I recently read this paper End-To-End Memory Networks which uses three different embeddings layers for sentence embeddings. Now I am trying to reproduce this architecture in keras.
But I am not sure how to create three different embeddings. These are exactly same dimension based on same corpus but should have different values for the embeddings. So to implement this layers, should I just use Embedding Layers with kernel_initializer =random_uniform?
I know pre-trained embeddings like Word2Vec, but currently pre-trained model is not important, is it?