Abstract

In many entity extraction applications, the entities to be
recognized are constrained to be from a list of “target entities”.
In many cases, these target entities are (i) ad-hoc,
i.e., do not exist in a knowledge base and (ii) homogeneous
(e.g., all the entities are IT companies). We study the following
novel disambiguation problem in this unique setting:
given the candidate mentions of all the target entities, determine
which ones are true mentions of a target entity. Prior
techniques only consider target entities present in a knowledge
base and/or having a rich set of attributes. In this paper,
we develop novel techniques that require no knowledge
about the entities except their names. Our main insight is to
leverage the homogeneity constraint and disambiguate the
candidate mentions collectively across all documents. We
propose a graph-based model, called MentionRank, for that
purpose. Furthermore, if additional knowledge is available
for some or all of the entities, our model can leverage it to
further improve quality. Our experiments demonstrate the
effectiveness of our model. To the best of our knowledge,
this is the first work on targeted entity disambiguation for
ad-hoc entities.