Disambiguation
of People’s Web Appearances
My
primary interest is automatic identification of web presence of particular
people. The problem appears to have a trivial solution when the name of the
person is unique – we can just google this person’s
name. However, the problem complexity grows significantly with the level of
commonness of the personal name. Indeed, given a common name such as “Tom
Mitchell”, we find hundreds of different people called Tom Mitchell. The
main complication is that a web mining system cannot a priori know how ambiguous
one or another personal name is. Yet another complication is that the person we
are looking for may not have any web
presence, but his or her namesakes have. And even if the person is well
presented in the web, an existence of his of her famous namesake can make the
search practically impossible (compare “Julia
Roberts” and "Julia
Roberts" Professor WKU).
Well,
given just a personal name, the
problem of finding the person’s web presence cannot be theoretically resolved.
Some additional information about the person is required. When we search for a
person in the web, we usually construct a query that provides some kind of
summary of the person’s activities, e.g. "Tom
Mitchell" Professor CMU. However, if our query is too common (such as "Tom
Mitchell" Professor), our first hit will be Tom Mitchell the UChicago Professor, and if our query is too narrow (as "Tom
Mitchell" "Professor CMU"), we may not find anyone. Since
modern search engines are so sensitive to small variations of the same query, it
is hard to require an automatic system to construct queries of high quality.
We
noticed however that the problem of personal name disambiguation becomes much
easier if we are given a few names of
people who are known to be related to each other. They can for example be
co-authors of scientific publications, share a profession or a hobby. Even if
their names are totally ambiguous, when we google two
names together, like "Tom
Mitchell" "William Cohen", we are almost guaranteed to find
the right people. Or accidentally to find nothing, if web
presence of at least one of the two people is sparse enough.
I
have built a web mining system that disambiguates web appearances of a group of
people. I test the system on a list of 12 personal names. These names are taken
from headers of email messages from one folder of a CALO participant’s email
directory. For each name, I retrieved100 first Google
hits and labeled them manually. After removing some empty pages and error
statements, the resulting dataset consists of 1085 web pages that refer to 187
different people, from which 420 pages are relevant (they refer to the 12 particular
people we are looking for). Some statistics of the dataset can be found below.
|
Personal
name |
Position |
Number of
retrieved pages |
Number of
namesakes |
Number of
relevant pages |
|
SRI Manager |
97 |
2 |
96 |
|
|
CMU Professor |
88 |
10 |
6 |
|
|
Steve Hardt |
SRI Engineer |
81 |
6 |
64 |
|
SRI Manager |
92 |
19 |
20 |
|
|
MIT Professor |
89 |
2 |
88 |
|
|
SRI Manager |
94 |
8 |
11 |
|
|
UMass Professor |
94 |
16 |
54 |
|
|
CMU Professor |
92 |
37 |
15 |
|
|
David Mulford |
Stanford Undergrad |
94 |
13 |
1 |
|
Stanford Professor |
87 |
29 |
32 |
|
|
UPenn Professor |
88 |
19 |
32 |
|
|
Lynn Voss |
SRI Engineer |
89 |
26 |
1 |
The dataset is ready
for download. Download the preprocessed dataset (4.8 Mb
tarred, gzipped). Makeup language was removed
from all the pages. At the end of each page there is a list of its hyperlinks,
starting with the URL of the page.
Publication. We proposed two unsupervised methods for the web
appearance disambiguation. One is based on link structure of the pages, another
one on distributional clustering of the pages. The hybrid of the two methods
achieves above 80% F-measure on our dataset. More information can be found in:
|
|
Disambiguating Web
Appearances of People in a Social Network. Joint work with A. McCallum. In Proceedings of WWW 2005
ps pdf . |
