Toward Intelligent Writing Support Beyond Completing Sentences
Restricted (Penn State Only)
- Author:
- Huang, Chieh Yang
- Graduate Program:
- Informatics
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- September 28, 2023
- Committee Members:
- Dongwon Lee, Professor in Charge/Director of Graduate Studies
Syed Billah, Major Field Member
C Lee Giles, Major Field Member
Rebecca Passonneau, Outside Unit & Field Member
Kenneth Huang, Chair & Dissertation Advisor - Keywords:
- Writing Support
Story Generation
Novel Writing
Large Language Model
Crowd Writing
Language Generation
Story Understanding
Creative Writing - Abstract:
- Writing is challenging, and even accomplished writers often encounter “writer’s block,” where they are not certain about what to write next. While some approaches utilize language generation models to offer sentence continuations to mitigate this issue, they often overlook a critical aspect: writers struggling with writer’s block typically lose sight of the overall goal and idea, not merely the next line. Therefore, a broader range of support is required, addressing everything from high-level plot suggestions to auto-completing paragraphs. This dissertation moves beyond traditional tools that focus on post-writing reviews (e.g., grammar checks), and offers in-situ assistance to meet diverse writer needs. We conducted four studies to explore these needs and the potential interventions. Our first study introduced HETEROGLOSSIA, a Google Docs add-on. Heteroglossia engaged crowd workers to role-play story characters, suggesting potential story continuations based on the character’s perspective. While this human-in-the-loop approach delivered inspiring ideas, it was not scalable for longer narratives. Next, we aimed to understand the potential developments in lengthy novels by employing semantic frame representation, a conceptual vector reflecting the activities that occurred in the context. By forecasting the semantic frame representation for the subsequent block, we could infer what would happen next in the story. Our third study involved a comparison of three distinct approaches for generating story plots, concluding that the semantic frame enhanced GPT-2 showed superior performance in storability and consistency. Finally, to synthesize what we have learned, we introduce INSPO, a standalone text editor that offers users access to various support types, including crowd plot ideation, AI plot ideation, AI plot ideation with guidance, and auto-completion. A week-long study with eight writers using Inspo revealed the strengths, weaknesses, and desired use cases for each function. We applied discourse community theory to investigate users’ “norm aligning” behavior, where they used Inspo to validate their writing plan. Such a sociocultural lens highlights the potential risks of employing artificial intelligence to learn community norms. In conclusion, this research offers an exploration of intelligent writing support, extending beyond mere sentence completion and paving the way for the development of more intuitive and diverse writing assistance tools.