I'm developing a method to document and query representation as concept vectors (bag-of-concepts). I want to train a machine learning model on ranking (learning to rank a task). So I have document vector V1 and query vector V2. How should I use these two numerical vectors in learning the way to rank a task? What are the possible scenarios?
Do I calculate relevance (similarity) by cosine and then enter the result as a single feature into a neural network? Is it correct to apply Hadamard to produce a single vector representing the features of a document and query pair, and then train a neural network with it? Can two vectors (document and query vector) be entered into the Siamese network in order to evaluate the relevance? One told me this is not possible because the network only take raw text as input and extracts features. Hence, it is useless to enter a vector that was generated by my vectorization method.
