|OCCHIUTO DANIELE||Cycle: XXXII |
Section: Computer Science and Engineering
Tutor: PERNICI BARBARA
Advisor: GARZOTTO FRANCA Major Research topic
:Levels of representation as bridge in the interaction continuum to prevent the rise of uncanny valley in conversationsAbstract:
My thesis highlights the need for a shared knowledge between humans and machine in conversational settings. It explores the differences between natural and artificial languages and captures the notion of abstraction under levels of representation to prevent the uncanny valley in conversations.
My Goal is to minimize the perceived eeriness in dialogues with conversational agents. In other words, minimize the gap in the interaction continuum between users and machines. As a consequence, represent the prelinguistically available “abstractions
” that demonstrate how utterances give rise to meaning, i.e., investigate the bond between syntax and semantics in natural language.
My first contribution consists of emphasizing the existence of an unspoken conversational agreement where both parties retain the concepts they can share. Machines prefer to represent a semantic abstraction by defining its properties. This idea is not new in information technology; ontologies best describe the modeling of entities with given properties and the relations among them. I am not satisfied with ontologies since there is no dynamic way to alter the entities and relationships.
My second contribution resides in determining how shared knowledge is structured. Researchers highlighted the presence of the so-called “conversational referents”  in human dialogues. Conversational referents indicate where the user lingers during the conversation and what they are most likely to produce as utterances. Referents encompass the voice tone, the body movements, and the like. But what conversational referents can be identified only from the dialogue act? I analyze referents coming from:
- the language of the dialogue,
- the contents of the dialogue,
- the flow of the dialogue.
In the end, conversational referents establish patterns in the dialogue, and when linked to the levels of representation they signal which utterance is more appropriate to the discussion. The agent can then produce a correct utterance because it understands the concept at hand from the shared knowledge. The understanding of a concept should be perceived as less eerie from the user, rather than having a response not respecting the conversational agreement. As a consequence, in my research, I expect to witness a shrinkage in the interaction continuum gap, possibly preventing the rise of the uncanny valley in natural language interaction.