We propose a semi-supervised method for expanding Frame Semantic corpus
resources with the goal of complementing manual annotation efforts and
providing better training corpora to Semantic Role Labeling methods.
Data expansion is performed by similarity-based projection of annotations
to sentences of a large unannotated corpus. The method consists of a
syntactic comparison step and a semantic step incorporating
a novel selectional preference model which can cope with situations of
sparse evidence.
Evaluation shows that in comparison to a purely syntactic baseline our
method significantly increases accuracy. In spite of lower coverage it
acquires relevant new information from the unannotated corpus.