diff --git a/sent_rating_feature.py b/sent_rating_feature.py
index df297d019f1b3ad505126b661d9127bc3b78c91f..16c68cfa0dd7fb0a2b1f0daf31b616b9b94eec0a 100644
--- a/sent_rating_feature.py
+++ b/sent_rating_feature.py
@@ -9,12 +9,12 @@ def extract(corpus_instance):
Returns numpy array of size 1.
"""
review = corpus_instance["REVIEW"]
- stars = corpus_instance["STARS"]
+ stars = float(corpus_instance["STARS"])
#sent = get_sent_vader(review)
sent = get_sent_textblob(review)
- if (sent <= 0.0 and stars == "5.0") or (sent > 0.0 and stars == "1.0"):
+ if (sent <= 0.0 and stars > 3.0) or (sent > 0.0 and stars < 3.0):
return np.array([1])
else:
return np.array([0])
diff --git a/training_testing.py b/training_testing.py
index 35adee9cbc5e0dbd55e82f882b09908aa0552207..c3c006535d30ed27bd526ce1e3a0c941a706bf8e 100644
--- a/training_testing.py
+++ b/training_testing.py
@@ -10,6 +10,36 @@ from sklearn import tree
from sklearn import naive_bayes
from sklearn import linear_model
from sklearn.model_selection import cross_val_score
+import time
+import pickle
+
+def extract_features(training_set, test_set):
+
+ # vocabularies
+ unigram_vocab = ngram_feature.get_vocabulary(train_set, 'REVIEW', 1)
+ bigram_vocab = ngram_feature.get_vocabulary(train_set, 'REVIEW', 2)
+ trigram_vocab = ngram_feature.get_vocabulary(train_set, 'REVIEW', 3)
+ pos_bigram_vocab = pos_feature.get_pos_vocabulary(train_set)
+ surface_bigram_vocab = ngram_feature.get_vocabulary(train_set, 'SURFACE_PATTERNS', 2)
+
+ # inputs:
+ print("------Feature Extraction------\n")
+ train_inputs = [create_vector(el, unigram_vocab, pos_bigram_vocab, surface_bigram_vocab)
+ for el in train_set] # 1000 vectors
+ test_inputs = [create_vector(el, unigram_vocab, pos_bigram_vocab, surface_bigram_vocab)
+ for el in test_set] # 254 vectors
+
+ # stats
+ print("Number of train samples: {}".format(len(train_inputs)))
+ print("Unigram vocab size: {}".format(len(unigram_vocab)))
+ print("Bigram vocab size: {}".format(len(bigram_vocab)))
+ print("Trigram vocab size: {}".format(len(trigram_vocab)))
+ print("POS-Bigram vocab size: {}".format(len(pos_bigram_vocab)))
+ print("SP-Bigram vocab size: {}".format(len(surface_bigram_vocab)))
+ print("Total features per train sample: {}".format(len(train_inputs[0])))
+ print("---> Duration Feature Extraction: {} sec.\n".format(int(time.time()-start_time)))
+
+ return train_inputs, test_inputs
def create_vector(corpus_instance, vocabulary=None, pos_vocabulary=None, surface_vocabulary=None):
@@ -33,16 +63,20 @@ def train_multiple(classifiers, train_input, train_labels):
classifier.fit(train_input, train_labels)
-def score_multiple(classifiers, train_input, train_labels):
- scores = []
+def validate_multiple(classifiers, train_input, train_labels):
+ print("\n------Cross Validation------")
+
for classifier in classifiers:
+ print("\n{}".format(classifier))
+
accuracy = cross_val_score(classifier, train_inputs, train_labels, cv=5, scoring='accuracy').mean()
f1 = cross_val_score(classifier, train_inputs, train_labels, cv=5, scoring='f1').mean()
- scores.append(accuracy, f1)
- return scores
+
+ print("\nAccuracy: {}, F1-Score: {}\n".format(accuracy, f1))
if __name__ == '__main__':
+ start_time = time.time()
corpus = corpus.read_corpus("corpus_shuffled.csv")
extended_corpus = surface_patterns.extract_surface_patterns(corpus, 1000)
@@ -51,64 +85,48 @@ if __name__ == '__main__':
train_set = extended_corpus[:1000]
test_set = extended_corpus[1000:]
- # vocabularies
- unigram_vocab = ngram_feature.get_vocabulary(train_set, 'REVIEW', 1)
- bigram_vocab = ngram_feature.get_vocabulary(train_set, 'REVIEW', 2)
- pos_bigram_vocab = pos_feature.get_pos_vocabulary(train_set)
- surface_bigram_vocab = ngram_feature.get_vocabulary(train_set, 'SURFACE_PATTERNS', 2)
+ train_inputs, train_labels = [], []
+ test_inputs, test_labels = [], []
- # inputs:
- train_inputs = [create_vector(el, bigram_vocab, pos_bigram_vocab, surface_bigram_vocab)
- for el in train_set] # 1000 vectors
- #test_inputs = [create_vector(el, bigram_vocab, pos_bigram_vocab, surface_bigram_vocab)
- # for el in test_set] # 254 vectors
+ re_extract = True # change to False if features are unchanged since previous run
- # labels
- train_labels = np.array([int(el['LABEL']) for el in train_set]) # 1000 labels
- test_labels = np.array([int(el['LABEL']) for el in test_set]) # 254 labels
+ if re_extract == True:
- print("Number of train samples: {}".format(len(train_inputs)))
- print("Number of features per train sample: {}".format(len(train_inputs[0])))
- print("Unigram vocab size: {}".format(len(unigram_vocab)))
- print("Bigram vocab size: {}".format(len(bigram_vocab)))
- print("POS-Bigram vocab size: {}".format(len(pos_bigram_vocab)))
- print("Surface Patterns-Bigram vocab size: {}".format(len(surface_bigram_vocab)))
+ # inputs (x)
+ train_inputs, test_inputs = extract_features(train_set, test_set)
- # TODO Pickle/outsource
+ # labels (y)
+ train_labels = np.array([int(el['LABEL']) for el in train_set]) # 1000 labels
+ test_labels = np.array([int(el['LABEL']) for el in test_set]) # 254 labels
- # ML
+ # save to pickle
+ pickle.dump([train_inputs, train_labels, test_inputs, test_labels], open("vectors.pickle", "wb"))
+
+ else:
+ # load from pickle
+ v = pickle.load(open("vectors.pickle", "rb"))
+ train_inputs, train_labels = v[0], v[1]
+ test_inputs, test_labels = v[2], v[3]
+
+
+ # Machine Learning
# init
- svm_clf = svm.SVC(C=200.0, kernel='linear') # large C: smaller-margin hyperplane
+ svm_clf = svm.SVC(C=500.0, kernel='linear') # large C: smaller-margin hyperplane
tree_clf = tree.DecisionTreeClassifier()
nb_clf = naive_bayes.MultinomialNB()
lr_clf = linear_model.LogisticRegression()
# training
- train_multiple([svm_clf, tree_clf, nb_clf, lr_clf], train_inputs, train_labels)
+ train_multiple([svm_clf, tree_clf], train_inputs, train_labels) #, nb_clf, lr_clf
+ print("---> Duration Training: {} sec.\n".format(int(time.time()-start_time)))
# validation
- svm_acc = cross_val_score(svm_clf, train_inputs, train_labels, cv=5, scoring='accuracy').mean()
- tree_acc = cross_val_score(tree_clf, train_inputs, train_labels, cv=5, scoring='accuracy').mean()
- nb_acc = cross_val_score(nb_clf, train_inputs, train_labels, cv=5, scoring='accuracy').mean()
- lr_acc = cross_val_score(lr_clf, train_inputs, train_labels, cv=5, scoring='accuracy').mean()
-
- svm_f1 = cross_val_score(svm_clf, train_inputs, train_labels, cv=5, scoring='f1').mean()
- tree_f1 = cross_val_score(tree_clf, train_inputs, train_labels, cv=5, scoring='f1').mean()
- nb_f1 = cross_val_score(nb_clf, train_inputs, train_labels, cv=5, scoring='f1').mean()
- lr_f1 = cross_val_score(lr_clf, train_inputs, train_labels, cv=5, scoring='f1').mean()
-
- print("\n--Cross Validation Scores-- ")
- print("\nSVM: Accuracy: {}, F1-Score: {}".format(svm_acc, svm_f1))
- print("\nTree: Accuracy: {}, F1-Score: {}".format(tree_acc, tree_f1))
- print("\nN. Bayes: Accuracy: {}, F1-Score: {}".format(nb_acc, nb_f1))
- print("\nLog. Regression: Accuracy: {}, F1-Score: {}".format(lr_acc, lr_f1))
+ validate_multiple([svm_clf, tree_clf], train_inputs, train_labels) #, nb_clf, lr_clf
+ print("---> Duration CV: {} sec.".format(int(time.time()-start_time)))
+
# testing
# print("\nSVM: Score on test Data:")
# print(svm_clf.score(test_inputs, test_labels))
-
- # print("\nDTree: Score on test Data:")
- # print(tree_clf.score(test_inputs, test_labels))
-
# predictions = svm_classifier.predict(train_inputs)