Machine-Learning Aided Annotation Transfer from Sentence- To Token-Level

Master Thesis

Many tasks require annotations on sentence AND on token level. This can e.g. be overall sentiment and phrase sentiment, overal l argument stance and pro/con argument spans or intent and slots. A large chunk of time during token- or character-level annotation is spent on deciding for the correct span, and inputting that span into an annotation system. Furthermore, reading, understanding, and remembering annotation guidelines which explain the correct annotation of spans imposes considerable mental load on annotators. The goal of this thesis is training machine learning models that help human annotators with making token-level annotations given already made sentence level annotations