Machine Learning of Language from Distributional Evidence
10/19/2007 10:45 AM Wong AuditoriumChris Manning, Associate Professor of Computer Science and Linguistics, Stanford UniversityDescription: Christopher Manning thinks linguistics went astray in the 20th century when it searched "for homogeneity in language, under the misguided assumption that only homogeneous systems can be structured." In the face of human creativity with language, rigid categories of linguistic use just don't help explain how people actually talk and what they choose to say. For every hard and fast rule linguists find, other linguists can determine an exception. Categorical constraints rise, then come crashing down.
Manning argues for acceptance of variable systems of language, and for searching for structure in these systems using probabilistic methods. Manning applies quantitative techniques to sentence structure, digging for the frequency, probability and likelihood that people will use specific turns of phrase in certain real"world contexts. Looking at distributions in the ways people express ideas in a language "can give a much richer description of how language is used." Indeed, Manning finds that certain typical constraints on sentence structure in one language "show up as softer constraints and preferences in other languages."
Manning looks at raw data, like sentences from the Wall Street Journal, and gleans such information as typical word associations that begin to "tell us about the dependencies of verbs and arguments." He looks for dependencies between words, the distance between them, and at a sentence's flow from left to right. Classes of words emerge, and clusters, yielding distributionally learned categories. Certain classes of syntax naturally fall together. Manning builds nested phrase structure trees, and branching structures, and derives simple probabilistic models that help explain "gradual learning and robustness in acquisition, non"homogeneous grammars of individuals, and gradual language change over time." Manning says computational linguistics is also proving useful in such applied fields as information retrieval, machine translation, and text mining.
About the Speaker(s): Christopher Manning's research concentrates on probabilistic models of language and statistical natural language processing, information extraction, text understanding and text mining, and other topics in computational linguistics and machine learning. Together with Dan Klein, he received the ACL 2003 best paper award.
He received a B.A. in mathematics, computer science and linguistics from the Australian National University in 1989. He earned a Ph.D. from Stanford in Linguistics in 1995. He previously served as an Assistant Professor at Carnegie Mellon University in the Computational Linguistics Program and as a lecturer in the University of Sydney Department of Linguistics.
Manning's "bestseller" is Foundations of Statistical Natural Language Processing, Manning and Schôtze,(MIT Press, 1999). Host(s): School of Engineering, Laboratory for Information and Decision Systems
MIT World -- special events and lectures