Reasoning about Structured Story Representations

当我们以人类的身份阅读一个故事时，我们会对故事所描述的内容建立一个心理模型。这种思维模式对阅读理解至关重要。它们使我们能够将故事与我们以前的经历联系起来，需要结合不同句子的信息进行推断，并正确解释模棱两可的句子。至关重要的是，心理模型捕捉到的信息比故事中真正提到的要多。它们是所描述的情境的呈现，而不是文本本身，它们是通过将故事文本与我们对世界运行方式的常识理解结合起来而构建的。自然语言处理（NLP）领域在过去几年中取得了迅速的进展，但重点主要集中在句子级表示上。新闻文章、社交媒体帖子或医疗病例报告等故事，基本上都是以句子的集合为模型的。因此，当前的系统与语言的模糊性作斗争，因为对单词或句子的正确解释通常只能通过考虑其更广泛的故事背景来推断。在解决需要将不同句子的信息组合起来的问题时，他们的能力也受到严重限制。最后一个例子是，当前的系统很难识别相关故事之间的对应关系（例如，关于同一事件的不同新闻文章），特别是如果它们是从不同的角度编写的。为了解决这些基本挑战，我们需要一种方法来学习故事层面的表征，这种表征可以作为心理模型的类比。从直觉上看，学习这种故事表示有两个步骤：首先，我们需要对故事中真正提到的内容进行建模，然后我们需要某种形式的常识性推理来填补空白。然而，在实践中，这两个步骤是密切相关的：解释故事中提到的内容需要一个故事情境的模型，而构建这个模型则需要对故事中提到的内容进行解释。我在这篇论文中提出的解决方案是基于故事图的表示。这些故事图对发生的事件、涉及的实体以及这些实体和事件之间的关系进行编码。然后，故事可以被视为故事图的不完整说明，类似于符号知识库对应于可能世界的不完整说明。基于这种观点，我们将依赖（加权）逻辑编码来表示我们对给定故事的了解。这些编码将特别充当对可能的故事图进行排序的紧凑表示，即对故事的可能解释进行排序。为了推理故事图，我提出了神经网络与系统推理的创新结合。关键思想是使用编码为图神经网络的集中推理模式。这些神经网络的预测将在本质上扮演与符号人工智能框架中的规则应用相同的角色。通过这种方式，我们的方法将神经网络的泛化能力和灵活性与具有原则性和可解释性的高级推理过程的优势紧密结合起来。提出的框架将允许我们以有原则的方式对文本信息进行推理。在需要对所描述的情况有常识性理解，或者需要将来自多个句子或文档的信息组合在一起的NLP任务中，它将导致显著的改进。它还将进一步改变直接依赖结构化文本表示的应用程序，如情景理解、法律、医疗和新闻领域的信息检索系统，以及从新闻故事和社交媒体提要中推断商业见解的工具。

When we read a story as a human, we build up a mental model of what is described. Such mental models are crucial for reading comprehension. They allow us to relate the story to our earlier experiences, to make inferences that require combining information from different sentences, and to interpret ambiguous sentences correctly. Crucially, mental models capture more information than what is literally mentioned in the story. They are representations of the situations that are described, rather than the text itself, and they are constructed by combining the story text with our commonsense understanding of how the world works.The field of Natural Language Processing (NLP) has made rapid progress in the last few years, but the focus has largely been on sentence-level representations. Stories, such as news articles, social media posts or medical case reports, are essentially modelled as collections of sentences. As a result, current systems struggle with the ambiguity of language, since the correct interpretation of a word or sentence can often only be inferred by taking its broader story context into account. They are also severely limited in their ability to solve problems where information from different sentences needs to be combined. As a final example, current systems struggle to identify correspondences between related stories (e.g. different news articles about the same event), especially if they are written from a different perspective.To address these fundamental challenges, we need a method to learn story-level representations that can act as an analogue to mental models. Intuitively, there are two steps involved in learning such story representations: first we need to model what is literally mentioned in the story, and then we need some form of commonsense reasoning to fill in the gaps. In practice, however, these two steps are closely interrelated: interpreting what is mentioned in the story requires a model of the story context, but constructing this model requires an interpretation of what is mentioned. The solution I propose in this fellowship is based on representations called story graphs. These story graphs encode the events that occur, the entities involved, and the relationships that hold between these entities and events. A story can then be viewed as an incomplete specification of a story graph, similar to how a symbolic knowledge base corresponds to an incomplete specification of a possible world. Based on this view, we will rely on (weighted) logical encodings to represent what we know about a given story. These encodings will in particular serve as a compact representation of a ranking over possible story graphs, i.e. a ranking over possible interpretations of the story. To reason about story graphs, I propose an innovative combination of neural networks with systematic reasoning. The key idea is to use focused inference patterns that are encoded as graph neural networks. The predictions of these neural networks will essentially play the same role as rule applications in symbolic AI frameworks. In this way, our method will tightly integrate the generalisation abilities and flexibility of neural networks with the advantages of having a principled and interpretable high-level reasoning process. The proposed framework will allow us to reason about textual information in a principled way. It will lead to significant improvements in NLP tasks where a commonsense understanding is required of the situations that are described, or where information from multiple sentences or documents needs to be combined. It will furthermore enable a step change in applications that directly rely on structured text representations, such as situational understanding, information retrieval systems for the legal, medical and news domains, and tools for inferring business insights from news stories and social media feeds.