Question

Hey I am trying to use a Naive Bayes classifier to classify some text. I am using NLTK. Whenever I test the classifier using the classify() method it always returns the correct classification for the first item, and the same classification for every other line of text I classify. The following is my code:

from nltk.corpus import movie_reviews
from nltk.tokenize import word_tokenize
import nltk
import random
import nltk.data

documents = [(list(movie_reviews.words(fileid)), category)
for category in movie_reviews.categories()
for fileid in movie_reviews.fileids(category)]
random.shuffle(documents)

all_words = nltk.FreqDist(w.lower() for w in movie_reviews.words())
word_features = all_words.keys()[:2000] 

def bag_of_words(words):
    return dict([word,True] for word in words)

def document_features(document): 
    document_words = set(document) 
    features = {}
    for word in word_features:
        features['contains(%s)' % word] = (word in document_words)
    return features

featuresets = [(document_features(d), c) for (d,c) in documents]
train_set, test_set = featuresets[100:], featuresets[:100]
classifier = nltk.NaiveBayesClassifier.train(train_set)

text1="i love this city"
text2="i hate this city"


feats1=bag_of_words(word_tokenize(text1))
feats2=bag_of_words(word_tokenize(text2))


print classifier.classify(feats1)
print classifier.classify(feats2)

This code will print pos twice where as if I flipped the last 2 lines of the code it will print neg twice. Can anyone help?

Answer 1

Change

features['contains(%s)' % word] = (word in document_words)

to

features[word] = (word in document)

Otherwise the classifier only knows about “words” of the form “contains(…)”, and is therefore clueless about the words in "i love this city"

---

import nltk.tokenize as tokenize
import nltk
import random
random.seed(3)

def bag_of_words(words):
    return dict([word, True] for word in words)

def document_features(document): 
    features = {}
    for word in word_features:
        features[word] = (word in document)
        # features['contains(%s)' % word] = (word in document_words)
    return features

movie_reviews = nltk.corpus.movie_reviews

documents = [(set(movie_reviews.words(fileid)), category)
             for category in movie_reviews.categories()
             for fileid in movie_reviews.fileids(category)]
random.shuffle(documents)

all_words = nltk.FreqDist(w.lower() for w in movie_reviews.words())
word_features = all_words.keys()[:2000] 

train_set = [(document_features(d), c) for (d, c) in documents[:200]]

classifier = nltk.NaiveBayesClassifier.train(train_set)

classifier.show_most_informative_features()
for word in ('love', 'hate'):
    # No hope in passing the tests if word is not in word_features
    assert word in word_features
    print('probability {w!r} is positive: {p:.2%}'.format(
        w = word, p = classifier.prob_classify({word : True}).prob('pos')))

tests = ["i love this city",
         "i hate this city"]

for test in tests:
    words = tokenize.word_tokenize(test)
    feats = bag_of_words(words)
    print('{s} => {c}'.format(s = test, c = classifier.classify(feats)))

yields

Most Informative Features
                   worst = True              neg : pos    =     15.5 : 1.0
              ridiculous = True              neg : pos    =     11.5 : 1.0
                  batman = True              neg : pos    =      7.6 : 1.0
                   drive = True              neg : pos    =      7.6 : 1.0
                   blame = True              neg : pos    =      7.6 : 1.0
                terrible = True              neg : pos    =      6.9 : 1.0
                  rarely = True              pos : neg    =      6.4 : 1.0
                 cliches = True              neg : pos    =      6.0 : 1.0
                       $ = True              pos : neg    =      5.9 : 1.0
               perfectly = True              pos : neg    =      5.5 : 1.0
probability 'love' is positive: 61.52%
probability 'hate' is positive: 36.71%
i love this city => pos
i hate this city => neg