Since the acquisition of rich, domain-independent semantic lexica has
been identified as a crucial step in improving many NLP applications,
considerable effort has been put into creating these kind of
resources. In the SALSA and SALSA II projects here in Saarbruecken, a
semantically annotated text corpus was created manually. However,
since manual annotation is a slow and expensive process, ways of
automatically expanding this highly reliable core data are necessary.

In this talk I first want to present some of my ideas about the
possibilities of improving NLP tasks in general and automatical
semantic annotation in particular by a notion of text types. Then I
will point out why this approach doesn't seem to be feasible for
semantic annotation in its present state. Finally I will talk about
how I am now trying to automatically expand the existing annotation
data by using large corpora and variation techniques.