Add test for trained relation extraction modules

We should have a test that runs the current (trained) relation extraction modules and makes sure that they produce reasonable output on a sample sentence or two. This should avoid any obvious regressions when committing new, updated models.