def __init__(self,
concept2idx,
tag2idx,
word2concept,
lstm_dim=50,
hidden_dim=50):
self.concept2idx = concept2idx,
self.tag2idx = tag2idx,
self.word2concept = word2concept
c_i = Input(shape=(dataset.MAX_SEQ_LEN, ), name="char_input")
w_i = Input(shape=(dataset.MAX_SEQ_LEN, ), name="word_input")
c_emb = dataset.get_char_emb_layer()(c_i)
w_emb = dataset.get_word_emb_layer()(w_i)
x = Bidirectional(LSTM(lstm_dim, return_sequences=True,
name="lstm"))(concatenate([c_emb, w_emb]))
x = TimeDistributed(Dense(hidden_dim, activation="sigmoid"))(x)
o = TimeDistributed(Dense(1, activation="sigmoid", name="output"))(x)
self.model = Model(inputs=[c_i, w_i], outputs=o)
self.model.compile(loss=losses.binary_crossentropy,
optimizer="adam",
metrics=[seq_accuracy, "accuracy"])
def __init__(self,
concept2idx,
tag2idx,
word2concept,
lstm_dim=50,
hidden_dim=50):
self.concept2idx = concept2idx,
self.tag2idx = tag2idx,
self.word2concept = word2concept
c_i = Input(shape=(dataset.MAX_SEQ_LEN, ), name="char_input")
con_i = Input(shape=(dataset.MAX_SEQ_LEN, ), dtype="int32")
t_i = Input(shape=(dataset.MAX_SEQ_LEN, ), dtype="int32")
c_emb = dataset.get_char_emb_layer()(c_i)
con_emb = Lambda(K.one_hot,
arguments={'num_classes': len(concept2idx) + 2},
output_shape=(dataset.MAX_SEQ_LEN,
len(concept2idx) + 2))(con_i)
t_emb = Lambda(K.one_hot,
arguments={'num_classes': len(tag2idx) + 2},
output_shape=(dataset.MAX_SEQ_LEN,
len(tag2idx) + 2))(t_i)
x = Bidirectional(LSTM(lstm_dim, return_sequences=True, name="lstm"))(
concatenate([c_emb, con_emb, t_emb]))
x = TimeDistributed(Dense(hidden_dim, activation="sigmoid"))(x)
o = TimeDistributed(Dense(1, activation="sigmoid", name="output"))(x)
self.model = Model(inputs=[c_i, con_i, t_i], outputs=o)
self.model.compile(loss=losses.binary_crossentropy,
optimizer="adam",
metrics=[seq_accuracy, "accuracy"])