German Tagsets

 

The point of reference for all tests described in this document is the German specification of the EAGLES morphosyntax encoding proposal, ELM-DE. The following needs to be taken into account when interpreting the test setups and the results.

• The tests take into account the technical state of the art of automatic tagging and thus concentrate on the part of speech annotation proposed in ELM-DE, leaving aside the more detailed and more complex set of morphosyntactic features. Thus, only the higher levels of the ELM-DE specification have been used here. This part corresponds to the STTS tagset for German, which distinguishes 54 part of speech tags (cf. Appendix C).
• The tagset used at RXRC is a derivative of STTS, but it omits the distinction between proper names and common nouns, and it also contains a few modifications, with respect to STTS, as far as the classification of function words is concerned. The latter differences are not particularly relevant for the present tests, and the former one will be kept track of.
• The 1995 version of STTS had a predecessor, the IMS/TUE-Tagset, of 1993. STTS is mappable to ELM-DE (it only has a few finer grained distinctions in the domain of indefinite pronouns). In the redesigned phase which led from IMS/TUE to STTS, a number of interesting questions came up which concern tagset design and its impact on tagging results. Some of these questions have been analyzed in more detail in the experiments we will report on in this document, the others have been collected in Appendix B. Thus the tagset evaluation is based on concrete questions which came up during tagset design and not just on arbitrary or otherwise motivated questions. A detailed definition of the modifications tested in the tagset evaluation work reported here is given in Appendix A.