Abstract |
Human-machine interaction and communication is one of the “hottest” topics in Computer
Science. As technological advancements are getting closer to the non-expert human, the
need for implementing free-form verbal communication between a user and an intelligent
system is becoming more important, due to its simplicity and naturalness of interaction.
One method of human-computer communication appropriate for non-expert users is
“robotic conversation agents”, the job of which is to engage in conversation with the
human: the user asks questions in a natural language (e.g., English), and the chatbot
answers the user’s question in the same natural language as the question itself.
In this thesis, we designed a chatbot architecture (which we will call P-Chat), which uses
the Event Calculus, a formal action language, in order to represent the chatbot’s
knowledge. It relies on the Clingo reasoner, which implements ASP (Answer Set
Programming) rules, in order to infer new knowledge about the robot’s world and about the
agents inhabiting it, exploiting causal, temporal and epistemic reasoning.
Our architecture is adaptable to diverse domains that require chatbots for human
interaction with machines. It enables users to ask questions not only about the chatbot’s
environment, but also about other agents’ beliefs (epistemic questions) or about
retrospective events, thus enabling users to form a more complete picture about the world
and the events that occur and modify it. Finally, when it comes to training our system, we
present a rigorous training method, so that a question can be reused with different
agents/objects/domains, reducing the cost of retraining from scratch for each individual
query.
We present the implementation of a dialogue system for the interactions between two
agents – a human and a robot – which interact with other objects in the environment. All
participating agents have partial observability of the world, therefore may possess
incomplete or erroneous beliefs about the current world state or about the events that took
place. The implemented system supports a wide variety of query types, ranging from polar
to wh- questions, being either non-epistemic (i.e., about the state of the world), 1st order
epistemic (i.e., about what the robot believes the state of the world may be), or 2nd order
epistemic (i.e., about what the robot believes that other agents believe the state of the
world is).
|