Part-of-Speech Tagging

Important Notions

• The n-class model
  1. Lexical probabilities: P(w|t)
  2. Contextual probabilities: P(t_i|t_i-n+1...t_i-1)
  3. Hidden Markov Models (HMM)
• Parameter estimation
  1. Tagged training data (MLE)
  2. Untagged training data (EM)
  3. Sparse data and smoothing
• Algorithms for HMMs
  1. String probability: Forward, Backward, Forward-Backward
  2. Optimal state sequence: Viterbi
  3. Unsupervised parameter estimation: Baum-Welch

Slides for lecture 6

Suggested Reading

• Krenn, B. & Samuelsson, C. (1997) The Linguist's Guide to Statistics. Sections 2.1 and 5.1.
• Manning, C. D. & Schütze, H. (1999) Foundations of Statistical Natural Language Processing. MIT Press. Chapter 9, sections 10.1-10.2.

Project

• Part-of-speech tagging (Swedish data)