Abstract:
Historical newspaper archives provide a wealth of information. They are of particular interest
to genealogists, historians and scholars for People Search.
In this thesis, we design a People Gazetteer from the noisy OCR text of historical newspapers
and identify \in
uential" people from it. A People Gazetteer is a dictionary of personal names;
each entry of the gazetteer is a tuple containing a person name and a list of articles in which
his name occurs along with the corresponding topic associated with each article.
To build the People Gazetteer, we rst spell correct the noisy text using an edit distance based
algorithm. A novel N-gram based evaluation algorithm is designed for measuring the perfor-
mance of the spell corrector. Next, a Named Entity Recognizer is run on the text of each article
to identify person entities and an LDA-based topic detector to assign categories to articles. To
identify in
uential people across each category of People Gazetteer, we de ne the notion of an
In
uential Person Index (IPI) and rank based on it.
Our corpus is a sample of 14020 OCR newspaper articles (roughly two months' data) obtained
from \The Sun" newspaper in the Chronicling America project. We present results on the top-K
in
uential people obtained from our algorithm by varying its parameters and verify results using
Wikipedia.
