/
main.py
246 lines (209 loc) · 9.95 KB
/
main.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
import sys
import json
import numpy as np
import gensim
from keras.models import Model
from keras.models import Sequential
from keras.layers import Input,Bidirectional,LSTM,TimeDistributed,Dense,Conv2D,MaxPooling2D,Flatten
from keras.optimizers import SGD,Adam
from keras_contrib.layers.crf import CRF
from keras.utils import to_categorical
import random
from sklearn.model_selection import train_test_split
from sklearn.metrics import precision_recall_fscore_support
from keras.callbacks import ModelCheckpoint, EarlyStopping
label4node = [ "null" ,"value", "agent", "condition", "theme", "theme_mod", "quant_mod", "co_quant", "location", "whole", "source", "reference_time", "quant", "manner", "time", "cause"]
label4edge = [ "equivalence" , "fact" , "analogy" ]
def getFeature(tokens,model): #tokens of data are stored as dictionary
feature = np.zeros((1,len(tokens),300))
# sentence to 2d feature vectors
for s in range(len(tokens)):
if tokens[s] in model.vocab:
feature[0,s,:] = model[tokens[s]]
else:
feature[0,s,:] = model[random.choice(model.wv.index2entity)]
return feature
def getAns(nodes,token_len):
label = np.zeros((1,token_len,1))
for n in nodes:
scope = n[0]
for indx in range(scope[0],scope[1]):
for key in n[1]:
label[0,indx,0] = label4node.index(key)
break
return label
def load_w2v():
print('load word2vec model with GoogleNews as corpus...')
model = gensim.models.KeyedVectors.load_word2vec_format('./GoogleNews-vectors-negative300.bin', binary=True)
return model
def json2dic(data):
return json.loads(data)
def extract_token_node_edges(filename): # Extract tokens,nodes,edges information,stored by dic, from json file
print('read file and store the data with tokens,nodes,edges...')
tokens , nodes , edges = [] , [] , []
with open(filename) as f:
data = f.readlines()
for i in range(len(data)):
d = json2dic(data[i])
tokens.append(d["tokens"])
nodes.append(d["nodes"])
edges.append(d["edges"])
return tokens,nodes,edges
def text_preprocessing(t):
return t
def read_data(filename,w2v_model):
tokens,nodes,edges = extract_token_node_edges(filename)
tokens = text_preprocessing(tokens)
return tokens,nodes,edges
def task1_data_generator(rawdata,model):
x , y = rawdata[0],rawdata[1]
i = 0
while(True):
sentence , label = x[i] , y[i]
Input , Label = np.zeros((1,len(sentence),300)),np.zeros((1,len(sentence),1))
# sentence to 2d feature vectors
for s in range(len(sentence)):
if sentence[s] in model.vocab:
Input[0,s,:] = model[sentence[s]]
else:
Input[0,s,:] = model[random.choice(model.wv.index2entity)]
# node label to label vector
for n in label:
scope = n[0]
for indx in range(scope[0],scope[1]):
for key in n[1]:
Label[0,indx,0] = label4node.index(key)
break
i = i + 1
if i == len(x):
i=0
yield Input, to_categorical(Label, num_classes= len(label4node))
def task2_data_generator_w2v(rawdata,w2v_model,lstm_model):
x , y = rawdata[0],rawdata[1]
Input , Label = [] , []
for i in range(len(x)):
sentence , edges = x[i] , y[i]
for info in edges:
node1,node2 = sentence[info[0][0]:info[0][1]],sentence[info[1][0]:info[1][1]]
vec1,vec2 = np.zeros((300,)),np.zeros((300,))
for w in node1:
if w in w2v_model.vocab:
vec1 = vec1 + w2v_model[w]
else:
vec1 = vec1 + w2v_model[random.choice(w2v_model.wv.index2entity)]
for w in node2:
if w in w2v_model.vocab:
vec2 = vec2 + w2v_model[w]
else:
vec2 = vec2 + w2v_model[random.choice(w2v_model.wv.index2entity)]
Input.append(np.concatenate((vec1, vec2)))
for key in info[2]:
Label.append(label4edge.index(key))
break
return np.array(Input),to_categorical(np.array(Label), num_classes= len(label4edge))
def task2_data_generator_w2lstm(rawdata,w2v_model,lstm_model):
x , y = rawdata[0],rawdata[1]
Input , Label = [] , []
for i in range(len(x)):
sentence , edges = x[i] , y[i]
lstm_vector = np.zeros((1,len(sentence),300))
# sentence to 2d feature vectors
for s in range(len(sentence)):
if sentence[s] in w2v_model.vocab:
lstm_vector[0,s,:] = w2v_model[sentence[s]]
else:
lstm_vector[0,s,:] = w2v_model[random.choice(w2v_model.wv.index2entity)]
feature_vec = lstm_model.predict(lstm_vector)
for info in edges:
vec1,vec2 = np.zeros((300,)),np.zeros((300,))
for i in range(info[0][0],info[0][1]):
vec1 = vec1 + feature_vec[0,i,:]
for i in range(info[1][0],info[1][1]):
vec2 = vec2 + feature_vec[0,i,:]
Input.append(np.concatenate((vec1, vec2)))
for key in info[2]:
Label.append(label4edge.index(key))
break
return np.array(Input),to_categorical(np.array(Label), num_classes= len(label4edge))
def task1_build_model(d_dim,l_dim):
print('task1 model building...')
INPUT = Input(shape = (None,d_dim))
BILSTM1 = Bidirectional(LSTM(300, return_sequences=True),merge_mode='sum',input_shape=(None,d_dim))(INPUT)
BILSTM2 = Bidirectional(LSTM(300, return_sequences=True),merge_mode='sum',input_shape=(None,d_dim))(BILSTM1)
TD = TimeDistributed(Dense(l_dim))(BILSTM2)
FC1 = Dense(units = l_dim, activation = 'softmax')(TD)
model = Model(INPUT,FC1)
high_feature = Model(INPUT,BILSTM2)
return model,high_feature
def task2_build_model(d_dim,l_dim):
print('task2 model building...')
INPUT = Input(shape = (d_dim,))
FC1 = Dense(units = 300, activation = 'relu')(INPUT)
FC2 = Dense(units = 300, activation = 'relu')(FC1)
FC3 = Dense(units = l_dim, activation = 'softmax')(FC2)
model = Model(INPUT,FC3)
return model
if __name__ == "__main__":
argv = sys.argv
if len(argv) < 2:
raise Exception('please enter train/test filename respectively')
train_filename = argv[1]
testing_filename = argv[2]
w2v_model = load_w2v()
tokens_train ,node_train ,edge_train = read_data(train_filename,w2v_model)
tokens_test ,node_test ,edge_test = read_data(testing_filename,w2v_model)
########################## Node label prediction given Tokens################################
l_dim,d_dim = len(label4node), 300
X_train, X_valid, y_train, y_vaild =train_test_split(tokens_train, node_train, test_size=0.33, random_state=42)
Num_traindata ,Num_validdata = len(X_train),len(X_valid)
g1 = task1_data_generator((X_train,y_train),w2v_model)
g2 = task1_data_generator((X_valid,y_vaild),w2v_model)
model4node , w2lstm_model = task1_build_model(d_dim,l_dim)
model4node.summary()
model4node.compile(loss='categorical_crossentropy',optimizer='adam',metrics=['accuracy'])
with open('model4node.json','w') as f: # save the model
f.write(model4node.to_json())
ckpt = ModelCheckpoint('model4node',monitor='val_acc',save_best_only=True,save_weights_only=True,verbose=1)
cb= [ckpt]
model4node.fit_generator(g1,steps_per_epoch=Num_traindata,
epochs=5,
verbose=1,
validation_data=g2,
max_queue_size = 15,
validation_steps=Num_validdata,callbacks=cb,
)
########################## edge label prediction given nodes################################
l_dim, d_dim = len(label4edge), 600
X_train, X_valid, y_train, y_vaild =train_test_split(tokens_train, edge_train, test_size=0.33, random_state=42)
Num_traindata ,Num_validdata = len(X_train),len(X_valid)
X_train,y_train = task2_data_generator_w2lstm((X_train,y_train),w2v_model,w2lstm_model)
X_valid,y_vaild = task2_data_generator_w2lstm((X_valid,y_vaild),w2v_model,w2lstm_model)
model4edge = task2_build_model(d_dim,l_dim)
model4edge.summary()
model4edge.compile(loss='categorical_crossentropy',optimizer='adam',metrics=['accuracy'])
with open('model4edge.json','w') as f: # save the model
f.write(model4edge.to_json())
ckpt = ModelCheckpoint('model4edge',monitor='val_acc',save_best_only=True,save_weights_only=True,verbose=1)
cb= [ckpt]
model4edge.fit(X_train,y_train,validation_data=(X_valid,y_vaild),callbacks=cb,batch_size=100,epochs=20)
########################## make prediction on testing data################################
y_true , y_pred = [] , []
for i in range(len(tokens_test)):
feature_vec = getFeature(tokens_test[i],w2v_model)
ans = getAns(node_test[i],len(tokens_test[i]))
ans = np.reshape(ans, (ans.shape[1],)).astype('int')
y_true = y_true + ans.tolist()
pred = model4node.predict(feature_vec)
pred = np.reshape(pred, (pred.shape[1], pred.shape[2]))
pred = np.argmax(pred, axis=1)
y_pred = y_pred + pred.tolist()
evaluation = precision_recall_fscore_support(y_true, y_pred, average='macro')
print("evaluation on node label prediction...")
print(evaluation)
X_test,y_true = task2_data_generator_w2lstm((tokens_test,edge_test),w2v_model,w2lstm_model)
y_true = np.argmax(y_true, axis=1)
y_pred = model4edge.predict(X_test)
y_pred = np.argmax(y_pred, axis=1)
evaluation = precision_recall_fscore_support(y_true, y_pred, average='macro')
print("evaluation on node label prediction...")
print(evaluation)