Nowadays deep learning models can solve many challenging tasks like machine translation, image captioning, or text understanding, but results are still not completely satisfying and new ideas are needed to make further steps toward full Natural Language Processing. In this research area, modeling dialog is surely one of the most challenging application; systems need “to understand” arguments and dialog history, and to generate coherent and aligned text. The research activity will aim at studying new ways to leverage deep learning and deep reinforcement learning techniques to improve on dialog generation obtaining more precise and sensible outputs. The starting point will be to solve the question about how to include into these pipelines external source of information like knowledge bases through specialized retrieval systems, increasing model performance, reliability, and generation accuracy. Specialized libraries like Pytorch will be used to implement and prove such new approaches.


Contacts:
Rossella Cancelliere – rossella.cancelliere-at-unito.it
Giovanni Bonetta – giovanni.bonetta-at-unito.it