Disambiguation Algorithm for People Search on the Web

1
Disambiguation Algorithm for People Search on
the Web

• Dmitri V. Kalashnikov, Sharad Mehrotra, Zhaoqi
  Chen,
• Rabia Nuray-Turan, Naveen Ashish
• For questions visit
• http//www.ics.uci.edu/dvk
• Computer Science Department
• University of California, Irvine

2
Entity (People) Search
Person2
Person1
Top-K Webpages
Person3
Unknown beforehand
3
Standard Approach to Entity Resolution
4
Key Observation More Info is Available

5
RelDC Framework
6
Where is the Graph here? Use Extraction!
7
Overall Algorithm Overview

• User Input. A user submits a query to the
  middleware via a web-based interface.
• Web page Retrieval. The middleware queries a
  search engines API, gets top-K Web pages.
• Preprocessing. The retrieved Web pages are
  preprocessed
• TF/IDF. Preprocessing steps for computing TF/IDF
  are carried out.
• Ontology. Ontologies are used to enrich the
  Webpage content.
• Extraction. Named entities, and web related
  information is extracted from the Webpages.
• Graph Creation. The Entity-Relationship Graph is
  generated
• Enhanced TF/IDF. Ontology-enhanced TF/IDF values
  are computed
• Clustering. Correlation clustering is applied
• Cluster Processing. Each resulting cluster is
  then processed as follows
• Sketches. A set of keywords that represent the
  web pages within a cluster is computed for each
  cluster. The goal is that the user should be able
  to find the person of interest by looking at the
  sketch.
• Cluster Ranking. All cluster are ranked by a
  choosing criteria to be presented in a certain
  order to the user
• Web page Ranking. Once the user hones in on a
  particular cluster, the Web pages in this cluster
  are presented in a certain order, computed on
  this step.
• Visualization of Results. The results are
  presented to the user in the form of clusters
  (and their sketches) corresponding to namesakes
  and which can be explored further.

8
Correlation Clustering

• In CC, each pair of nodes (u,v) is labeled
• with or - edge
• labeling is done according to a similarity
  function s(u,v)
• Similarity function s(u,v)
• if s(u,v) believes u and v are similar, then
  label
• else label -
• s(u,v) is typically trained from past data
• Clustering
• looks at edges
• tries to minimize disagreement
• disagreement for element x placed in cluster C,
  is a number of - edges that connect x and other
  elements in C

9
Similarity Function

• Connection strength between u and v
• where ck the number of u-v paths of type k
• and wk the weigh of u-v paths of type k
• Similarity s(u,v) is a combination

10
Training s(u,v) on pre-labeled data
11
Experiments Quality of Disambiguation
By Artiles, et al. in SIGIR05
By Bekkerman McCallum in WWW05
12
Experiments Effect on Search