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.

181Chapter 8Future directionsIn th<strong>is</strong> final chapter we suggest a number <strong>of</strong> potential continuations <strong>of</strong> the work described in th<strong>is</strong><strong>d<strong>is</strong>sertation</strong>. Probabil<strong>is</strong>tic language modeling has certainly proven to be a theoretically and practically usefulapproach, including the model merging paradigm for structural learning studied here. It has given r<strong>is</strong>e toconcrete applications (in the case <strong>of</strong> HMMs) and gives new perspectives on some lingu<strong>is</strong>tic <strong>is</strong>sues, as we havetried to show mainly in our d<strong>is</strong>cussion <strong>of</strong> the SCFG and PAG modeling frameworks.However, and not surpr<strong>is</strong>ingly, the work to date ra<strong>is</strong>es more questions than it provides definitiveanswers. Apart from model-specific problems d<strong>is</strong>cussed in the preceding chapters, a number <strong>of</strong> morefundamental <strong>is</strong>sues can be identified. Invariably, these lead to worthwhile avenues for future work.8.1 Formal characterization <strong>of</strong> learning dynamicsIt would be extremely useful to be able to predict the sample complexity <strong>of</strong> a given target grammarand merging-based learning algorithm, i.e., the number <strong>of</strong> samples required to reliably learn the targetgrammar. As a first step, th<strong>is</strong> would include a prediction <strong>of</strong> how many samples are needed to make the targetgrammar (or one equivalent to it) the one with globally optimal posterior probability. A further, more difficultproblem <strong>is</strong> to predict when a limited search procedure <strong>of</strong> specified type would be able to actually find thatglobal optimum.Within th<strong>is</strong> general problem <strong>of</strong> sample complexity it <strong>is</strong> also interesting to draw a d<strong>is</strong>tinctionbetween sample structure and sample d<strong>is</strong>tribution. In other words, how sensitive <strong>is</strong> a learner to a sample that<strong>is</strong> representative regarding the possible strings <strong>of</strong> the language, but exhibits different frequency stat<strong>is</strong>tics? 11 Th<strong>is</strong> question was pointed out by Robert Wilensky.

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

Saved successfully!

Ooh no, something went wrong!