Approaches based on topic signatures are
similar in spirit to our approach. A topic signature
is a list of topically related words (Agirre et al.,
2000). There are many topic signature approaches
(e.g., Agirre et al., 2000; Zhou et al., 2007). Zhou
et al. (2007) propose a Topic Signature Language
Model that is used to perform semantic smoothing
to increase retrieval performance. They create
topic signatures for each concept defined in a
domain specific ontology using a highly relevant
document collection. The topic signature terms are
found by collocation. They assume the concepts
are unique and consequently circumvent the
problem of word disambiguation. For general
domains where no ontology exists, they propose to
use multiword expressions as topic signatures. The
multiword expressions contain context and are
consequently mostly unambiguous.