Inference mechanisms using pragmatics and detailed knowledge for Commonsense reasoning

常识推理（CSR）是任何真正的人工智能代理的关键组成部分；此外，解释自然语言（NL）问题的能力对人工智能的许多应用都很重要。该项目旨在调查并在这一重要但具有挑战性的领域取得进展。在问答任务中，企业社会责任的一个关键方面是能够利用丰富的知识，以及对世界的共同假设和以可行的方式解释术语的能力。构建这样的知识是一个巨大的挑战，也是一个活跃的研究领域。包括语用学在内的认知理论已经发展得很好，涉及对术语的共同假设和解释；然而，关于如何从这些理论中提取规则却鲜有提及。接下来的问题是，我们如何利用现有的知识库，同时保持来自语用学的想法，以解决问题回答基准？国家语言可以是非常密集的信息，允许说话者在几个单词中传达很多信息。关于语篇理论以及人类在交际中如何处理这种不规范的语义，已有大量的研究。只有通过语用推理，句子才能转化为说话人想要表达的命题的形式表征。因此，任何试图解释自然语言的系统都应该包含推断这些隐藏信息的机制。此外，能够解释这些类型的推论在AI的CSR中非常重要，特别是在Winograd Schema Challenge （WSC）中。然而，这项任务的难度意味着在形式化这些实用知识或将其整合到推理系统上的工作很少。尽管有人试图将语用学的某些方面形式化，但正如Bunt和Black所指出的那样，语用学的某些部分“并没有丰富的表征形式主义”。从语用学的角度来看，一个有趣的途径是如何处理原型。正如我在论文中所讨论的，吸引原型是CSR的一个重要方面。这就提出了两个特别的问题，首先，我们如何识别原型？其次，我们如何识别何时使用原型是不合适的，在这些情况下，我们如何改变/放松我们正在使用的定义？例如，当一个人对“奖杯装不进手提箱，因为它太大了”这句话进行推理时，什么是太大了？没有必要担心“大”概念的精确语义承诺，而是考虑满足大多数“大”概念的“大”解释来评估句子。还有一个有趣的关于空间介词的例子。像“on”这样的空间介词是严重不明确的；它可以用来表示各种各样的空间配置——桌子上的玻璃，墙上的画等等。虽然它似乎有一个原型定义-在上面和接触-它通常只是表示两个物体之间的显著空间关系，然后从两个物体和上下文的知识中我们推断出更精确的配置。理解何时使用原型“on”和“large”是一个类似的问题——解释说话者的意图和世界知识是必要的。本研究的目的是开发口译系统，同时结合详细的知识和使用语用学原则来指导这一点。WSC有助于激发这种讨论，然而，将推理的必要方面形式化以解决WSC并将它们集成到一个系统中是出了名的困难。因此，WSC不一定是一个适合解决的问题。由于我们将试图提取复杂的语用信息，因此在语义表示得到很好发展和理解的领域工作是有意义的，因此我打算限制在空间领域的问题上工作。

Commonsense reasoning (CSR) is a key component of any truly artificially intelligent agent; moreover the ability to interpret questions posed in natural language (NL) is important for many applications of AI. This project aims to investigate and make progress in this important but challenging area.A key aspect of CSR involved in question answering tasks is the ability to draw on a wealth of knowledge as well as appeal to common assumptions about the world and interpret terms in a defeasible manner. The task of structuring such knowledge is a huge challenge and is an active area of research. Well-developed cognitive theories exist involving pragmatics, dealing with common assumptions and interpretations of terms; however little is said on how to extract rules from these theories. The question then is, how can we leverage existing knowledge bases while maintaining ideas from pragmatics in order to solve question answering benchmarks?NL can be very informationally dense, allowing speakers to convey a lot of information in only a few words. There is a wealth of research relating to theories of discourse and how humans manage to deal with this semantic under specification in communication. Only with pragmatic inferences can a sentence be transformed into a formal representation of the intended proposition of the speaker. Therefore any system attempting to interpret NL should incorporate mechanisms for inferring this hidden information. Further, being able to account for these sorts of inferences can be very important in CSR for AI, in particular in the Winograd Schema Challenge (WSC). However, the difficulty of the task means that little work exists on formalising this pragmatic knowledge or integrating it into reasoning systems.Though there exist attempts to formalize some aspects of pragmatics, as pointed out by Bunt and Black some parts of pragmatics "do not enjoy a wealth of representational formalisms". One interesting avenue from the perspective of pragmatics is how to deal with prototypes. As I discuss in my paper, appealing to prototypes is an important aspect of CSR. This raises two particular questions, firstly how do we identify prototypes? Secondly, how do we identify when using prototypes is inappropriate and in these cases how do we change/loosen the definitions we are using? For instance, when one reasons about the sentence "The trophy doesn't fit into the suitcase because it is too large", what is too large? It is not necessary to worry about a precise semantic commitment for the notion of 'large', but instead to evaluate the sentence considering an interpretation of large which satisfies most notions of large. There is also the interesting case of spatial prepositions. A spatial preposition like 'on' is heavily underspecified; it can be used to denote a variety of spatial configurations - the glass on the table, the picture on the wall etc. Though it appears to have a prototypical definition - being above and in contact with - it often just denotes a salient spatial relationship between two objects and then from knowledge of the two objects and context we infer a more precise configuration. Understanding when to use a prototypical 'on' is a similar problem to 'large' - interpreting the intention of the speaker and world knowledge is required.The aim of this research is to develop systems for interpreting NL while incorporating detailed knowledge and using the principles of pragmatics to guide this. The WSC helps motivates this discussion, however formalizing the necessary aspects of reasoning to tackle the WSC and integrating them into one system is notoriously hard. Therefore, the WSC is not necessarily a suitable problem to tackle. As we will be trying to extract complex pragmatic information it makes sense to work in an area where semantic representations are well developed and understood, therefore I intend to restrict to work on problems in the spatial domain.