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 6. EFFICIENT PARSING WITH STOCHASTIC CONTEXT-FREE GRAMMARS 153Its chart <strong>is</strong> d<strong>is</strong>joint from the one used by the parent instance. It cannot use states from the parent¡chart, except those explicitly passed to it (see below). Conversely, when fin<strong>is</strong>hed, the parent has accessonly to those states returned explicitly by the child instance. 16 (<strong>The</strong> first restriction prevents parsing <strong>of</strong>constituents that cross the left phrase boundary, while the second restriction prevents a violation <strong>of</strong> theright phrase boundary.)<strong>The</strong> chart <strong>of</strong> the child <strong>is</strong> initialized with all incomplete states from the parent’s state set at the start¡position <strong>of</strong> the substring.<strong>The</strong> child returns to the parent all (and only) the complete states from its last state set. <strong>The</strong> parent adds¡the returned states to the state set at the position immediately following the end <strong>of</strong> the substring, usingit as the input for its own completion procedure.Thus the recursive parser invocation and the following completion step replaces the usual predictionscanning-completioncycle for the entire bracketed substring. After the child returns, the¡parentcontinues processing regular input symbols, or other bracketed substrings.Needless to say, the child parser instance may itself call on recursive instances to deal with nested¡bracketings.Th<strong>is</strong> recursion scheme <strong>is</strong> efficient in that it never explicitly rejects a parse that would be incons<strong>is</strong>tentwith the bracketing. Instead it only considers those parses that are cons<strong>is</strong>tent with the bracketing, whilecontinuing to make use <strong>of</strong> top-down information like a standard Earley parser.Processing bracketed strings requires no modification to the computation <strong>of</strong> probabilities. Probabilitiesare passed between parent and child as part <strong>of</strong> states, and processed as before. <strong>The</strong> recursive controlstructure simply constrains the set <strong>of</strong> Earley paths considered by the parser, thereby affecting the probabilitiesindirectly. For example, ambiguous strings may end up with lower inner probabilities because somederivations are incons<strong>is</strong>tent with the bracketing.Only the forward pass <strong>is</strong> directly affected by the bracketing. Both the Viterbi procedure (Section6.5.1) and the reverse completion pass (Section 6.5.2) only examine the states already in the chart. <strong>The</strong>yare therefore automatically constrained by the bracketing.Complexity To assess the complexity benefit <strong>of</strong> bracketing we can extend the analys<strong>is</strong> <strong>of</strong> Section 6.4.9,making use <strong>of</strong> the recursive structure <strong>of</strong> the algorithm.In the standard parsing scheme, the time ð G3complexity <strong>is</strong> 0 for an input <strong>of</strong> length . Hence, in,

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

Saved successfully!

Ooh no, something went wrong!