LARGA: Learning Argumentation Axioms from Monological and Dialogical Texts

在许多现实生活中，从政治辩论到论文写作，论证的有效性不仅取决于选择最佳论点，还取决于遵循最佳策略来提供它们。因此，我们要问：在一个文本中，论证应该以什么样的顺序呈现？什么样的规则或惯例指导这些排序决策？具体的线性化以何种方式提高或降低作者观点的可接受性？是否存在适用于不同文本类型的原则？这样的组织知识的一个例子是“具有轶事类型单位的论点应该先于具有统计类型单位的论点。”这是一种常见于社论中的模式。现有的计算论证研究主要集中在论证挖掘和论证评估方面，而对于哪种文本体裁或模式的排序策略有效以及如何识别这类策略知识的实证研究还很缺乏。我们希望通过引入一种公理化方法来解决这个问题，该方法基于“主题不可知”属性对参数排列偏好进行建模。我们将这些在最近的研究过程中一直在编辑的属性分配给三个抽象层次：论元层次、论元层次和话语层次。基于足够的数据集，这必须在这三个层次上进行适当的注释，我们试图发现和分析可解释的公理知识，从不同体裁，模式和水平的写作专业知识的议论文。我们的项目提供了一个具体的计划，以获得必要的数据集，归纳所描述的那种在独白和对话设置的公理，并分析这些公理之间的有趣的关系。例如，专家论证文本与非专家论证文本在这些公理方面有何不同？此外，我们计划研究我们基于公理的方法对当前文本结构理论的影响，我们设计了实验来证明在两个下游应用中操作公理论证策略知识的好处：增强写作和对话辅助。作为本项目的一部分开发的资源，包括注释、代码、公理知识和原型工具，将免费提供，有助于RATIO优先方案和一般论证研究。

In many real-life situations, from political debates to paper writing, the effectiveness of argumentation depends not only on picking the best arguments, but also on following the best strategy to deliver them. We hence ask: In what order should arguments be presented in a text? What rules or conventions guide these ordering decisions? In what way does a specific linearisation improve or diminish the acceptability of the author's standpoint? Do principles exist that apply across text genres?An example of such organisation knowledge is ``Arguments with units of anecdote type should precede those with units of statistical type.'', which is a pattern often found in editorials. With our research, we want to systematise the identification and the analysis of such patterns (or axioms, as they are called here).Most existing work on computational argumentation concentrates on argument mining and argumentation assessment, while the empirical knowledge about what arrangement strategies are effective for which text genre or mode as well as how to identify such kind of strategy knowledge is missing so far. We want to address this gap by introducing an axiomatic approach for modeling argument arrangement preferences on the basis of ``topic-agnostic'' attributes. We assign these attributes, which we have been compiling in the course of our recent research, to three abstraction levels: the argument unit level, the argument level, and the discourse level. Based on adequate datasets, which have to be properly annotated at these three levels, we seek to uncover and analyse interpretable axiomatic knowledge from argumentative texts of different genres, modes, and levels of writing expertise.Our project provides a concrete plan to acquire the necessary datasets, to induce axioms of the described kind both in monological and dialogical settings, and to analyse interesting relations among these axioms. For instance, how do expert argumentative texts differ from non-expert argumentative texts in terms of these axioms? In addition, we plan to investigate the consequences that our axiom-based approach has on current theories of text structure, and we have devised experiments to demonstrate the benefit of operationalising axiomatic argumentation strategy knowledge in two downstream applications: augmented writing and dialog assistance.The resources developed as part of this project, including the annotations, code, axiomatic knowledge, and prototypical tools, will be made freely available, contributing to the RATIO priority programme and research on argumentation in general.