From Fluency to Communicative Competence: Designing Conversational Agents for High-Stakes Dialogue

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


PhD Thesis Defence


Title: "From Fluency to Communicative Competence: Designing Conversational 
Agents for High-Stakes Dialogue"

By

Mr. Dingdong LIU


Abstract:

Recent large language models have given conversational agents (CAs) 
open-domain fluency, expanding the range of dialogue tasks they can take on. 
Yet fluency in producing language does not by itself amount to communicative 
competence: the ability to coordinate turns, sustain mutual engagement, and 
elicit what users do not explicitly say. This thesis takes up the challenge 
of moving CAs from fluency to communicative competence.

We focus on a class of interactions in which the agent is taking on, or 
working alongside, portions of work normally performed by trained 
practitioners. Taking healthcare dialogue as the empirical setting and 
drawing on theories of dialogue grounding, we argue that communicative 
competence operates at two levels: a coordination level (turn-taking, floor 
management, engagement) and a content level (eliciting latent intent, 
scaffolding self-report, aligning mental models). The two levels form a 
coupled hierarchy rather than a pipeline: progress at one level introduces 
new demands at the other.

The thesis develops this argument through five studies. Papers 1 and 2 
address coordination-level competence: Paper 1 presents an inter-pausal- unit 
and behavioral-cue architecture for floor coordination in humanoid- robot 
patient interviews, with interaction rules derived from clinician co-design; 
Paper 2 generalizes it with multimodal LLM-driven dynamic prompting for 
stance estimation, and in evaluation interrupted users less than a native 
audio LLM. Papers 3 and 4 advance content-level competence: Paper 3 designs 
scaffolding for self-report in hospital admission interviews, grounded in 
expert input; Paper 4 formalizes information prematureness in patient 
narratives from large-scale telehealth data and derives scaffolding from 
model reasoning rather than encoded rules, improving information delivery 
quality in a between-subjects study. Paper 5 contributes a diagnostic 
taxonomy that traces communication breakdowns in automated neurocognitive 
disorder screening to four structural origins. Throughout, we combine 
clinician co-design, system building, and mixed- methods user studies. Across 
both levels, the studies also trace a progression from expert-grounded 
designs, encoded from clinician practice, to reasoning-grounded designs 
derived from model capability; we argue this progression is a sequence rather 
than a rivalry, with empirical groundwork supplying the domain formalizations 
that model reasoning then scales.

The thesis offers a view of human-CA communication in which coordination and 
content competence are coupled, and in which the central difficulty at both 
levels is grounding what is not explicitly said. It contributes systems for 
participating in such dialogue and a framework for understanding where 
competence can and cannot yet be expected.


Date:                   Monday, 10 August 2026

Time:                   2:00pm - 4:00pm

Venue:                  Room 5501
                        Lifts 25/26

Chairman:               Dr. June Zijun SHI (MARK)

Committee Members:      Dr. Xiaojuan MA (Supervisor)
                        Prof. Fugee TSUNG (IEDA)
                        Dr. Sehi L'YI
                        Prof. Qian ZHANG
                        Prof. Bertram SHI (ECE)
                        Dr. Chun YU (Tsinghua University)