def calculate(self, w, seed):
# Step 2: calculate P(w) and P(seed)
w_count = 0
seed_count = 0
for i, sent in enumerate(self.train_X):
if w in sent:
w_count += 1 # count the occurrence of word w
if seed in sent:
seed_count += 1 # count the occurrence of word seed
p_w = w_count / self.train_size
p_seed = seed_count / self.train_size
# Step 3: calculate P(w,seed)
w_seed_count = 0
for i, sent in enumerate(self.train_X):
if (w in sent) and (seed in sent):
w_seed_count += 1 # count the occurrence of word seed
p_w_seed = w_seed_count / self.train_size / self.train_size
# Step 4: calculate PMI
value = math.log2(p_w_seed / (p_w * p_seed))
return value
if __name__ == '__main__':
data = SST(fine_grained=False)
pmi = PMI(data, 0.5)
pmi_1 = pmi.calculate("movie", "good")
pmi_2 = pmi.calculate("movie", "bad")
lex_w = pmi_1 - pmi_2
def test(self):
self.model.eval()
X = self._get_feature_vec(self.data.test["sents"])
for i in range(X[0].shape[0]):
outputs = []
for j in range(self.class_size):
outputs.append(self.model(torch.FloatTensor(X[j][i])).squeeze()) # wv+b for all class in pytorch
probabilities = list(torch.softmax(torch.stack(outputs), dim=0)) # get the probability for each class
max_prob = max(probabilities)
max_prob_idx = probabilities.index(max_prob) # the max output as the predicted class c
c = self.class_dict[max_prob_idx]
if c == 0:
c_string = "Very negative"
elif c == 1:
c_string = "Negative"
elif c == 2:
c_string = "Neutural"
elif c == 3:
c_string = "Positive"
else:
assert c == 4
c_string = "Very positive"
print(' '.join(self.data.test["sents"][i]),": ",c_string)
if __name__ == '__main__':
data = SST(fine_grained=True)
classifier = Multiclass_log_linear(data,0.2)
classifier.train()
classifier.test()