ReMLAV: Relational Machine Learning for Argument Validation

我们关注论证机的一个重要方面——论证验证，并将问题形式化为我们所说的论证链接的验证：一个论证中的两个连续句子。当且仅当两个句子构成一个论证的有效部分时，一个论证链接被归类为有效的，例如，因为第一个句子需要或逻辑上暗示第二个句子。我们提出了一种新的参数链接验证方法，该方法基于三个先前工作的主体：(i)关系机器学习，（ii）嵌入和（iii）子空间分析。(i)关系机器学习涉及对象之间关系的模型；我们在这里通过将句子建模为对象和关系（在有效性上）或（在更细的粒度上）作为蕴涵、因果关系、矛盾等来应用它。（ii）嵌入是语言对象的高维表示，并支持强大的模型，可以很好地推广到新数据。我们在这里用它们来表示论点中的句子。（iii）嵌入的缺点是难以检查和理解。但是这种能力对于人类可以使用的论证机器来说是至关重要的。我们开发了嵌入空间的子空间模型，支持人类理解句子嵌入空间和论证机决策。

We focus on argument validation, one important aspect of argumentation machines, and formalize the problem as the validation of what we call an argument link: two consecutive sentences in an argument. An argument link is classified as valid if and only if the two sentences form a valid part of an argument, e.g., because the first entails or logically implies the second. We propose a new argument link validation method that is based on three bodies of prior work: (i) relational machine learning, (ii) embeddings and(iii) subspace analysis. (i) Relational machine learning is concerned with models of relations between objects; we apply it here by modeling sentences as objects and relations as (in)validity or (on a finer grain) as entailment, causality, contradiction etc. (ii) Embeddings are high-dimensional representations of linguistic objects and support powerful models that generalize well to new data. We apply them here to representing sentences in arguments. (iii) Embeddings have the drawback that they are difficult to inspect andunderstand. But this ability is critical for argumentation machines that are usable by humans. We develop subspace models of the embedding space that support humans in understanding sentence embedding spaces and argumentation machine decisions.