SAD: A Large-Scale Strategic Argumentative Dialogue Dataset

TL;DR

The paper introduces SAD, a large-scale, strategy-aware multi-turn argumentative dialogue dataset derived from CMV, with 392,822 dialogue instances and 722,812 utterances across 20,619 topics. It annotates each utterance with stance and one or more of five strategies (Question, Causality, Example, Analogy, Statement) and formalizes a strategy-conditioned generation task $P(A|T,H,S,[R])$, complemented by a learned persuasiveness evaluator based on user Likes. Extensive analyses reveal clear evolutionary patterns in strategy usage across dialogue phases and demonstrate that incorporating explicit strategies improves generation quality across fluency, coherence, relevance, and persuasiveness, especially when combined with fine-tuning and strategy-aware prompts. The dataset and evaluators offer a foundation for building more persuasive, strategic, and controllable argumentation dialogue systems, with public release planned for academic and non-commercial research while acknowledging biases and ethical considerations.

Abstract

Argumentation generation has attracted substantial research interest due to its central role in human reasoning and decision-making. However, most existing argumentative corpora focus on non-interactive, single-turn settings, either generating arguments from a given topic or refuting an existing argument. In practice, however, argumentation is often realized as multi-turn dialogue, where speakers defend their stances and employ diverse argumentative strategies to strengthen persuasiveness. To support deeper modeling of argumentation dialogue, we present the first large-scale \textbf{S}trategic \textbf{A}rgumentative \textbf{D}ialogue dataset, SAD, consisting of 392,822 examples. Grounded in argumentation theories, we annotate each utterance with five strategy types, allowing multiple strategies per utterance. Unlike prior datasets, SAD requires models to generate contextually appropriate arguments conditioned on the dialogue history, a specified stance on the topic, and targeted argumentation strategies. We further benchmark a range of pretrained generative models on SAD and present in-depth analysis of strategy usage patterns in argumentation.

Figures (11)

• Figure 1: An example of SAD. The speakers on the left are opponents of the topic, and the speakers on the right are supporters of the topic. The argumentative strategies used are highlighted in blue. Strategic sentences are emphasized using other colors.
• Figure 2: Cumulative distribution of strategies. The x-axis denotes the relative position in dialogue, and the y-axis denotes the cumulative proportion.
• Figure 3: Visualization of the strategy flow patterns.
• Figure 4: Instruction for Evaluator (Qwen3-8B). A self-trained evaluator is used for persuasiveness assessment.
• Figure 5: Instruction for GPT-4.1 Evaluator (Relevance).