-
Notifications
You must be signed in to change notification settings - Fork 0
/
legal_att.py
251 lines (199 loc) · 10.2 KB
/
legal_att.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
247
248
249
250
251
import tensorflow as tf
class LegalAtt:
def __init__(self, config, embedding_matrix, is_training):
self.accu_num = config.accu_num
self.art_num = config.art_num
self.impr_num = config.impr_num
self.top_k = config.top_k
self.threshold = config.threshold
self.max_seq_len = config.sequence_len
self.kernel_size = config.kernel_size
self.filter_dim = config.filter_dim
self.att_size = config.att_size
self.fc_size = config.fc_size_s
self.embedding_matrix = tf.get_variable(
initializer=tf.constant_initializer(embedding_matrix),
shape=embedding_matrix.shape,
trainable=config.embedding_trainable,
dtype=tf.float32,
name='embedding_matrix'
)
self.embedding_size = embedding_matrix.shape[-1]
self.lr = config.lr
self.optimizer = config.optimizer
self.dropout = config.dropout
self.l2_rate = config.l2_rate
self.use_batch_norm = config.use_batch_norm
self.is_training = is_training
self.w_init = tf.truncated_normal_initializer(stddev=0.1)
self.b_init = tf.constant_initializer(0.1)
if self.l2_rate > 0.0:
self.regularizer = tf.keras.regularizers.l2(self.l2_rate)
else:
self.regularizer = None
self.batch_size = tf.placeholder(dtype=tf.int32, shape=[], name='batch_size')
self.fact = tf.placeholder(dtype=tf.int32, shape=[None, self.max_seq_len], name='fact')
self.fact_len = tf.placeholder(dtype=tf.int32, shape=[None], name='fact_len')
self.art = tf.placeholder(dtype=tf.int32, shape=[None, self.art_num, self.max_seq_len], name='art')
self.art_len = tf.placeholder(dtype=tf.int32, shape=[None, self.art_num], name='art_len')
self.accu = tf.placeholder(dtype=tf.float32, shape=[None, self.accu_num], name='accu')
self.relevant_art = tf.placeholder(dtype=tf.float32, shape=[None, self.art_num], name='relevant_art')
self.impr = tf.placeholder(dtype=tf.float32, shape=[None, self.impr_num], name='impr')
with tf.variable_scope('fact_embedding'):
fact_em = self.embedding_layer(self.fact)
with tf.variable_scope('fact_encoder'):
fact_enc = self.cnn_encoder(fact_em)
with tf.variable_scope('article_extractor'):
art_score, top_k_score, top_k_indices = self.get_top_k_indices(fact_enc)
top_k_art, top_k_art_len = self.get_top_k_articles(top_k_indices)
with tf.variable_scope('article_embedding'):
top_k_art_em = self.embedding_layer(top_k_art)
with tf.variable_scope('article_encoder'):
shared_layers = {}
for kernel_size in self.kernel_size:
shared_layers['conv_' + str(kernel_size)] = tf.keras.layers.Conv1D(
self.filter_dim,
kernel_size,
padding='same',
kernel_regularizer=self.regularizer,
name='conv_' + str(kernel_size)
)
if self.use_batch_norm:
shared_layers['norm_' + str(kernel_size)] = tf.keras.layers.BatchNormalization(name='norm_' + str(kernel_size))
top_k_art_enc = self.art_encoder(top_k_art_em, shared_layers)
with tf.variable_scope('attention'):
key = tf.keras.layers.Dense(
self.att_size,
tf.nn.tanh,
use_bias=False,
kernel_regularizer=self.regularizer
)(fact_enc)
ones = tf.ones_like(top_k_score, dtype=tf.float32)
zeros = tf.zeros_like(top_k_score, dtype=tf.float32)
relevant_score = tf.where(top_k_score >= self.threshold, ones, zeros)
legal_atts = []
dense_layer = tf.keras.layers.Dense(
self.att_size,
tf.nn.tanh,
use_bias=False,
kernel_regularizer=self.regularizer
)
for i in range(self.top_k):
art_enc = top_k_art_enc[i]
art_len = top_k_art_len[:, i]
score = relevant_score[:, i]
query = dense_layer(art_enc)
att_matrix = tf.reshape(score, [-1, 1, 1]) * self.get_attention(query, key, art_len, self.fact_len)
legal_atts.append(tf.reduce_sum(att_matrix, axis=-2))
# prevent dividing by zero
att_num = tf.reshape(tf.reduce_sum(relevant_score, axis=-1), [-1, 1])
ones = tf.ones_like(att_num, dtype=tf.float32)
att_num = tf.where(att_num > 0, att_num, ones)
fact_enc_with_att = [tf.reduce_max(tf.expand_dims(att, axis=-1) * fact_enc, axis=-2) for att in legal_atts]
fact_enc_with_att = tf.add_n(fact_enc_with_att) / att_num + tf.reduce_max(fact_enc, axis=-2)
with tf.variable_scope('output'):
self.task_1_output, task_1_loss = self.output_layer(fact_enc_with_att, self.accu, layer='sigmoid')
ones = tf.ones_like(art_score, dtype=tf.float32)
zeros = tf.zeros_like(art_score, dtype=tf.float32)
self.task_2_output = tf.where(tf.nn.sigmoid(art_score) >= self.threshold, ones, zeros)
with tf.variable_scope('loss'):
task_2_loss = tf.reduce_mean(tf.reduce_sum(
tf.nn.sigmoid_cross_entropy_with_logits(labels=self.relevant_art, logits=art_score),
axis=-1
))
self.loss = task_1_loss + task_2_loss
if self.regularizer is not None:
l2_loss = tf.add_n(tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES))
self.loss += l2_loss
if not is_training:
return
self.global_step, self.train_op = self.get_train_op()
def embedding_layer(self, inputs):
inputs_em = tf.nn.embedding_lookup(self.embedding_matrix, inputs)
if self.is_training and self.dropout < 1.0:
inputs_em = tf.nn.dropout(inputs_em, rate=self.dropout)
return inputs_em
def cnn_encoder(self, inputs):
enc_output = []
for kernel_size in self.kernel_size:
conv = tf.keras.layers.Conv1D(
self.filter_dim,
kernel_size,
padding='same',
kernel_regularizer=self.regularizer,
name='conv_' + str(kernel_size)
)(inputs)
if self.use_batch_norm:
conv = tf.keras.layers.BatchNormalization(name='norm_' + str(kernel_size))(conv)
conv = tf.nn.relu(conv)
enc_output.append(conv)
enc_output = tf.concat(enc_output, axis=-1)
return enc_output
def get_top_k_indices(self, inputs):
inputs = tf.reduce_max(inputs, axis=-2)
scores = tf.keras.layers.Dense(self.art_num, kernel_regularizer=self.regularizer)(inputs)
if self.is_training:
top_k_score, top_k_indices = tf.math.top_k(self.relevant_art, k=self.top_k)
else:
top_k_score, top_k_indices = tf.math.top_k(tf.nn.sigmoid(scores), k=self.top_k)
return scores, top_k_score, top_k_indices
def get_top_k_articles(self, top_k_indices):
top_k_art = tf.batch_gather(self.art, indices=top_k_indices)
top_k_art_len = tf.batch_gather(self.art_len, indices=top_k_indices)
return top_k_art, top_k_art_len
def art_encoder(self, top_k_art_em, shared_layers):
top_k_art_enc = []
for i in range(self.top_k):
art_enc = []
art_em = top_k_art_em[:, i, :, :]
for kernel_size in self.kernel_size:
conv = shared_layers['conv_' + str(kernel_size)](art_em)
if self.use_batch_norm:
conv = shared_layers['norm_' + str(kernel_size)](conv)
conv = tf.nn.relu(conv)
art_enc.append(conv)
art_enc = tf.concat(art_enc, axis=-1)
top_k_art_enc.append(art_enc)
return top_k_art_enc
def get_attention(self, query, key, query_len, key_len):
att = tf.matmul(query, key, transpose_b=True)
query_mask = tf.sequence_mask(query_len, maxlen=self.max_seq_len, dtype=tf.float32)
key_mask = tf.sequence_mask(key_len, maxlen=self.max_seq_len, dtype=tf.float32)
mask = tf.matmul(tf.expand_dims(query_mask, axis=-1), tf.expand_dims(key_mask, axis=-2))
inf = 1e10 * tf.ones_like(att, dtype=tf.float32)
# set masked value to -inf to prevent gradient
masked_att = tf.where(mask > 0.0, att, -inf)
masked_att = tf.nn.softmax(masked_att, axis=-1)
masked_att = mask * masked_att
return masked_att
def output_layer(self, inputs, labels, layer):
fc_output = tf.keras.layers.Dense(self.fc_size, kernel_regularizer=self.regularizer)(inputs)
if self.is_training and self.dropout < 1.0:
fc_output = tf.nn.dropout(fc_output, rate=self.dropout)
logits = tf.keras.layers.Dense(labels.shape[-1], kernel_regularizer=self.regularizer)(fc_output)
if layer == 'softmax':
output = tf.nn.softmax(logits)
ce_loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels=labels, logits=logits))
elif layer == 'sigmoid':
output = tf.nn.sigmoid(logits)
ce_loss = tf.reduce_mean(tf.reduce_sum(
tf.nn.sigmoid_cross_entropy_with_logits(labels=labels, logits=logits),
axis=-1
))
else:
assert False
return output, ce_loss
def get_train_op(self):
global_step = tf.Variable(0, trainable=False, name='global_step')
if self.optimizer == 'Adam':
optimizer = tf.train.AdamOptimizer(learning_rate=self.lr)
elif self.optimizer == 'Adadelta':
optimizer = tf.train.AdadeltaOptimizer(learning_rate=self.lr)
elif self.optimizer == 'Adagrad':
optimizer = tf.train.AdagradOptimizer(learning_rate=self.lr)
elif self.optimizer == 'SGD':
optimizer = tf.train.GradientDescentOptimizer(learning_rate=self.lr)
else:
assert False
train_op = optimizer.minimize(self.loss, global_step=global_step)
return global_step, train_op