CAREER: Learning Coherent Concepts: Theory and Applications to Natural Language

这是为期4年的持续奖的第一年资助。本研究考察了多个学习者共存的学习情景，这些学习者可能在相同的输入上学习不同的功能，但他们的结果存在相互兼容性的限制。PI将为这些情况开发一种学习理论，并将研究在自然语言推理中利用它们的算法方法。理论研究将集中于开发连贯条件的语义学，并从学习理论的角度对其进行研究，目的是了解在这些情况下学习如何变得更容易和更健壮。算法研究将集中于开发利用一致性的方法，并将有一个重要的实验组成部分，将浅层句法分析问题作为研究一致性分类器和依赖于几个分类器的结果的推理的试验床。这项研究将对学习中的理论研究和我们在自然语言中执行更高级别推理的能力产生重大影响。这将有助于解决预测的学习难度和认知系统学习的明显轻松之间的反差。此外，它将提供对如何利用一致性的理解，以便为这些情况制定更好的学习和推理方法，并将导致各种浅层分析任务的综合学习方法，并使用SNOW学习体系结构以大规模方式实施和演示。将对相互作用的分类器的理解以及执行依赖于几个分类器的推理的方法纳入SNOW将直接适用于该领域的各种其他任务。Http://L2R.cs.uiuc.edu/~danr/Grants/eareerOO.html

This is the first year of funding of a 4-year continuing award. This research investigates learning scenarios in which multiple learners co-exist and may learn different functions on the same input, but there are mutual compatibility constraints on their outcomes. The PI will develop a learning theory for these situations, and will study algorithmic ways to exploit them in natural language inference. The theoretical study will concentrate, on developing a semantics for the coherency conditions and study it from a learning theory point of view with a goal of understanding in what ways learning becomes easier and more robust in these situations. The algorithmic study will concentrate on developing ways to exploit coherency and will have a significant experimental component, using the problem of shallow parsing as a testbed for investigating chaining of coherent classifiers and inferences that rely on the outcomes of several classifiers This research will have a significant impact on theoretical research in learning and on our ability to perform higher level inference in natural language. It will help to resolve the contrast between the predicted hardness of learning and the apparent ease at which cognitive systems learn. Moreover, it will provide an understanding of how to exploit coherency in order to develop better learning and inference methods for these situations, and will result in an integrated learning approach to a variety of shallow parsing tasks, implemented and demonstrated in a large scale manner using the SNoW learning architecture. Incorporating the understanding of interacting classifiers, as well as methods to perform inferences that rely on several classifiers, into SNoW will be directly applicable to a variety of other tasks in this domain. http://L2R.cs.uiuc.edu/~danr/Grants/eareerOO.html