Between the Lines - Knowledge-enhanced Argument Analysis in a Formal Argumentation Reasoning System

议论文广泛应用于政治语篇和一般交际中。论据的力度和性质各不相同，人们所理解的，明确地说出的话只反映了构成论据的知识和推理的一部分。本项目将研究计算分析和验证论点的方法，以便i)用自动获取的提供缺失解释链接的知识来补充公开表达的论点，ii)构建对论点的正式的、知识增强的分析，并在此基础上，iii)使用机器学习和形式推理的组合来建立和验证扩展的论点结构。我们的目标是通过开发一个知识增强的形式论证推理系统来改进现有的论证分析方法，该系统可以对论证文本进行语义分析。我们通过分析论元中语句之间的语义连贯来实现这一点。我们将把给定陈述中的实体和概念与知识库联系起来，并学习重建隐含理解的背景知识，以增强论点的语义连贯性。作为这个过程的结果，我们将获得一个抽象的论证知识图，它构建在显性的论证文本上，并丰富了从现有知识库中获得的相关隐性知识或从文本来源中获得的相关隐性知识。我们应用有监督的机器学习来检测关联知识中的抽象模式，以确定和评分相关(类型)的关联知识。然后，形式推理过程将基于论证图中每个潜在边的语义连贯性，基于显式语句的类型、数量和连通性以及添加的隐含知识，共同建立形式论点结构并确定论点的强度。结果将是链接到可扩展知识源的论点的语义丰富的形式表示：结构化知识库是从具有各种背景和领域特定信息的文本源动态丰富的。我们将使用事实知识的链接知识库(例如，DBpedia，或gene Ontology)，语言本体(WordNet)，以及常识知识的大型知识库(ConceptNet，OpenCyc)，以及从相关文本或网络语料库自动获取的背景知识。我们将我们的方法应用于不同流派的一般领域和特殊领域的论证，并针对现有系统进行组件评估和基准测试。在SPP健壮论证机器上，我们的项目侧重于验证：知识增强分析将使论点的潜在逻辑和隐含假设变得明确。除其他外，该项目将通过提供关于辩论中隐含知识的新颖数据集和一个评价平台，为社区行动、共同任务和评价工作作出贡献。

Argumentation is widely used in political discourse and general communication. Arguments vary in strength and quality, and what is explicitly said reflects only part of the knowledge and reasoning that underlies an argument, as understood by humans. This project will investigate methods to computationally analyze and validate arguments in order to i) complement the overtly expressed argument with automatically acquired knowledge that provides missing explanatory links, to ii) construct a formal, knowledge-enhanced analysis of the argument, and on this basis, iii) establish and verify the extended argument structure using a combination of machine learning and formal reasoning. We aim to advance current methods in argument analysis by developing a knowledge-enhanced formal argument reasoning system that analyses argumentative texts semantically. We achieve this by analyzing the semantic coherence between statements in an argument. We will link entities and concepts in the given statements to knowledge bases and learn to reconstruct implicitly understood background knowledge that enhances the argument's semantic coherence. As a result of this process we will obtain an abstract argumentation knowledge graph, constructed over explicit argumentative text and enriched with relevant implicit knowledge obtained from existing knowledge bases or harvested from textual sources. We apply supervised machine learning to detect abstract patterns in the linked knowledge, in order to determine and score relevant (types of) connecting knowledge. A formal reasoning process will then jointly establish the formal argument structure and determine the strength of the argument, based on the semantic coherence of each potential edge in the argumentation graph based on the type, the amount and the connectivity of the explicit statements and the added implicit knowledge.The outcome will be a semantically enriched formal representation of arguments linked to extensible knowledge sources: structured knowledge bases that are dynamically enriched from textual sources with various kinds of background and domain-specific information. We will employ linked knowledge repositories of factual knowledge (eg. DBpedia, or Gene Ontology), linguistic ontologies (WordNet), and large repositories of common sense knowledge (ConceptNet, OpenCyc) in conjunction with automatically harvested background knowledge, from relevant text or web corpora. We apply our methods to argumentation in general and special domains in different genres, and perform both component evaluations and benchmarking against existing systems.Within the SPP Robust Argumentation Machines our project focuses on Validation: knowledge-enhanced analysis will make the underlying logics and implicit assumptions of arguments explicit. The project will contribute to community actions, shared tasks and evaluation efforts by providing novel data sets on implied knowledge in arguments and an evaluation platform, among others.