VISAR: A Human-AI Argumentative Writing Assistant with Visual Programming and Rapid Draft Prototyping

Document Title

VISAR: A Human-AI Argumentative Writing Assistant with Visual Programming and Rapid Draft Prototyping

Document Information

• Subject Area: Human-AI collaborative writing assistance, visual programming, and rapid draft generation technology
• Keywords: Human-AI collaboration, writing assistance, creative support, large language models, visualization tools, visual programming, rapid prototyping, AI writing, argumentative writing, logical writing

Research Background and Problem

• What problems or challenges did the authors identify?

  1. The process of argumentative writing is complex, involving the construction of multi-level goals, enhancement of persuasiveness, and iterative refinement across stages.
  2. Current interactive writing assistance tools, such as systems based on large language models (LLMs), often lack the capability to seamlessly connect different stages and levels of abstraction in writing.
  3. Generation models based on chat interfaces (e.g., ChatGPT) fail to fully support users' needs for controlling writing intent, revising plans, and explaining across levels of abstraction.

• Why is this problem important?

  • Argumentative writing requires students, academics, and even professional writers to effectively construct structured, logical, and persuasive texts, yet current writing tools fall short in comprehensively supporting this task.
  • The complexity arising from the highly iterative and non-linear nature of the writing process remains inadequately addressed.

• Research Motivation and Related Work

  • The authors propose the VISAR system, aiming to address the shortcomings of existing tools by integrating visual programming and the rapid prototyping capabilities of AI models to assist users in planning, organizing, and improving their argumentative writing process.

Solution

• Method or Solution: The authors developed a Human-AI collaborative writing assistance tool called VISAR, which supports the following core functionalities:

  1. Hierarchical Writing Goal Recommendation: Utilizes the "chain of thought" technique in large language models to generate primary and secondary hierarchical writing goals.
  2. Synchronized Text and Visual Planning: Designs a system that synchronizes a text editor with an interactive visual programming tree diagram.
  3. Argumentative Sparks: Provides suggestions based on logical fallacies, counterarguments, and supporting evidence.
  4. Rapid Draft Generation: Automatically generates drafts based on user-defined writing plans for further modification and refinement.

• Innovations:

  1. Innovatively combines visual programming and rapid prototyping methods for the planning phase of argumentative writing, bridging the semantic gap between AI language models and users.
  2. Allows users to freely switch between text and visual planning interfaces, leveraging the advantages of both technologies.
  3. Introduces a technique for hierarchical, progressive recommendations of writing goals to align with users' non-linear, iterative thinking processes.

• Implementation Steps and Key Technologies:

  1. After users input a core writing topic into VISAR, the system provides step-by-step goal expansion suggestions.
  2. Constructs logical structures and dynamically adjusts them using synchronized text and visual editing interfaces.
  3. Employs VISAR's "Argumentative Sparks" feature to offer counterarguments, logical error prompts, and supporting evidence suggestions for the current draft.
  4. Finally, assists users in organizing their thoughts and refining their writing plans through rapid generation of preliminary drafts.

Research Outcomes

• Specific Outcomes:

  1. Experimental results demonstrate that VISAR effectively supports users in writing planning and significantly improves the logical structure and coherence of their texts.
  2. Provides a user-friendly interactive platform, receiving positive feedback from participants, particularly for its argumentative sparks and rapid draft generation features.

• Advantages Compared to Existing Solutions:

  • Compared to traditional LLM-based text generation tools, VISAR offers greater possibilities for user interaction and control by combining visualization and text-based regulation.
  • Focuses not only on final manuscript generation but also on the planning and revision stages of writing.

• Experimental or Evaluation Results:

  • In a user experiment involving 12 participants, VISAR significantly outperformed traditional text editors and GPT Playground with auto-completion features.
  • Qualitative feedback indicates that VISAR's interactive design and tool features significantly reduce the complexity of planning writing and improve writing efficiency.

• Limitations and Future Directions:

  1. The current VISAR system is limited to tree-structured hierarchical writing planning; future work should enhance support for more complex relationships such as graph-based and recursive logic structures.
  2. Drafts generated by VISAR exhibit issues such as uniform language style and redundant content, which are related to the underlying implementation of its large language model and require further optimization.
  3. Expand VISAR's potential applications to broader writing domains, such as education and collaborative team writing.
  4. Address ethical concerns related to AI-generated content, such as bias or copyright infringement.

Appendix

• Provides detailed experimental design, statistical results of user surveys, descriptions of research task topics, as well as specific interaction flowcharts and case text demonstrations for key system tasks.