image.png
image.png
image.png
image.png
-
比较：其 [a¯i,βi][a¯i,βi] 就是把两部分concat起来
（句子中第 i 个词, 和另一个句子里所有词与它比较的加权词向量）
image.png -
总计aggregate：
就是分别du把对v1,v2v1,v2 中的每个元素求和，得到两个值之后做分类。
image.png
def _compare(self, sentence, attention, sentence_length):
# sent_and_alignment has shape [batch, time_steps, num_units]
inputs = [sentence, attention, sentence - attention, sentence * attention]
features = tf.concat(axis=2, values=inputs)
return self._feedforward(features, self._m_config['compare']['dim'])
def _aggregate(self, v1, v2):
# sum over time steps; resulting shape is [batch, num_units]
v1 = self._mask3d(v1, self._m_ph_sent1_size, 0, 1)
v2 = self._mask3d(v2, self._m_ph_sent2_size, 0, 1)
v1_sum = tf.reduce_sum(v1, 1)
v2_sum = tf.reduce_sum(v2, 1)
v1_max = tf.reduce_max(v1, 1)
v2_max = tf.reduce_max(v2, 1)
aggregate_features = tf.concat([v1_sum, v2_sum, v1_max, v2_max], axis=1)
return self._feedforward(aggregate_features, self._m_config['aggregate']['dim'])
def _feedforward(self, inputs, hidden_units):
if isinstance(hidden_units, int):
hidden_units = [hidden_units]
last_inputs = inputs
for i, dim in enumerate(hidden_units):
with tf.variable_scope('layer_%d' % (i)):
inputs = tf.nn.dropout(last_inputs, self._m_ph_dropout_rate)
last_inputs = tf.layers.dense(inputs, dim, tf.nn.relu)
return last_inputs
def _mask3d(self, values, sentence_sizes, mask_value, axis=2):
if axis != 1 and axis != 2:
raise ValueError("'axis' must be 1 or 2")
if axis == 1:
values = tf.transpose(values, [0, 2, 1])
time_steps1 = tf.shape(values)[1]
time_steps2 = tf.shape(values)[2]
# [batch_size, time_step1, time_step2]
ones = tf.ones_like(values, dtype=tf.float32)
pad_values = mask_value * ones
# [batch_size, time_step2]
mask = tf.sequence_mask(sentence_sizes, time_steps2)
# [batch_size, 1, time_step2]
mask3d = tf.expand_dims(mask, 1)
# [batch_size, time_step1, time_step2]
mask3d = tf.tile(mask3d, (1, time_steps1, 1))
masked = tf.where(mask3d, values, pad_values)
if axis == 1:
masked = tf.transpose(masked, [0, 2, 1])
return masked
