def get_s(inputs, name):
with tf.name_scope('word_encode'):
inputs = RNN(inputs, sen_len, n_hidden=FLAGS.n_hidden, scope=FLAGS.scope + 'word_layer' + name)
inputs = tf.reshape(inputs, [-1, FLAGS.max_sen_len, sh2])
with tf.name_scope('word_attention'):
w1 = func.get_weight_varible('word_att_w1'+ name, [sh2, sh2])
b1 = func.get_weight_varible('word_att_b1'+ name, [sh2])
w2 = func.get_weight_varible('word_att_w2'+ name, [sh2, 1])
senEncode = func.att_var(inputs, sen_len, w1, b1, w2)
senEncode = tf.reshape(senEncode, [-1, FLAGS.max_doc_len, sh2])
return senEncode
def build_model(word_embedding,
pos_embedding,
x,
word_dis,
sen_len,
doc_len,
keep_prob1,
keep_prob2,
RNN=func.biLSTM):
x = tf.nn.embedding_lookup(word_embedding, x)
inputs = tf.reshape(x, [-1, FLAGS.max_sen_len, FLAGS.embedding_dim])
word_dis = tf.nn.embedding_lookup(pos_embedding, word_dis)
word_dis = tf.reshape(word_dis,
[-1, FLAGS.max_sen_len, FLAGS.embedding_dim_pos])
inputs = tf.concat([inputs, word_dis],
axis=2) # [-1, max_sen_len, dim + dim_pos]
inputs = tf.nn.dropout(inputs, keep_prob=keep_prob1)
sen_len = tf.reshape(sen_len, [-1])
with tf.name_scope('word_encode'):
lstm_wordEncode = RNN(inputs,
sen_len,
n_hidden=FLAGS.n_hidden,
scope=FLAGS.scope + 'word_layer')
lstm_wordEncode = tf.reshape(lstm_wordEncode,
[-1, FLAGS.max_sen_len, 2 * FLAGS.n_hidden])
with tf.name_scope('word_attention'):
sh2 = 2 * FLAGS.n_hidden
w1 = func.get_weight_varible('word_att_w1', [sh2, sh2])
b1 = func.get_weight_varible('word_att_b1', [sh2])
w2 = func.get_weight_varible('word_att_w2', [sh2, 1])
s_wordEncode = func.att_var(lstm_wordEncode, sen_len, w1, b1, w2)
s_senEncode = tf.reshape(s_wordEncode,
[-1, FLAGS.max_doc_len, 2 * FLAGS.n_hidden])
for i in range(FLAGS.clause_layer):
s_senEncode = RNN(s_senEncode,
doc_len,
n_hidden=FLAGS.n_hidden,
scope=FLAGS.scope + 'sentence_layer' + str(i))
n_feature = 2 * FLAGS.n_hidden
with tf.name_scope('softmax'):
s = tf.reshape(s_senEncode, [-1, n_feature])
s = tf.nn.dropout(s, keep_prob=keep_prob2)
w = func.get_weight_varible('softmax_w', [n_feature, FLAGS.n_class])
b = func.get_weight_varible('softmax_b', [FLAGS.n_class])
pred = tf.matmul(s, w) + b
pred = tf.nn.softmax(pred)
pred = tf.reshape(pred, [-1, FLAGS.max_doc_len, FLAGS.n_class],
name="pred")
reg = tf.nn.l2_loss(w) + tf.nn.l2_loss(b)
return pred, reg
def get_s(inputs, name): #xianzai的inputs 不是文档级 batchsize级别的
with tf.name_scope('word_encode'):
inputs = RNN(inputs,
sen_len,
n_hidden=FLAGS.n_hidden,
scope=FLAGS.scope + 'word_layer' + name) #30 200
with tf.name_scope('word_attention'):
sh2 = 2 * FLAGS.n_hidden
w1 = get_weight_varible('word_att_w1' + name, [sh2, sh2])
b1 = get_weight_varible('word_att_b1' + name, [sh2])
w2 = get_weight_varible('word_att_w2' + name, [sh2, 1])
s = att_var(inputs, sen_len, w1, b1, w2) #(?,200)
s = tf.reshape(
s, [-1, FLAGS.max_doc_len, 2 * FLAGS.n_hidden]) #(?,75,200)
return s
def build_model(x,
sen_len,
doc_len,
word_dis,
word_embedding,
pos_embedding,
keep_prob1,
keep_prob2,
RNN=func.biLSTM):
x = tf.nn.embedding_lookup(word_embedding, x) #选取wordembedding中x对应的元素
inputs = tf.reshape(x, [-1, FLAGS.max_sen_len, FLAGS.embedding_dim])
# print("word_dis:{}".format(word_dis))
word_dis = tf.nn.embedding_lookup(pos_embedding, word_dis)
sh2 = 2 * FLAGS.n_hidden
inputs = tf.nn.dropout(inputs, keep_prob=keep_prob1)
sen_len = tf.reshape(sen_len, [-1])
# print("sen_len:{}".format(sen_len))
with tf.name_scope('word_encode'):
wordEncode = RNN(inputs,
sen_len,
n_hidden=FLAGS.n_hidden,
scope=FLAGS.scope + 'word_layer')
wordEncode = tf.reshape(wordEncode, [-1, FLAGS.max_sen_len, sh2])
with tf.name_scope('attention'):
w1 = func.get_weight_varible('word_att_w1', [sh2, sh2])
b1 = func.get_weight_varible('word_att_b1', [sh2])
w2 = func.get_weight_varible('word_att_w2', [sh2, 1])
senEncode = func.att_var(wordEncode, sen_len, w1, b1, w2)
senEncode = tf.reshape(senEncode, [-1, FLAGS.max_doc_len, sh2])
word_dis = tf.reshape(word_dis[:, :, 0, :],
[-1, FLAGS.max_doc_len, FLAGS.embedding_dim_pos])
n_feature = 2 * FLAGS.n_hidden + FLAGS.embedding_dim_pos
out_units = 2 * FLAGS.n_hidden
for i in range(1, FLAGS.n_layers):
senEncode_dis = tf.concat([senEncode, word_dis], axis=2) # 距离拼在子句上
senEncode = trans_func(senEncode_dis, senEncode, n_feature, out_units,
'layer' + str(i))
pred, reg = senEncode_softmax(senEncode, 'softmax_w', 'softmax_b',
out_units, doc_len)
return pred, reg
def build_model(word_embedding,
pos_embedding,
word_dis,
x,
sen_len,
doc_len,
keep_prob1,
keep_prob2,
RNN=func.biLSTM):
x = tf.nn.embedding_lookup(word_embedding, x)
inputs = tf.reshape(x, [-1, FLAGS.max_sen_len, FLAGS.embedding_dim])
inputs = tf.nn.dropout(inputs, keep_prob=keep_prob1)
word_dis = tf.nn.embedding_lookup(pos_embedding, word_dis)
word_dis = tf.reshape(word_dis[:, :, 0, :],
[-1, FLAGS.max_doc_len, FLAGS.embedding_dim])
sen_len = tf.reshape(sen_len, [-1])
with tf.name_scope('word_encode'):
lstm_wordEncode = RNN(inputs,
sen_len,
n_hidden=FLAGS.n_hidden,
scope=FLAGS.scope + 'word_layer')
lstm_wordEncode = tf.reshape(lstm_wordEncode,
[-1, FLAGS.max_sen_len, 2 * FLAGS.n_hidden])
with tf.name_scope('word_attention'):
sh2 = 2 * FLAGS.n_hidden
w1 = func.get_weight_varible('word_att_w1', [sh2, sh2])
b1 = func.get_weight_varible('word_att_b1', [sh2])
w2 = func.get_weight_varible('word_att_w2', [sh2, 1])
s_wordEncode = func.att_var(lstm_wordEncode, sen_len, w1, b1, w2)
senEncode = tf.reshape(s_wordEncode,
[-1, FLAGS.max_doc_len, 2 * FLAGS.n_hidden])
n_feature = 2 * FLAGS.n_hidden
out_units = 2 * FLAGS.n_hidden
for i in range(1, FLAGS.n_layers):
senEncode_dis = senEncode + word_dis
senEncode = trans_func(senEncode_dis, senEncode, n_feature, out_units,
'layer' + str(i))
pred, reg = senEncode_softmax(senEncode, 'softmax_w', 'softmax_b',
out_units, doc_len)
return pred, reg
def build_model(x,
sen_len,
doc_len,
word_dis,
word_embedding,
pos_embedding,
keep_prob1,
keep_prob2,
RNN=func.biLSTM):
x = tf.nn.embedding_lookup(word_embedding, x)
inputs = tf.reshape(x, [-1, FLAGS.max_sen_len, FLAGS.embedding_dim])
word_dis = tf.nn.embedding_lookup(pos_embedding, word_dis)
sh2 = 2 * FLAGS.n_hidden
inputs = tf.nn.dropout(inputs, keep_prob=keep_prob1)
sen_len = tf.reshape(sen_len, [-1])
with tf.name_scope('word_encode'):
wordEncode = RNN(inputs,
sen_len,
n_hidden=FLAGS.n_hidden,
scope=FLAGS.scope + 'word_layer')
wordEncode = tf.reshape(wordEncode, [-1, FLAGS.max_sen_len, sh2])
with tf.name_scope('attention'):
w1 = func.get_weight_varible('word_att_w1', [sh2, sh2])
b1 = func.get_weight_varible('word_att_b1', [sh2])
w2 = func.get_weight_varible('word_att_w2', [sh2, 1])
senEncode = func.att_var(wordEncode, sen_len, w1, b1, w2)
senEncode = tf.reshape(senEncode, [-1, FLAGS.max_doc_len, sh2])
word_dis = tf.reshape(word_dis[:, :, 0, :],
[-1, FLAGS.max_doc_len, FLAGS.embedding_dim_pos])
senEncode_dis = tf.concat([senEncode, word_dis], axis=2) # 距离拼在子句上
n_feature = 2 * FLAGS.n_hidden + FLAGS.embedding_dim_pos
out_units = 2 * FLAGS.n_hidden
batch = tf.shape(senEncode)[0]
pred_zeros = tf.zeros(([batch, FLAGS.max_doc_len, FLAGS.max_doc_len]))
pred_ones = tf.ones_like(pred_zeros)
pred_two = tf.fill([batch, FLAGS.max_doc_len, FLAGS.max_doc_len], 2.)
matrix = tf.reshape((1 - tf.eye(FLAGS.max_doc_len)),
[1, FLAGS.max_doc_len, FLAGS.max_doc_len]) + pred_zeros
pred_assist_list, reg_assist_list, pred_assist_label_list = [], [], []
if FLAGS.n_layers > 1:
'''*******GL1******'''
senEncode = trans_func(senEncode_dis, senEncode, n_feature, out_units,
'layer1')
pred_assist, reg_assist = senEncode_softmax(senEncode,
'softmax_assist_w1',
'softmax_assist_b1',
out_units, doc_len)
pred_assist_label = tf.cast(
tf.reshape(tf.argmax(pred_assist, axis=2),
[-1, 1, FLAGS.max_doc_len]), tf.float32)
# masked the prediction at the current position
pred_assist_label = pred_assist_label * pred_two - pred_ones
pred_assist_label = (pred_assist_label + pred_zeros) * matrix
# feedforward
w_for = func.get_weight_varible('w_for1',
[FLAGS.max_doc_len, FLAGS.max_doc_len])
b_for = func.get_weight_varible('b_for1', [FLAGS.max_doc_len])
pred_assist_label = tf.tanh(
tf.matmul(tf.reshape(pred_assist_label, [-1, FLAGS.max_doc_len]),
w_for) + b_for)
pred_assist_label = tf.reshape(
pred_assist_label, [batch, FLAGS.max_doc_len, FLAGS.max_doc_len])
pred_assist_label_list.append(pred_assist_label)
pred_assist_list.append(pred_assist)
reg_assist_list.append(reg_assist)
'''*******GL n******'''
for i in range(2, FLAGS.n_layers):
senEncode_assist = tf.concat([senEncode, pred_assist_label], axis=2)
n_feature = out_units + FLAGS.max_doc_len
senEncode = trans_func(senEncode_assist, senEncode, n_feature,
out_units, 'layer' + str(i))
pred_assist, reg_assist = senEncode_softmax(
senEncode, 'softmax_assist_w' + str(i),
'softmax_assist_b' + str(i), out_units, doc_len)
pred_assist_label = tf.cast(
tf.reshape(tf.argmax(pred_assist, axis=2),
[-1, 1, FLAGS.max_doc_len]), tf.float32)
# masked the prediction at the current position
pred_assist_label = pred_assist_label * pred_two - pred_ones
pred_assist_label = (pred_assist_label + pred_zeros) * matrix
# feedforward
w_for = func.get_weight_varible('w_for' + str(i),
[FLAGS.max_doc_len, FLAGS.max_doc_len])
b_for = func.get_weight_varible('b_for' + str(i), [FLAGS.max_doc_len])
pred_assist_label = tf.tanh(
tf.matmul(tf.reshape(pred_assist_label, [-1, FLAGS.max_doc_len]),
w_for) + b_for)
pred_assist_label = tf.reshape(
pred_assist_label, [batch, FLAGS.max_doc_len, FLAGS.max_doc_len])
pred_assist_label_list.append(pred_assist_label)
pred_assist_label = tf.divide(
tf.reduce_sum(pred_assist_label_list, axis=0), i)
pred_assist_list.append(pred_assist)
reg_assist_list.append(reg_assist)
'''*******Main******'''
if FLAGS.n_layers > 1:
senEncode_dis_GL = tf.concat([senEncode, pred_assist_label], axis=2)
n_feature = out_units + FLAGS.max_doc_len
senEncode_main = trans_func(senEncode_dis_GL, senEncode, n_feature,
out_units, 'block_main')
else:
senEncode_main = trans_func(senEncode_dis, senEncode, n_feature,
out_units, 'block_main')
pred, reg = senEncode_softmax(senEncode_main, 'softmax_w', 'softmax_b',
out_units, doc_len)
return pred, reg, pred_assist_list, reg_assist_list
def build_model(x,
sen_len,
doc_len,
word_embedding,
clause_position,
embedding_pos,
keep_prob1,
keep_prob2,
RNN=func.biLSTM):
x = tf.nn.embedding_lookup(word_embedding, x)
inputs = tf.reshape(x, [-1, FLAGS.max_sen_len, FLAGS.embedding_dim])
n_hidden = 2 * FLAGS.n_hidden
inputs = tf.nn.dropout(inputs, keep_prob=keep_prob1)
sen_len = tf.reshape(sen_len, [-1])
with tf.name_scope('word_encode'):
wordEncode = RNN(inputs,
sen_len,
n_hidden=FLAGS.n_hidden,
scope=FLAGS.scope + 'word_layer')
wordEncode = tf.reshape(wordEncode, [-1, FLAGS.max_sen_len, n_hidden])
with tf.name_scope('attention'):
w1 = func.get_weight_varible('word_att_w1', [n_hidden, n_hidden])
b1 = func.get_weight_varible('word_att_b1', [n_hidden])
w2 = func.get_weight_varible('word_att_w2', [n_hidden, 1])
senEncode = func.att_var(wordEncode, sen_len, w1, b1, w2)
# (32*75,200)
senEncode = tf.reshape(senEncode,
[-1, FLAGS.max_doc_len, n_hidden]) #(32, 75, 200)
n_feature = 2 * FLAGS.n_hidden
out_units = 2 * FLAGS.n_hidden #200
batch = tf.shape(senEncode)[0] #32
pred_zeros = tf.zeros(([batch, FLAGS.max_doc_len,
FLAGS.max_doc_len])) #(32,75,75)
matrix = tf.reshape(
(1 - tf.eye(FLAGS.max_doc_len)),
[1, FLAGS.max_doc_len, FLAGS.max_doc_len]) + pred_zeros # 构造单位矩阵
pred_emotion_assist_list, reg_emotion_assist_list, pred_emotion_assist_label_list = [], [], []
pred_cause_assist_list, reg_cause_assist_list, pred_cause_assist_label_list = [], [], []
if FLAGS.assist_n_layers > 1:
'''******* emotion layer 1******'''
emotion_senEncode = trans_func(senEncode, senEncode, n_feature,
out_units,
'emotion_layer1') #(32,75,200)
pred_emotion_assist, reg_emotion_assist = senEncode_emotion_softmax(
emotion_senEncode, 'softmax_assist_w1', 'softmax_assist_b1',
out_units, doc_len)
#(32, 75,2)
pred_emotion_assist_label = tf.cast(
tf.reshape(tf.argmax(pred_emotion_assist, axis=2),
[-1, 1, FLAGS.max_doc_len]), tf.float32)
#(32, 75, 1)=>(32, 1, 75)
pred_emotion_assist_position = tf.cast(
tf.reshape(tf.argmax(pred_emotion_assist_label, axis=2), [-1, 1]) +
1, tf.int32) #emotion clause的所在位置，辅助clause的提取
pred_clause_relative_position = tf.cast(
tf.reshape(clause_position - pred_emotion_assist_position + 69,
[-1, FLAGS.max_doc_len]),
tf.float32) #基于emotion clause的相对位置 (32, 1, 75)
pred_clause_relative_position *= func.getmask(doc_len,
FLAGS.max_doc_len,
[-1, FLAGS.max_doc_len])
pred_clause_relative_position = tf.cast(pred_clause_relative_position,
tf.int32)
pred_clause_rep_embed = tf.nn.embedding_lookup(
embedding_pos, pred_clause_relative_position) #(32, 75, 50)
pred_emotion_assist_label = (pred_emotion_assist_label +
pred_zeros) * matrix # 屏蔽预测为1的标签
#matrix=>(32, 75, 75)
#pred_assist_label=>(32, 75, 75)
pred_emotion_assist_label_list.append(pred_emotion_assist_label)
pred_emotion_assist_list.append(pred_emotion_assist)
reg_emotion_assist_list.append(reg_emotion_assist)
'''******* cause layer 1******'''
cause_senEncode_assist = tf.concat([senEncode, pred_clause_rep_embed],
axis=2)
n_feature = out_units + FLAGS.embedding_dim_pos
cause_senEncode = trans_func(cause_senEncode_assist, senEncode,
n_feature, out_units, 'cause_layer')
pred_cause_assist, reg_cause_assist = senEncode_cause_softmax(
cause_senEncode, 'cause_softmax_assist_w1',
'cause_softmax_assist_b1', out_units, doc_len)
# (32, 75,2)
pred_cause_assist_label = tf.cast(
tf.reshape(tf.argmax(pred_cause_assist, axis=2),
[-1, 1, FLAGS.max_doc_len]), tf.float32)
# (32, 75, 1)=>(32, 1, 75)
pred_cause_assist_label = (pred_cause_assist_label +
pred_zeros) * matrix # 屏蔽预测为1的标签
# matrix=>(32, 75, 75)
# pred_assist_label=>(32, 75, 75)
pred_cause_assist_label_list.append(pred_cause_assist_label)
pred_cause_assist_list.append(pred_cause_assist)
reg_cause_assist_list.append(reg_cause_assist)
for i in range(2, FLAGS.assist_n_layers):
emotion_senEncode_assist = tf.concat([
emotion_senEncode, pred_emotion_assist_label,
pred_cause_assist_label
],
axis=2) # (32, 75, 275)
n_feature = out_units + 2 * FLAGS.max_doc_len # 275
emotion_senEncode = trans_func(emotion_senEncode_assist,
emotion_senEncode, n_feature, out_units,
'emotion_layer' + str(i)) # (32,75,200)
pred_emotion_assist, reg_emotion_assist = senEncode_emotion_softmax(
emotion_senEncode, 'emotion_softmax_assist_w' + str(i),
'emotion_softmax_assist_b' + str(i), out_units, doc_len)
pred_emotion_assist_label = tf.cast(
tf.reshape(tf.argmax(pred_emotion_assist, axis=2),
[-1, 1, FLAGS.max_doc_len]), tf.float32)
# pred_emotion_assist_position = tf.cast(tf.reshape(tf.argmax(pred_emotion_assist_label, axis=2), [-1, 1]),
# tf.float32) + 1 # emotion clause的所在位置，辅助clause的提取
# pred_clause_relative_position = tf.reshape(clause_position - pred_emotion_assist_position,
# [-1, FLAGS.max_doc_len]) # 基于emotion clause的相对位置 (32, 1, 75)
# pred_clause_relative_position *= func.getmask(doc_len, FLAGS.max_doc_len, [-1, FLAGS.max_doc_len])
# pred_clause_rep_embed = tf.nn.embedding_lookup(embedding_pos, pred_clause_relative_position) # (32, 75, 50)
pred_emotion_assist_label = (pred_emotion_assist_label +
pred_zeros) * matrix
pred_emotion_assist_label_list.append(pred_emotion_assist_label)
pred_emotion_assist_label = tf.reduce_sum(
pred_emotion_assist_label_list, axis=0)
# 不同layer加和 pred_assist_label=>(32,75,75)
pred_emotion_assist_list.append(pred_emotion_assist)
reg_emotion_assist_list.append(reg_emotion_assist)
cause_senEncode_assist = tf.concat([
cause_senEncode, pred_cause_assist_label, pred_emotion_assist_label
],
axis=2) #(32, 75, 275)
n_feature = out_units + 2 * FLAGS.max_doc_len #275
cause_senEncode = trans_func(cause_senEncode_assist, cause_senEncode,
n_feature, out_units,
'cause_layer' + str(i)) #(32,75,200)
pred_cause_assist, reg_cause_assist = senEncode_cause_softmax(
cause_senEncode, 'cause_softmax_assist_w' + str(i),
'cause_softmax_assist_b' + str(i), out_units, doc_len)
pred_cause_assist_label = tf.cast(
tf.reshape(tf.argmax(pred_cause_assist, axis=2),
[-1, 1, FLAGS.max_doc_len]), tf.float32)
pred_cause_assist_label = (pred_cause_assist_label +
pred_zeros) * matrix
pred_cause_assist_label_list.append(pred_cause_assist_label)
pred_cause_assist_label = tf.reduce_sum(pred_cause_assist_label_list,
axis=0)
#不同layer加和 pred_assist_label=>(32,75,75)
pred_cause_assist_list.append(pred_cause_assist)
reg_cause_assist_list.append(reg_cause_assist)
'''*******Main******'''
if FLAGS.main_n_layers > 1:
senEncode_main = tf.concat([emotion_senEncode, cause_senEncode],
axis=2)
n_feature = 2 * out_units
senEncode_main = trans_func(senEncode_main, senEncode_main, n_feature,
out_units, 'main_layer1')
senEncode_main = tf.concat([
senEncode_main, pred_emotion_assist_label, pred_cause_assist_label
],
axis=2)
n_feature = out_units + 2 * FLAGS.max_doc_len
senEncode_main = trans_func(senEncode_main, senEncode_main, n_feature,
out_units, 'main_layer2')
else:
senEncode_main = tf.concat([emotion_senEncode, cause_senEncode],
axis=2)
n_feature = 2 * out_units
senEncode_main = trans_func(senEncode_main, senEncode_main, n_feature,
out_units, 'main_layer1')
pred, reg = senEncode_main_softmax(senEncode_main, 'softmax_w',
'softmax_b', out_units, doc_len)
return pred, reg, pred_emotion_assist_list, reg_emotion_assist_list, pred_cause_assist_list, reg_cause_assist_list