基于机器学习的情感分析



王成军

wangchengjun@nju.edu.cn

计算传播网 http://computational-communication.com

情感分析(sentiment analysis)和意见挖掘(opinion mining)虽然相关,但是从社会科学的角度而言,二者截然不同。这里主要是讲情感分析(sentiment or emotion),而非意见挖掘(opinion, 后者通过机器学习效果更可信)。

classify emotion

Different types of emotion: anger, disgust, fear, joy, sadness, and surprise. The classification can be performed using different algorithms: e.g., naive Bayes classifier trained on Carlo Strapparava and Alessandro Valitutti’s emotions lexicon.

classify polarity

To classify some text as positive or negative. In this case, the classification can be done by using a naive Bayes algorithm trained on Janyce Wiebe’s subjectivity lexicon.

LIWC & TextMind

http://ccpl.psych.ac.cn/textmind/

“文心(TextMind)”中文心理分析系统是由中科院心理所计算网络心理实验室研发的,针对中文文本进行语言分析的软件系统,通过“文心”,您可以便捷地分析文本中使用的不同类别语言的程度、偏好等特点。针对中国大陆地区简体环境下的语言特点,参照LIWC2007和正體中文C-LIWC词库,我们开发了“文心(TextMind)”中文心理分析系统。“文心”为用户提供从简体中文自动分词,到语言心理分析的一揽子分析解决方案,其词库、文字和符号等处理方法专门针对简体中文语境,词库分类体系也与LIWC兼容一致。

Preparing the data

NLTK

Anaconda自带的(默认安装的)第三方包。http://www.nltk.org/

NLTK is a leading platform for building Python programs to work with human language data. It provides easy-to-use interfaces to over 50 corpora and lexical resources such as WordNet, along with a suite of text processing libraries for classification, tokenization, stemming, tagging, parsing, and semantic reasoning, wrappers for industrial-strength NLP libraries, and an active discussion forum.


In [1]:
import nltk

pos_tweets = [('I love this car', 'positive'),
    ('This view is amazing', 'positive'),
    ('I feel great this morning', 'positive'),
    ('I am so excited about the concert', 'positive'),
    ('He is my best friend', 'positive')]

neg_tweets = [('I do not like this car', 'negative'),
    ('This view is horrible', 'negative'),
    ('I feel tired this morning', 'negative'),
    ('I am not looking forward to the concert', 'negative'),
    ('He is my enemy', 'negative')]

In [2]:
tweets = []
for (words, sentiment) in pos_tweets + neg_tweets:
    words_filtered = [e.lower() for e in words.split() if len(e) >= 3]
    tweets.append((words_filtered, sentiment))
tweets[:2]


Out[2]:
[(['love', 'this', 'car'], 'positive'),
 (['this', 'view', 'amazing'], 'positive')]

In [3]:
test_tweets = [
    (['feel', 'happy', 'this', 'morning'], 'positive'),
    (['larry', 'friend'], 'positive'),
    (['not', 'like', 'that', 'man'], 'negative'),
    (['house', 'not', 'great'], 'negative'),
    (['your', 'song', 'annoying'], 'negative')]

Extracting Features

Then we need to get the unique word list as the features for classification.


In [4]:
# get the word lists of tweets
def get_words_in_tweets(tweets):
    all_words = []
    for (words, sentiment) in tweets:
        all_words.extend(words)
    return all_words

# get the unique word from the word list	
def get_word_features(wordlist):
    wordlist = nltk.FreqDist(wordlist)
    word_features = wordlist.keys()
    return word_features

word_features = get_word_features(get_words_in_tweets(tweets))
' '.join(word_features)


Out[4]:
'friend tired great like amazing this looking horrible forward excited best not concert enemy love view feel the car morning about'

To create a classifier, we need to decide what features are relevant. To do that, we first need a feature extractor.


In [5]:
def extract_features(document):
    document_words = set(document)
    features = {}
    for word in word_features:
        features['contains(%s)' % word] = (word in document_words)
    return features

In [30]:
help(nltk.classify.util.apply_features)


Help on function apply_features in module nltk.classify.util:

apply_features(feature_func, toks, labeled=None)
    Use the ``LazyMap`` class to construct a lazy list-like
    object that is analogous to ``map(feature_func, toks)``.  In
    particular, if ``labeled=False``, then the returned list-like
    object's values are equal to::
    
        [feature_func(tok) for tok in toks]
    
    If ``labeled=True``, then the returned list-like object's values
    are equal to::
    
        [(feature_func(tok), label) for (tok, label) in toks]
    
    The primary purpose of this function is to avoid the memory
    overhead involved in storing all the featuresets for every token
    in a corpus.  Instead, these featuresets are constructed lazily,
    as-needed.  The reduction in memory overhead can be especially
    significant when the underlying list of tokens is itself lazy (as
    is the case with many corpus readers).
    
    :param feature_func: The function that will be applied to each
        token.  It should return a featureset -- i.e., a dict
        mapping feature names to feature values.
    :param toks: The list of tokens to which ``feature_func`` should be
        applied.  If ``labeled=True``, then the list elements will be
        passed directly to ``feature_func()``.  If ``labeled=False``,
        then the list elements should be tuples ``(tok,label)``, and
        ``tok`` will be passed to ``feature_func()``.
    :param labeled: If true, then ``toks`` contains labeled tokens --
        i.e., tuples of the form ``(tok, label)``.  (Default:
        auto-detect based on types.)


In [36]:
training_set[0]


Out[36]:
({'contains(about)': False,
  'contains(amazing)': False,
  'contains(best)': False,
  'contains(car)': True,
  'contains(concert)': False,
  'contains(enemy)': False,
  'contains(excited)': False,
  'contains(feel)': False,
  'contains(forward)': False,
  'contains(friend)': False,
  'contains(great)': False,
  'contains(horrible)': False,
  'contains(like)': False,
  'contains(looking)': False,
  'contains(love)': True,
  'contains(morning)': False,
  'contains(not)': False,
  'contains(the)': False,
  'contains(this)': True,
  'contains(tired)': False,
  'contains(view)': False},
 'positive')

In [6]:
training_set = nltk.classify.util.apply_features(extract_features,\
                                                 tweets)
classifier = nltk.NaiveBayesClassifier.train(training_set)

In [26]:
# You may want to know how to define the ‘train’ method in NLTK here:
def train(labeled_featuresets, estimator=nltk.probability.ELEProbDist):
    # Create the P(label) distribution
    label_probdist = estimator(label_freqdist)
    # Create the P(fval|label, fname) distribution
    feature_probdist = {}
    model = NaiveBayesClassifier(label_probdist, feature_probdist)
    return model

In [12]:
tweet_positive = 'Harry is my friend'
classifier.classify(extract_features(tweet_positive.split()))


Out[12]:
'positive'

In [16]:
tweet_negative = 'Larry is not my friend'
classifier.classify(extract_features(tweet_negative.split()))


Out[16]:
'negative'

In [19]:
# Don’t be too positive, let’s try another example:
tweet_negative2 = 'Your song is annoying'
classifier.classify(extract_features(tweet_negative2.split()))


Out[19]:
'positive'

In [20]:
def classify_tweet(tweet):
    return classifier.classify(extract_features(tweet)) 
    # nltk.word_tokenize(tweet)

total = accuracy = float(len(test_tweets))

for tweet in test_tweets:
    if classify_tweet(tweet[0]) != tweet[1]:
        accuracy -= 1

print('Total accuracy: %f%% (%d/20).' % (accuracy / total * 100, accuracy))


Total accuracy: 80.000000% (4/20).

使用sklearn的分类器


In [12]:
# nltk有哪些分类器呢?
nltk_classifiers = dir(nltk)
for i in nltk_classifiers:
    if 'Classifier' in i:
        print(i)


ClassifierBasedPOSTagger
ClassifierBasedTagger
ClassifierI
ConditionalExponentialClassifier
DecisionTreeClassifier
MaxentClassifier
MultiClassifierI
NaiveBayesClassifier
PositiveNaiveBayesClassifier
SklearnClassifier
WekaClassifier

In [21]:
from sklearn.svm import LinearSVC
from nltk.classify.scikitlearn import SklearnClassifier
classif = SklearnClassifier(LinearSVC())
svm_classifier = classif.train(training_set)

In [22]:
# Don’t be too positive, let’s try another example:
tweet_negative2 = 'Your song is annoying'
svm_classifier.classify(extract_features(tweet_negative2.split()))


Out[22]:
'negative'

作业1:

使用另外一种sklearn的分类器来对tweet_negative2进行情感分析

作业2:

使用https://github.com/victorneo/Twitter-Sentimental-Analysis 所提供的推特数据进行情感分析,可以使用其代码 https://github.com/victorneo/Twitter-Sentimental-Analysis/blob/master/classification.py

Sentiment Analysis using TextBlob

安装textblob

https://github.com/sloria/TextBlob

pip install -U textblob

python -m textblob.download_corpora


In [17]:
from textblob import TextBlob

text = '''
The titular threat of The Blob has always struck me as the ultimate movie
monster: an insatiably hungry, amoeba-like mass able to penetrate
virtually any safeguard, capable of--as a doomed doctor chillingly
describes it--"assimilating flesh on contact.
Snide comparisons to gelatin be damned, it's a concept with the most
devastating of potential consequences, not unlike the grey goo scenario
proposed by technological theorists fearful of
artificial intelligence run rampant.
'''

blob = TextBlob(text)
blob.tags           # [('The', 'DT'), ('titular', 'JJ'),
                    #  ('threat', 'NN'), ('of', 'IN'), ...]

blob.noun_phrases   # WordList(['titular threat', 'blob',
                    #            'ultimate movie monster',
                    #            'amoeba-like mass', ...])

for sentence in blob.sentences:
    print(sentence.sentiment.polarity)
# 0.060
# -0.341

blob.translate(to="es")  # 'La amenaza titular de The Blob...'


0.06000000000000001
-0.34166666666666673
Out[17]:
TextBlob("La amenaza principal de The Blob siempre me ha parecido la mejor película
monstruo: una masa insaciablemente hambrienta, similar a una ameba capaz de penetrar
prácticamente cualquier salvaguardia, capaz de - como un doctor condenado escalofriante
lo describe - "asimilando carne en contacto.
Las malditas comparaciones con la gelatina pueden ser condenadas, es un concepto con la mayor cantidad de
devastador de posibles consecuencias, a diferencia del escenario gris goo
propuesto por teóricos tecnológicos temerosos de
la inteligencia artificial corre desenfrenada.")

Sentiment Analysis Using Turicreate

In this notebook, I will explain how to develop sentiment analysis classifiers that are based on a bag-of-words model. Then, I will demonstrate how these classifiers can be utilized to solve Kaggle's "When Bag of Words Meets Bags of Popcorn" challenge.

Code Recipe: Creating Sentiment Classifier

Using GraphLab Turicreate it is very easy and straight foward to create a sentiment classifier based on bag-of-words model. Given a dataset stored as a CSV file, you can construct your sentiment classifier using the following code:


In [ ]:
import turicreate as tc
train_data = tc.SFrame.read_csv(traindata_path,header=True, 
                                delimiter='\t',quote_char='"', 
                                column_type_hints = {'id':str, 
                                                     'sentiment' : int, 
                                                     'review':str } )
train_data['1grams features'] = tc.text_analytics.count_ngrams(
    train_data['review'],1)
train_data['2grams features'] = tc.text_analytics.count_ngrams(
    train_data['review'],2)
cls = tc.classifier.create(train_data, target='sentiment', 
                           features=['1grams features',
                                     '2grams features'])

In the rest of this notebook, we will explain this code recipe in details, by demonstrating how this recipe can used to create IMDB movie reviews sentiment classifier.

Set up

Before we begin constructing the classifiers, we need to import some Python libraries: turicreate (tc), and IPython display utilities.


In [2]:
import turicreate as tc
from IPython.display import display
from IPython.display import Image

IMDB movies reviews Dataset

Bag of Words Meets Bags of Popcorn

Throughout this notebook, I will use Kaggle's IMDB movies reviews datasets that is available to download from the following link: https://www.kaggle.com/c/word2vec-nlp-tutorial/data. I downloaded labeledTrainData.tsv and testData.tsv files, and unzipped them to the following local files.

DeepLearningMovies

Kaggle's competition for using Google's word2vec package for sentiment analysis

https://github.com/wendykan/DeepLearningMovies


In [5]:
traindata_path = "/Users/datalab/bigdata/cjc/kaggle_popcorn_data/labeledTrainData.tsv"
testdata_path = "/Users/datalab/bigdata/cjc/kaggle_popcorn_data/testData.tsv"

Loading Data

We will load the data with IMDB movie reviews to an SFrame using SFrame.read_csv function.


In [6]:
movies_reviews_data = tc.SFrame.read_csv(traindata_path,header=True, 
                                         delimiter='\t',quote_char='"', 
                                         column_type_hints = {'id':str, 
                                                              'sentiment' : str, 
                                                              'review':str } )


Finished parsing file /Users/datalab/bigdata/cjc/kaggle_popcorn_data/labeledTrainData.tsv
Parsing completed. Parsed 100 lines in 0.318532 secs.
Finished parsing file /Users/datalab/bigdata/cjc/kaggle_popcorn_data/labeledTrainData.tsv
Parsing completed. Parsed 25000 lines in 0.499892 secs.

By using the SFrame show function, we can visualize the data and notice that the train dataset consists of 12,500 positive and 12,500 negative, and overall 24,932 unique reviews.


In [7]:
movies_reviews_data


Out[7]:
id sentiment review
5814_8 1 With all this stuff going
down at the moment with ...
2381_9 1 "The Classic War of the
Worlds" by Timothy Hines ...
7759_3 0 The film starts with a
manager (Nicholas Bell) ...
3630_4 0 It must be assumed that
those who praised this ...
9495_8 1 Superbly trashy and
wondrously unpretentious ...
8196_8 1 I dont know why people
think this is such a bad ...
7166_2 0 This movie could have
been very good, but c ...
10633_1 0 I watched this video at a
friend's house. I'm glad ...
319_1 0 A friend of mine bought
this film for £1, and ...
8713_10 1 <br /><br />This movie is
full of references. Like ...
[25000 rows x 3 columns]
Note: Only the head of the SFrame is printed.
You can use print_rows(num_rows=m, num_columns=n) to print more rows and columns.

Constructing Bag-of-Words Classifier

One of the common techniques to perform document classification (and reviews classification) is using Bag-of-Words model, in which the frequency of each word in the document is used as a feature for training a classifier. GraphLab's text analytics toolkit makes it easy to calculate the frequency of each word in each review. Namely, by using the count_ngrams function with n=1, we can calculate the frequency of each word in each review. By running the following command:


In [8]:
movies_reviews_data['1grams features'] = tc.text_analytics.count_ngrams(movies_reviews_data ['review'],1)

By running the last command, we created a new column in movies_reviews_data SFrame object. In this column each value is a dictionary object, where each dictionary's keys are the different words which appear in the corresponding review, and the dictionary's values are the frequency of each word. We can view the values of this new column using the following command.


In [9]:
movies_reviews_data#[['review','1grams features']]


Out[9]:
id sentiment review 1grams features
5814_8 1 With all this stuff going
down at the moment with ...
{'just': 3, 'sickest': 1,
'smooth': 1, 'this': 11, ...
2381_9 1 "The Classic War of the
Worlds" by Timothy Hines ...
{'year': 1, 'others': 1,
'those': 2, 'this': 1, ...
7759_3 0 The film starts with a
manager (Nicholas Bell) ...
{'hair': 1, 'bound': 1,
'this': 1, 'when': 2, ...
3630_4 0 It must be assumed that
those who praised this ...
{'crocuses': 1, 'that':
7, 'batonzilla': 1, ...
9495_8 1 Superbly trashy and
wondrously unpretentious ...
{'unshaven': 1, 'just':
1, 'in': 5, 'when': 2, ...
8196_8 1 I dont know why people
think this is such a bad ...
{'harry': 3, 'this': 4,
'of': 2, 'hurt': 1, ' ...
7166_2 0 This movie could have
been very good, but c ...
{'acting': 1,
'background': 1, 'just': ...
10633_1 0 I watched this video at a
friend's house. I'm glad ...
{'photography': 1,
'others': 1, 'zapruder': ...
319_1 0 A friend of mine bought
this film for £1, and ...
{'just': 1, 'this': 2,
'when': 1, 'as': 5, 's': ...
8713_10 1 <br /><br />This movie is
full of references. Like ...
{'peter': 1, 'ii': 1,
'full': 1, 'others': 1, ...
[25000 rows x 4 columns]
Note: Only the head of the SFrame is printed.
You can use print_rows(num_rows=m, num_columns=n) to print more rows and columns.

We are now ready to construct and evaluate the movie reviews sentiment classifier using the calculated above features. But first, to be able to perform a quick evaluation of the constructed classifier, we need to create labeled train and test datasets. We will create train and test datasets by randomly splitting the train dataset into two parts. The first part will contain 80% of the labeled train dataset and will be used as the training dataset, while the second part will contain 20% of the labeled train dataset and will be used as the testing dataset. We will create these two dataset by using the following command:


In [10]:
train_set, test_set = movies_reviews_data.random_split(0.8, seed=5)

We are now ready to create a classifier using the following command:


In [11]:
model_1 = tc.classifier.create(train_set, target='sentiment', \
                               features=['1grams features'])


PROGRESS: Creating a validation set from 5 percent of training data. This may take a while.
          You can set ``validation_set=None`` to disable validation tracking.

PROGRESS: The following methods are available for this type of problem.
PROGRESS: LogisticClassifier, SVMClassifier
PROGRESS: The returned model will be chosen according to validation accuracy.
Logistic regression:
--------------------------------------------------------
Number of examples          : 19077
Number of classes           : 2
Number of feature columns   : 1
Number of unpacked features : 68246
Number of coefficients      : 68247
Starting L-BFGS
--------------------------------------------------------
+-----------+----------+-----------+--------------+-------------------+---------------------+
| Iteration | Passes   | Step size | Elapsed Time | Training Accuracy | Validation Accuracy |
+-----------+----------+-----------+--------------+-------------------+---------------------+
| 0         | 2        | 1.000000  | 1.111660     | 0.942182          | 0.860697            |
| 1         | 4        | 1.000000  | 1.253890     | 0.968444          | 0.865672            |
| 2         | 6        | 1.000000  | 1.390344     | 0.990040          | 0.897512            |
| 3         | 7        | 1.000000  | 1.474481     | 0.992923          | 0.899502            |
| 4         | 8        | 1.000000  | 1.563669     | 0.997379          | 0.891542            |
| 9         | 13       | 1.000000  | 2.052863     | 1.000000          | 0.867662            |
+-----------+----------+-----------+--------------+-------------------+---------------------+
SVM:
--------------------------------------------------------
Number of examples          : 19077
Number of classes           : 2
Number of feature columns   : 1
Number of unpacked features : 68246
Number of coefficients    : 68247
Starting L-BFGS
--------------------------------------------------------
+-----------+----------+-----------+--------------+-------------------+---------------------+
| Iteration | Passes   | Step size | Elapsed Time | Training Accuracy | Validation Accuracy |
+-----------+----------+-----------+--------------+-------------------+---------------------+
| 0         | 2        | 1.000000  | 0.125585     | 0.942182          | 0.860697            |
| 1         | 4        | 1.000000  | 0.268260     | 0.973738          | 0.875622            |
| 2         | 5        | 1.000000  | 0.348993     | 0.989411          | 0.881592            |
| 3         | 6        | 1.000000  | 0.433099     | 0.992976          | 0.884577            |
| 4         | 7        | 1.000000  | 0.519594     | 0.996016          | 0.881592            |
| 9         | 12       | 1.000000  | 0.923521     | 0.999685          | 0.886567            |
+-----------+----------+-----------+--------------+-------------------+---------------------+
PROGRESS: Model selection based on validation accuracy:
PROGRESS: ---------------------------------------------
PROGRESS: LogisticClassifier              : 0.8676616915422886
PROGRESS: SVMClassifier                   : 0.8865671641791045
PROGRESS: ---------------------------------------------
PROGRESS: Selecting SVMClassifier based on validation set performance.

We can evaluate the performence of the classifier by evaluating it on the test dataset


In [13]:
result1 = model_1.evaluate(test_set)

In order to get an easy view of the classifier's prediction result, we define and use the following function


In [14]:
def print_statistics(result):
    print( "*" * 30)
    print( "Accuracy        : ", result["accuracy"])
    print( "Confusion Matrix: \n", result["confusion_matrix"])
print_statistics(result1)


******************************
Accuracy        :  0.8710858072387149
Confusion Matrix: 
 +--------------+-----------------+-------+
| target_label | predicted_label | count |
+--------------+-----------------+-------+
|      0       |        1        |  374  |
|      1       |        0        |  260  |
|      1       |        1        |  2133 |
|      0       |        0        |  2151 |
+--------------+-----------------+-------+
[4 rows x 3 columns]

As can be seen in the results above, in just a few relatively straight foward lines of code, we have developed a sentiment classifier that has accuracy of about ~0.88. Next, we demonstrate how we can improve the classifier accuracy even more.

Improving The Classifier

One way to improve the movie reviews sentiment classifier is to extract more meaningful features from the reviews. One method to add additional features, which might be meaningful, is to calculate the frequency of every two consecutive words in each review. To calculate the frequency of each two consecutive words in each review, as before, we will use turicreate's count_ngrams function only this time we will set n to be equal 2 (n=2) to create new column named '2grams features'.


In [15]:
movies_reviews_data['2grams features'] = tc.text_analytics.count_ngrams(movies_reviews_data['review'],2)

In [16]:
movies_reviews_data


Out[16]:
id sentiment review 1grams features 2grams features
5814_8 1 With all this stuff going
down at the moment with ...
{'just': 3, 'sickest': 1,
'smooth': 1, 'this': 11, ...
{'alone a': 1, 'most
people': 1, 'hope he' ...
2381_9 1 "The Classic War of the
Worlds" by Timothy Hines ...
{'year': 1, 'others': 1,
'those': 2, 'this': 1, ...
{'slightest resemblance':
1, 'which is': 1, 'very ...
7759_3 0 The film starts with a
manager (Nicholas Bell) ...
{'hair': 1, 'bound': 1,
'this': 1, 'when': 2, ...
{'quite boring': 1,
'packs a': 1, 'small ...
3630_4 0 It must be assumed that
those who praised this ...
{'crocuses': 1, 'that':
7, 'batonzilla': 1, ...
{'but i': 1, 'is
represented': 1, 'opera ...
9495_8 1 Superbly trashy and
wondrously unpretentious ...
{'unshaven': 1, 'just':
1, 'in': 5, 'when': 2, ...
{'unpretentious 80': 1,
'sleazy black': 1, 'd ...
8196_8 1 I dont know why people
think this is such a bad ...
{'harry': 3, 'this': 4,
'of': 2, 'hurt': 1, ' ...
{'like that': 1, 'see
this': 1, 'is such': 1, ...
7166_2 0 This movie could have
been very good, but c ...
{'acting': 1,
'background': 1, 'just': ...
{'linked to': 1, 'way
short': 1, 'good but' ...
10633_1 0 I watched this video at a
friend's house. I'm glad ...
{'photography': 1,
'others': 1, 'zapruder': ...
{'curiously ends': 1,
'several clips': 1, ...
319_1 0 A friend of mine bought
this film for £1, and ...
{'just': 1, 'this': 2,
'when': 1, 'as': 5, 's': ...
{'bob thornton': 1, 'in
the': 1, 'taking a': 1, ...
8713_10 1 <br /><br />This movie is
full of references. Like ...
{'peter': 1, 'ii': 1,
'full': 1, 'others': 1, ...
{'in the': 1, 'is a': 1,
'lorre this': 1, 'much ...
[25000 rows x 5 columns]
Note: Only the head of the SFrame is printed.
You can use print_rows(num_rows=m, num_columns=n) to print more rows and columns.

As before, we will construct and evaluate a movie reviews sentiment classifier. However, this time we will use both the '1grams features' and the '2grams features' features


In [17]:
train_set, test_set = movies_reviews_data.random_split(0.8, seed=5)
model_2 = tc.classifier.create(train_set, target='sentiment', features=['1grams features','2grams features'])
result2 = model_2.evaluate(test_set)


PROGRESS: Creating a validation set from 5 percent of training data. This may take a while.
          You can set ``validation_set=None`` to disable validation tracking.

PROGRESS: The following methods are available for this type of problem.
PROGRESS: LogisticClassifier, SVMClassifier
PROGRESS: The returned model will be chosen according to validation accuracy.
Logistic regression:
--------------------------------------------------------
Number of examples          : 19077
Number of classes           : 2
Number of feature columns   : 2
Number of unpacked features : 1206694
Number of coefficients      : 1206695
Starting L-BFGS
--------------------------------------------------------
+-----------+----------+-----------+--------------+-------------------+---------------------+
| Iteration | Passes   | Step size | Elapsed Time | Training Accuracy | Validation Accuracy |
+-----------+----------+-----------+--------------+-------------------+---------------------+
| 0         | 3        | 0.500000  | 0.884358     | 0.999266          | 0.866667            |
| 1         | 5        | 0.500000  | 1.542838     | 0.999948          | 0.866667            |
| 2         | 6        | 0.625000  | 1.909261     | 1.000000          | 0.865672            |
| 3         | 8        | 0.625000  | 2.436618     | 1.000000          | 0.864677            |
| 4         | 10       | 0.625000  | 2.971373     | 1.000000          | 0.863682            |
| 9         | 18       | 0.976563  | 5.228981     | 1.000000          | 0.862687            |
+-----------+----------+-----------+--------------+-------------------+---------------------+
SVM:
--------------------------------------------------------
Number of examples          : 19077
Number of classes           : 2
Number of feature columns   : 2
Number of unpacked features : 1206694
Number of coefficients    : 1206695
Starting L-BFGS
--------------------------------------------------------
+-----------+----------+-----------+--------------+-------------------+---------------------+
| Iteration | Passes   | Step size | Elapsed Time | Training Accuracy | Validation Accuracy |
+-----------+----------+-----------+--------------+-------------------+---------------------+
| 0         | 2        | 1.000000  | 0.710178     | 0.999266          | 0.866667            |
| 1         | 4        | 1.000000  | 1.227603     | 1.000000          | 0.865672            |
| 2         | 5        | 1.000000  | 1.524246     | 1.000000          | 0.865672            |
| 3         | 6        | 1.000000  | 1.824261     | 1.000000          | 0.865672            |
| 4         | 13       | 0.001263  | 3.080125     | 1.000000          | 0.865672            |
| 9         | 26       | 0.262737  | 6.006328     | 1.000000          | 0.865672            |
+-----------+----------+-----------+--------------+-------------------+---------------------+
PROGRESS: Model selection based on validation accuracy:
PROGRESS: ---------------------------------------------
PROGRESS: LogisticClassifier              : 0.8626865671641791
PROGRESS: SVMClassifier                   : 0.8656716417910447
PROGRESS: ---------------------------------------------
PROGRESS: Selecting SVMClassifier based on validation set performance.

In [18]:
print_statistics(result2)


******************************
Accuracy        :  0.8816592110614071
Confusion Matrix: 
 +--------------+-----------------+-------+
| target_label | predicted_label | count |
+--------------+-----------------+-------+
|      0       |        1        |  343  |
|      1       |        0        |  239  |
|      1       |        1        |  2154 |
|      0       |        0        |  2182 |
+--------------+-----------------+-------+
[4 rows x 3 columns]

Indeed, the new constructed classifier seems to be more accurate with an accuracy of about ~0.9.

Unlabeled Test File

To test how well the presented method works, we will use all the 25,000 labeled IMDB movie reviews in the train dataset to construct a classifier. Afterwards, we will utilize the constructed classifier to predict sentiment for each review in the unlabeled dataset. Lastly, we will create a submission file according to Kaggle's guidelines and submit it.


In [19]:
traindata_path = "/Users/datalab/bigdata/cjc/kaggle_popcorn_data/labeledTrainData.tsv"
testdata_path = "/Users/datalab/bigdata/cjc/kaggle_popcorn_data/testData.tsv"
#creating classifier using all 25,000 reviews
train_data = tc.SFrame.read_csv(traindata_path,header=True, delimiter='\t',quote_char='"', 
                                column_type_hints = {'id':str, 'sentiment' : int, 'review':str } )
train_data['1grams features'] = tc.text_analytics.count_ngrams(train_data['review'],1)
train_data['2grams features'] = tc.text_analytics.count_ngrams(train_data['review'],2)

cls = tc.classifier.create(train_data, target='sentiment', features=['1grams features','2grams features'])
#creating the test dataset
test_data = tc.SFrame.read_csv(testdata_path,header=True, delimiter='\t',quote_char='"', 
                               column_type_hints = {'id':str, 'review':str } )
test_data['1grams features'] = tc.text_analytics.count_ngrams(test_data['review'],1)
test_data['2grams features'] = tc.text_analytics.count_ngrams(test_data['review'],2)

#predicting the sentiment of each review in the test dataset
test_data['sentiment'] = cls.classify(test_data)['class'].astype(int)

#saving the prediction to a CSV for submission
test_data[['id','sentiment']].save("/Users/datalab/bigdata/cjc/kaggle_popcorn_data/predictions.csv", format="csv")


Finished parsing file /Users/datalab/bigdata/cjc/kaggle_popcorn_data/labeledTrainData.tsv
Parsing completed. Parsed 100 lines in 0.282738 secs.
Finished parsing file /Users/datalab/bigdata/cjc/kaggle_popcorn_data/labeledTrainData.tsv
Parsing completed. Parsed 25000 lines in 0.507212 secs.
PROGRESS: Creating a validation set from 5 percent of training data. This may take a while.
          You can set ``validation_set=None`` to disable validation tracking.

PROGRESS: The following methods are available for this type of problem.
PROGRESS: LogisticClassifier, SVMClassifier
PROGRESS: The returned model will be chosen according to validation accuracy.
Logistic regression:
--------------------------------------------------------
Number of examples          : 23750
Number of classes           : 2
Number of feature columns   : 2
Number of unpacked features : 1407914
Number of coefficients      : 1407915
Starting L-BFGS
--------------------------------------------------------
+-----------+----------+-----------+--------------+-------------------+---------------------+
| Iteration | Passes   | Step size | Elapsed Time | Training Accuracy | Validation Accuracy |
+-----------+----------+-----------+--------------+-------------------+---------------------+
| 0         | 2        | 1.000000  | 0.772874     | 0.998821          | 0.896000            |
| 1         | 4        | 1.000000  | 1.443709     | 0.999916          | 0.894400            |
| 2         | 6        | 0.648072  | 2.077022     | 0.999958          | 0.895200            |
| 3         | 8        | 0.648072  | 2.769055     | 0.999958          | 0.894400            |
| 4         | 10       | 0.648072  | 3.420360     | 0.999958          | 0.894400            |
| 9         | 22       | 0.486054  | 6.816458     | 1.000000          | 0.892800            |
+-----------+----------+-----------+--------------+-------------------+---------------------+
SVM:
--------------------------------------------------------
Number of examples          : 23750
Number of classes           : 2
Number of feature columns   : 2
Number of unpacked features : 1407914
Number of coefficients    : 1407915
Starting L-BFGS
--------------------------------------------------------
+-----------+----------+-----------+--------------+-------------------+---------------------+
| Iteration | Passes   | Step size | Elapsed Time | Training Accuracy | Validation Accuracy |
+-----------+----------+-----------+--------------+-------------------+---------------------+
| 0         | 2        | 1.000000  | 0.724382     | 0.998821          | 0.896000            |
| 1         | 4        | 1.000000  | 1.284643     | 0.999916          | 0.896000            |
| 2         | 5        | 1.000000  | 1.634216     | 0.999958          | 0.896000            |
| 3         | 6        | 1.000000  | 2.002875     | 0.999958          | 0.896000            |
| 4         | 13       | 0.000338  | 3.462338     | 1.000000          | 0.895200            |
| 9         | 38       | 4.080042  | 9.019969     | 1.000000          | 0.895200            |
+-----------+----------+-----------+--------------+-------------------+---------------------+
PROGRESS: Model selection based on validation accuracy:
PROGRESS: ---------------------------------------------
PROGRESS: LogisticClassifier              : 0.8928
PROGRESS: SVMClassifier                   : 0.8952
PROGRESS: ---------------------------------------------
PROGRESS: Selecting SVMClassifier based on validation set performance.
Finished parsing file /Users/datalab/bigdata/cjc/kaggle_popcorn_data/testData.tsv
Parsing completed. Parsed 100 lines in 0.313905 secs.
Finished parsing file /Users/datalab/bigdata/cjc/kaggle_popcorn_data/testData.tsv
Parsing completed. Parsed 25000 lines in 0.560208 secs.

We then submitted the predictions.csv file to the Kaggle challange website and scored AUC of about 0.88.

Further Readings

Further reading materials can be found in the following links:

http://en.wikipedia.org/wiki/Bag-of-words_model

https://dato.com/products/create/docs/generated/graphlab.SFrame.html

https://dato.com/products/create/docs/graphlab.toolkits.classifier.html

https://www.kaggle.com/c/word2vec-nlp-tutorial/details/part-1-for-beginners-bag-of-words

Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. (2011). "Learning Word Vectors for Sentiment Analysis." The 49th Annual Meeting of the Association for Computational Linguistics (ACL 2011).