def _generate_mft_features(model, mft_axes, word_data):
mft_features = {
k: {
'framing-bias': 0,
'intensity': 0
}
for k, _ in mft_axes.items()
}
total_count = 0
for category in mft_axes.keys():
for word, attribs in word_data.items():
if word not in model.wv: continue
count = attribs['count']
mft_features[category]['framing-bias'] += (
count * cos_sim(mft_axes[category], model.wv.get_vector(word)))
total_count += count
if total_count == 0:
mft_features[category]['framing-bias'] = 0
else:
mft_features[category]['framing-bias'] /= total_count
for word, attribs in word_data.items():
if word not in model.wv: continue
count = attribs['count']
mft_features[category]['intensity'] += (count * (
(cos_sim(mft_axes[category], model.wv.get_vector(word)) -
mft_features[category]['framing-bias'])**2))
if total_count == 0:
mft_features[category]['intensity'] = 0
else:
mft_features[category]['intensity'] /= total_count
return mft_features
def emit_new_weight(self, inp_h, weight_h, memory_h): #context addressing key = T.dot(inp_h, self.key_w)+self.key_b beta = T.nnet.softplus(T.dot(inp_h, self.beta_w )+self.beta_b) g = T.nnet.sigmoid(T.dot(inp_h, self.g_w)+self.g_b) #gamma = T.nnet.softplus(T.dot(inp_h,self.gamma_w)+self.gamma_b) weight_c = tools.vector_softmax(beta*tools.cos_sim(key, memory_h)) #location addressing #interpolating weight_g = g*weight_c+ (1-g)*weight_h weight_location = T.tanh(T.dot(weight_g, self.location_w)+self.location_b) weight_new = weight_location #erase and add erase = T.nnet.sigmoid(T.dot(inp_h,self.erase_w)+self.erase_b) add = T.dot(inp_h,self.add_w)+self.add_b #if test: # return key, beta,weight_c,g,weight_g,shift,weight_shift,gamma,weight_gamma,weight_new return weight_new, erase, add
from tools import df_split, cos_sim
import random
import statsmodels.api as sm
import pandas as pd
import numpy as np
if __name__ == "__main__":
# set random seed for consistency, load data
random.seed(0)
df_q = pd.read_csv('data/quora_duplicate_questions.tsv', sep="\t")
q1_split, q1_full = df_split(df_q, 'question1')
q2_split, q2_full = df_split(df_q, 'question2')
# get cos sim, get difference in length between the two questions (both features in our model)
cos = cos_sim(q1_split, q2_split)
len_diff = (q1_full.str.len() - q2_full.str.len()).values
# prepare features
x = np.vstack([cos, len_diff]).T
y = df_q['is_duplicate'].values
x = sm.add_constant(x)
# randomly generate n / 2 indices. we then use these indices to create a train/test set (50/50 split here)
n = y.shape[0]
train_ind = random.sample(range(0, n), int(n / 2))
test_ind = list(set(range(0, n)) - set(train_ind))
xtrain, xtest = x[train_ind], x[test_ind]
ytrain, ytest = y[train_ind], y[test_ind]
# train logistic regression, find accuracy on test set
model = sm.Logit(ytrain, xtrain).fit()
ypred = (model.predict(xtest) > 0.5).astype(int)
accuracy = (ypred == ytest).sum() / len(ypred)
print(accuracy)
def emit_new_weight(self, inp_h, weight_h, memory_h): #context addressing key = T.dot(inp_h, self.key_w)+self.key_b beta = T.nnet.softplus(T.dot(inp_h, self.beta_w )+self.beta_b) g = T.nnet.sigmoid(T.dot(inp_h, self.g_w)+self.g_b) #shift = tools.vector_softmax(T.dot(inp_h,self.shift_w)+self.shift_b) #shift = shift.dimshuffle((0,'x')) #gamma = T.nnet.softplus(T.dot(inp_h,self.gamma_w)+self.gamma_b)+1. ''' key_normal = key/(T.sqrt(T.sum(key**2)) + tools.mini) memory_mod = T.sqrt(T.sum(memory**2, axis = 1).dimshuffle((0,'x'))) memory_normal = memory/(memory_mod + tools.mini) weight_c = T.exp(beta*T.dot(memory_normal, key_normal)) weight_cnormal = weight_c/T.sum(weight_c + tools.mini) ''' weight_c = tools.vector_softmax(beta*tools.cos_sim(key, memory_h)) #location addressing #interpolating #weight_g = g*weight_c+ (1-g)*weight_h #weight_conv = theano.tensor.signal.conv(weight_g.reshape(memory.shape[0],1), # shift_normal.reshape(self.shift_width, 1)) #code from shaw #shift_normal = shift/T.sum(shift) ''' wlength = weight_g.shape[0] shift_sidewidth = self.shift_width/2 weight_shift = weight_h def cal_shift(pos): weight_pos = 0. for j in range(0,self.shift_width): pos = i+j-shift_sidewidth if pos < 0: pos += wlength if pos >= wlength: pos -= wlength weight_shift += shift[j]*weight_g[pos] return weight_pos weight_shift ''' #weight_shift = T.sum(shift*weight_g[self.shift_conv], axis = 0) #sharpening ''' weight_gamma = weight_shift ** gamma #weight_gamma = weight_g weight_new = weight_gamma/T.sum(weight_gamma) ''' weight_new = weight_c #erase and add erase = T.nnet.sigmoid(T.dot(inp_h,self.erase_w)+self.erase_b) add = T.dot(inp_h,self.add_w)+self.add_b #if test: # return key, beta,weight_c,g,weight_g,shift,weight_shift,gamma,weight_gamma,weight_new return weight_new, erase, add
