We determine criteria for modelling

concession signalled via plan-based

“but” in task-oriented dialogue (TOD)

(following (Thomas, 2003)) based

on some examples from the corpora,

analysing where Speech Act (SA) and

planning information fails to predict

concession. In our approach we focus

on cases involving cross-speaker

concession, where the speaker accepts

part of the previous speaker’s turn but

signals contrast via “but”, and we argue

that this contrast can (in the examples

shown) be modelled as concessive.

Then given a representation of task-plan

history in the Information State (IS)

model of the dialogue, we present an

initial framework for an algorithm that

predicts concessive interpretation across

speakers in two situations encountered

in the corpora. We motivate this work

by showing how it updates beliefs in the

PTT (Poesio and Traum, 1998) model

of dialogue and can be used to facilitate

recognition of planning mismatches and

potentially avoid misunderstandings,

and more generally improve discourse

understanding and natural language

generation (NLG).