Legend:

• pruned datasets are reduced in size to contain only examples in which the source paragraph and the target paragraph are not identical
• hard datasets have their training and test split created from separate pools of books with no overlap (so all paragraphs from a given book are contained in only a single split)
• transduced datasets have their training split processed by a rule-based normalizer

Models were accordingly created based on the 4 dataset variants.

Evaluation repositories:

https://github.com/kedudzic/pl-normalizer-evaluation (private)

https://github.com/kedudzic/pl-normalizer-evaluation-just-results (public)