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...

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

CHAPTER 5. PROBABILISTIC ATTRIBUTE GRAMMARS 110Notice that the features cannot be attached to the lexical productions created, since that would implyknowing the semantic contributions <strong>of</strong> each word. Hence the features are first attached to the sentence as awhole, corresponding to rote learning <strong>of</strong> sentence-meaning associations.5.3.2 Nonterminal merging and chunking<strong>The</strong> nonterminal merging and chunking operators introduced in Chapter 4 can be applied to PAGrules with a few modifications. Merging in the CFG part <strong>of</strong> productions operates as usual, by renamingnonterminals and possibly merging productions. Feature equations are also affected by renaming nonterminals:Whenever two or more feature equations become identical as the result <strong>of</strong> nonterminal merging, theirassociated counts are added as usual.By way <strong>of</strong> example, suppose the we merge nonterminals N1 and N2, starting with productionsNP --> Det N1 (2)NP.f = N1.f1 (2)NP --> Det N2 (2)NP.f = N2.f2 (2)<strong>The</strong> result <strong>of</strong> merge(N1, N2) = N <strong>is</strong>NP --> Det N (4)NP.f = N.f1 (2)NP.f = N.f2 (2)Notice that the feature names f1 and f2 are not automatically merged; that will be accompl<strong>is</strong>hed by a separateoperator.Nonterminal chunking operates as before, but may require the creation <strong>of</strong> new ad hoc featuresattached to the newly created nonterminal in order to preserve any features found on the chunked nonterminals.Th<strong>is</strong> too <strong>is</strong> most easily seen by example:S --> Det N V A NS.tr = N(1).f (2)S.lm = N(2).f (2)S.rel = V.hWhen applying chunk(ANP:N) = NP the f feature must be preserved by creating a corresponding feature onS --> NP V NPS.tr = NP(1).g (2)S.lm = NP(2).g (2)S.rel = V.hNP --> Det NNP.g = N.f (4)

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

Saved successfully!

Ooh no, something went wrong!