Towards Human-AI Collaborative Information Synthesis

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


PhD Thesis Defence


Title: "Towards Human-AI Collaborative Information Synthesis"

By

Mr. Chengbo ZHENG


Abstract:

Synthesizing information into a summative and coherent representation is a 
common necessity. For example, researchers need to synthesize relevant 
literature to ground their studies. The information synthesis process typically 
involves gathering and evaluating information from various sources, 
transforming information, and organizing it coherently. Although foundation 
models, such as large language models (LLMs), offer significant potential for 
automating this process, they still exhibit deficiencies like bias and 
instability. Additionally, users often seek control over the synthesis process 
to integrate their unique contexts or to learn from the process.

This thesis investigates the human-AI collaboration approach to information 
synthesis, aiming to optimally leverage the capabilities of foundation models 
while maintaining user agency. In the first part of the thesis, we adopt a 
user-centered design approach to explore how human-AI collaboration can be 
integrated into common information synthesis tasks. First, we designed 
DiscipLink, an LLM-powered interactive system that supports knowledge 
integration for interdisciplinary research, focusing on guiding the orientation 
of information gathering and rapid screening of relevant literature. Through 
two user studies, we found DiscipLink supports researchers more comprehensively 
in performing interdisciplinary information-seeking and better learning from 
the search process. Our second work presents NB2Slides, an AI-assisted 
presentation slide creation system that supports data scientists in 
synthesizing their technical work for communication with stakeholders. In our 
exploratory user study, data scientists found NB2Slides easy to use and highly 
preferred its human-AI collaboration workflow.

The second part of this thesis examines how to enable novel information 
synthesis workflows through human-AI collaboration. Specifically, we explore 
how to synthesize a self-assessment report from students' use of AI tools in 
course projects. Lacking the empirical understanding of the needs and 
challenges in this synthesis task, we approached the problem guided by 
research-through-design (RtD). We initially conducted co-design workshops to 
explore students' motivations and attitudes. Based on the insights gathered, we 
designed an intelligent tool, SelfGauge, to help students synthesize their AI 
usage and project traces into self-assessment reports, thus supporting their 
reflection on their AI-enhanced learning journeys.

In summary, this thesis contributes design knowledge on forming human-AI 
workflows for information synthesis and provides an empirical understanding of 
human-AI interaction to inform future interactive AI system design.


Date:                   Monday, 12 August 2024

Time:                   3:00pm - 5:00pm

Venue:                  Room 4472
                        Lifts 25/26

Chairman:               Dr. Song LIN (MARK)

Committee Members:      Dr. Xiaojuan MA (Supervisor)
                        Prof. Pedro SANDER
                        Prof. Qian ZHANG
                        Prof. Hongbo FU (AMC)
                        Prof. Uichin LEE (KAIST)