Can Machines Learn Common-Sense Reasoning?

常识推理（CSR）最近在人工智能社区中重新引起了人们的兴趣，这是由于某些技术的出现（例如，对话系统、故事理解软件和推荐工具），在缺乏常识的情况下，这些工具似乎非常不智能。拟议研究的目标是使机器学习（ML）模型能够实现常识推理的学习，以帮助它们推广到真实的世界中经常遇到的复杂场景。具体来说，该研究将首先尝试量化各种自然语言理解应用中常识性问题的普遍性并对其复杂性进行分类。接下来，该研究将继续研究开发评估协议的策略，这些协议将精确识别深度学习模型如何利用训练实例来解决测试时的问题。当与研究计划的第一步相结合时，这些评估协议将能够为我们提供对模型实例难度的更细致入微的理解。此外，该研究将探讨在这些协议中纳入对当今语言生成模型中存在的性别，种族和其他形式的偏见的纠正措施。受软件工程原理的启发，还将探索和推广使用通用行为测量，例如不变性或最小功能测试，作为当前和未来NLP系统的重要新基准测量。下一步的研究将致力于开发新的深度学习模型，基于最近提出的Transformer架构，通过修改和增强各种模型组件，包括注意力机制，解码器/编码器层和训练方法，这些模型可能具有CSR。该研究将解决的一些基本和技术问题包括：i）我们如何开发一个统一的通用系统，能够同时处理一些困难的CSR任务，以及具有高覆盖率和可解释性的更一般的下游对应任务; ii）我们如何确定在各种自然语言处理（NLP）基准上的性能改进是否代表了向常识支持系统的真正进展; iii）我们如何设计更具挑战性、更现实和更大规模的基准; iv）我们如何使用常识推理来帮助弥合机器人和人类在真实的世界中的性能差距;以及v）常识推理是否可以用作开发端到端去偏见技术的基础，该技术可以对抗当前预训练语言模型和语料库中普遍存在的各种形式的偏见（性别/种族/文化）？

Common-sense reasoning (CSR) has recently garnered a significant renewal of interest in the Artificial Intelligence community, resulting from the advent of certain technologies (e.g., dialogue systems, story understanding software, and recommendation tools) which can seem strikingly unintelligent in the absence of common sense. The goal of the proposed research is to make the learning of common-sense reasoning achievable for Machine Learning (ML) models, in order to help them generalize to complex scenarios that are encountered frequently in  the real world. Specifically, the research will first attempt to quantify the prevalence and classify the complexities of common-sense demanding problems in a variety of natural language understanding applications. The research will next proceed with studying strategies for the development of evaluation protocols that precisely identify how deep learning models make use of train instances to resolve problems at test time. When coupled with the first step of the research plan, these evaluation protocols will be able to provide us with a more nuanced understanding of the difficulty of instances for models. In addition, the research will explore incorporating in these protocols corrective measures for gender, racial and other forms of bias present in today's language generation models. Inspired by principles in software engineering, the use of all-purpose behavioural measures, such as invariance or minimum functionality tests, will also be explored and promoted as an important new benchmark measure for the current and future NLP systems. The next step of research will be dedicated to the development of novel deep learning models, based on the recently proposed Transformer architecture, which are potentially CSR-endowed, through modifying and augmenting various model components, including the attention mechanism, the decoder/encoder layers, and the training method. Some of the fundamental and technological questions that the research will address include: i) How can we develop a unified, generalized system that is capable of simultaneously tackling a number of difficult CSR tasks, as well as their more general, downstream counterparts with high coverage and interpretability; ii) How can we determine whether an improved performance on various Natural Language Processing (NLP) benchmarks represents a genuine progress towards common-sense-enabled systems; iii) How can we devise benchmarks that are more challenging, realistic and large-scale; iv) How can we use common-sense reasoning to help bridge the performance gap between robots and humans in the real world; and v) Can common-sense reasoning be used as a basis for developing end-to-end debiasing techniques that can combat various forms of biases prevalent in current pre-trained language models and corpora (gender/race/cultural)?