Alafate Abulimiti

Docteurs récents

Status : Doctorant

Address :

LLF, CNRS – UMR 7110
Université Paris Diderot-Paris 7
Case 7031 – 5, rue Thomas Mann,
75205 Paris cedex 13

E-mail : nynsngr.nohyvzvgv@tznvy.pbz

Thèse

Title : The Role of Socio-conversational Strategies in Task-Oriented Dialogues in the Case of Peer-Tutoring Interactions: A Focus on Off-Task Talk and Hedges

Supervision :
  Jonathan Ginzburg

PhD Defense : 2023-12-14

Inscription : 2020 à Université Paris-Cité

Jury :

  • Prof. Justine Cassell (INRIA Paris, Carnegie Mellon University) — Thesis Director
  • Prof. Chloé Clavel (LTCI, Institut Polytechnique de Paris, Telecom Paris) — Thesis Co-Director
  • Prof. Magalie Ochs (LIS UMR 7020 CNRS / AMU / UTLN) — Reporter
  • Prof. Laurent Prevot (CNRS, AMU) — Reporter
  • Prof. Gaël Guibon (University of Lorraine, LORIA) — Examiner
  • Prof. Salvador MASCARENHAS (Ecole Normale Supérieure) — Examiner

Abstract :

This thesis explores how social language enhances learning in peer tutoring conversations between teenagers. It focuses on two key conversational phenomena - off-task talk and hedging. Off-task talk refers to casual asides about random topics during tutoring. Although temporarily distracting, this thesis shows off-task exchanges might improve student learning. Using machine learning, high-performing models were developed to automatically detect off-task talk. A computational model represents how opportunistically triggering off-task exchanges can balance educational goals and rapport. Hedging means softening speech to avoid embarrassment or discomfort. Appropriate hedging makes directive feedback gentler. This thesis exposes limitations of modern language model for nuanced social skills like hedging. Novel techniques, including re-ranking model outputs, were validated to enhance hedge generation capabilities. Feature analysis of hedge prediction models provided data-driven insights into contextual factors, like gaze patterns, influencing human hedge usage. Overall, the analytical and technical advancements expand understanding of how social language subtly shapes productive tutoring interaction.