Use character level augmentation and evaluate approach

[djaszak]

In previous issues 6 different augmentation methods were implemented which were used to augment the imdb dataset available on HuggingFace.co. The next step should be to use some standard model, also available on HuggingFace and fine-tune it with the different augemented datasets, checking which approach works best. Different combinations should be implemented and tuned.

• complete_randomizer
• keyboard_replacer
• mid_randomizer
• replacer
• remover
• misspeller All of this will be trained and evaluated on their own, augmented eval datasplit as well as on the correct eval datasplit. Also all of them should be evaluated after being trained on a purely augmented train datasplit and the augmented train datasplit concatenated to the correct imdb train datasplit.
• [x] only augmentented
• [x] augemented with non-augmented eval
• [x] augmented + non-augmented test & non-augmented eval

[djaszak]

Base was implemented in b21b1b7fd8a2446925671b7e469ac8bffd0d664a