Fanfiction platforms become very popular. However, since fan fiction stories can also contain content
that can be disturbing to readers, it is important to assign appropriate warnings to them. The automatic
assignment of 32 trigger labels to fanfiction works is addressed by the Trigger Detection task of PAN’23
in terms of a multi-label document classification. This paper presents a two-stage approach in which
the final multi-label classification was preceded by a pre-classifier that predicted newly formed upper
classes of trigger warnings. For both classifiers, features based on fastText word-embeddings and semisupervised topic modeling were used. Multi-Layer-Perceptron (MLP) was used as classifier in both stages,
and its performance was compared with Random Forest (RF) for the first classifier. The applicability
of the two-step approach was shown in a comparison with a traditional one-step procedure. The best
model achieved a micro F1-score of 0.54.