CODEWITHSUNDEEP
pythonmachine-learningnlpscikit-learnnltk
tumbleweed
tumbleweed

Reputation: 4630

Problems classifiying labeled text, wrong prediction?

I'm playing with the different classifiers and vectorizers that scikit-learn provide so let's say I have the following:

training = [["this was a good movie, 'POS'"],
      ["this was a bad movie, 'NEG'"],
      ["i went to the movies, 'NEU'"], 
      ["this movie was very exiting it was great, 'POS'"], 
      ["this is a boring film, 'NEG'"]
        ,........................,
          [" N-sentence, 'LABEL'"]]

#Where each element of the list is another list that have documents, then.

splitted = [#remove the tags from training]

from sklearn.feature_extraction.text import HashingVectorizer
X = HashingVectorizer(
    tokenizer=lambda  doc: doc, lowercase=False).fit_transform(splitted)

print X.toarray()

Then I have this vector representation:

[[ 0.  0.  0. ...,  0.  0.  0.]
 [ 0.  0.  0. ...,  0.  0.  0.]
 [ 0.  0.  0. ...,  0.  0.  0.]
 [ 0.  0.  0. ...,  0.  0.  0.]
 [ 0.  0.  0. ...,  0.  0.  0.]]

The problem with this is that I dont know if I vectorized right the corpus, then:

#This is the test corpus:
test = ["I don't like this movie it sucks it doesn't liked me"]

#I vectorize the corpus with hashing vectorizer
Y = HashingVectorizer(
    tokenizer=lambda  doc: doc, lowercase=False).fit_transform(test)

Then I print the Y:

[[ 0.  0.  0. ...,  0.  0.  0.]]

Then

y = [x[-1]for x in training]

#import SVM and classify
from sklearn.svm import SVC
svm = SVC()
svm.fit(X, y)
result = svm.predict(X)
print "\nThe opinion is:\n",result

And here's the problem, I got the following insted of [NEG] which is actually the right prediction:

["this was a good movie, 'POS'"]

I guess I am not vectorizing right training or y target is wrong, could anybody help me to understand what's happening and how should I vectorize training in order to have a right prediction?

Upvotes: 2

Views: 219

Answers (1)

elyase
elyase

Reputation: 40963

I will leave it to you to get the training data into the expected format:

training = ["this was a good movie",
            "this was a bad movie",
            "i went to the movies",
            "this movie was very exiting it was great", 
            "this is a boring film"]

labels = ['POS', 'NEG', 'NEU', 'POS', 'NEG']

Feature extraction

>>> from sklearn.feature_extraction.text import HashingVectorizer
>>> vect = HashingVectorizer(n_features=5, stop_words='english', non_negative=True)
>>> X_train = vect.fit_transform(training)
>>> X_train.toarray()
[[ 0.          0.70710678  0.          0.          0.70710678]
 [ 0.70710678  0.70710678  0.          0.          0.        ]
 [ 0.          0.          0.          0.          0.        ]
 [ 0.          0.89442719  0.          0.4472136   0.        ]
 [ 1.          0.          0.          0.          0.        ]]

With bigger corpus you should increase n_features to avoid collisions, I used 5 so that the resulting matrix can be visualized. Also note that I used stop_words='english', I think with so few examples it is important to get rid of stopwords, otherwise you could confuse the classifier.

Model training

from sklearn.svm import SVC

model = SVC()
model.fit(X_train, labels)

Prediction

>>> test = ["I don't like this movie it sucks it doesn't liked me"]
>>> X_pred = vect.transform(test)
>>> model.predict(X_pred)
['NEG']

>>> test = ["I think it was a good movie"]
>>> X_pred = vect.transform(test)
>>> model.predict(X_pred)
['POS']

EDIT: Note that the correct classification of the first test example is just a fortunate coincidence as I don't see any word that could have been learned from the training set as negative. In the second example the word good could have triggered the positive classification.

Upvotes: 2

Related Questions