This statement of purpose is provided as required for the ICS PhD Portfolio. It summarizes my goals, interests, and progress towards degree completion.
Note: This statement of purpose represents my intentions in 2022 and is not current with my present research. See my Projects page for more details.
Modern Artificial Intelligence systems have demonstrated incredible feats in recent years, including advances in AI generated art and text as well as contributions to scientific advancement in other fields. These successes have been accomplished primarily using modern variants of neural networks, trained on vast pools of training data. AI systems now permeate our lives, and it is likely that AI’s influence on society will only continue to expand for the foreseeable future. As impressive as the accomplishments of modern neural networks have been, their ubiquity in society has also highlighted several shortcomings of purely sub-symbolic AI systems. In most cases, a neural network is not able to justify or explain its decisions to a human being. Additionally, it can be difficult to modify a neural network’s behavior without vast quantities of new training data. Neural networks can be particularly difficult to control because at their core they are pattern-matching algorithms which do not rely on traditional logical inferences and declarations. Although sub-symbolic AI systems have had many impressive successes, we should not give up on re-incorporating logical reasoning into AI. Logical reasoning has two primary benefits within the context of AI. First, logical reasoning is more comprehensible to human beings than the workings of neural networks and statistical algorithms. Second, logical reasoning is required or at least beneficial for solving highly complex problems that cannot be reduced to problems of classification, perception, or mimicry.
There are several promising avenues for pursuing the goal of merging symbolic and sub-symbolic reasoning. My research background is in the field of robotics, which has long relied on modular software architectures which include both symbolic and sub-symbolic components working in tandem. Another promising area of research is Natural Language Processing. Human language is too vague and too varied to be described by a formal language, making purely logical interpretation of natural language texts impossible. However, human communication still relies on a degree of logical consistency that is difficult to enforce in a sub-symbolic system. My recent research as been into ways of interpreting and representing the contents of stories. Stories are a particular promising area for study because narrative is a core component of human communication, so understanding and modelling narrative could aid in human-computer interaction. Stories also share certain elements and motifs, such as characters, settings, and a generally linear plot. These are elements which may or may not be present in other common source materials for NLP research, such as news articles, social media posts, or textbooks. Developing AI which can create representational models of stories could be an important step towards the future development of AI that more fully understands the people and environment it is interacting with.
Over the past year I have been investigating existing techniques for an NLP system which can start with a story in un-annotated plain-text and produce a representational model of the characters, setting, and plot points which occur within the story. I have not found any extant literature on any system capable of accomplishing this goal, and I must conclude that this remains an unsolved problem. I am currently developing a system that will solve this problem by isolating areas of uncertainty within a semantic parse of the text and applying sub-symbolic techniques used in modern question-answering systems to answer a series of automatically generated questions about the text. The answers to these questions will then be combined into a complete model of the events of the story represented in formal language. Full details of the proposed system will be included in my dissertation proposal, which I hope to present in Spring 2023. Details of my research into current systems that accomplish similar goals, and an analysis of the strengths and weaknesses of those systems towards the goal of story modelling, is provided in the literature review provided with this portfolio.
Over the next year, I plan to develop the described story-modelling software and publish at least one peer-reviewed paper on the results. After completing my PhD my goal is to work at a teaching-focused university, since my desire to teach has always been my primary motivation for seeking a PhD. In addition to working as a TA at UH for ICS211, ICS361, and ICS461, I have also taught summer courses on programming to middle schoolers, been a mentor for FIRST robotics club for high schoolers, and mentored several professional colleagues through learning or improving their programming skills in an office setting. If I do not achieve my initial goal of an academic position, I will continue to seek other ways to spread computer science education and to use my skills as a programmer and researcher to benefit my community in whatever ways are available to me.