[paarthn]

One possible improvement can be training the text embeddings along with the entire model (Instead of using the pretrained embeddings like skip-thought-vectors). It is on my to-do list, I'll try it out.