def build_x_y_t(args):
time_t = build_time_t(args)
base_x_t = build_base_x_t(args)
x_t = extend_function.extend_f_t(time_t,
base_x_t,
args.start_t,
args.end_t,
extension_args=args)
time_t = build_time_t(args)
base_y_t = build_base_y_t(args)
y_t = extend_function.extend_f_t(time_t,
base_y_t,
args.start_t,
args.end_t,
extension_args=args)
reg_x_t, reg_y_t = regularization.regularize(x_t, y_t, args)
x_scale = args.x_scale
def scale_x_t(t):
return reg_x_t(t) * x_scale
y_scale = args.y_scale
def scale_y_t(t):
return reg_y_t(t) * y_scale
return scale_x_t, scale_y_t
def test_capped_linear_function(self):
"""
Tests the result of regularization.regularize for a capped linear regularization
"""
reg_args = Namespace(
regularization_linear_cap=5.0,
regularization_method=regularization.Regularization.CAPPED_LINEAR)
x_t = lambda t: t
y_t = lambda t: 0
reg_x_t, reg_y_t = regularization.regularize(x_t, y_t, reg_args)
self.assertAlmostEqual(2.0, reg_x_t(2))
self.assertAlmostEqual(0.0, reg_y_t(2))
self.assertAlmostEqual(5.0, reg_x_t(20))
self.assertAlmostEqual(0.0, reg_y_t(20))
x_t = lambda t: t * 3 / 5
y_t = lambda t: t * 4 / 5
reg_x_t, reg_y_t = regularization.regularize(x_t, y_t, reg_args)
self.assertAlmostEqual(1.2, reg_x_t(2))
self.assertAlmostEqual(1.6, reg_y_t(2))
self.assertAlmostEqual(3.0, reg_x_t(20))
self.assertAlmostEqual(4.0, reg_y_t(20))
def test_hyperbolic_function(self):
"""
Tests the result of regularization.regularize for a hyperbolic regularization
"""
reg_args = Namespace(
regularization_x_trans=-1.0,
regularization_y_trans=-2.0,
regularization_slope=2,
regularization_method=regularization.Regularization.HYPERBOLIC)
self.assertEqual(
"Hyperbolic regularization: y = 2 (0.5 (x + 1.0) + (0.25 + 0.25 (x + 1.0)^{2})^{0.5}) - 2.0",
regularization.hyperbolic_function_string(reg_args))
x_t = lambda t: t
y_t = lambda t: 0
reg_x_t, reg_y_t = regularization.regularize(x_t, y_t, reg_args)
self.assertAlmostEqual(2.236067977, reg_x_t(1))
self.assertAlmostEqual(0, reg_y_t(1))
x_t = lambda t: t * 3 / 5
y_t = lambda t: t * 4 / 5
reg_x_t, reg_y_t = regularization.regularize(x_t, y_t, reg_args)
self.assertAlmostEqual(2.236067977 * 3 / 5, reg_x_t(1))
self.assertAlmostEqual(2.236067977 * 4 / 5, reg_y_t(1))
def main(messages_test):
#tokenize all messages
tokens_test = tokenize(messages_test)
#compute pos tags for all messages
pos_tags_test = arktagger.pos_tag_list(messages_test)
#compute pos tag bigrams
pos_bigrams_test = getBigrams(pos_tags_test)
#compute pos tag trigrams
pos_trigrams_test = getTrigrams(pos_tags_test)
now = time.time()
#load scores
pos_tags_scores_neutral, pos_tags_scores_positive, pos_tags_scores_negative, pos_bigrams_scores_neutral, pos_bigrams_scores_positive, pos_bigrams_scores_negative, pos_trigrams_scores_neutral, pos_trigrams_scores_positive, pos_trigrams_scores_negative, mpqaScores = loadScores(
)
#load lexicons
negationList, slangDictionary, lexicons, mpqa_lexicons = loadLexiconsFromFile(
)
#load clusters
clusters = loadClustersFromFile()
print "Resources loaded"
#load Glove embeddings
d = 25
glove = loadGlove(d)
#Subjectivity Detection Features
#SD1 features
features_test_1 = features.getFeatures(
messages_test, tokens_test, pos_tags_test, slangDictionary, lexicons,
mpqa_lexicons, pos_bigrams_test, pos_trigrams_test,
pos_bigrams_scores_negative, pos_bigrams_scores_positive,
pos_trigrams_scores_negative, pos_trigrams_scores_positive,
pos_tags_scores_negative, pos_tags_scores_positive, mpqaScores,
negationList, clusters, pos_bigrams_scores_neutral,
pos_trigrams_scores_neutral, pos_tags_scores_neutral)
#SD2 features
features_test_2 = []
for i in range(0, len(messages_test)):
features_test_2.append(glove.findCentroid(tokens_test[i]))
features_test_2 = np.array(features_test_2)
#regularize features
print "After Reg"
features_test_1 = regularization.regularize(features_test_1)
print features_test_1
features_test_2 = regularization.regularizeHorizontally(features_test_2)
print features_test_2
#load SD classifiers
with open('resources/sd_models.pkl', 'rb') as input:
sd1 = pickle.load(input)
sd2 = pickle.load(input)
#get confidence scores
test_confidence_1 = sd1.decision_function(features_test_1)
test_confidence_2 = sd2.decision_function(features_test_2)
#normalize confidence scores
softmax = lambda x: 1 / (1. + math.exp(-x))
test_confidence_1 = [softmax(conf) for conf in test_confidence_1]
test_confidence_2 = [softmax(conf) for conf in test_confidence_2]
test_confidence_1 = np.array(test_confidence_1)
test_confidence_2 = np.array(test_confidence_2)
#Sentiment Polarity Features (append confidence scores to SD features)
#SP1 features
features_test_1 = np.hstack(
(features_test_1,
test_confidence_1.reshape(test_confidence_1.shape[0], 1)))
#SP2 features
features_test_2 = np.hstack(
(features_test_2,
test_confidence_2.reshape(test_confidence_2.shape[0], 1)))
#load SP classifiers
with open('resources/sp_models.pkl', 'rb') as input:
sp1 = pickle.load(input)
sp2 = pickle.load(input)
#get confidence scores of every system
confidence1 = sp1.decision_function(features_test_1)
confidence2 = sp2.decision_function(features_test_2)
for i in range(0, confidence1.shape[0]):
for j in range(0, confidence1.shape[1]):
confidence1[i][j] = softmax(confidence1[i][j])
for i in range(0, confidence2.shape[0]):
for j in range(0, confidence2.shape[1]):
confidence2[i][j] = softmax(confidence2[i][j])
#ensemble confidence scores with weight W
W = 0.66
confidence = confidence1 * W + confidence2 * (1 - W)
print "confidence"
print confidence
#get final prediction
prediction = [np.argmax(x) - 1 for x in confidence]
prediction = np.array(prediction)
print "Prediction\n"
for i in range(0, prediction.shape[0]):
if prediction[i] == -1:
pol = "Negative"
elif prediction[i] == 0:
pol = "Neutral"
elif prediction[i] == 1:
pol = "Positive"
print "Message : " + messages_test[i] + "Polarity : " + pol + "\n"
#accuracy and number of wrong line
count_t = 0
num_f = []
num_f1 = []
num_f2 = []
num_f3 = []
num_f4 = []
num_f5 = []
num_f6 = []
senti_t = []
prediction_f = []
for j in range(0, senti.shape[0]):
if senti[j] == prediction[j]:
count_t = count_t + 1
else:
num_f.append(j)
senti_t.append(senti[j])
prediction_f.append(prediction[j])
print count_t * 100.00 / count
plt.scatter(num_f, senti_t, c='r')
plt.scatter(num_f, prediction_f, c='b')
plt.show()
#compare value of sentiment -1 0 1
for j in range(0, senti.shape[0]):
if senti[j] == 1:
if prediction[j] == 0:
num_f1.append(j)
elif prediction[j] == -1:
num_f2.append(j)
if senti[j] == 0:
if prediction[j] == 1:
num_f3.append(j)
elif prediction[j] == -1:
num_f4.append(j)
if senti[j] == -1:
if prediction[j] == 1:
num_f5.append(j)
elif prediction[j] == 0:
num_f6.append(j)
print num_f1, len(num_f1)
print num_f2, len(num_f2)
print num_f3, len(num_f3)
print num_f4, len(num_f4)
print num_f5, len(num_f5)
print num_f6, len(num_f6)
def main(messages_test):
#tokenize all messages
tokens_test = tokenize(messages_test)
#compute pos tags for all messages
pos_tags_test = arktagger.pos_tag_list(messages_test)
#compute pos tag bigrams
pos_bigrams_test = getBigrams(pos_tags_test)
#compute pos tag trigrams
pos_trigrams_test = getTrigrams(pos_tags_test)
now = time.time()
#load scores
pos_tags_scores_neutral, pos_tags_scores_positive, pos_tags_scores_negative, pos_bigrams_scores_neutral, pos_bigrams_scores_positive, pos_bigrams_scores_negative, pos_trigrams_scores_neutral, pos_trigrams_scores_positive, pos_trigrams_scores_negative, mpqaScores = loadScores()
#load lexicons
negationList, slangDictionary, lexicons, mpqa_lexicons = loadLexiconsFromFile()
#load clusters
clusters = loadClustersFromFile()
print "Resources loaded"
#load Glove embeddings
d = 200
glove = loadGlove(d)
#Subjectivity Detection Features
#SD1 features
features_test_1 = features.getFeatures(messages_test,tokens_test,pos_tags_test,slangDictionary,lexicons,mpqa_lexicons,pos_bigrams_test,pos_trigrams_test,pos_bigrams_scores_negative,pos_bigrams_scores_positive,pos_trigrams_scores_negative,pos_trigrams_scores_positive,pos_tags_scores_negative,pos_tags_scores_positive,mpqaScores,negationList,clusters,pos_bigrams_scores_neutral,pos_trigrams_scores_neutral,pos_tags_scores_neutral)
#SD2 features
features_test_2=[]
for i in range(0,len(messages_test)):
features_test_2.append(glove.findCentroid(tokens_test[i]))
features_test_2 = np.array(features_test_2)
#regularize features
features_test_1=regularization.regularize(features_test_1)
features_test_2 = regularization.regularizeHorizontally(features_test_2)
#load SD classifiers
with open('resources/sd_models.pkl', 'rb') as input:
sd1 = pickle.load(input)
sd2 = pickle.load(input)
#get confidence scores
test_confidence_1 = sd1.decision_function(features_test_1)
test_confidence_2 = sd2.decision_function(features_test_2)
#normalize confidence scores
softmax = lambda x: 1 / (1. + math.exp(-x))
test_confidence_1 = [softmax(conf) for conf in test_confidence_1]
test_confidence_2 = [softmax(conf) for conf in test_confidence_2]
test_confidence_1 = np.array(test_confidence_1)
test_confidence_2 = np.array(test_confidence_2)
#Sentiment Polarity Features (append confidence scores to SD features)
#SP1 features
features_test_1 = np.hstack((features_test_1,test_confidence_1.reshape(test_confidence_1.shape[0],1)))
#SP2 features
features_test_2 = np.hstack((features_test_2,test_confidence_2.reshape(test_confidence_2.shape[0],1)))
#load SP classifiers
with open('resources/sp_models.pkl', 'rb') as input:
sp1 = pickle.load(input)
sp2 = pickle.load(input)
#get confidence scores of every system
confidence1 = sp1.decision_function(features_test_1)
confidence2 = sp2.decision_function(features_test_2)
for i in range(0,confidence1.shape[0]):
for j in range(0,confidence1.shape[1]):
confidence1[i][j] = softmax(confidence1[i][j])
for i in range(0,confidence2.shape[0]):
for j in range(0,confidence2.shape[1]):
confidence2[i][j] = softmax(confidence2[i][j])
#ensemble confidence scores with weight W
W=0.66
confidence = confidence1*W + confidence2*(1-W)
#get final prediction
prediction = [np.argmax(x)-1 for x in confidence]
prediction = np.array(prediction)
print "Prediction\n"
for i in range(0, prediction.shape[0]):
if prediction[i] == -1:
pol = "Negative"
elif prediction[i] == 0:
pol = "Neutral"
else:
pol = "Positive"
print "Message : " + messages_test[i]+"Polarity : "+pol+"\n"
def classify(messages_train,labels_train,messages_test,process_messages_train,process_messages_test,tokens_train,tokens_test,process_tokens_train,process_tokens_test,pos_tags_train,pos_tags_test,negationList,clusters,slangDictionary,lexicons,mpqa_lexicons):
# 0 - negative messages
# 1 - positives messages
labels_train = [0 if x=="negative" else 1 for x in labels_train]
#compute pos tag bigrams for all messages
pos_bigrams_train = getBigrams(pos_tags_train)
pos_bigrams_test = getBigrams(pos_tags_test)
#compute pos tag trigrams for all messages
pos_trigrams_train = getTrigrams(pos_tags_train)
pos_trigrams_test = getTrigrams(pos_tags_test)
#get the unique pos bigrams and trigrams from training set
unique_pos_tags = getPosTagsSet(pos_tags_train)
unique_bigrams = getBigramsSet(pos_bigrams_train)
unique_trigrams= getTrigramsSet(pos_trigrams_train)
#calculate pos bigrams score for all categories
#both dictionaries will be used for training and testing (cannot create new for testing because we don't know the labels of the new messages)
pos_tags_scores_negative = posTagsScore(unique_pos_tags,0,pos_tags_train,labels_train)
pos_tags_scores_positive = posTagsScore(unique_pos_tags,1,pos_tags_train,labels_train)
#calculate pos bigrams score for all categories
#both dictionaries will be used for training and testing (cannot create new for testing because we don't know the labels of the new messages)
pos_bigrams_scores_negative = posBigramsScore(unique_bigrams,0,pos_bigrams_train,labels_train)
pos_bigrams_scores_positive = posBigramsScore(unique_bigrams,1,pos_bigrams_train,labels_train)
#calculate pos bigrams score for all categories
#both dictionaries will be used for training and testing (cannot create new for testing because we don't know the labels of the new messages)
pos_trigrams_scores_negative = posTrigramsScore(unique_trigrams,0,pos_trigrams_train,labels_train)
pos_trigrams_scores_positive = posTrigramsScore(unique_trigrams,1,pos_trigrams_train,labels_train)
#assign a precision and F1 score to each word of to all mpqa lexicons
mpqaScores = getScores(mpqa_lexicons,process_messages_train,labels_train)
#get features from train messages
features_train = features.getFeatures(messages_train,process_messages_train,tokens_train,process_tokens_train,pos_tags_train,slangDictionary,lexicons,mpqa_lexicons,pos_bigrams_train,pos_trigrams_train,pos_bigrams_scores_negative,pos_bigrams_scores_positive,pos_trigrams_scores_negative,pos_trigrams_scores_positive,pos_tags_scores_negative,pos_tags_scores_positive,mpqaScores,negationList,clusters)
#regularize train features
features_train=regularization.regularize(features_train)
#get features from test messages
features_test = features.getFeatures(messages_test,process_messages_test,tokens_test,process_tokens_test,pos_tags_test,slangDictionary,lexicons,mpqa_lexicons,pos_bigrams_test,pos_trigrams_test,pos_bigrams_scores_negative,pos_bigrams_scores_positive,pos_trigrams_scores_negative,pos_trigrams_scores_positive,pos_tags_scores_negative,pos_tags_scores_positive,mpqaScores,negationList,clusters)
#regularize test features
features_test=regularization.regularize(features_test)
#feature selection
#features_train, features_test = selection.feature_selection(features_train,labels_train,features_test,1150)
#C parameter of SVM
C = 0.001953125
#C = 19.3392161013
#train classifier and return trained model
#model = LogisticRegression.train(features_train,labels_train)
model = SVM.train(features_train,labels_train,c=C,k="linear")
#predict labels
#prediction = LogisticRegression.predict(features_test,model)
prediction = SVM.predict(features_test,model)
return prediction
def main(messages_test):
#tokenize all messages
tokens_test = tokenize(messages_test)
#compute pos tags for all messages
pos_tags_test = arktagger.pos_tag_list(messages_test)
#compute pos tag bigrams
pos_bigrams_test = getBigrams(pos_tags_test)
#compute pos tag trigrams
pos_trigrams_test = getTrigrams(pos_tags_test)
now = time.time()
#load scores
pos_tags_scores_neutral, pos_tags_scores_positive, pos_tags_scores_negative, pos_bigrams_scores_neutral, pos_bigrams_scores_positive, pos_bigrams_scores_negative, pos_trigrams_scores_neutral, pos_trigrams_scores_positive, pos_trigrams_scores_negative, mpqaScores = loadScores(
)
#load lexicons
negationList, slangDictionary, lexicons, mpqa_lexicons = loadLexiconsFromFile(
)
#load clusters
clusters = loadClustersFromFile()
print "Resources loaded"
#load Glove embeddings
d = 200
glove = loadGlove(d)
#Subjectivity Detection Features
#SD1 features
features_test_1 = features.getFeatures(
messages_test, tokens_test, pos_tags_test, slangDictionary, lexicons,
mpqa_lexicons, pos_bigrams_test, pos_trigrams_test,
pos_bigrams_scores_negative, pos_bigrams_scores_positive,
pos_trigrams_scores_negative, pos_trigrams_scores_positive,
pos_tags_scores_negative, pos_tags_scores_positive, mpqaScores,
negationList, clusters, pos_bigrams_scores_neutral,
pos_trigrams_scores_neutral, pos_tags_scores_neutral)
#SD2 features
features_test_2 = []
for i in range(0, len(messages_test)):
features_test_2.append(glove.findCentroid(tokens_test[i]))
features_test_2 = np.array(features_test_2)
#regularize features
features_test_1 = regularization.regularize(features_test_1)
features_test_2 = regularization.regularizeHorizontally(features_test_2)
#load SD classifiers
with open('resources/sd_models.pkl', 'rb') as input:
sd1 = pickle.load(input)
sd2 = pickle.load(input)
#get confidence scores
test_confidence_1 = sd1.decision_function(features_test_1)
test_confidence_2 = sd2.decision_function(features_test_2)
#normalize confidence scores
softmax = lambda x: 1 / (1. + math.exp(-x))
test_confidence_1 = [softmax(conf) for conf in test_confidence_1]
test_confidence_2 = [softmax(conf) for conf in test_confidence_2]
test_confidence_1 = np.array(test_confidence_1)
test_confidence_2 = np.array(test_confidence_2)
#Sentiment Polarity Features (append confidence scores to SD features)
#SP1 features
features_test_1 = np.hstack(
(features_test_1,
test_confidence_1.reshape(test_confidence_1.shape[0], 1)))
#SP2 features
features_test_2 = np.hstack(
(features_test_2,
test_confidence_2.reshape(test_confidence_2.shape[0], 1)))
#load SP classifiers
with open('resources/sp_models.pkl', 'rb') as input:
sp1 = pickle.load(input)
sp2 = pickle.load(input)
#get confidence scores of every system
confidence1 = sp1.decision_function(features_test_1)
confidence2 = sp2.decision_function(features_test_2)
for i in range(0, confidence1.shape[0]):
for j in range(0, confidence1.shape[1]):
confidence1[i][j] = softmax(confidence1[i][j])
for i in range(0, confidence2.shape[0]):
for j in range(0, confidence2.shape[1]):
confidence2[i][j] = softmax(confidence2[i][j])
#ensemble confidence scores with weight W
W = 0.66
confidence = confidence1 * W + confidence2 * (1 - W)
#get final prediction
prediction = [np.argmax(x) - 1 for x in confidence]
prediction = np.array(prediction)
print "Prediction\n"
for i in range(0, prediction.shape[0]):
if prediction[i] == -1:
pol = "Negative"
elif prediction[i] == 0:
pol = "Neutral"
else:
pol = "Positive"
print "Message : " + messages_test[i] + "Polarity : " + pol + "\n"
def main(f):
print "System training started"
#load training dataset
dataset_train = f
ids, labels_train, messages_train = tsvreader.opentsv(dataset_train)
print "Train data loaded"
#labels for subjectivity detection (2 categories)
temp_labels_train = [0 if x == "neutral" else 1 for x in labels_train]
#labels for polarity detection (3 categories)
labels_train = [
0 if x == "neutral" else -1 if x == "negative" else 1
for x in labels_train
]
#convert labels to numpy arrays
temp_labels_train = np.array(temp_labels_train)
labels_train = np.array(labels_train)
#load word clusters
clusters = loadClusters()
print "Clusters loaded"
#load Lexicons
negationList, slangDictionary, lexicons, mpqa_lexicons = loadLexicons()
print "Lexicons loaded"
#tokenize all messages
tokens_train = tokenize(messages_train)
print "Messages tokenized"
#compute pos tags for all messages
pos_tags_train = arktagger.pos_tag_list(messages_train)
print "Pos tags computed"
#compute pos tag bigrams
pos_bigrams_train = getBigrams(pos_tags_train)
#compute pos tag trigrams
pos_trigrams_train = getTrigrams(pos_tags_train)
#get the unique pos bigrams from training set
unique_pos_tags = getPosTagsSet(pos_tags_train)
unique_bigrams = getBigramsSet(pos_bigrams_train)
unique_trigrams = getTrigramsSet(pos_trigrams_train)
#compute POS tag scores
pos_tags_scores_neutral = posTagsScore(unique_pos_tags, 0, pos_tags_train,
labels_train)
pos_tags_scores_positive = posTagsScore(unique_pos_tags, 1, pos_tags_train,
labels_train)
pos_tags_scores_negative = posTagsScore(unique_pos_tags, -1,
pos_tags_train, labels_train)
pos_bigrams_scores_neutral = posBigramsScore(unique_bigrams, 0,
pos_bigrams_train,
labels_train)
pos_bigrams_scores_positive = posBigramsScore(unique_bigrams, 1,
pos_bigrams_train,
labels_train)
pos_bigrams_scores_negative = posBigramsScore(unique_bigrams, -1,
pos_bigrams_train,
labels_train)
pos_trigrams_scores_neutral = posTrigramsScore(unique_trigrams, 0,
pos_trigrams_train,
labels_train)
pos_trigrams_scores_positive = posTrigramsScore(unique_trigrams, 1,
pos_trigrams_train,
labels_train)
pos_trigrams_scores_negative = posTrigramsScore(unique_trigrams, -1,
pos_trigrams_train,
labels_train)
#compute mpqa scores
mpqaScores = getScores(mpqa_lexicons,
messages_train,
labels_train,
neutral=True)
#save scores and other resources for future use
savePosScores(pos_tags_scores_neutral, pos_tags_scores_positive,
pos_tags_scores_negative, pos_bigrams_scores_neutral,
pos_bigrams_scores_positive, pos_bigrams_scores_negative,
pos_trigrams_scores_neutral, pos_trigrams_scores_positive,
pos_trigrams_scores_negative, mpqaScores)
#save lexicons
saveLexicons(negationList, slangDictionary, lexicons, mpqa_lexicons)
#save clusters
saveClusters(clusters)
#load Glove embeddings
d = 200
glove = GloveDictionary.Glove(d)
#save Glove embeddings for future use
saveGlove(glove)
#Subjectivity Detection Features
#SD1 features
features_train_1 = features.getFeatures(
messages_train, tokens_train, pos_tags_train, slangDictionary,
lexicons, mpqa_lexicons, pos_bigrams_train, pos_trigrams_train,
pos_bigrams_scores_negative, pos_bigrams_scores_positive,
pos_trigrams_scores_negative, pos_trigrams_scores_positive,
pos_tags_scores_negative, pos_tags_scores_positive, mpqaScores,
negationList, clusters, pos_bigrams_scores_neutral,
pos_trigrams_scores_neutral, pos_tags_scores_neutral)
#SD2 features
features_train_2 = []
#for message in tokens_train :
for i in range(0, len(messages_train)):
features_train_2.append(glove.findCentroid(tokens_train[i]))
features_train_2 = np.array(features_train_2)
#regularize features
features_train_1 = regularization.regularize(features_train_1)
features_train_2 = regularization.regularizeHorizontally(features_train_2)
#Penalty parameter C of the error term for every SD system
C1 = 0.001953125
C2 = 1.4068830572470667
#get confidence scores
train_confidence_1 = getConfidenceScores(features_train_1,
temp_labels_train, C1)
train_confidence_2 = getConfidenceScores(features_train_2,
temp_labels_train, C2)
#normalize confidence scores
softmax = lambda x: 1 / (1. + math.exp(-x))
train_confidence_1 = [softmax(conf) for conf in train_confidence_1]
train_confidence_2 = [softmax(conf) for conf in train_confidence_2]
train_confidence_1 = np.array(train_confidence_1)
train_confidence_2 = np.array(train_confidence_2)
#train SD classifiers
sd1 = SVM.train(features_train_1, temp_labels_train, c=C1, k="linear")
sd2 = SVM.train(features_train_2, temp_labels_train, c=C2, k="linear")
#Sentiment Polarity Features (append confidence scores to SD features)
#SP1 features
features_train_1 = np.hstack(
(features_train_1,
train_confidence_1.reshape(train_confidence_1.shape[0], 1)))
#SP1 features
features_train_2 = np.hstack(
(features_train_2,
train_confidence_2.reshape(train_confidence_2.shape[0], 1)))
#Penalty parameter C of the error term for every SP system
C1 = 0.003410871889693192
C2 = 7.396183688299606
#train SP classifiers
sp1 = SVM.train(features_train_1, labels_train, c=C1, k="linear")
sp2 = SVM.train(features_train_2, labels_train, c=C2, k="linear")
#save trained models
saveModels(sd1, sd2, sp1, sp2)
print "System training completed!"
def main(f):
print "System training started"
#load training dataset
dataset_train = f
ids,labels_train,messages_train=tsvreader.opentsv(dataset_train)
print "Train data loaded"
#labels for subjectivity detection (2 categories)
temp_labels_train = [0 if x=="neutral" else 1 for x in labels_train]
#labels for polarity detection (3 categories)
labels_train = [0 if x=="neutral" else -1 if x =="negative" else 1 for x in labels_train]
#convert labels to numpy arrays
temp_labels_train=np.array(temp_labels_train)
labels_train=np.array(labels_train)
#load word clusters
clusters = loadClusters()
print "Clusters loaded"
#load Lexicons
negationList, slangDictionary, lexicons, mpqa_lexicons = loadLexicons()
print "Lexicons loaded"
#tokenize all messages
tokens_train = tokenize(messages_train)
print "Messages tokenized"
#compute pos tags for all messages
pos_tags_train = arktagger.pos_tag_list(messages_train)
print "Pos tags computed"
#compute pos tag bigrams
pos_bigrams_train = getBigrams(pos_tags_train)
#compute pos tag trigrams
pos_trigrams_train = getTrigrams(pos_tags_train)
#get the unique pos bigrams from training set
unique_pos_tags = getPosTagsSet(pos_tags_train)
unique_bigrams = getBigramsSet(pos_bigrams_train)
unique_trigrams= getTrigramsSet(pos_trigrams_train)
#compute POS tag scores
pos_tags_scores_neutral = posTagsScore(unique_pos_tags,0,pos_tags_train,labels_train)
pos_tags_scores_positive = posTagsScore(unique_pos_tags,1,pos_tags_train,labels_train)
pos_tags_scores_negative = posTagsScore(unique_pos_tags,-1,pos_tags_train,labels_train)
pos_bigrams_scores_neutral = posBigramsScore(unique_bigrams,0,pos_bigrams_train,labels_train)
pos_bigrams_scores_positive = posBigramsScore(unique_bigrams,1,pos_bigrams_train,labels_train)
pos_bigrams_scores_negative = posBigramsScore(unique_bigrams,-1,pos_bigrams_train,labels_train)
pos_trigrams_scores_neutral = posTrigramsScore(unique_trigrams,0,pos_trigrams_train,labels_train)
pos_trigrams_scores_positive = posTrigramsScore(unique_trigrams,1,pos_trigrams_train,labels_train)
pos_trigrams_scores_negative = posTrigramsScore(unique_trigrams,-1,pos_trigrams_train,labels_train)
#compute mpqa scores
mpqaScores = getScores(mpqa_lexicons,messages_train,labels_train,neutral=True)
#save scores and other resources for future use
savePosScores(pos_tags_scores_neutral, pos_tags_scores_positive,pos_tags_scores_negative,pos_bigrams_scores_neutral,pos_bigrams_scores_positive,pos_bigrams_scores_negative,pos_trigrams_scores_neutral,pos_trigrams_scores_positive,pos_trigrams_scores_negative,mpqaScores)
#save lexicons
saveLexicons(negationList,slangDictionary,lexicons,mpqa_lexicons)
#save clusters
saveClusters(clusters)
#load Glove embeddings
d = 200
glove = GloveDictionary.Glove(d)
#save Glove embeddings for future use
saveGlove(glove)
#Subjectivity Detection Features
#SD1 features
features_train_1 = features.getFeatures(messages_train,tokens_train,pos_tags_train,slangDictionary,lexicons,mpqa_lexicons,pos_bigrams_train,pos_trigrams_train,pos_bigrams_scores_negative,pos_bigrams_scores_positive,pos_trigrams_scores_negative,pos_trigrams_scores_positive,pos_tags_scores_negative,pos_tags_scores_positive,mpqaScores,negationList,clusters,pos_bigrams_scores_neutral,pos_trigrams_scores_neutral,pos_tags_scores_neutral)
#SD2 features
features_train_2 = []
#for message in tokens_train :
for i in range(0,len(messages_train)):
features_train_2.append(glove.findCentroid(tokens_train[i]))
features_train_2 = np.array(features_train_2)
#regularize features
features_train_1 = regularization.regularize(features_train_1)
features_train_2 = regularization.regularizeHorizontally(features_train_2)
#Penalty parameter C of the error term for every SD system
C1=0.001953125
C2=1.4068830572470667
#get confidence scores
train_confidence_1 = getConfidenceScores(features_train_1, temp_labels_train, C1)
train_confidence_2 = getConfidenceScores(features_train_2, temp_labels_train, C2)
#normalize confidence scores
softmax = lambda x: 1 / (1. + math.exp(-x))
train_confidence_1 = [softmax(conf) for conf in train_confidence_1]
train_confidence_2 = [softmax(conf) for conf in train_confidence_2]
train_confidence_1 = np.array(train_confidence_1)
train_confidence_2 = np.array(train_confidence_2)
#train SD classifiers
sd1 = SVM.train(features_train_1,temp_labels_train,c=C1,k="linear")
sd2 = SVM.train(features_train_2,temp_labels_train,c=C2,k="linear")
#Sentiment Polarity Features (append confidence scores to SD features)
#SP1 features
features_train_1 = np.hstack((features_train_1,train_confidence_1.reshape(train_confidence_1.shape[0],1)))
#SP1 features
features_train_2 = np.hstack((features_train_2,train_confidence_2.reshape(train_confidence_2.shape[0],1)))
#Penalty parameter C of the error term for every SP system
C1=0.003410871889693192
C2=7.396183688299606
#train SP classifiers
sp1 = SVM.train(features_train_1,labels_train,c=C1,k="linear")
sp2 = SVM.train(features_train_2,labels_train,c=C2,k="linear")
#save trained models
saveModels(sd1,sd2,sp1,sp2)
print "System training completed!"
#get the unique pos bigrams from training set
unique_bigrams = getBigramsSet(pos_bigrams_train)
#calculate pos bigrams score for all categories
#both dictionaries will be used for training and testing (cannot create new for testing because we don't know the labels of the new messages)
pos_bigrams_scores_objective = posBigramsScore(unique_bigrams,0,pos_bigrams_train,labels_train)
pos_bigrams_scores_subjective = posBigramsScore(unique_bigrams,1,pos_bigrams_train,labels_train)
#assign a precision and F1 score to each word of to all mpqa and semeval_13 lexicons
mpqaScores = getScores(mpqa_lexicons,process_messages_train,labels_train)
#get features from train messages
features_train = features_subjectivity.getFeatures(messages_train,process_messages_train,tokens_train,process_tokens_train,pos_tags_train,slangDictionary,lexicons,mpqa_lexicons,pos_bigrams_train,pos_bigrams_scores_objective,pos_bigrams_scores_subjective,mpqaScores,negationList,clusters)
#regularize train features
features_train=regularization.regularize(features_train)
#get features from test messages
features_test = features_subjectivity.getFeatures(messages_test,process_messages_test,tokens_test,process_tokens_test,pos_tags_test,slangDictionary,lexicons,mpqa_lexicons,pos_bigrams_test,pos_bigrams_scores_objective,pos_bigrams_scores_subjective,mpqaScores,negationList,clusters)
#regularize test features
features_test=regularization.regularize(features_test)
else:
# 0 - negative messages
# 1 - positives messages
labels_train = [0 if x=="negative" else 1 for x in labels_train]
labels_test = [0 if x=="negative" else 1 for x in labels_test]
#compute pos tag bigrams for all messages
pos_bigrams_train = getBigrams(pos_tags_train)
pos_bigrams_test = getBigrams(pos_tags_test)