Annotation as trade-off between generalizations and specializations

Annotation is based on the hermeneutic process of generalizing from individual instances to interpretative classes, for example from prenominal "both" to the part-of-speech class "determiner". I will discuss the trade-off between modeling a phenomenon accurately in terms of annotation classes and the accuracy of assigning correct classes to individual instances (both manually as well as automatically) given that many words and structures are ambiguous. I will exemplify this by work done on part-of-speech annotation including a recent community effort on adapting the Stuttgart Tübingen Tagset (STTS) to non-standard language (http://www.linguistics.ruhr-uni-bochum.de/stts).