Skipgram: neural network instead of lookup table

Follow the full discussion on Reddit.
I'm looking for papers which use the skipgram model but instead of a lookup table they use a neural network. The use case is instead of sentences of words I want to use sequences of human behavior where additional information is available, e.g. think sequences of visited Amazon products. Cold-start also happens to be very common and I'm thinking that using a neural network instead of lookup embeddings table would be better.