- Save tokenization rules alongside the serialized structure.
- Generalize the rule search method to a beam search (with beam=1 by default)
- Create a sensible stopping criterion when further refining list iteration rules.
- When applying exhaustive rules, allow disjunctions to be used to yield the maximal number of results
- How well would successive generalisation work?
- Make use of multiple processors when learning rules.
- Fix Ariel so it works when there are no Wildcards