Grounded and Ungrounded Referring Expressions in Human Dialogues: Language Mirrors Different Grounding Conditions

We study how language use differs between dialogue partners in a visually grounded reference task when a referent is mutually identifiable by both interlocutors vs. when it is only available to one of them. In the latter case, the addressee needs to disconfirm a proposed description – a skill largely neglected by both the theoretical and the computational linguistics communities. We consider a number of linguistic features that we expect to vary across conditions. We then analyze their effectiveness in distinguishing among the two conditions by means of statistical tests and a feature-based classifier. Overall, we show that language mirrors different grounding conditions, paving the way to future deeper investigation of referential disconfirmation