HCII Ph.D. Proposal: Andrew Kuznetsov
When
-
Description
AI-Mediated Distributed Sensemaking
Time and Location:
April 23rd, 2024 - 11:30am ET
Gates & Hillman Centers (GHC) 6501
Zoom Link link
Thesis Committee:
Aniket Kittur (Co-Chair), HCII, CMU
Anita Williams Woolley (Co-Chair), Organizational Behavior and Theory, CMU
Ken Koedinger, HCII, CMU
Ryen W. White, Microsoft Research
Abstract:
While the research community has proposed many systems to help people extract, organize, and annotate snippets of information, there is still a significant gap between what exists and the vision of “distributed sensemaking”: where collections and experiences are merged together in a way that is useful to others. Recent advances in natural language generation have enabled the creation of many helpful task ‘Copilots’ that operate on top of existing wiki, workspace, and even web browsing data. However, these tools are designed to perform intelligent information retrieval and aggregation at interactive speeds - their need to keep things ‘easy to read’ results in much of the rich provenance data from the original authoring process being tucked away. This creates a missed opportunity, as prior research has repeatedly demonstrated that this deprioritized information is often sought and relied on to make decisions: e.g. who wrote this, why, how long ago? Without this data, it is difficult to imagine scaling sensemaking beyond the individual - in many domains, the inability to make informed decisions regarding information trustworthiness reduces the usage of the respective resource and prevents the development of collective intelligence, especially in situations where users may not be familiar with one another apriori (e.g. temporary teams or asynchronous work) or with the task at hand (e.g. non-experts or one-off tasks).
This thesis outlines an approach to scaling sensemaking called “AI-Mediated Distributed Sensemaking” (AIMDS) that builds upon existing sensemaking and organizational behavior theory. In this approach, AI systems mediate the presentation of provenance data from the authoring process, providing users with a bridge between artifacts created during individual sensemaking and a compiled collective resource. By surfacing this information in a lightweight way, this approach offers a scaffold from which users can inspect individual artifacts, form judgements regarding artifact origins (e.g. authors), and ultimately learn to trust collective resources, thus developing collective intelligence.
To begin, the thesis outlines the core ideas of AI-Mediated Distributed Sensemaking, and introduces a prototype implementation through a framework named THESEUS (Towards Helpful and Effective Sensemaking: Enabling User-Centric Synthesis). The framework outlines four key phases where provenance data interacts with user sensemaking: provenance capture (foraging), organization (structuring), interpretation (evaluation), and contextualization (adaptation). The rest of the thesis describes specific research contributions that explore the AI-mediation of each phase. To start, the thesis introduces Fuse, a prototype system that enables users to capture and organize provenance data in a lightweight way during information foraging and structuring. Next in the section regarding provenance data interpretation, the thesis describes a mixed method study of the impact of surfacing provenance data on user trust perceptions in an existing general-purpose knowledge platform. In the contextualization section, the thesis presents a multi-day mixed methods study of the common contexts found in everyday complex sensemaking tasks, and describes how the THESEUS framework can be applied to contextualize task-data outside of typical knowledge work such as informal home care. To complete my dissertation, I propose the iterative development and study of an AI-mediated contextualization system for the informal home care setting, and demonstrate its utility in supporting the development of collective intelligence in caregiving teams.
Proposal Document:
https://drive.google.com/drive/u/2/folders/12EYpmeInACW_CVpLpjElbasYnja28R-p
Time and Location:
April 23rd, 2024 - 11:30am ET
Gates & Hillman Centers (GHC) 6501
Zoom Link link
Thesis Committee:
Aniket Kittur (Co-Chair), HCII, CMU
Anita Williams Woolley (Co-Chair), Organizational Behavior and Theory, CMU
Ken Koedinger, HCII, CMU
Ryen W. White, Microsoft Research
Abstract:
While the research community has proposed many systems to help people extract, organize, and annotate snippets of information, there is still a significant gap between what exists and the vision of “distributed sensemaking”: where collections and experiences are merged together in a way that is useful to others. Recent advances in natural language generation have enabled the creation of many helpful task ‘Copilots’ that operate on top of existing wiki, workspace, and even web browsing data. However, these tools are designed to perform intelligent information retrieval and aggregation at interactive speeds - their need to keep things ‘easy to read’ results in much of the rich provenance data from the original authoring process being tucked away. This creates a missed opportunity, as prior research has repeatedly demonstrated that this deprioritized information is often sought and relied on to make decisions: e.g. who wrote this, why, how long ago? Without this data, it is difficult to imagine scaling sensemaking beyond the individual - in many domains, the inability to make informed decisions regarding information trustworthiness reduces the usage of the respective resource and prevents the development of collective intelligence, especially in situations where users may not be familiar with one another apriori (e.g. temporary teams or asynchronous work) or with the task at hand (e.g. non-experts or one-off tasks).
This thesis outlines an approach to scaling sensemaking called “AI-Mediated Distributed Sensemaking” (AIMDS) that builds upon existing sensemaking and organizational behavior theory. In this approach, AI systems mediate the presentation of provenance data from the authoring process, providing users with a bridge between artifacts created during individual sensemaking and a compiled collective resource. By surfacing this information in a lightweight way, this approach offers a scaffold from which users can inspect individual artifacts, form judgements regarding artifact origins (e.g. authors), and ultimately learn to trust collective resources, thus developing collective intelligence.
To begin, the thesis outlines the core ideas of AI-Mediated Distributed Sensemaking, and introduces a prototype implementation through a framework named THESEUS (Towards Helpful and Effective Sensemaking: Enabling User-Centric Synthesis). The framework outlines four key phases where provenance data interacts with user sensemaking: provenance capture (foraging), organization (structuring), interpretation (evaluation), and contextualization (adaptation). The rest of the thesis describes specific research contributions that explore the AI-mediation of each phase. To start, the thesis introduces Fuse, a prototype system that enables users to capture and organize provenance data in a lightweight way during information foraging and structuring. Next in the section regarding provenance data interpretation, the thesis describes a mixed method study of the impact of surfacing provenance data on user trust perceptions in an existing general-purpose knowledge platform. In the contextualization section, the thesis presents a multi-day mixed methods study of the common contexts found in everyday complex sensemaking tasks, and describes how the THESEUS framework can be applied to contextualize task-data outside of typical knowledge work such as informal home care. To complete my dissertation, I propose the iterative development and study of an AI-mediated contextualization system for the informal home care setting, and demonstrate its utility in supporting the development of collective intelligence in caregiving teams.
Proposal Document:
https://drive.google.com/