Open Argument Mining

公开辩论包含了如此多的论点，以至于合理的决策超出了感兴趣的公众或负责任的专家的认知能力。不断有新的论点被提出(挑战C1)，经常是不完整的(C2)，需要关于共同事实或先前论点的知识才能理解它们(C3)。本项目旨在研究计算方法，i)不断提高他们在正在进行的辩论中识别论点的能力，ii)将不完整的论点与先前的论点对齐，并用自动获取的背景知识来丰富它们，iii)不断地用理解论点所需的信息来扩展语义知识库。我们通过结合和推进论点挖掘和知识图构建这两个研究领域的现有最新算法来实现这一点。为了处理正在进行的辩论中的概念漂移，我们的目标是用一种知识感知的终身学习方法来推进论点挖掘方法。我们将研究新的神经结构来学习话题不变论点特征以及论点和辩论话题之间的关系，使用知识图嵌入向神经网络注入语义知识，并利用自我训练来不断扩展训练数据。为了处理不完整的论点，检索到的论点将与已知论点对齐，并以背景知识丰富。我们将通过链接发现和关键字搜索相结合的方式将论点实体与背景知识联系起来。这种关联的背景知识将被合并到增量聚类方法中，以将类似的论点分组到论点簇中。这些论据簇之间的论辩支持和攻击关系将使用监督学习来确定。我们的目标是通过结合包含百科全书和常识知识的当代语义知识库(Babelnet和ConceptNet)和从非结构化Web语料库中提取重点知识(Common Crawl)来自动获取所需的背景知识。为了将这些背景知识整合到机器学习模型中，我们将采用现有的知识嵌入技术来支持增量训练。此外，该项目专注于开发新的标注方案和新的基准语料库，使我们能够跨主题、文本类型和不同的时间戳来评估我们的挖掘和对齐方法，其结果将是获得一个开放论证图的新方法，该图包括来自与支持和攻击关系相关联的多个文本来源的语义丰富的相似论点组。为了确保辩论风格的广泛覆盖，我们将把我们的方法应用于在线新闻和Twitter消息中经常讨论的不同主题，并使用带注释的GOLD数据进行组成部分评估和基于人群的事后评估。

Open debates include so many arguments that sound decision making exceeds cognitive capabilities of the interested public or responsible experts. New arguments are continuously contributed (challenge C1), are oftenincomplete (C2), and knowledge about common facts or previous arguments is needed to understand them (C3).This project aims at investigating computational methods that i) continuously improve their capability to recognize arguments in ongoing debates, ii) align incomplete arguments with previous arguments and enrichthem with automatically acquired background knowledge, and iii) constantly extend semantic knowledge bases with information required to understand arguments.We achieve this by combining and advancing current state-of-the-art algorithms from the two research fields argument mining and knowledge graph construction. To deal with concept drifts in ongoing debates, we aim to advance argument mining methods with a knowledge-aware lifelong learning approach. We will investigate novel neural architectures for learning topic invariant argument features and the relation between arguments and debate topics, inject semantic knowledge into the neural network using knowledge graph embeddings and leverage self-training to continuously extend the training data. To cope with incomplete arguments, the retrieved arguments will be aligned with known arguments and enriched with background knowledge. We will link the entities of arguments to background knowledge by combining link discovery and keyword search. This linked background knowledge will be incorporated into incremental clustering methods for grouping similar arguments into argument clusters. Argumentative support and attack relations between these argument clusters will be determined using supervised learning. We aim to automatically acquire the required background knowledge by combining contemporary semantic knowledge bases containing encyclopedic and commonsense knowledge (Babelnet and ConceptNet) and focused knowledge extraction from unstructured Web corpora (Common Crawl). To integrate this background knowledge into machine learning models, we are going to adopt existing knowledge embedding techniques to support incremental training. Furthermore, this project focuses on developing novel annotation schemes and new benchmark corpora allowing us to evaluate our mining and alignment methods across topics, text types, and varying timestamps.The outcome will be novel methods for obtaining an Open Argumentation Graph including semantically enriched groups of similar arguments from multiple textual sources linked with support and attack relations. To ensure a wide coverage of argumentation styles, we will apply our methods to different topics frequently discussed in online news and Twitter messages and conduct both component evaluation using annotated gold data and crowd-based post-hoc evaluations.