Specification of dialogue functionality for the Demo-2 demo

We hereby make the first step towards the design of four dialogues that we will present as part of the demonstration of Demo-2. We will determine the final dialogues by trial and error, that is, we try utterances and follow-up utterances until we find a sequence that the system handles correctly. The precise sequence depends on details of system behaviour, for example, the questions that the qa's give useable answers to. So, at that moment, we will need to freeze the Demo-2 version that we designed the dialogues on. For this reason, we will delay the determination of the dialogues themselves until the final version of Demo-2, which will be due february.

To make the final determination of dialogues easy, we will come up with two things now: a precise specification of the types of follow-up user utterances and system reactions that will be demonstrable in the demo, and a set of cases of such utterances. These cases will be taken from a corpus of "second utterances" collected from potential users. The collection of this corpus is in progress (see at the bottom of the page). The relevant follow-up questions in the corpus will be annotated with the corresponding re-written self-contained question.

Handling of well-formed follow-up questions

We will support a doable subset of follow-up questions. They are dealt with by rewriting them to self-contained questions. The system will reliably support: In addition, the system will have less reliable support for:

Handling of ill-formed follow-up questions and negative user feedback

What's left is the description of the negative user feedback that the system understands. This feedback will be dealt with either by the problematic user question detection procedure, or by a simple prompt for a reformulation of the question. The negative feedback that is understood is:

Dealing with perception and interpretation problems

In some cases, the system may not be able to construct a reliable intepretation of the user utterance, due to errors signalled during the processes of either speech recognition, syntactic analysis, semantic tagging or dialogue act recognition.

The system will respond by prompting the user to rephrase or repeat his/her utterance (question). The prompts will based on the generation of negative auto-feedback dialogue acts that are mapped directly to fixed system prompts.

User-1: wat is er te doen aan een ....
{Word is not understood}
Sys-1: ik heb de vraag niet begrepen, kunt u hem herformuleren?
 

User-1: .....
{system cannot understand what the user says}
Sys-1: ik heb u niet kunnen verstaan, kunt u het herhalen?

Following social conventions

The system will respond to greetings/openings, thanks and valedictions/closings in a reactive manner.

User-1: Hallo, ik wilde graag weten wat [...] wat is er te doen aan een [...]
{User asks normal question, but starts with greeting}
Sys-1: Goedendag. Ik heb de volgende antwoorden gevonden: ...
 

User-n: Oke, bedankt.
{user thanks system after system reply}
Sys-n: Graag gedaan. U kunt met de knoppen afsluiten of een nieuwe dialoog beginnen.

Annotated corpus

Part of the corpus that is being collected has been annotated. The part of this which is relevant as examples to the above specification is found here:



Boris van Schooten en Simon Keizer, 16 december 2005