def _bcnn_semantic_feature_layer(self,
seq_left,
seq_right,
dpool_index=None,
granularity="word"):
name = self.model_name + granularity
seq_len = self.params["max_seq_len_%s" % granularity]
# [batch, s, d] => [batch, s, d, 1]
seq_left = tf.expand_dims(seq_left, axis=-1)
seq_right = tf.expand_dims(seq_right, axis=-1)
left_ap_list = [None] * (self.params["bcnn_num_layers"] + 1)
right_ap_list = [None] * (self.params["bcnn_num_layers"] + 1)
left_ap_list[0] = self._all_ap(x=seq_left,
seq_len=seq_len,
name=name + "global_pooling_input_left")
right_ap_list[0] = self._all_ap(x=seq_right,
seq_len=seq_len,
name=name +
"global_pooling_input_right")
x1 = seq_left
x2 = seq_right
d = self.params["embedding_dim"]
outputs = []
for layer in range(self.params["bcnn_num_layers"]):
x1, left_ap_list[layer + 1], x2, right_ap_list[
layer + 1], att_pooled = self._bcnn_cnn_layer(
x1=x1,
x2=x2,
seq_len=seq_len,
d=d,
name=name + "cnn_layer_%d" % (layer + 1),
dpool_index=dpool_index,
granularity=granularity)
d = self.params["bcnn_num_filters"]
if self.params["bcnn_mp_att_pooling"] and att_pooled is not None:
outputs.append(att_pooled)
for l, r in zip(left_ap_list, right_ap_list):
outputs.append(
metrics.cosine_similarity(
l, r, self.params["similarity_aggregation"]))
outputs.append(
metrics.dot_product(l, r,
self.params["similarity_aggregation"]))
outputs.append(
metrics.euclidean_distance(
l, r, self.params["similarity_aggregation"]))
return tf.concat(outputs, axis=-1)
def _get_matching_features(self):
with tf.name_scope(self.model_name):
tf.set_random_seed(self.params["random_seed"])
with tf.name_scope("word_network"):
if self.params["attend_method"] == "context-attention":
emb_seq_word_left, enc_seq_word_left, att_seq_word_left, sem_seq_word_left, \
emb_seq_word_right, enc_seq_word_right, att_seq_word_right, sem_seq_word_right = \
self._interaction_semantic_feature_layer(
self.seq_word_left,
self.seq_word_right,
self.seq_len_word_left,
self.seq_len_word_right,
granularity="word")
else:
emb_seq_word_left, enc_seq_word_left, att_seq_word_left, sem_seq_word_left = \
self._semantic_feature_layer(
self.seq_word_left,
self.seq_len_word_left,
granularity="word", reuse=False)
emb_seq_word_right, enc_seq_word_right, att_seq_word_right, sem_seq_word_right = \
self._semantic_feature_layer(
self.seq_word_right,
self.seq_len_word_right,
granularity="word", reuse=True)
#### matching
# match score
sim_word = tf.concat(
[
metrics.cosine_similarity(
sem_seq_word_left, sem_seq_word_right,
self.params["similarity_aggregation"]),
metrics.dot_product(
sem_seq_word_left, sem_seq_word_right,
self.params["similarity_aggregation"]),
metrics.euclidean_distance(
sem_seq_word_left, sem_seq_word_right,
self.params["similarity_aggregation"]),
# self._canberra_score(sem_seq_word_left, sem_seq_word_right),
],
axis=-1)
# match pyramid
match_matrix_word = self._get_match_matrix(self.seq_word_left,
emb_seq_word_left,
enc_seq_word_left,
self.seq_word_right,
emb_seq_word_right,
enc_seq_word_right,
granularity="word")
mp_word = self._mp_semantic_feature_layer(
match_matrix_word,
self.dpool_index_word,
granularity="word")
# esim
esim_word = self._esim_semantic_feature_layer(
emb_seq_word_left,
emb_seq_word_right,
self.seq_len_word_left,
self.seq_len_word_right,
granularity="word")
# bcnn
bcnn_word = self._bcnn_semantic_feature_layer(
emb_seq_word_left, emb_seq_word_right, granularity="word")
# dense
deep_in_word = tf.concat(
[sem_seq_word_left, sem_seq_word_right], axis=-1)
deep_word = mlp_layer(
deep_in_word,
fc_type=self.params["fc_type"],
hidden_units=self.params["fc_hidden_units"],
dropouts=self.params["fc_dropouts"],
scope_name=self.model_name + "deep_word",
reuse=False,
training=self.training,
seed=self.params["random_seed"])
with tf.name_scope("char_network"):
if self.params["attend_method"] == "context-attention":
emb_seq_char_left, enc_seq_char_left, att_seq_char_left, sem_seq_char_left, \
emb_seq_char_right, enc_seq_char_right, att_seq_char_right, sem_seq_char_right = \
self._interaction_semantic_feature_layer(
self.seq_char_left,
self.seq_char_right,
self.seq_len_char_left,
self.seq_len_char_right,
granularity="char")
else:
emb_seq_char_left, enc_seq_char_left, att_seq_char_left, sem_seq_char_left = \
self._semantic_feature_layer(
self.seq_char_left,
self.seq_len_char_left,
granularity="char", reuse=False)
emb_seq_char_right, enc_seq_char_right, att_seq_char_right, sem_seq_char_right = \
self._semantic_feature_layer(
self.seq_char_right,
self.seq_len_char_right,
granularity="char", reuse=True)
# match score
sim_char = tf.concat(
[
metrics.cosine_similarity(
sem_seq_char_left, sem_seq_char_right,
self.params["similarity_aggregation"]),
metrics.dot_product(
sem_seq_char_left, sem_seq_char_right,
self.params["similarity_aggregation"]),
metrics.euclidean_distance(
sem_seq_char_left, sem_seq_char_right,
self.params["similarity_aggregation"]),
# self._canberra_score(sem_seq_char_left, sem_seq_char_right),
],
axis=-1)
# match pyramid
match_matrix_char = self._get_match_matrix(self.seq_char_left,
emb_seq_char_left,
enc_seq_char_left,
self.seq_char_right,
emb_seq_char_right,
enc_seq_char_right,
granularity="char")
mp_char = self._mp_semantic_feature_layer(
match_matrix_char,
self.dpool_index_char,
granularity="char")
# esim
esim_char = self._esim_semantic_feature_layer(
emb_seq_char_left,
emb_seq_char_right,
self.seq_len_char_left,
self.seq_len_char_right,
granularity="char")
# bcnn
bcnn_char = self._bcnn_semantic_feature_layer(
emb_seq_char_left, emb_seq_char_right, granularity="char")
# dense
deep_in_char = tf.concat(
[sem_seq_char_left, sem_seq_char_right], axis=-1)
deep_char = mlp_layer(
deep_in_char,
fc_type=self.params["fc_type"],
hidden_units=self.params["fc_hidden_units"],
dropouts=self.params["fc_dropouts"],
scope_name=self.model_name + "deep_char",
reuse=False,
training=self.training,
seed=self.params["random_seed"])
with tf.name_scope("matching_features"):
matching_features_word = tf.concat(
[
sim_word,
mp_word,
esim_word,
bcnn_word,
deep_word, # sem_seq_word_left, sem_seq_word_right,
],
axis=-1)
matching_features_char = tf.concat(
[
sim_char,
mp_char,
esim_char,
bcnn_char,
deep_char, # sem_seq_char_left, sem_seq_char_right,
],
axis=-1)
return matching_features_word, matching_features_char
def _get_matching_features(self):
with tf.name_scope(self.model_name):
tf.set_random_seed(self.params["random_seed"])
with tf.name_scope("word_network"):
if self.params["attend_method"] == "context-attention":
emb_seq_word_left, enc_seq_word_left, att_seq_word_left, sem_seq_word_left, \
emb_seq_word_right, enc_seq_word_right, att_seq_word_right, sem_seq_word_right = \
self._interaction_semantic_feature_layer(
self.seq_word_left,
self.seq_word_right,
self.seq_len_word_left,
self.seq_len_word_right,
granularity="word")
else:
# 这里应该对应了dssm的双塔模型
# semantic layer第一层为针对seq_word_left的embedding
# 然后输入embeding和training，做dropout，默认配置embedding_dropout = 0.3
# 注意这里一个语义特征组合层的输入，从basemode中通过self.trainning变量输入第一层
# 第二层为编码层，输入维度为embedding的维度，dssm使用fast text编码, 改编码方式的输出维度等于embeded的维度, reuse参数对fast text无作用。为何fast text直接返回输入值？？
# 第三层为注意力层，dssm使用了多个注意力方法，所以注意力层有多个输出, 然后被tf.concat了
# 第四层为普通的多层FC，会根据params["fc_hidden_units"]元素个数，配合params["fc_dropouts"]来构建多层全连接神经网络
emb_seq_word_left, enc_seq_word_left, att_seq_word_left, sem_seq_word_left = \
self._semantic_feature_layer(
self.seq_word_left, # 这里是tf.Variable,由baseModel初始化
self.seq_len_word_left, # 这里也是tf.Variable
granularity="word", reuse=False) # reuse 为false的应该是Query塔
emb_seq_word_right, enc_seq_word_right, att_seq_word_right, sem_seq_word_right = \
self._semantic_feature_layer(
self.seq_word_right,
self.seq_len_word_right,
granularity="word", reuse=True) # reuse 为true的应该是doc塔（一个query对应多个塔）
# match score, 原论文中的余弦相似度。作者这里还加入了点积和欧几里得距离，相似度是一个三维向量
sim_word = tf.concat(
[
metrics.cosine_similarity(
sem_seq_word_left, sem_seq_word_right,
self.params["similarity_aggregation"]),
metrics.dot_product(
sem_seq_word_left, sem_seq_word_right,
self.params["similarity_aggregation"]),
metrics.euclidean_distance(
sem_seq_word_left, sem_seq_word_right,
self.params["similarity_aggregation"]),
# self._canberra_score(sem_seq_word_left, sem_seq_word_right),
],
axis=-1)
with tf.name_scope("char_network"):
if self.params["attend_method"] == "context-attention":
emb_seq_char_left, enc_seq_char_left, att_seq_char_left, sem_seq_char_left, \
emb_seq_char_right, enc_seq_char_right, att_seq_char_right, sem_seq_char_right = \
self._interaction_semantic_feature_layer(
self.seq_char_left,
self.seq_char_right,
self.seq_len_char_left,
self.seq_len_char_right,
granularity="char")
else:
emb_seq_char_left, enc_seq_char_left, att_seq_char_left, sem_seq_char_left = \
self._semantic_feature_layer(
self.seq_char_left,
self.seq_len_char_left,
granularity="char", reuse=False)
emb_seq_char_right, enc_seq_char_right, att_seq_char_right, sem_seq_char_right = \
self._semantic_feature_layer(
self.seq_char_right,
self.seq_len_char_right,
granularity="char", reuse=True)
# match score
sim_char = tf.concat(
[
metrics.cosine_similarity(
sem_seq_char_left, sem_seq_char_right,
self.params["similarity_aggregation"]),
metrics.dot_product(
sem_seq_char_left, sem_seq_char_right,
self.params["similarity_aggregation"]),
metrics.euclidean_distance(
sem_seq_char_left, sem_seq_char_right,
self.params["similarity_aggregation"]),
# self._canberra_score(sem_seq_char_left, sem_seq_char_right),
],
axis=-1)
with tf.name_scope("matching_features"):
matching_features_word = sim_word
matching_features_char = sim_char
return matching_features_word, matching_features_char
def _get_matching_features(self):
with tf.name_scope(self.model_name):
tf.set_random_seed(self.params["random_seed"])
with tf.name_scope("word_network"):
if self.params["attend_method"] == "context-attention":
emb_seq_word_left, enc_seq_word_left, att_seq_word_left, sem_seq_word_left, \
emb_seq_word_right, enc_seq_word_right, att_seq_word_right, sem_seq_word_right = \
self._interaction_semantic_feature_layer(
self.seq_word_left,
self.seq_word_right,
self.seq_len_word_left,
self.seq_len_word_right,
granularity="word")
else:
emb_seq_word_left, enc_seq_word_left, att_seq_word_left, sem_seq_word_left = \
self._semantic_feature_layer(
self.seq_word_left,
self.seq_len_word_left,
granularity="word", reuse=False)
emb_seq_word_right, enc_seq_word_right, att_seq_word_right, sem_seq_word_right = \
self._semantic_feature_layer(
self.seq_word_right,
self.seq_len_word_right,
granularity="word", reuse=True)
# match score
sim_word = tf.concat(
[
metrics.cosine_similarity(
sem_seq_word_left, sem_seq_word_right,
self.params["similarity_aggregation"]),
metrics.dot_product(
sem_seq_word_left, sem_seq_word_right,
self.params["similarity_aggregation"]),
metrics.euclidean_distance(
sem_seq_word_left, sem_seq_word_right,
self.params["similarity_aggregation"]),
# self._canberra_score(sem_seq_word_left, sem_seq_word_right),
],
axis=-1)
with tf.name_scope("char_network"):
if self.params["attend_method"] == "context-attention":
emb_seq_char_left, enc_seq_char_left, att_seq_char_left, sem_seq_char_left, \
emb_seq_char_right, enc_seq_char_right, att_seq_char_right, sem_seq_char_right = \
self._interaction_semantic_feature_layer(
self.seq_char_left,
self.seq_char_right,
self.seq_len_char_left,
self.seq_len_char_right,
granularity="char")
else:
emb_seq_char_left, enc_seq_char_left, att_seq_char_left, sem_seq_char_left = \
self._semantic_feature_layer(
self.seq_char_left,
self.seq_len_char_left,
granularity="char", reuse=False)
emb_seq_char_right, enc_seq_char_right, att_seq_char_right, sem_seq_char_right = \
self._semantic_feature_layer(
self.seq_char_right,
self.seq_len_char_right,
granularity="char", reuse=True)
# match score
sim_char = tf.concat(
[
metrics.cosine_similarity(
sem_seq_char_left, sem_seq_char_right,
self.params["similarity_aggregation"]),
metrics.dot_product(
sem_seq_char_left, sem_seq_char_right,
self.params["similarity_aggregation"]),
metrics.euclidean_distance(
sem_seq_char_left, sem_seq_char_right,
self.params["similarity_aggregation"]),
# self._canberra_score(sem_seq_char_left, sem_seq_char_right),
],
axis=-1)
with tf.name_scope("matching_features"):
matching_features_word = sim_word
matching_features_char = sim_char
return matching_features_word, matching_features_char