Word Sense Disambiguation: Decision Trees on Bigrams

• Pedersen (2001) exploits the "one sense per collocation" property, but also relies on features that occur in the context of an ambiguous word
  • Word sense disambiguation within the SENSEVAL competition, starting with a manually sense-tagged corpus (senses are drawn from WordNet, which tends to be quirky and make overly fine sense distinctions); this is therefore supervised WSD
  • One goal: "establish clear upper bounds on the accuracy of diambiguation" without resorting to "substantial pre-processing" (i.e., parsing and sematnic interpretation)
  • In conjunction with this goal, discover good feature sets that can distinguish word senses
  • Issues to be dealt with:
    • selecting bigrams that are good disambiguators; hence, choosing good statistical measures of significance("power dievergence"--like Chi-square, or Dice coefficient); note that the bigrams may not include the ambiguous word iteself, though of course many useful ones do!
    • designing the decision trees (sequences of decisions based on the bigrams) to choose a sense for a new, untagged occurrence of an ambiguous word; some features count for more than others in this model, unlike Bayesian classifier models; a machine-learning algorithm constructs the decision trees, using the most informative bigrams for distinguishing among senses
  • Three benchmark methods for comparison:
    • always choose the most frequent sense ("majority classifier")
    • Bayesian classifiers, using every word in the tagged corpus as a feature
    • decision stump: the simplest decision tree, with a single criterion (the most informative) for making the choice among senses
  • Results and analysis:
    • Both statistical methods performed well in finding significant bigrams
    • Overall, decision trees outperformed the benchmark methods, but decision stumps, using just the most informative bigram, did fairly well
    • Most of the features identified initially as statistically significant don't need to be used in the decision trees (i.e., adding them wouldn't improve accuracy)
    • Decision trees are more easily interpreted by people than Bayesian classifiers

(2)

(on to syntactic ambiguity resolution)

(return to syllabus)