Improving morphology induction by learning spelling rules, ACL 2009

Presented by: Loganathan Ramasamy
Report by: Eduard Bejček

• The paper describes morphology induction using Bayesian approach
  • It is based on the Minimum Description Length (MDL) principle
• Baseline: Goldwater et al., 2006: Interpolating between types and tokens by estimation power-law generators
  • only stem & suffix
  • Dirichlet priors over the multinominal distributions for word class, stem and suffix

• Improvements to the baseline system
  • introduces spelling rules (context/change, e.g.: “ut_i”/ε→t (in “shut.ing”) or “ke_i”/e→ε (in “take.ing”)), which are simultaneously learned with morphological analysis
  • Dirichlet priors first set by hand: to prefer empty rules to deletion/insertion

• Experiments are done only for English and only for verbs – that's constrained too much
• Results are not convincing enough – F-score and underlying form accuracy outperform baseline only for stems (not for suffixes) and precision doesn't outperform baseline at all

• Loganathan has the code (although he couldn't compile it, yet)
• Spelling rules are also simultaneously learned along with morphological analysis
• It's unsupervised and clever: using just a couple of (hyper)parameters (some of them are learned automatically), they can describe a wide range of morphological rules - within one elegant framework.