Perspectivized Argument Knowledge Graphs for Deliberation Support

在日常生活和政治或社会背景下，许多决定都很难解决，需要考虑赞成或反对采取行动的论点。 在这一过程中，需要平衡有关各方相互冲突的利益，以实现最佳结果、普遍接受和公平。可以支持人类在审议过程中的工具，从辩论门户网站与预先结构的赞成和反对的论点广泛讨论的问题，以论点搜索引擎与不断增长，但不完善的能力，论点结构注释。然而，真实的论证远远超出预先陈述的赞成和反对论点的粗粒度结构。需要深入了解辩论的问题，这些问题带来i）决策的潜在影响; ii）这些影响如何影响相关方，以及iii）如何权衡潜在后果，以便找到最佳和广泛接受的解决方案。 重要的是，iv）我们需要能够解决以前没有讨论过的新问题，因此，需要建立能够访问知识资源并提供推理能力的论证系统。计算论证迄今为止还远远没有达到对论证的深入理解。 例如，在一个示例中，支持-攻击关系的分类工具在话语环境中表现良好，但我们的工作表明，它们主要依赖于话语标记，如果这些标记被从输入中屏蔽，则表现不佳。类似地，论证推理挑战旨在对论证进行更深入的推理。但是，基于BERT的模型的良好表现依赖于表面线索，一旦这些线索被消除，就会下降到随机行为。在之前的工作中，我们通过以可解释的方式整合背景知识，使用先进的深度学习方法和揭示隐含信息来解决这些弱点。但需要更多的研究来赋予系统更深层次的论点理解，包括对论点的推理和对行动后果的预测，以支持审议。在这个项目中，我们的目标是超越我们在基于知识的论点分析方面的成就：i）我们的目标是扩展我们的神经符号方法，以实现对背景知识支持的论点的更深层次的理解。这包括归纳论证成分之间更深层次的关系，从而增强论证的显化。ii）我们在一个多因素论证知识图谱中表示、情境化和丰富论证，该图谱超越了当前的工作，包括利益相关者的观点及其利益、价值观和目标。iii）基于论证图，我们开发了从多个角度进行推理以得出新颖结论的方法，将分解和重组论证以执行透视论证图完成，并从多个角度分析有争议的问题，并学习向替代前提推理，以提供可解释的审议支持。

Many decisions in daily life and in political or societal contexts are difficult to resolve and require deliberation of arguments in favor or against actions to be taken. In this process, conflicting interests of involved parties need to be balanced to achieve optimal outcomes, general acceptance and fairness. Tools that can support humans in deliberation processes range from debating portals with pre-structured pro and con arguments on widely discussed issues to argument search engines with growing but imperfect capacities for argument structure annotation. Real argumentation, however, goes far beyond coarse-grained structuring of pre-stated pro and con arguments. Deep understanding of debated issues is needed that brings about i) potential implications of decisions; ii) how these affect interested parties, and iii) how to weight potential consequences, in order to find optimal and widely accepted solutions. Importantly, iv) we need to be able to address novel issues not discussed before, and thus, need to build argumentation systems that have access to knowledge resources and offer reasoning capabilities.Computational argumentation is to date still far from achieving deep understanding of arguments. E.g., tools for classifying support–attack relations perform well in a discourse setting, but our work shows that they crucially rely on discourse markers and perform poorly if these are masked from the input. Similarly, the Argument Reasoning Challenge was designed to enforce deeper reasoning over arguments. But the good performance of BERT-based models turned out to rely on surface cues, and dropped to random behaviour once these cues were eliminated. In prior work we have addressed these weaknesses by integrating background knowledge in an interpretable way, using advanced deep learning methods and uncovering implicit information. But more research is needed to endow systems with deep argument understanding, which includes reasoning about arguments and foreseeing consequences of actions, to support deliberation.In this project we aim to move beyond our achievements in knowledge-based argument analysis: i) We aim to extend our neural-symbolic methods to achieve deeper understanding of arguments supported by background knowledge. This includes the induction of deeper relations between argument components leading to enhanced argument explicitation. ii) We represent, contextualize and enrich arguments in a multi-factorial argument knowledge graph that goes beyond current work by including stakeholder perspectives and their interests, values and goals. iii) Based on the argument graph we develop methods that perform reasoning to derive novel conclusions from many perspectives, will de- and recompose arguments to perform perspectivized argument graph completion, and analyze debated issues from multipleperspectives and learn to reason towards alternative premises, to offer interpretable deliberation support.