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 3. HIDDEN MARKOV MODELS 60(a)0.26a1.000.210.570.650.530.510.350.270.00Start0.00a0.43b0.23a bEnd(b)0.270.00 0.210.39 0.250.570.000.471.000.520.610.43Start ab a b0.28EndFigure 3.9: Case study II: BW-derived HMM structures that fail on generalization.Baum-Welch studies As in the previous case study, we looked at various model structures found byBaum-Welch estimation. All examples in th<strong>is</strong> section are from training on the 20 random samples.Figure 3.9(a) shows a structure that <strong>is</strong> overly general: it generates , %§E#&#%0have an HMM that partly overgeneralizes, but at the same time exhibits a rather peculiar case <strong>of</strong> overfitting: itw¤07#& # . In (b), weexcludes strings <strong>of</strong> the >#&6#&# form where"<strong>is</strong> even. No such cases happened to be present in the trainingset.<strong>The</strong> accurate model structures <strong>of</strong> 10 states found by the Baum-Welch method again tended to berather convoluted. Figure 3.10 shows as case in point.Merging studies We also repeated the experiment examining the levels <strong>of</strong> generalization by the mergingalgorithm, as the value <strong>of</strong> the global prior weight was increased over three orders <strong>of</strong> magnitude.Figure 3.11 shows the progression <strong>of</strong> models for. 6 0£ 018 0£ 18, and 1£ 0. <strong>The</strong> pattern <strong>is</strong> similar tothat in in the first case study (Figure 3.11). <strong>The</strong> resulting models range from a simple merged representation <strong>of</strong>the samples to a plausible overgeneralization from the training data ( , B#&#%07# ). <strong>The</strong> target model <strong>is</strong> obtainedfor.values between these two extremes.

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

Saved successfully!

Ooh no, something went wrong!