Enabling Common Sense Reasoning in Natural Language Processing Systems

在自然语言处理系统中实现常识推理

• 批准号: RGPIN-2020-04871
• 负责人: Cheung, Jackie
• 金额: $ 2.11万
• 依托单位: McGill University
• 依托单位国家: 加拿大
• 项目类别: Discovery Grants Program - Individual
• 财政年份: 2022
• 资助国家: 加拿大
• 起止时间: 2022-01-01 至 2023-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=750343
• 关键词: Enabling; Common; Sense; Reasoning; Natural

Artificial intelligence systems that understand and generate natural language have made great strides in recent years, leading to successful commercial deployments. Nevertheless, there is ample evidence that current natural language processing (NLP) systems fail in systematic and un-human-like ways, despite high levels of average performance. Much of this failure is due to the lack of knowledge about typical causes and effects in the world, as well as a limited ability to reason about them; in other words, they lack common sense knowledge about the world. My research program aims to develop NLP systems that are capable of such common sense reasoning skills, such that they can communicate and collaborate with human users without requiring explicit training for every new task or topic. The key idea of the proposed research is to build an explicit formal structure that connects the textual elements to each other, on the one hand, and also to relevant background information that the system has already built up knowledge about, on the other. On the technical side, we will propose new neural network architectures and models, which are promising because they provide easy and natural methods to transfer knowledge between multiple sources and tasks via network parameter sharing in multi-task learning and transfer learning setups. The objectives of the project will be organized by the type of common sense and world knowledge structure that is required for a particular inference task. The first one is to make use of the extensive structured knowledge base and resource about common entities and event sequences, such as Wikipedia, and ConceptNet. Here, the target NLP applications we aim to improve are entity and event coreference resolution, which involve determining which textual elements point to the same real-world entities and events, respectively. A second objective domain is in reasoning about emotions, social interactions, and the inner mental states of others, which is known in cognitive science as possessing a theory of mind. Being able to reason about the likely reactions of others is a crucial step in generating contextually appropriate responses in conversational agents, and in ensuring coherence in longer generated passages. A third, challenging domain is that of fictional scenarios and situations, in which the typical expectation of the world which might be learned by reading large amounts of web text might be subverted in interesting and predictable ways. A system that is truly capable of common sense reasoning should be able to imagine the potential consequences even in such unlikely scenarios; e.g., that humans living on Mars would need a spacesuit to breathe or to terraform it so that there is oxygen in the atmosphere.

近年来，理解和生成自然语言的人工智能系统取得了长足的进步，导致了成功的商业部署。然而，有充分的证据表明，当前的自然语言处理（NLP）系统以系统和非人类的方式失败，尽管平均性能很高。这种失败在很大程度上是由于缺乏对世界上典型的因果关系的知识，以及对它们的推理能力有限;换句话说，他们缺乏关于世界的常识知识。我的研究项目旨在开发能够具备这种常识推理技能的NLP系统，这样它们就可以与人类用户进行交流和协作，而无需对每个新任务或主题进行明确的培训。所提出的研究的关键思想是建立一个明确的形式化结构，一方面将文本元素相互连接，另一方面也将系统已经建立的知识相关的背景信息连接起来。在技术方面，我们将提出新的神经网络架构和模型，这是有前途的，因为它们提供了简单而自然的方法，通过多任务学习和迁移学习设置中的网络参数共享在多个源和任务之间传输知识。该项目的目标将由一个特定的推理任务所需的常识和世界知识结构的类型来组织。第一种是利用广泛的结构化知识库和资源的共同实体和事件序列，如维基百科，和ConceptNet。在这里，我们要改进的目标NLP应用程序是实体和事件共指解析，这涉及到确定哪些文本元素分别指向相同的真实世界实体和事件。 第二个客观领域是关于情感、社会互动和他人内在心理状态的推理，这在认知科学中被称为拥有心理理论。能够推理他人可能的反应是在会话主体中生成上下文适当的反应以及确保较长生成段落的连贯性的关键一步。第三个具有挑战性的领域是虚构的场景和情况，在这些场景和情况中，通过阅读大量网络文本可以了解到的对世界的典型期望可能会以有趣和可预测的方式被颠覆。一个真正能够进行常识推理的系统应该能够想象出潜在的后果，即使是在这种不太可能的情况下;例如，生活在火星上的人类需要宇航服来呼吸或改造火星，使大气中有氧气。