With current version (3.9.2) analysis against a test corpus showed much better results as with version 4.0 dropped in. It needs more investigation than just bumping versions.