Semantic Alignment for Agent Interactions: making communication meaningful in open environments

The fact that the meaning of words depends on the context in which they are used is evident for any speaker: if someone asks for chips in a cafeteria, she will unlikely be expecting to get electronic circuits. In human dialogues this kind of semantic alignment happens permanently and has been extensively studied. In this talk I will discuss how these ideas can also be applied to help achieve meaningful communication in artificial multi-agent systems, in which heterogeneous interlocutors will likely use different vocabularies.

I will start by presenting a notion of context that is based on the formal specifications of the tasks performed by agents. I will then show how this context can be used by the agents to align their vocabularies dynamically, by learning mappings from the experience of previous interactions. In doing so, we will also rethink the traditional approach to semantic matching and its evaluation, tackling the following questions: What does it mean for agents to "understand each other"? When is an alignment good for a particular application? How can the interaction context help interoperability?