Comprehensibility

Determining comprehensibility for something that classifies 95 classes is somewhat difficult. For the above cases, very large trees are generated - with an average with J48 of 637 nodes. PART generates rules; and for the above case with all the bells and whistles turned on it still averages 233 rules. However, this is only 2.45 rules per class; which is reasonable.

A subjective correlation of the rules produced by PART and the Auslan dictionary proved largely unsuccessful. This was thought to be due to the large number of classes. Hence, to simplify the learning task and produce better definitions, a simple approach was taken: turn the classification problem into a binary one and see if the definitions come out clearer.

And indeed these do. Figure 6.27 shows a definition of the sign thank. The gloss for the sign thank is shown in Figure 1.1. The definition looks rather strange; however, it is the result of the way the learner works. The first two rules are ``quick reject'' rules to eliminate most possibilities (for example, the first rule eliminates approximately 93 per cent of all training instances). The first rule says that if there is no movement toward the body early in the sign (midtime=0.19) and then no movement away from the body late in the sign and there is no local minimum in the movement it can't be the sign thank. This is encouraging, since of course, the sign thank is exactly the opposite: a movement towards the body then away from the body.

The second rule says that if the middle finger isn't closing towards the end of the sign, and that the z value doesn't change towards the end of the sign, then it is not thank. The y value check is probably an oddity of the data, but it doesn't matter since the first two parts of the rule will probably prevent checking of the third part. Again, since the sign is open-handed, this means that the middle finger stays stable.

The third rule says that if the the first two rules haven't eliminated the possibility and the hand moves towards the centre of the body, then it is the sign thank.

The fourth and fifth rules can also be explained, but it is omitted here for brevity since they cover less than 1 per cent of the training set.

**Figure 6.27:** A decision list produced by *TClass* on the Nintendo sign data for the sign `thank`.
$\begin{figure}\footnotesize\begin{boxedverbatim}PART decision list -----------... ...: thank (3.0)Number of Rules : 5\end{boxedverbatim}\normalsize\par\end{figure}$

**Figure 6.28:** Events referred to in Figure 6.27 with bounds.
$\begin{figure}\footnotesize\begin{boxedverbatim}Event index ----------- *1: in... ....10,0.00] duration=0.11 r=[0.03,0.14]\end{boxedverbatim}\normalsize\end{figure}$