Proir Probabilities

• Use Dirichlet priors

  • Conjugate to multinomial distribution (i.e., the posterior = prior × likelihood is again multinomial)
  • Amounts to including fake counts in re-estimation routines
    • +=2 for word-to-word rewrites
    • +=1 for stem-identity rewrites
    • +=1 for pure identity rewrites

• Also use this to initialize parameters