🧑‍🔬 Andreas Stolcke

Andreas Stolcke works on probabilistic models of language and speech. The papers here cover two related problems: how to compute probabilities efficiently over the structures of a grammar, and how to recognise what a speaker is doing with each utterance in spontaneous conversation. Both rely on statistical models, stochastic grammars in one case and hidden Markov models in the other, and both concern how sequential structure in language can be captured and put to use in recognition. The topics touched include Speech act theory and Dialogue systems.

The most cited work, Dialogue Act Modeling for Automatic Tagging and Recognition of Conversational Speech (2000), assigns each utterance in Switchboard telephone conversations a dialogue act such as statement, question, backchannel or agreement. It treats the acts as hidden states, draws evidence from the words, whether transcribed or recognised, and from prosody, and uses an n-gram over acts to model their order. This lets the paper measure how much knowing the sequence of acts improves tagging. The earlier paper, An Efficient Probabilistic Context-Free Parsing Algorithm that Computes Prefix Probabilities (1995), deals with stochastic context-free grammars. It addresses the standard tasks for these grammars, including the probability of a string or of a prefix, the most likely parse, and the estimation of rule probabilities. Together the two papers move from grammar-level parsing toward conversation-level labelling.

The current institution is the University of Illinois Urbana-Champaign. Earlier bylines include SRI International, the International Computer Science Institute and the University of California, Berkeley, followed by industrial research at Microsoft and Amazon. Co-authors on the papers here include Dan Jurafsky, Elizabeth Shriberg and Klaus Ries.

Andreas Stolcke has 2 papers in our library, in 1995–2000, most of them on Speech act theory, Dialogue systems; often with Dan Jurafsky, Elizabeth Shriberg, Klaus Ries.