The parser was crashing when run with the both non-projective and second order options. The cause was incomplete serialization of DependencyInstances; this bug has been squashed. Download release 0.4.3b to get the fix.
The last version (0.4.2) had a lot of extra feature turned on which led to reduced performance for a number of data sets. (Those features were for discourse parsing, not sentential parsing.) Things are now cleanly separated, so the old results can be obtained again, modulo a few minor differences.
The deadline for joining the CoNLL 2007 shared task on dependency parsing is approaching (January 20). Check out details here:
You may also find other useful dependency parsing resources there, especially other parsers and also data for languages other than English.
Also,as a further resource, the site for last years competition is here:
I have added two releases. The first is Ryan's v0.2 code that is also available at the MSTParser web site:
The second is v0.4, a release that I have modified in various ways, though nothing too drastic. The main difference is that the parser now supports the CoNLL format for dependencies, and it is much easier to add new features. There are also some handy Python scripts, like mst_experiment.py, which takes care of n-folding and such. See the Change Log for details.