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.

<strong>The</strong> forward and inner probabilities <strong>of</strong> the states thus created are those <strong>of</strong> the first state )First, the unit-production relationhas to be extended to allow for unit-production chains due to null productions. A rule ) ¸\¸ £=1Ó‘Œ\XE@ Ó‘Œ 6­¸CHAPTER 6. EFFICIENT PARSING WITH STOCHASTIC CONTEXT-FREE GRAMMARS 1426.4.7.2 Prediction with null productionsPrediction <strong>is</strong> mediated by the left-corner relation. For each )generate states for all= that are reachable from ) by way <strong>of</strong> the )occurring to the right <strong>of</strong> a dot, weTh<strong>is</strong> reachability criterionhas to be extended in the presence <strong>of</strong> null productions. Specifically, if ) has a production)¸=1 £££+=@A?then=@<strong>is</strong> a left corner <strong>of</strong> ) iff=1£££=@A?<strong>of</strong> such a production to the left-corner probability +-, )1 all have a non-zero probability<strong>of</strong> expanding toA.<strong>The</strong> contribution£=@0 <strong>is</strong>¸relation.11=@.@A?+-,1=@.B0) ¸=1 £££+=@A?6X<strong>The</strong> old prediction procedure can now be modified in two steps. First, replace the old relation+by the one that takes into account null productions, as sketched above. From the resulting compute thereflexive transitive closure + , and use it to generate predictions as before.©Second, when predicting a left corner= with a production=¸£££˜=@A?=1dot positions up to the first RHS nonterminal that cannot expand toA, say ) ¸from £££˜=@A?£=1) ¸=1 £££+=@A?1=@., add states for all1=@.through1£=@.. We will call th<strong>is</strong> procedure ‘spontaneous dot shifting.’ It accounts prec<strong>is</strong>ely for thosederivations that expand the RHS prefix=1 £££+=@A?1 without consuming any <strong>of</strong> the input symbols.£££=@C?£=1d6X ¬1Ó‘Œ,1=@., multiplied by factors that account for the impliedA-expansions. Th<strong>is</strong> factor <strong>is</strong> just the productwhere <strong>is</strong> the dot position.6.4.7.3 Completion with null productionsModification <strong>of</strong> the completion step follows a similar pattern.if all other nonter-£££+=@A?1=@=@#=1 £££+=d 1 can effectively act like a unit production that ) and=@linksminals on the RHS can expand toA. Its contribution to the unit production relation ) ¸ =@ +9,be0 will thenFrom the resulting rev<strong>is</strong>ed + matrix we compute the closure ©§ as usual.<strong>The</strong> second modification <strong>is</strong> another instance <strong>of</strong> spontaneous dot shifting. When completing a state, additional states have to be added, obtained by moving the˜) ¸ +-,£££=@A?1=@=@=1#1 £££+=df0dot further over any nonterminals in ˜ that have non-zeroA-expansion probability. As in prediction, forward) ¸.M£=§˜ and moving the dot to get ) ¸ .V=i£and inner probabilities are multiplied by the correspondingA-expansion probabilities.6.4.7.4 Eliminating null productionsGiven these added complications one might consider simply eliminating allA-productions in apreprocessing step. Th<strong>is</strong> <strong>is</strong> mostly straightforward and analogous to the corresponding procedure for non-

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

Saved successfully!

Ooh no, something went wrong!