Andrea Zugarini (DIISM, University of Siena)
March 22, 2018 – 9:30 AM
DIISM, Artificial Intelligence laboratory (room 201), Siena SI
Current task-oriented dialogue systems are limited to operate in a pre-defined domain. They provide good performances, but they lack of generalization capabilities because of their bound with the specific task.
On the other hand, opendomain chat systems can capture interesting aspects of dialogue.
However there is not structured knowledge, which is a crucial aspect for any task-oriented application.
A bridge between the two approaches could lead to better conversational agents.
In He at al. they present a symmetric collaborative dialogue setting in which two agents communicate to achieve a common goal.
Each agent has its own private knowledge, and, as humans do, they cooperate exploiting language as a communication mean. During the conversation both KBs are updated from unstructured language utterances.
Agents are trained exploiting a dataset of 11k human-human dialogues.