My initial reaction would be that pagination is a secondary concern, and that the list of questions is the root of the matter, so I'd start with an array of hashes -- or, if the index-naming of the questions is "non-linear" or "semantic" in some way (i.e. not just an ordered list, but a set of distinctly named entities), then make it a hash of hashes.

In any case, the "outermost, primary" unit of organization is the "question", and features of each question are simply:

• its position in the sequence (or its ID/name, which is presumably sortable in some way)
• which page it should be on
• what its label is
• what its possible answers are

so the structure could be:

my @questions =
 [
   { page => 'p1', idstr => '4B',
     label => "... ice cream?", answertype => 'single',
     answerset => { 1 => 'yes', 2 => 'no' }
   }
   ...
  ];

# or ...
my %questions =
  { '4B' => { page => 'p1', label => '... ice cream?',
              ... }
    '4C' => { page => 'p1', label => '... whipped cream?',
              ... }
    ...
  };
[download]

As for parsing the input text to fill that structure, there are numerous ways, and Parse::RecDescent would certainly do it (but it might be overkill -- other ways would suffice and be easier if you're really green with P::RD).