def __init__(self, trainset=[]):
print ('Loading training modules')
self.bag_of_words = []
self.vectorizer = DictVectorizer(dtype=int, sparse=True)
self.encoder = LabelEncoder()
self.lexicon_classifier = LexiconClassifier()
self.classifier = LinearSVC(C=0.005)
self.train(trainset)
def __init__(self, tweets=[]):
# initialize internal variables
self.rules_classifier = RulesClassifier()
self.lexicon_classifier = LexiconClassifier()
self.ml_classifier = None
# if the ML model has been generated, load the model from model.pkl
if sys.version_info >= (3,0):
if os.path.exists('model_python3.pkl'):
print ('Reading the model from model_python3.pkl')
self.ml_classifier = pickle.load(open('model_python3.pkl','rb'))
else:
if os.path.exists('model_python2.pkl'):
print ('Reading the model from model_python2.pkl')
self.ml_classifier = pickle.load(open('model_python2.pkl','rb'))
if self.ml_classifier == None:
# Preprocess the data and train a new model
print ('Preprocessing the training data')
tweet_messages = [tweet_message for tweet_message,label in tweets]
tweet_labels = [label for tweet_message,label in tweets]
# preproces all the tweet_messages (Tokenization, POS and normalization)
tweet_tokens = pre_process(tweet_messages)
# compile a trainset with tweek_tokens and labels (positive,
# negative or neutral)
trainset = [(tweet_tokens[i],tweet_labels[i]) for i in range(len(tweets))]
# initialize the classifier and train it
classifier = MachineLearningClassifier(trainset)
# dump the model into de pickle
python_version = sys.version_info[0]
model_name = 'model_python' + str(python_version) + '.pkl'
print ('Saving the trained model at ' + model_name)
pickle.dump(classifier, open(model_name, 'wb'))
self.ml_classifier = classifier
def __init__(self, trainset=[]):
self.rules_classifier = RulesClassifier()
self.lexicon_classifier = LexiconClassifier()
self.ml_classifier = MachineLearningClassifier(trainset)
class TwitterHybridClassifier(object):
def __init__(self, trainset=[]):
self.rules_classifier = RulesClassifier()
self.lexicon_classifier = LexiconClassifier()
self.ml_classifier = MachineLearningClassifier(trainset)
# Apply the classifier over a tweet message in String format
def classify(self,tweet_text):
# 0. Pre-process the text (emoticons, misspellings, tagger)
tweet_text = pre_process(tweet_text)
# 1. Rule-based classifier. Look for emoticons basically
positive_score,negative_score = self.rules_classifier.classify(tweet_text)
rules_score = positive_score + negative_score
# 1. Apply the rules, If any found, classify the tweet here. If none found, continue for the lexicon classifier.
if rules_score != 0:
if rules_score > 0:
sentiment = 'positive'
else:
sentiment = 'negative'
return sentiment
# 2. Lexicon-based classifier
positive_score, negative_score = self.lexicon_classifier.classify(tweet_text)
lexicon_score = positive_score + negative_score
# 2. Apply lexicon classifier, If the lexicon score is
# 0 (strictly neutral), >3 (positive with confidence) or
# <3 (negative with confidence), classify the tweet here. If not,
# continue for the SVM classifier
if lexicon_score == 0:
sentiment = 'neutral'
return sentiment
if lexicon_score >= 3:
sentiment = 'positive'
return sentiment
if lexicon_score <= -3:
sentiment = 'negative'
return sentiment
# 3. Machine learning based classifier - used the training set to define the best features to classify new instances
scores = self.ml_classifier.classify(tweet_text)
positive_conf = scores[0][1]
negative_conf = scores[1][1]
neutral_conf = scores[2][1]
# 3. Apply machine learning classifier, If positive or negative
# confidence (probability) is >=0.3, classify with the sentiment.
# Otherwise, classify as neutral
if positive_conf >= 0.3 and negative_conf < positive_conf:
sentiment = 'positive'
elif negative_conf >= 0.3:
sentiment = 'negative'
else:
sentiment = 'neutral'
return sentiment
class TwitterHybridClassifier(object):
def __init__(self, tweets=[]):
# initialize internal variables
self.rules_classifier = RulesClassifier()
self.lexicon_classifier = LexiconClassifier()
self.ml_classifier = None
# if the ML model has been generated, load the model from model.pkl
if sys.version_info >= (3,0):
if os.path.exists('model_python3.pkl'):
print ('Reading the model from model_python3.pkl')
self.ml_classifier = pickle.load(open('model_python3.pkl','rb'))
else:
if os.path.exists('model_python2.pkl'):
print ('Reading the model from model_python2.pkl')
self.ml_classifier = pickle.load(open('model_python2.pkl','rb'))
if self.ml_classifier == None:
# Preprocess the data and train a new model
print ('Preprocessing the training data')
tweet_messages = [tweet_message for tweet_message,label in tweets]
tweet_labels = [label for tweet_message,label in tweets]
# preproces all the tweet_messages (Tokenization, POS and normalization)
tweet_tokens = pre_process(tweet_messages)
# compile a trainset with tweek_tokens and labels (positive,
# negative or neutral)
trainset = [(tweet_tokens[i],tweet_labels[i]) for i in range(len(tweets))]
# initialize the classifier and train it
classifier = MachineLearningClassifier(trainset)
# dump the model into de pickle
python_version = sys.version_info[0]
model_name = 'model_python' + str(python_version) + '.pkl'
print ('Saving the trained model at ' + model_name)
pickle.dump(classifier, open(model_name, 'wb'))
self.ml_classifier = classifier
# Apply the classifier over a tweet message in String format
def classify(self,tweet_text):
# 0. Pre-process the teets (tokenization, tagger, normalizations)
tweet_tokens_list = []
print ('Preprocessing the string')
# pre-process the tweets
tweet_tokens_list = pre_process([tweet_text])
predictions = []
total_tweets = len(tweet_tokens_list)
# iterate over the tweet_tokens
for index, tweet_tokens in enumerate(tweet_tokens_list):
# 1. Rule-based classifier. Look for emoticons basically
positive_score,negative_score = self.rules_classifier.classify(tweet_tokens)
# 1. Apply the rules, If any found, classify the tweet here. If none found, continue for the lexicon classifier.
if positive_score >= 1 and negative_score == 0:
sentiment = ('positive','RB')
predictions.append(sentiment)
continue
elif positive_score == 0 and negative_score <= -1:
sentiment = ('negative','RB')
predictions.append(sentiment)
continue
# 2. Lexicon-based classifier
positive_score, negative_score = self.lexicon_classifier.classify(tweet_tokens)
lexicon_score = positive_score + negative_score
# 2. Apply lexicon classifier,
# If in the threshold classify the tweet here. If not, continue for the ML classifier
if positive_score >= 1 and negative_score == 0:
sentiment = ('positive','LB')
predictions.append(sentiment)
continue
elif negative_score <= -2:
sentiment = ('negative','LB')
predictions.append(sentiment)
continue
# 3. Machine learning based classifier - used the Train+Dev set sto define the best features to classify new instances
result = self.ml_classifier.classify(tweet_tokens)
positive_conf = result['positive']
negative_conf = result['negative']
neutral_conf = result['neutral']
if negative_conf >= -0.4:
sentiment = ('negative','ML')
elif positive_conf > neutral_conf:
sentiment = ('positive','ML')
else:
sentiment = ('neutral','ML')
predictions.append(sentiment)
return predictions
# Apply the classifier in batch over a list of tweet messages in String format
def classify_batch(self,tweet_texts):
# 0. Pre-process the teets (tokenization, tagger, normalizations)
tweet_tokens_list = []
if len(tweet_texts) == 0:
return tweet_tokens_list
print ('Preprocessing the test data')
# pre-process the tweets
tweet_tokens_list = pre_process(tweet_texts)
predictions = []
total_tweets = len(tweet_tokens_list)
# iterate over the tweet_tokens
for index, tweet_tokens in enumerate(tweet_tokens_list):
print('Testing for tweet n. {}/{}'.format(index+1,total_tweets))
# 1. Rule-based classifier. Look for emoticons basically
positive_score,negative_score = self.rules_classifier.classify(tweet_tokens)
# 1. Apply the rules, If any found, classify the tweet here. If none found, continue for the lexicon classifier.
if positive_score >= 1 and negative_score == 0:
sentiment = ('positive','RB')
predictions.append(sentiment)
continue
elif positive_score == 0 and negative_score <= -1:
sentiment = ('negative','RB')
predictions.append(sentiment)
continue
# 2. Lexicon-based classifier
positive_score, negative_score = self.lexicon_classifier.classify(tweet_tokens)
lexicon_score = positive_score + negative_score
# 2. Apply lexicon classifier,
# If in the threshold classify the tweet here. If not, continue for the ML classifier
if positive_score >= 1 and negative_score == 0:
sentiment = ('positive','LB')
predictions.append(sentiment)
continue
elif negative_score <= -2:
sentiment = ('negative','LB')
predictions.append(sentiment)
continue
# 3. Machine learning based classifier - used the Train+Dev set sto define the best features to classify new instances
result = self.ml_classifier.classify(tweet_tokens)
positive_conf = result['positive']
negative_conf = result['negative']
neutral_conf = result['neutral']
if negative_conf >= -0.4:
sentiment = ('negative','ML')
elif positive_conf > neutral_conf:
sentiment = ('positive','ML')
else:
sentiment = ('neutral','ML')
predictions.append(sentiment)
return predictions
# Output Individual scores for each method
def output_individual_scores(self,tweets):
tweet_texts = [tweet_message for tweet_message,label in tweets]
tweet_labels = [label for tweet_message,label in tweets]
# write the log
fp = codecs.open('individual_scores.tab','w',encoding='utf8')
line = 'pos_score_rule\tneg_score_rule\tpos_score_lex\tneg_score_lex\tpos_conf\tneg_conf\tneutral_conf\tclass\tmessage\n'
fp.write(line)
# 0. Pre-process the text (emoticons, misspellings, tagger)
tweet_tokens_list = None
tweet_tokens_list = pre_process(tweet_texts)
predictions = []
for index,tweet_tokens in enumerate(tweet_tokens_list):
line = ''
# 1. Rule-based classifier. Look for emoticons basically
positive_score,negative_score = self.rules_classifier.classify(tweet_tokens)
line += str(positive_score) + '\t' + str(negative_score) + '\t'
# 2. Lexicon-based classifier (using url_score obtained from RulesClassifier)
positive_score, negative_score = self.lexicon_classifier.classify(tweet_tokens)
lexicon_score = positive_score + negative_score
line += str(positive_score) + '\t' + str(negative_score) + '\t'
# 3. Machine learning based classifier - used the training set to define the best features to classify new instances
result = self.ml_classifier.decision_function(tweet_tokens)
line += str(result['positive']) + '\t' + str(result['negative']) + '\t' + str(result['neutral']) + '\t'
line += tweet_labels[index] + '\t"' + tweet_texts[index].replace('"','') + '"\n'
fp.write(line)
print('Indivual score saved in the file: individual_scores.tab')
class MachineLearningClassifier(object):
# Constructor
def __init__(self, trainset=[]):
print ('Loading training modules')
self.bag_of_words = []
self.vectorizer = DictVectorizer(dtype=int, sparse=True)
self.encoder = LabelEncoder()
self.lexicon_classifier = LexiconClassifier()
self.classifier = LinearSVC(C=0.005)
self.train(trainset)
# Extract features for ML process
# Some insights from http://aclweb.org/anthology/S/S13/S13-2053.pdf
def extract_features(self, tweet_tokens):
if len(self.bag_of_words) == 0:
print('Bag-of-Words empty!')
unigrams = [w.lower() for w,t in tweet_tokens]
tokens = unigrams
tokens += ['_'.join(b) for b in bigrams(unigrams)]
tokens += ['_'.join(t) for t in trigrams(unigrams)]
tokens += [t1 + '_*_' + t3 for t1,t2,t3 in trigrams(unigrams)]
tweet_tags = [tag for token, tag in tweet_tokens]
feature_set = {}
# 1st set of features: bag-of-words
for token in set(tokens).intersection(self.bag_of_words):
feature_set['has_'+token] = True
# 2nd set of features: the count for each tag type present in the message
# Tweet_nlp taget. Info:
# http://www.ark.cs.cmu.edu/TweetNLP/annot_guidelines.pdf
for tag in ['N','O','^','S','Z','V','A','R','!','D','P','&','T','X','#','@','~','U','E','$',',','G','L','M','Y']:
feature_set['num_'+tag] = sum([1 for t in tweet_tags if t == tag])
# 3rd feature: negation is present?
negators = set(LexiconClassifier().read_negation_words())
if len(negators.intersection(set(tokens))) > 0:
feature_set['has_negator'] = True
# 4th feature: character ngrams
regexp = re.compile(r"([a-z])\1{2,}")
feature_set['has_char_ngrams'] = False
for token,tag in tweet_tokens:
if regexp.search(token):
feature_set['has_char_ngrams'] = True
break
# 5th feature: punctuation ngrams
regexp = re.compile(r"([!\?])\1{2,}")
feature_set['has_punct_ngrams'] = False
for token,tag in tweet_tokens:
if regexp.search(token):
feature_set['has_punct_ngrams'] = True
break
# 6th feature: the number of all upper cased words
feature_set['num_all_caps'] = sum([1 for token,tag in tweet_tokens if token.isupper() and len(token)>=3])
# 7th and 8th feature: the positive and negative score from lexicon
# classifier (i.e., number of positive and negative words from lexicon)
positive_score, negative_score = self.lexicon_classifier.classify(tweet_tokens)
feature_set['pos_lexicon'] = positive_score
feature_set['neg_lexicon'] = -1 * negative_score
return feature_set
# train the classifier
# Tweets argument must be a list of dictionaries. Each dictionary must
# have the keys ['MESSAGE'] and ['SENTIMENT'] with the message string and
# the classificationclass, respectively.
def train(self,tweets):
# 1st step: build the bag-of-words model
tweet_tokens_list = [tweet_tokens for tweet_tokens,label in tweets]
tokens = []
print('Computing the trainset vocabulary of n-grams')
for tweet_tokens in tweet_tokens_list:
unigrams = [w.lower() for w,t in tweet_tokens]
tokens += unigrams
tokens += ['_'.join(b) for b in bigrams(unigrams)]
tokens += ['_'.join(t) for t in trigrams(unigrams)]
tokens += [t1 + '_*_' + t3 for t1,t2,t3 in trigrams(unigrams)]
# build the bag-of-words list using all the tokens
self.bag_of_words = set(tokens)
data = list()
total_tweets = len(tweets)
features_list = list()
for index,(tweet_tokens,label) in enumerate(tweets):
print('Training for tweet n. {}/{}'.format(index+1,total_tweets))
features_list.append(self.extract_features(tweet_tokens))
# Train a SVM classifier
#data = self.vectorizer.fit_transform([features for features,label in self.train_set_features])
print('Vectorizing the features')
data = self.vectorizer.fit_transform(features_list)
target = self.encoder.fit_transform([label for tweet_tokens,label in tweets])
print('Building the model')
self.classifier.fit(data, target)
# classify a new message. Return the scores (probabilities) for each
# classification class
def classify(self, tweet_tokens):
data = self.vectorizer.transform(self.extract_features(tweet_tokens))
probs = self.classifier.decision_function(data)
classes = self.encoder.classes_
return {classes.item(i): probs.item(i) for i in range(len(classes))}
# return the probability of classification into one of the three classes
def decision_function(self, tweet_tokens):
data = self.vectorizer.transform(self.extract_features(tweet_tokens))
probs = self.classifier.decision_function(data)
classes = self.encoder.classes_
return {classes.item(i): probs.item(i) for i in range(len(classes))}
