ToDo

• empty file for evaluation [x] fixed not only for evaluation, but for every task
• every prediction is NN for evaluation [x]
• alpha jede iteration minimieren [x]
• token bei jeder iteration random durchlaufen [x]
• mutual information bei affixes
• optimierungsverfahren: hill climbing, ...
• viterbi + perceptron?
• maxent (https://web.stanford.edu/class/cs124/lec/Maximum_Entropy_Classifiers.pdf)