Symposium on Mathematical Statistics in Natural Language, San Francisco, CA, February 2001

2001年美国科学促进会（AAAS）会议将举行一次专题讨论会，主题是自然语言分析和处理的统计方面。本次研讨会起源于语言数学协会（MOL），该协会是计算语言学协会的一个特殊兴趣小组，每两年举行一次会议，已有10多年的历史。本次研讨会的更大目标是提高对自然语言结构和处理的数学研究的普遍科学意识。语言的形式和计算研究借鉴了广泛的数学领域（其中包括形式语言理论；复杂性；经典和非经典逻辑；代数和范畴论；博弈论），并在其中一些领域激发了创新和扩展。在美国科学促进会的研讨会上，很难以一种既能体现其广度又能体现其深度的方式来讨论这项工作的全部范围。因此，研讨会将集中在一个单一的、更明确的主题上——自然语言分析的统计方面——以一种将语言工程、语言结构和处理方面的技术进步联系起来的方式。语言能力取决于生物禀赋和经验经验的相互作用，以这样一种方式，面对说话者之间的广泛差异，在不断变化的环境中，随着时间的推移，对物理对象的稳定解释是可能的。存储、搜索和处理计算方法的进步，使得以一种适合定量方法的方式，对禀赋和经验的动态相互作用进行定性建模成为可能。本次研讨会的重点是在统计学和信息论的基础上解决这一系列问题的方法。演讲将以基于统计的语音处理工作的调查开始（这是成功和可用的计算语音到文本处理的基础）。然后将研究这些思想在语法分析和词法歧义解决中的应用。演讲将强调不同统计方法对自然语言分析和处理中不同问题的适当性的重要性，并展示当前模型之间的区别如何与诸如监督学习和无监督学习之间的区别等经典问题联系起来。

A symposium will be held at the 2001 meeting of the American Association for the Advancement of Science (AAAS) on the topic of statistical aspects of natural language analysis and processing.This symposium originates with the Association for the Mathematics of Language (MOL), a special interest group of the Association for Computational Linguistics which has held bi-annual meetings for over 10 years. The bigger goal of the proposed symposium is to heighten general scientific awareness of mathematical investigations of natural language structure and processing. Formal and computational studies of language draw on a broad range of mathematical fields (among them formal language theory; complexity; classical and non-classical logic; algebra and category theory; game theory) and have inspired innovations and extensions in a number of them. It would be difficult in a AAAS Symposium to address the full range of this work in a way that would do justice to both its breadth and its depth. The symposium will therefore concentrate on a single, more sharply delineated topic -- statistical aspects of natural language analysis -- in a way that connects technological advances in language engineering, linguistic structure and processing.Capacity for language depends on the interaction of biological endowment and empirical experience, in such a way that the stable interpretation of physical objects is possible in the face of broad variation across speakers, in changing contexts, and over time. Advances in computational methods of storage, search, and processing make it possible to model the dynamic interaction of endowment and experience qualitatively, in a way amenable to quantitative methods. This symposium focuses on approaches to this range of problems based on statistics and information theory. The presentations will begin with a survey of statistically-based work on speech processing (which underlies successful and available computational speech-to-text processing). It will then examine applications of these ideas to syntactic parsing and lexical ambiguity resolution. The presentations will be designed to stress the importance of the appropriateness of different statistical methodsto different problems in natural language analysis and processing and to show how distinctions among current models connect with such classical problems as the distinction between supervised and unsupervised learning.