HCII Ph.D. Defense: Amber Horvath

Andrew Macvean - Google
Elena Glassman - Harvard

Abstract

Software development requires developers to juggle and balance many information-seeking and understanding tasks. From understanding how a bug was introduced, to choosing what application programming interface (API) method to use to resolve the bug, to determining how to properly integrate this change, even the smallest implementation tasks can lead to many questions. These questions may range from hard-to-answer questions about the rationale behind the original code to common questions such as how to use an API. Software development, in contrast to other sensemaking domains, has the unique property that many information artifacts are created at different points during the development process (e.g., output data). Once this challenging sensemaking is done, this rich thought history is often lost given the high cost of externalizing these details, despite potentially being useful to future developers.

In this thesis, I explore different systems and methods for authoring and using this rich thought history as meta-information about code. Specifically, I have developed systems for annotating to support developers’ natural sensemaking when understanding information-dense sources such as software documentation and source code. I then demonstrated how this meta-information can be captured and harnessed for new tasks, including for assessing the trustworthiness of documentation, for capturing design rationale and provenance data of code, and for supporting developer implementation tasks such as debugging.

This thesis begins by exploring methods for externalizing developers' thoughts in a form that is lightweight yet contextualized. We explore annotating as a method for simultaneously allowing developer's to offload their mental processes, while presenting that information in-context for later developers to utilize. We developed 2 prototype annotation systems, Adamite and Catseye, which showed the promise of annotating for assisting developers both in overcoming issues with using developer documentation and debugging code. The dynamic nature of code and its connection to annotated materials introduced unique design challenges in which information can quickly become outdated and disconnected, leading both to changes in the Catseye user interface and to the insight that leveraging the (lack of) connectivity between annotation and documentation can support other documentation-related tasks, which inspired Sodalite. The final two systems, Meta-Manager and MMAI, explore capturing other forms of already-authored meta-information, such as edit traces and log data, for question-answering support, with MMAI utilizing large language models to make that querying possible in natural language.

The series of work introduced in this thesis points to the need to treat user-thought and intent as a first-class entity and that meta-information is a way of presenting that information. I show that developers' thought histories can be represented in the form of code-related meta-information and, through proper tooling, can be used by later developers to accelerate their sensemaking of code.