claim_input = Input(shape=(train_max_length,), name='claim_input')
claim_emb = Embedding(output_dim=50, input_dim=num_words,
input_length=train_max_length)(claim_input)
claim_out = LSTM(128, dropout=0.3, recurrent_dropout=0.3,
kernel_regularizer=regularizers.l2(0.01))(claim_emb)

evidence_input = Input(shape=(evid_max_length,), name='evidence_input')
evidence_emb = Embedding(output_dim=50, input_dim=num_words,
input_length=evid_max_length)(evidence_input)
evidence_out = LSTM(128, dropout=0.3, recurrent_dropout=0.3,
kernel_regularizer=regularizers.l2(0.01))(evidence_emb)

x = concatenate([claim_out, evidence_out])
x = Dense(1, activation='sigmoid')(x)
sum_model = Model(inputs=[claim_input , evidence_input], outputs=[x])
# summarize layers
print(sum_model.summary())

optimizer = Adam(lr=1e-3)
sum_model.compile(loss='binary_crossentropy',
          optimizer=optimizer,
          metrics=['accuracy'])
earlystopping = EarlyStopping(monitor='val_loss', patience=3, verbose=1)
lr_reduction = ReduceLROnPlateau(monitor='val_loss',
                                   factor=0.1, min_lr=1e-5, patience=0,
                                   verbose=1)
callbacks = [
    earlystopping,
    checkpoint,
    lr_reduction
]

sum_model.fit([claim_input , evidence_input],
          new_label,
          epochs=5,
          batch_size=128,
          steps_per_epoch=6000)