How to build a character-level siamise network using Keras

Question

I am trying to build a Siamese neural network on characters-level using Keras, to learn if two names are similars or not.

So my two inputs X1 and X2 are a 3-D matrices:
X[number_of_cases, max_length_of_name, total_number_of_chars_in_DB]

In the real case:

• number_of_cases = 5000
• max_length_of_name = 50
• total_number_of_chars_in_DB = 38

I have one output binary matrix of size y[number_of_cases].

So for example: print(X1[:3, :2])

Will give the following result:

[[[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
   0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0.]
  [0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
   0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]]

 [[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
   0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0.]
  [0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
   0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]]

 [[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0.
   0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
  [0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0.
   0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]]]

I use the following code to build my model:

from keras.models import Sequential
from keras.layers import Dense
from keras.layers import LSTM, SimpleRNN
from keras.models import Model
import keras
from keras import backend as k

input_1 = Input(shape=(X1.shape[1], X1.shape[2],))
input_2 = Input(shape=(X2.shape[1], X2.shape[2],))

lstm1 = Bidirectional(LSTM(256, input_shape=(X1.shape[1], X1.shape[2],), return_sequences=False))
lstm2 = Bidirectional(LSTM(256, input_shape=(X1.shape[1], X1.shape[2],), return_sequences=False))

l1_norm = lambda x: 1 - k.abs(x[0] - x[1])

merged = Lambda(function=l1_norm, output_shape=lambda x: x[0], name='L1_distance')([lstm1, lstm2])

predictions = Dense(1, activation = 'sigmoid', name='classification_layer')(merged)

model = Model([input_1, input_2], predictions)
model.compile(loss = 'binary_crossentropy', optimizer="adam", metrics=["accuracy"])

model.fit([X1, X2], validation_split=0.1, epochs = 20,shuffle=True, batch_size = 256)

I am getting the following error:

Layer L1_distance was called with an input that isn't a symbolic tensor.

I think that the error is that I need to tell the L1_distance layer to use the output of the two precedent LSTM layers, but I do not know how to do it.

The second question, is, am I obliged to add an embedding layer before the LSTM, even in the scenario of character level network?

Thank you.