Greg Kochanski

Hilary Term 2004. Tuesdays, 12-1, Centre for Language and Philology Common Room. [Syllabus]

This course explores

• Why one needs quantitative models: a historical view of major goofs in other fields.
• Probability theory for Linguists.
• Converting hypotheses about language into mathematical models.
• Testing models against data.
• Evaluating models for:
  • robustness (is it resistant to a small amount of bad data?),
  • stability (does a small change in the data lead to a small change in the result?), and
  • sensitivity (are some data much more important than others?).
• Choosing and customising machine learning systems so they match known behaviours of the system.

We will use modern statistical techniques like the Bootstrap and Jackknife, Markov Chain Monte Carlo, and Monte-Carlo simulations. We may also touch on Bayesian Belief Networks as a way of thinking about uncertainty in grammars and other discrete models. The goal is to help the researcher decide on a good model and estimate how important a result might be.

The beginning of the course will review probability, conditional probabilities and Bayes' Theorem. It will also look at the historical record to study well-documented examples of pathological science in other fields to look for parallels in linguistics. Topics discussed may include spam filtering, attribution of authorship and style, reconstructing language groups in historical linguistics, and models of stress and duration, with some attention to perceptual and articulatory constraints on speech.

The course will be taught primarily via lectures, with class discussions, and some numerical experiments. A part of the course will involve discussing research problems that the members are working on. Calculations will be either accessible on the Web, or download-able to either Windows or Linux machines. Eight sessions, 1 hour each.

Prerequisites: General understanding of probability, basic statistical concepts, algebra. Some programming experience is valuable, but not essential. Students who have completed Formal Foundations for Linguistics (S. Pulman) will meet the prerequisites.

Target Audience: Linguistics and Philology. It could have application inside Experimental Psychology.