Hotfix data generation

[jm-glowienke]

• Added a tokenizer (Moses) and subword processing (Byte-Pair Encondig, subword-nmt) to the pre-processing.
• The encoding of SPARQL was adapted and the names/IDs in the queries are surrounded by quotation marks.
• The source files and directories were cleaned. Furthermore, the directory structure was adapted

This PR closes #16 and the subword processing might also fix the corresponding issue. The latter has to be checked after training the model