12.07.2015 Views

The dissertation of Andreas Stolcke is approved: University of ...

The dissertation of Andreas Stolcke is approved: University of ...

The dissertation of Andreas Stolcke is approved: University of ...

SHOW MORE
SHOW LESS
  • No tags were found...

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

CHAPTER 5. PROBABILISTIC ATTRIBUTE GRAMMARS 118<strong>The</strong> result <strong>is</strong> that probabilities have to be split (the exact probabilities depend on the sample stat<strong>is</strong>tics), therebylowering the grammar likelihood. Th<strong>is</strong> alternative syntactic structure <strong>is</strong> therefore less preferable than the first,but only due to the attached semantic features.5.5 Limitations and Extensions<strong>The</strong> PAG framework as introduced in th<strong>is</strong> chapter has a number <strong>of</strong> more or less obvious shortcomings,mainly as a result <strong>of</strong> our desire to keep the probabil<strong>is</strong>tic model simple enough so that various techniquesfamiliar from earlier models could be used (combinations <strong>of</strong> multinomials with associated priors, Viterbiapproximations, simple merging operators, etc.) Below we mention the most important limitations andpossible extensions to remedy them.5.5.1 More expressive feature constraintsDerivation probabilities for PAGs were defined with carefully chosen conditional independencestipulations in order to make them computationally tractable.<strong>The</strong> marginal probability <strong>of</strong> the context-free aspect <strong>of</strong> a derivation, –EÃù6 W ±?r+9,–Ã NP VPNP.number = VP.numberNP --> Det NNP.number = N.numberVP --> V NOVP.number = V.number...where number <strong>is</strong> assigned in the lexical productions for both N and V. Although the feature equation inthe first production <strong>is</strong> highly intuitive, it would effectively require stating a marginal probability for the jointevent <strong>of</strong> value assignments, as opposed to conditional probabilitiy <strong>of</strong> one value given the other. However, the6 <strong>The</strong>refore the strict bottom-up feature format could be relaxed, e.g., by using the notion <strong>of</strong> L-attributed feature specifications (Ahoet al. 1986).

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!