Is there someone interested in goal-oriented dialog investigation ? I would like to add a “rasa-nlg” to the system. Seems the RASA architecture is quite flexible, we could create a NLG module that would complete and improve the global system : The state-of-the-art in G-O dialog systems are achieved by using both memory networks [Weston et al,2014] for the dialog policy and generative adversarial networks (GAN) [Goodfellow et al. ,2014] for the answer generation. Instead of a retrieval-based system, I think this could improve a lot the framework. Finally, as there is already an active learning implementation (interactive learning), maybe we could go a step further and build a lifelong learning system (like the LILITH projetct) in order to give more “cognition” to our bot.
Anyway, if someone is interested, tell me, so we can code together, because it will be a LOT of work.