First prototypes of hybrid combination models

Risk assessment: The model does not show significant improvements in evaluation. This model is aimed to improve upon the more basic cascade-based model in M5.1. If the performance is still too weak, we have 6 months to improve the hybrid translator before the final version in M5.3.

Verification: The methods are implemented and evaluated on a specific test set. Reported in D5.4.

(THESE IDs, taken from Annex, DO NOT MAKE SENSE)