Enhancing Hypernym Extraction for Named Entities Using Machine
             Learning Based ClassificationArndt Faulhaber


                            Abstract

In this thesis I will discuss the construction and evaluation of a
learning System that aims at extracting hypernyms for arbitrary named
entities. The presented approach tries to accomplish this by using the
web as a source of information. Namely querying Google for a list of
named entities and creating a corpus by retrieving the pages
returned. From the corpus a set of snippets, i.e. sequences of words,
is extracted that contains the named entity at the middle. A set of
vectors is constructed, that is based on structural information
conveyed by the words in the snippets and data extracted from the
titles of the web pages. These vectors build the foundation for
learning patterns that indicate a correct hypernym of the named
entity. Only nouns are taken into account as hypernym candidates. To
create a supervised learning scenario all extracted nouns are
annotated äs either representing a hypernym to a given named entity or
not. Regarding these annotations the built System is evaluated. By
comparing the results of the evaluation to those of a previously
defined baseline method, the viability of the proposed approach can be
assessed.