Plan Dynamically, Express Rhetorically: A Debate-Driven Rhetorical Framework for Argumentative Writing

Xueguan Zhao, Wenpeng Lu, Chaoqun Zheng, Weiyu Zhang, Jiasheng Si, Deyu ZhouEMNLP 2025

TL;DR

This paper introduces DARE, a novel framework for argumentative essay generation that tackles the limitations of current models by dynamically planning arguments through simulated debates and incorporating rhetorical situation theory for more persuasive and diverse writing. DARE significantly improves logical depth, argumentative diversity, and rhetorical persuasiveness, offering a more human-like approach to generating complex argumentative texts.

What Problem Does This Paper Address?

Argumentative essay generation (AEG) is a challenging task that requires not only understanding complex topics and reasoning logically but also organizing diverse perspectives and expressing them persuasively. Existing methods often rely on static planning, which can lead to limited argument diversity and a lack of rhetorical flexibility. They tend to prioritize logical rigor over the dynamic, hierarchical nature of human argumentation, potentially resulting in "perspective collapse" where only a single viewpoint is explored. Furthermore, current approaches often overlook the importance of adapting rhetorical strategies to the specific context of the writing, a crucial aspect of human persuasive communication.

Methods

DARE employs a two-stage, debate-driven framework:

  1. Debate-Guided Outline Tree Planning:
    • Host Agent: A central agent initiates the process by analyzing the topic and establishing a core argument, forming the root of an outline tree.
  2. Depth-then-Breadth Strategy: The outline tree is dynamically expanded. "In-depth extension" elaborates on existing arguments by adding supporting sub-arguments and evidence. "Breadth extension" explores new, diverse perspectives from a parent node.
  3. Multi-Agent Debate: To ensure logical coherence and refine arguments, simulated debates are conducted between "Defender" and "Attacker" agents. This iterative process of presenting, rebutting, and responding helps to uncover deeper issues and strengthen arguments.
  4. Debate Record: The interactions from the debate are recorded and used to inform subsequent planning and writing.
  5. Rhetorical Argumentative Writing:
    • Outline Tree Extension: The planning process continues until the host agent deems the outline sufficiently developed in both depth and breadth.
  6. Rhetorical Situation Optimization: Inspired by Bitzer's rhetorical situation theory (Exigence, Audience, Constraints), DARE dynamically selects appropriate rhetorical strategies (e.g., metaphor, rhetorical questions, value appeals) to enhance the persuasiveness and clarity of each argument.
  7. Essay Generation: The LLM synthesizes the structured outline tree and rhetorical optimizations into a cohesive introduction, body, and conclusion, forming the final argumentative essay.

Datasets

The authors evaluated DARE on four datasets:

  • ArgEssay: 11,282 topic-argumentative text pairs from international standardized English writing tests (IELTS, TOEFL).
  • NYT-Editorial: 9,178 topic-argumentative text pairs from the "Room for Debate" section of The New York Times, featuring professional writers' perspectives on current events.
  • CHE-Essay: 3,750 prompt-essay pairs from the Chinese national college entrance examination (Gaokao), representing exam-style writing.
  • CHN-Editorial: 2,998 argumentative essays from Chinese news websites (PengPai, GuangMingWang), covering various current events.

These datasets cover both exam essays and news editorials, in both English and Chinese.

Key Insights and Ideas

  • Dynamic Planning is Crucial: Static planning in AEG can lead to a lack of diversity and rhetorical depth. DARE's debate-driven, dynamic approach allows for more nuanced exploration of arguments.
  • Simulated Debate Enhances Coherence: The iterative defender-attacker interaction effectively simulates human debate, leading to more logically sound and well-reasoned arguments by forcing critical examination and refinement.
  • Rhetorical Situation Theory Matters: Explicitly incorporating Bitzer's theory helps tailor rhetorical strategies to the specific context, significantly boosting persuasiveness and overall essay quality, rather than just applying rhetorical devices arbitrarily.
  • Depth and Breadth are Complementary: The depth-then-breadth strategy ensures that arguments are not only logically developed but also explore a wide range of perspectives, avoiding redundancy.
  • Human-like Argumentation: DARE's framework mimics human debate and writing processes, leading to outputs that are not only factually correct but also more engaging and persuasive.

Key Results

  • Superior Performance: DARE consistently outperformed strong baselines (DD-KW, E2E, CoT, ToT, DPE, D2W) across all five automatic evaluation metrics (Relevance, Logic, Complexity, Persuasiveness, Rhetoric) and on all four datasets.
  • Significant Improvements: DARE showed notable gains in Logic, Argument Complexity, and Rhetoric, indicating its effectiveness in structuring content and deepening reasoning.
  • Human Evaluation Alignment: Human evaluations corroborated automatic metrics, showing DARE's superiority in generating high-quality arguments, with significant gains in Rhetoric and Persuasiveness.
  • Outline Tree Depth and Breadth: Analysis showed that DARE's outline trees consistently increased in depth and breadth, leading to improved semantic coherence and logical reasoning. BARTScore and NLI-based non-contradiction scores showed upward trends with greater tree depth.
  • Rhetorical Strategy Effectiveness: Manual analysis revealed that DARE effectively balances logical appeals (Logos) with emotional (Pathos) and credibility-building (Ethos) strategies, enhancing overall persuasiveness.

Limitations and Future Work

  • Capturing Hierarchy: A key challenge is helping LLMs fully grasp the hierarchical structure of the outline tree. While sub-claims are generated, logical relations are textually annotated. Future work could explore discourse markers or rhetorical cues for more implicit hierarchy conveyance.
  • Handling Long Contexts: As outline depth and breadth grow, arguments become longer and harder to process. The authors suggest hierarchical summarization, selective filtering, or memory-augmented mechanisms to address this.
  • Topic Drift: The paper acknowledges that topic drift can occur during debates, even with explicit prompts, potentially leading to semantic mismatches between claim layers.
  • Rhetoric Undermining Coherence: In some cases, rhetorical optimization might unintentionally disrupt logical flow, making it harder to preserve structural coherence. Highly abstract or figurative rewrites can obscure relations.
  • Future Work: The authors suggest exploring more effective ways to convey hierarchy implicitly and developing methods to handle longer contexts more efficiently.

Summary updated

BibTeX
@inproceedings{zhao2025plan,
  author = {Xueguan Zhao and Wenpeng Lu and Chaoqun Zheng and Weiyu Zhang and Jiasheng Si and Deyu Zhou},
  title = {{Plan Dynamically, Express Rhetorically: A Debate-Driven Rhetorical Framework for Argumentative Writing}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure