Add functionality:

Create a layer similar to the semantic layer, but instead of using the normal (mikolov) embeddings uses the recent context embeddings. This new type of embeddings could be produced by the popular transform-based architectures like BERT.

Also, take into consideration a static version of this layer, that will take an already precomputed similarity matrix since feedforward of models such as BERT are extremely expensive

Add functionality:

Create a Exact interaction layer, where the interactions are computed based on the equality of the query and documents tokens, i.e, if q_i==d_j: then 1 otherwise: 0 for all i and j

Add functionality:

Create a Semantic interaction layer, where the interactions are computed based on the cosine similarity of the query and document embeddings.

Additionally, this layer would also be responsible to encode the query and document tokens, using an embedding lookup

Add functionality:

Create a generic masking mechanism that can be easily extended by all the interaction layers.

This masking is accomplished by multiplying all the padded entries (query and document) by zero in the final interaction matrice.

Add functionality:

This layer should also consider the term importance of each query and document term. More precisely it should accept an importance vector, like TF-IDF weights, that will be added to the final interaction matrix.

Add functionality:

This layer should add the ability to compute the term importance of each query and document term as a projection of its embedding representation.