Language modeling is crucial to all applications that process human language with less than complete knowledge. This includes speech recognition, machine translation, optical character recognition, handwriting recognition, spelling and grammar correction, and others. Formal theories of grammar have so far failed to account adequately for actual natural usage of language. Stochastic versions of formal grammars are still less successful (as measured by cross entropy of their predictions) than simple Markovian models (ngrams) which are estimated from larger amounts of data. With the advent of huge textual corpora, a breakthrough in language modeling will come when we successfully integrate linguistic knowledge with statistical estimation techniques.

This workshop will bring together researchers who are working on various aspects of language modeling (stochastic grammars, clustering, maximum entropy models) with mathematicians with interest in these and related problems (Bayesian methods, clustering, information theory). The first 2 days will consist of an overview of the field and existing techniques, followed by presentations of ongoing research. The overall objective is to encourage interaction and collaborations between mathematicians and practitioners to pursue the next generation of solutions to language modeling problems. The specific goals are:

• Familiarize the mathematicians with the language technology, outline underlying fundamental problems and currently popular/successful solutions.

• Present novel models, ideas or approaches currently being pursued in the language modeling community.

• Present recent advances in mathematics which may be relevant to the language modeling community.

Talk Abstracts

2000-2001 Program: Mathematics in Multimedia