ORCID ID

https://orcid.org/0000-0003-2498-705X

Date Awarded

2023

Document Type

Dissertation

Degree Name

Doctor of Philosophy (Ph.D.)

Department

Computer Science

Advisor

Denys Poshyvanyk

Committee Member

Oscar Chaparro

Committee Member

Adwait Nadkarni

Committee Member

Huajie Shao

Committee Member

Robert M Lewis

Abstract

Software has eaten the world with many of the necessities and quality of life services people use requiring software. Therefore, tools that improve the software development experience can have a significant impact on the world such as generating code and test cases, detecting bugs, question and answering, etc. The success of Deep Learning (DL) over the past decade has shown huge advancements in automation across many domains, including Software Development processes. One of the main reasons behind this success is the availability of large datasets such as open-source code available through GitHub or image datasets of mobile Graphical User Interfaces (GUIs) with RICO and ReDRAW to be trained on. Therefore, the central research question my dissertation explores is: In what ways can the software development process be improved through leveraging DL techniques on the vast amounts of unstructured software engineering artifacts? We coin the approaches that leverage DL to automate or augment various software development task as Intelligent Software Tools. To guide our research of these intelligent software tools, we performed a systematic literature review to understand the current landscape of research on applying DL techniques to software tasks and any gaps that exist. From this literature review, we found code generation to be one of the most studied tasks with other tasks and artifacts such as impact analysis or tasks involving images and videos to be understudied. Therefore, we set out to explore the application of DL to these understudied tasks and artifacts as well as the limitations of DL models under the well studied task code completion, a subfield in code generation. Specifically, we developed a tool for automatically detecting duplicate mobile bug reports from user submitted videos. We used the popular Convolutional Neural Network (CNN) to learn important features from a large collection of mobile screenshots. Using this model, we could then compute similarity between a newly submitted bug report and existing ones to produce a ranked list of duplicate candidates that can be reviewed by a developer. Next, we explored impact analysis, a critical software maintenance task that identifies potential adverse effects of a given code change on the larger software system. To this end, we created Athena, a novel approach to impact analysis that integrates knowledge of a software system through its call-graph along with high-level representations of the code inside the system to improve impact analysis performance. Lastly, we explored the task of code completion, which has seen heavy interest from industry and academia. Specifically, we explored various methods that modify the positional encoding scheme of the Transformer architecture for allowing these models to incorporate longer sequences of tokens when predicting completions than seen during their training as this can significantly improve training times.

DOI

https://dx.doi.org/10.21220/s2-23n9-zs76

Rights

© The Author

Download

Share Feedback

Share

COinS