ABScribe: Rapid Exploration & Organization of Multiple Writing Variations in Human-AI Co-Writing Tasks using Large Language Models

Document Title

ABScribe: Rapid Exploration & Organization of Multiple Writing Variations in Human-AI Co-Writing Tasks using Large Language Models

Document Information

• Domain: Collaborative writing interface design and evaluation in Human-Computer Interaction (HCI) using Large Language Models (LLMs)
• Keywords: LLMs, human-AI collaborative writing, text variation management, user interface, task workload, human-computer interaction, text editing, design framework

Research Background and Problem Statement

• Identified Problems or Challenges:

  1. Conventional text editors typically rely on linear revision histories, which are not suited to support nonlinear, iterative, and fine-grained needs during text rewriting.
  2. AI-generated text variations often lead to document clutter, making it difficult to efficiently manage multiple text variations simultaneously.
  3. While AI's rapid expansion capabilities are convenient, the lack of systematic methods to organize and compare these variations can hinder writing fluency and creativity.

• Research Importance:

  • Revision and rewriting are core processes in text creation, not only enhancing precision in expression but also helping authors better explore and refine their ideas.
  • As LLMs improve their ability to generate variations, supporting authors in efficiently managing and comparing multiple variations becomes increasingly critical.

• Research Motivation and Related Work:

  • Previous studies have focused on the performance of generative pre-trained models (e.g., ChatGPT, GPT-4) in creative writing but have paid less attention to user interface design for managing multiple text variations.
  • Related literature suggests that simultaneously comparing multiple options (at least five variations) can prevent design fixation and improve creativity.

Solution

• Method and Solution: The authors propose a collaborative document editing interface called ABScribe, specifically designed for human-AI collaborative text creation tasks, enabling rapid exploration and organization of multiple variations.

• Innovative Features of the Solution:

  1. Nonlinear variation organization: Through "variation fields" and "popup toolbars," the system enables nonlinear revision history and rapid comparison of text.
  2. Reusable AI instruction buttons: By converting user-input LLM modification instructions into reusable buttons, the system reduces repetitive input workload.
  3. Tightly integrated interaction design: Users can directly insert or modify text in the document using "@ai <prompt>", enhancing workflow fluidity.

• Implementation Steps and Key Technologies:

  1. Variation Fields: Users can select text paragraphs and create independent variation fields for each, storing variations side-by-side without overwriting the original text.
  2. Popup Toolbar: Allows users to quickly compare different variations by hovering, without interrupting the writing flow.
  3. Variation Sidebar: Provides structured management through collapsible navigation for organizing variations.
  4. AI Modifiers: Stores AI instructions as buttons, enabling flexible application across multiple text segments.
  5. AI Drafter Module: Allows users to directly insert AI-generated text using natural language prompts.

Research Outcomes

• Specific Results:

  1. Reduction in User Workload:
     • Compared to baseline interfaces, ABScribe significantly reduced subjective task workload (NASA-TLX score decreased by 1.20, p < 0.001).
  2. Positive User Perception of the Revision Process:
     • User evaluations indicated that ABScribe improved satisfaction with the revision process (agreement increased by 2.41, p < 0.001).
  3. User Study Results (based on 12 writers):
     • ABScribe reduced document clutter, enabling users to more freely create and explore diverse variations.
     • The button-based AI instruction design encouraged users to write clearer and more concise prompts.

• Advantages Compared to Existing Solutions:

  • ABScribe's nonlinear management of multiple variations overcomes the limitations of traditional linear writing tools, making writing more efficient.
  • Compared to conventional conversational interfaces, ABScribe's UI components promote more explicit command issuance and reduce cognitive load during task switching.

• Experiment and Evaluation Results:

  1. In tasks involving the generation and management of 24 text variations, users found ABScribe more effective for managing and comparing fine-grained text modifications.
  2. When using ABScribe, users described the revision process as "reducing choice pressure," "more exploratory," and "easier to edit small text segments."

• Limitations and Future Directions:

  1. The current study is limited to English writing environments; further validation is needed for other languages.
  2. User studies were limited to short-term experiments, and the tool's adaptability to long-term writing tasks has not been examined.
  3. Future work could incorporate interface data logging features, such as tracking word count distribution and user behavior frequency in human-AI collaborative texts, to further analyze user operation habits.