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.

description. Again, we ignore th<strong>is</strong> technicality in our terminology. <strong>The</strong> term <strong>is</strong> motivated by the fact that´,,1,661,00derives@‰yy,,<strong>is</strong>inCHAPTER 6. EFFICIENT PARSING WITH STOCHASTIC CONTEXT-FREE GRAMMARS 150Definition 6.9 Given a string 1 , .1the probabilities <strong>of</strong> all paths that.6{G , the outer probability´@_,6)¸.—£ ˜M0 <strong>of</strong> an Earley state <strong>is</strong> the sum <strong>of</strong>¡start with the initial state,¡generate the prefix 1 0 £££1 ,16?¡pass through6)¸ £j@$˜ , for some@,generate the suffix 1 @ £££1ED ?¡1 starting with state6)¸@>£ ˜ ,end in the final state.¡Outer probabilities complement inner probabilities in that they refer to prec<strong>is</strong>ely to those parts<strong>of</strong> complete paths generating not covered by the corresponding inner probabilityŠ@7,6)¸ .—£ ˜M0 . A1potentially confusing aspect <strong>of</strong> th<strong>is</strong> definition <strong>is</strong> that the choice <strong>of</strong> the production ) ¸outer probability associated with state6)¸a.<strong>of</strong> the RHS: all states sharing the ) same",˜ and will identical´@have ..>˜ <strong>is</strong> not part <strong>of</strong> the.M£ ˜ . In fact, the definition makes no reference to the first part.—£ ˜M0 <strong>is</strong> the probability that an Earley parser operating as a string generatoryields the prefix 1 0å…å…å6?¸Intuitively,´@_,6)1 and theD ?å…å…å suffix 1, while passing state6)¸through<strong>is</strong> independent <strong>of</strong>.). As was the case for 1 forward <strong>is</strong> actually an expectation <strong>of</strong> the number@ probabilities,´.M£ ˜ at position H (which<strong>of</strong> such states in set H , as unit production cycles can lead to paths that have more than one state fitting th<strong>is</strong>reduces to the ‘outer probability’ <strong>of</strong> )as defined in Baker (1979) if the dot <strong>is</strong> in final position.6.5.2.1 Computing expected production countsBefore going into the details <strong>of</strong> computing outer probabilities we briefly describe their use inobtaining the expected rule counts needed for the E-step in grammar estimation.Let ¸ .%. )string . Alternatively, 1complete Earley path generating . Let production 1 ¸.along a path‰.)denote the expected number <strong>of</strong> uses <strong>of</strong> production ) ¸ 0) ¸ .w. 0 <strong>is</strong> the expected number <strong>of</strong> times that ) ¸ . .) ¸the derivation <strong>of</strong>used for prediction in a.%.‰ë0 be the number <strong>of</strong> occurrences <strong>of</strong> predicted states with0_.%.‰ë0) ¸.w.1) ¸y 06derives+-,‰K.9 1C 1 ‰1C10_.%.‰ë0) ¸1+-,9 1 +-,‰ ?91)z@ C1@C?0å…å…å <strong>of</strong> all paths passing through H : @ ) ¸£.. Inner and outer probabilities have been defined such that th<strong>is</strong>1)z@+-,9 0å…å…å@A?C10F£C1C1Summation <strong>is</strong> over all predicted states using ) ¸.. +-,9 :^0 <strong>is</strong> the sum <strong>of</strong> the probabilities”' å/+-,9C1C1

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

Saved successfully!

Ooh no, something went wrong!