Towards meaningful, grounded conversations with intelligent agents
This addresses the need for more meaningful and grounded conversations with intelligent agents that can deeply understand and evolve with the world, though it appears incremental as it builds on existing paradigms.
The paper tackles the problem of conversational agents reaching performance bottlenecks with current data-intensive models, arguing for a shift from slot filling to a more realistic conversation paradigm that handles concrete and abstract entities with evolving relations.
As conversational agents become integral parts of many aspects of our lives, current approaches are reaching bottlenecks of performance that require increasing amounts of data or increasingly powerful models. It is also becoming clear that such agents are here to stay and accompany us for long periods of time. If we are, therefore, to design agents that can deeply understand our world and evolve with it, we need to take a step back and revisit the trade-offs we have made in the current state of the art models. This paper argues that a) we need to shift from slot filling into a more realistic conversation paradigm; and b) that, to realize that paradigm, we need models that are able to handle concrete and abstract entities as well as evolving relations between them.