.
Moreover, the aliases extracted using the
proposed method improve recall by 
in a relation-detection task.
1
.
Moreover, the aliases extracted using the
proposed method improve recall by in a relation-detection task.
H.3.3Information SystemsInformation Search and Retrieval Algorithms Name alias extraction, Semantic Web, Web Mining
different people are listed among the top 
results returned by Google for the name.
On the other hand, referential ambiguity occurs because people use different names to refer to
the same entity on the web. For example, the American movie star Will Smith is often
called the fresh prince in web contents.
Although lexical ambiguity, particularly ambiguity related to personal names, has been explored extensively in the previous studies of name disambiguation [1], the problem of referential
ambiguity of entities on the web has received much less attention.
In this paper, we specifically examine on the problem of automatically extracting the various references
on the web to a particular entity. In contrast to the real name of an entity, we use the term alias
to describe words or multi-word expressions that are used to refer to the entity
(e.g., fresh prince is an alias for Will Smith).
pairs of names and aliases as seeds, we extracted over 
lexical patterns.
Patterns are ranked according to their 
-scores on training data. Top ranking patterns
are shown in Table 1.
Considering the noise in web-snippets, candidates extracted using a set of shallow lexical patterns
might include some invalid aliases. Therefore, it is imperative that we identify the candidates
which are most likely to be correct aliases of a given name. We model this problem of alias
recognition as one of ranking candidate aliases with respect to a given name such that the
candidates which are most likely to be correct aliases are assigned a higher rank. We define
various ranking scores to measure the association between a name and a candidate alias using
two approaches: hyperlink structure on the web and page-counts retrieved from a search
engine. Inbound anchor texts of a url provide useful semantic clues related to the resource
represented by the url. We define two words 
and 
as co-occurring, if they
appear in at least two different inbound anchor texts 
and 
, respectively, in a url 
.
Using this definition, we compute 
popular co-occurrence measures. Hubs in the hyperlink
graph are automatically detected and co-occurrences in hubs are appropriately adjusted.
Moreover, four page-count-based association measures are computed [2].
All ranking scores are integrated using raking support vector machines to leverage
a robust ranking function. Because of the limited availability of space, we omit the details
of the ranking algorithm.
| pattern | -score |
||
| * | aka | [NAME] |  |
| [NAME] | aka | * |  |
| [NAME] | better known as | * |  |
| [NAME] | alias | * |  |
| [NAME] | also known as | * |  |
| * | nee | [NAME] |  |
| [NAME] | nickname | * |  |
| * | whose real name is | [NAME] |  |
names), Japanese personal names ( names) and English place names ( U.S. states).
Personal names datasets include people from various fields of cinema, sports, politics and science.
As shown in Table 2, the proposed method reports high mean reciprocal rank (MRR)
and average precision on all datasets. Moreover, the proposed method outperforms a previously
proposed alias extraction algorithm by Hokama and Kitagawa [3]
(MRR=
on Japanese names dataset) which uses manually created patterns
specific to Japanese names. Top ranking aliases as extracted by the proposed method
for some names are shown in Table 3. Overall, in Table 3 the proposed
method extracts most aliases assigned in the manually created gold standard (shown in bold).
| Dataset | MRR | Average Precision |
| English Personal Names |  |
 |
| English Place Names |  |
 |
| Japanese Personal Names | |
 |
| Real Name | Extracted Aliases |
| David Hasselhoff | hoff, michael knight, michael |
| Courteney Cox | dirt lucy, lucy, monica |
| Al Pacino | michael corleone |
| Teri Hatcher | susan mayer, susan, mayer |
| Texas | lone star state, lone star, lone |
| Vermont | green mountain state, green, |
| Wyoming | equality state, cowboy state |
| Hideki Matsui | Godzilla, nishikori, matsui |
| Real name only | Real name and top alias | ||||
| Precision | Recall | F | Precision | Recall | F |
 |
 |
 |
 |
 |
 |
We evaluate the extracted aliases on a relation detection task.
First, we manually classify people
in the English personal names dataset into four categories:
music, politics, movies, and sports. Then
we measure the association between two people using the WebPMI [2]
and perform group average agglomerative clustering to form four clusters.
Clustering accuracies with and without using aliases are shown in Table 4.
The use of aliases significantly improves recall (ca. ) and consequently the score.
By considering not only real names but also aliases, it is possible to discover relations that are
unidentifiable solely using real names.
