
- different link search (inference) for terminals and for non-terminals
  terminals: allow many-to-many links
  non-terminals: only 1:1 links

- handle berkeley parser format

- SEARN - without reading through training corpus several times
  (too little to gain?)

- multi-threaded alignment? is it worth it?
  (--> simpler: run on cluster!)
  http://www.mathematik.uni-ulm.de/help/perl5/doc/perlthrtut.html


..... and many other things ....
