Locating strongly informative utterances in conversation using multimodal cues
Résumé
Interaction theories argue that mutual understanding between speakers in natural conversations arises from building shared knowledge (common ground), but no model specifies what information is retained or under what conditions. Previous studies have used Information Theory metrics to quantify the dynamics of information exchanged between participants but lack an efficient way to identify which information becomes common ground. These attempts furthermore limited themselves to the study of conversation transcripts, overlooking nonverbal cues like visuals and intonation. To address this, we propose a method for annotating new corpora using models trained on a subset of annotated utterances. Results show a fair applicability (κ 0.3) across corpora, though this is strongly modulated by the conversational task being investigated.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |