I tend to agree with apl on this point (and I also agree with the person who suggested changing your indent/line-break style). Apart from that, I was struck by the variety of ways in which you phrase similar conditions, and by the seemingly arbitrary ordering of conditions.

If your "split_token()" sub is returning a 3-element list whose members vary significantly among all of: /"undef", "defined but empty string", "non-empty string but evaluates to 0", "non-empty, non-zero value"/, I think you should be making that more clear. When I see conditions like:

    if ( $atail and $ntail ne '' ) {
    ...
    elsif ( !$atail and $ntail ne '' ) {
    ...
    elsif ( not ($atail eq '' and $ntail eq '' )) {
[download]

I can't help thinking that it should probably be:

    if ( $atail and $ntail ) {
    ...
    elsif ( $atail ) {
    ...
    elsif ( $ntail ) {
    ...
[download]

If it needs to be different from that, you should be really clear about why it needs to be different.

I was puzzled by the last elsif condition:

    elsif($checkPrefix and 
          $unprefBase and    # Set in an earlier test, above
          $savgrp = &group_of($unprefBase, 1) and # This time, try to 
+derive a group for the base.
          &member($metagroup{$savgrp}, 'NVRB', 'NOUN','VERB','ADJECTIV
+E','ADVB', 'GERUND')){}
[download]

When I looked for "$unprefBase" above that point, I found it only once, near the top of the sub:

    my($sing, $unprefBase, $savgrp, $savbase);
[download]

Nothing is ever assigned to $unprefBase, so the following two parts of that last elsif condition will never be evaluated -- and the block itself is empty anyway, so why have that condition at all? (That's the first thing I'd get rid of, if the idea is to shorten the list of elsif blocks.) You might want to set up some test data, such that you believe it should exercise every block, and then see whether it really does get into every block.

I think the issue is not so much to shorten the list of elsif conditions, but rather to organize them into something like a coherent order and structure, so that sequential dependencies, and the extra assignment steps that affect outcomes, are more easily (and logically) identifiable as such.

Another point: at the top you say "driveBase returns true for success...", but then it contains recursive calls to driveBase where the return value is not checked. In fact, since "wordCache{$tok}" is the last statement in the sub, it will only return false if this hash element is 0, empty string, or undef. And what would that mean, exactly?