The Parser is still using old and deprecated APIs to train the POS Tagger and Chunker. Especially setting the beam size of the POS Tagger and Chunker doesn't work in those deprecated code paths and lead to a performance difference between the current 1.6.0 RC2 and the 1.5.3 release.
The Parser should be updated to use the new 1.6.0 API for training. The new code path was broken too, but should be in a good state since the fixes in OPENNLP-762.