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 112After attributing ‘square’ to the new feature SQUARE.f the productions becomeS --> A CIRCLE IS BELOW A SQUARES.tr = ‘circle’S.lm = SQUARE.fS.rel = ‘below’S --> A SQUARE TOUCHES A TRIANGLES.tr = SQUARE.fS.lm = ‘triangle’S.rel = ‘below’SQUARE --> squareSQUARE.f = ‘square’<strong>The</strong> choice <strong>of</strong> which feature value to attribute to which nonterminal occurrence on the RHS <strong>is</strong>nondetermin<strong>is</strong>tic, but heur<strong>is</strong>tics can be used in practice (see below).5.3.4 Efficient search for feature operationsCombining operatorsIt <strong>is</strong> convenient to conceptualize the four operators described above as separateinduction steps. However, in practice <strong>is</strong> more efficient to combine several <strong>of</strong> these operators due to theirtypical interactions. In the experiments reported below th<strong>is</strong> was done so the ex<strong>is</strong>ting two-level best-first searchstrategy could be used, instead <strong>of</strong> a more extensive search method.<strong>The</strong> feature attribution and merging operators, in particular, <strong>of</strong>ten produce an improved posteriorprobability score in conjunction with certain nonterminal merging operations. Consider the rulesNP --> Det N1 [0.5]NP.f = N1.f1--> Det N2 [0.5]NP.f = ‘circle’N1 --> squareN1.f1 = ‘square’N2 --> circleSimply applying the syntactic merging operation merge(N1, N2) = N would result inNP --> Det N [1.0]NP.f = N.f [0.5]NP.f = ‘circle’ [0.5]N --> square [0.5]N1.f = ‘square’--> circle [0.5]which cuts the probability <strong>of</strong> all samples using one <strong>of</strong> these productions in half. Alternatively we might precedethe merging operation by ‘appropriate’ feature operations, namely fattrib(N2, circle) = f2, followed byfmerge(f1, f2) = f. Th<strong>is</strong> gives the grammarNP --> Det N [1.0]NP.f = N.f [1.0]

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

Saved successfully!

Ooh no, something went wrong!