🧑‍🔬 John D. Lafferty

John D. Lafferty works on probabilistic models of language and text. Their papers here cover two settings: speech recognition, where a grammar has to predict the words still to come, and large document collections whose themes shift over the years. The approach is mathematical throughout, from exact derivations for probabilistic grammars to latent-variable models that need approximate posterior inference.

The most cited of these papers is Dynamic Topic Models (2006), written with David M. Blei. It adapts topic modelling to archives such as journals, news or email whose contents change over time. A corpus is cut into time slices, and each topic is allowed to evolve smoothly from one slice to the next. Dirichlet priors link poorly across time, so the model places Gaussian dynamics on the topics' natural parameters instead. The earlier paper, Computation of the Probability of Initial Substring Generation by Stochastic Context-Free Grammars (1991), written with Jelinek, is purely theoretical. It derives the Left-to-Right Inside algorithm, which gives, for a grammar in Chomsky Normal Form, the chance that a generated sentence starts with a particular string of words. That prefix quantity is what a recognizer requires when a grammar acts as its language model, and the paper falls under Syntax & grammar formalisms.

Between the two papers here, the direction runs from grammar-based language modelling for speech toward statistical models of how text collections evolve. Lafferty is currently at Yale University. Earlier affiliations include Carnegie Mellon University, where much of the work was done, as well as University of Chicago and, from 1990 to 1994, IBM Research - Thomas J. Watson Research Center.

John D. Lafferty has 2 papers in our library, in 1991–2006, most of them on Syntax & grammar formalisms; often with David M. Blei.