Natural Language (NL) is by far the easiest and most powerful communication device we possess, so it is reasonable to require an intelligent machine to be able to communicate through language . Since the early days of Artificial Intelligence, researchers have tried to design systems able to communicate with humans, by relying on complex hand-engineered features (i.e., dialogue acts and slots). Although these handcrafted approaches can be effective for specific domains, it is clear that this kind of supervision will not allow these systems to scale due to the high variability of Natural Language utterances and the high cost and effort of data annotation for each new application.
In recent years, Deep Reinforcement Learning techniques have demonstrated their effectiveness in different challenging games such as Go  and Poker . In particular, they allow agents to directly learn a mapping between states and actions without the need for hand-engineered features, by means of Deep Neural Networks. In the same way, these techniques can be exploited to train a conversational agent able to interact with humans by learning to compose meaningful sentences. The composition of Natural Language sequences can be realised as a sequence generation problem in which the agent learns how to put together symbols of language in a sensible way for the task that it needs to solve, similar to a child learning to speak by putting together simple words and by receiving feedback from his/her relatives.
The results from my PhD research program will try to outline possible ways to answer some of the following research questions:
- Is it possible to train a conversational agent to interact within real world contexts consisting of embodied agents and situated objects?
- Is it possible to train a conversational agent to generate accurate contextualised responses for the user?
- Is the system able to adapt easily to different domains by exploiting what it has previously learned?
: Mikolov, Tomas, Armand Joulin, and Marco Baroni. "A roadmap towards machine intelligence." arXiv preprint arXiv:1511.08130 (2015).
: Silver, David, et al. "Mastering the game of Go with deep neural networks and tree search." Nature 529.7587 (2016): 484-489.
: Moravčík, Matej, et al. "Deepstack: Expert-level artificial intelligence in no-limit poker." arXiv preprint arXiv:1701.01724 (2017).
I am a first year PhD student under the supervision of Prof. Oliver Lemon at the CDT of “Robotics and Autonomous Systems” at the Edinburgh Centre for Robotics. I am mainly interested in developing Deep Learning models for conversational agents that are able to continuously learn by interacting with the user and by exploiting features coming from the environment in which they are deployed. I received an MSc degree in "Knowledge Engineering and Machine Intelligence" from the University of Bari and before starting my PhD I was an NLP consultant for Plusimple, a health care startup. I was mainly responsible for developing a search engine able to retrieve personalised contents for the user as well as for developing an analytics platform aimed at extracting knowledge from raw data that were extracted from users sessions.