Say Something Smart – Looking for a credible dialogue system

From HLT@INESC-ID

Date

  • 15:00, Friday, February 1st, 2013
  • Room 336

Speaker

  • David Ameixa

Abstract

In this report we target out of domain interactions in dialogue systems. Dialogue systems are always faced with users that try to cross the limits of their knowledge, by unawareness of its domain limitation or simply to test the capacity of the system. These interactions are considered out of domain since they cannot be handled by the back-end system. Edgar, a virtual butler whose goal is to give information about Monserrate’s Palace, is no exception, and its users often ask questions that are considered to be out of domain. In order to keep the users engaged, a module able to address this kind of interactions, based on the Artificial Intelligence Markup Language and Program D was created. Moreover, several other modules are being developed with the same goal (Talkpedia, Just.Ask and Just.Chat), although these are not yet integrated in Edgar. This work intends to enrich the knowledge base of one of the aforementioned modules – Just.Chat – with real human interactions based on subtitles or books’ dialogues to give the system more credibility, by making the dialogues closer to a human conversation. Also, we intent to integrate the previously implemented modules into a general framework that deals with out of domain interactions, named Say Something Smart. This report describes related work, proposes Say Something Smart architecture and points to our future work on this subject.