len(neg_ids) 
1000, 1000
>>> print movie_reviews.raw(pos_ids[0])[:100]
films adapted from comic books have had plenty of success , 
whether they're about superheroes ( batm 
>>> movie_reviews.words(pos_ids[0])
['films', 'adapted', 'from', 'comic', 'books', 'have', ...]
 Common data cleanup operations
● Switch to lower case: s.lower()
● Remove accentuated chars:
  import unicodedata
  s = ''.join(c for c in unicodedata.normalize('NFD', s)
                if unicodedata.category(c) != 'Mn')

● Extract only word tokens of at least 2 chars

● Using NLTK tokenizers & stemmers
● Using a simple regexp:

     re.compile(r"\b\w\w+\b", re.U).findall(s)
 Feature Extraction with NLTK
● Simple word binary occurrence features:
def word_features(words):
    return dict((word, True) for word in words)

● Word Bigrams occurrence features:

from nltk.collocations import BigramCollocationFinder
from nltk.metrics import BigramAssocMeasures as BAM
from itertools import chain

def bigram_word_features(words, score_fn=BAM.chi_sq, n=200):
    bigram_finder = BigramCollocationFinder.from_words(words)
    bigrams = bigram_finder.nbest(score_fn, n)
    return dict((bg, True) for bg in chain(words, bigrams))
 The NLTK - Naïve Bayes Classifier
from nltk.classify import NaiveBayesClassifier
mr = movie_reviews
neg_examples = [(features(mr.words(i)), 'neg')
                for i in neg_ids]
pos_examples = [(features(mr.words(i)), 'pos')
                for i in pos_ids]
train_set = pos_examples + neg_examples
classifier = NaiveBayesClassifier.train(train_set)

# later on a previously unseed document 
predicted_label = classifier.classify(new_doc_features)
 Most informative features
>>> classifier.show_most_informative_features()
         magnificent = True              pos : neg    =     15.0 : 1.0
         outstanding = True              pos : neg    =     13.6 : 1.0
           insulting = True              neg : pos    =     13.0 : 1.0
          vulnerable = True              pos : neg    =     12.3 : 1.0
           ludicrous = True              neg : pos    =     11.8 : 1.0
              avoids = True              pos : neg    =     11.7 : 1.0
         uninvolving = True              neg : pos    =     11.7 : 1.0
          astounding = True              pos : neg    =     10.3 : 1.0
         fascination = True              pos : neg    =     10.3 : 1.0
             idiotic = True              neg : pos    =      9.8 : 1.0
scikit-learn
 Features Extraction in scikit-learn
from scikits.learn.features.text import *

text = u"J'ai mang\xe9 du kangourou  ce midi, c'\xe9tait pas 
tr\xeas bon."
print WordNGramAnalyzer(min_n=1, max_n=2).analyze(text)
[u'ai', u'mange', u'du', u'kangourou', u'ce', u'midi', 
u'etait', u'pas', u'tres', u'bon', u'ai mange', u'mange du', 
u'du kangourou', u'kangourou ce', u'ce midi', u'midi etait', 
u'etait pas', u'pas tres', u'tres bon']
char_ngrams = CharNGramAnalyzer(min_n=3, max_n=6)
print char_ngrams[:5] + char_ngrams[­5:]
[u"j'a", u"'ai", u'ai ', u'i m', u' ma', u's tres', u' tres 
', u'tres b', u'res bo', u'es bon']
 TF-IDF features & SVMs
from scikits.learn.features.text import *
from scikits.learn.sparse.svm import LinearSVC

hv = SparseHashingVectorizer(dim=1000000, analyzer=)
hv.vectorize(list_of_documents)
features = hv.get_tfidf()
clf = SparseLinearSVC(C=10, dual=false)
clf.fit(features, labels)

# later with the same clf instance
predicted_labels = clf.predict(features_of_new_docs) 
 Typical performance results
● Naïve Bayesian Classifier with unigram
occurrences on movie reviews: ~ 70%
● Same as above selecting the top 10000 most
informative features only: ~ 93%
● TF-IDF unigram features + Linear SVC on 20
newsgroups ~93% (with 20 target categories)
● Language guessing with character ngram
frequencies features + Linear SVC: almost
perfect if document is long enough
 Confusion Matrix (20 newsgroups)
00 alt.atheism
01 comp.graphics
02 comp.os.ms-windows.misc
03 comp.sys.ibm.pc.hardware
04 comp.sys.mac.hardware
05 comp.windows.x
06 misc.forsale
07 rec.autos
08 rec.motorcycles
09 rec.sport.baseball
10 rec.sport.hockey
11 sci.crypt
12 sci.electronics
13 sci.med
14 sci.space
15 soc.religion.christian
16 talk.politics.guns
17 talk.politics.mideast
18 talk.politics.misc
19 talk.religion.misc
 Handling many possible outcomes
● Example: possible outcomes are all the
categories of Wikipedia (565,108)
● Document Categorization becomes Information
Retrieval
● Instead of building one linear model for each
outcome build a fulltext index and perform TF-
IDF similarity queries
● Smart way to find the top 10 search keywords
● Use Apache Lucene / Solr MoreLikeThisQuery
NLTK – Online demos
 NLTK – REST APIs
% curl -d "text=Inception is the best movie ever" \
http://text-processing.com/api/sentiment/

{
"probability": {
"neg": 0.36647424288117808,
"pos": 0.63352575711882186
},
"label": "pos"
}
Google Prediction API
 Some pointers
● http://www.nltk.org (Code & Doc & PDF Book)
● http://scikit-learn.sf.net (Doc & Examples)
http://github.com/scikit-learn (Code)
● http://www.slideshare.net/ogrisel (These slides)

● http://streamhacker.com/(Blog on NLTK & APIs)

● http://github.com/hmason/tc (Twitter classifier –
work in progress)