Semantic Search

1
Semantic Search

• R. Guha IBM Almaden
• Rob McCool Stanford Univ.
• Eric Miller W3C/MIT

2
Outline

• Introduction
• Assumptions
• Aspects of the Semantic Web
• TAP
• GetData
• Semantic Search
• Denotation
• What to Show
• Formatting
• Text Search
• Evaluation
• Conclusion

3
Introduction

• The Semantic Search uses well-defined meaning
  given to information on the Semantic Web to
  augment traditional Information Retrieval (IR)
  searches
• The focus of this paper is to demonstrate the
  implementation of the Semantic Search
• The scope of the paper is a subset of the web as
  it pertains to IR
• Two kinds of searches, Navigational and Research

4
Assumptions

• The Semantic Web has well-defined meaning for the
  information it contains
• The resources on the Semantic Web have
  relationships between each other
• The Semantic Web is modeled as a directed labeled
  graph
• RDF is used to describe resources and
  relationships

5
Assumptions
A segment of the Semantic Web pertaining to Eric
Miller
6
Aspects of the Semantic Web

• Documents vs. Real World Objects
• Human information vs. Machine information
• The relationship between HTML and the Semantic
  Web
• Distributed Extensibility
• Trustworthiness of the data

7
TAP

• TAP enables sites to publish data onto the
  Semantic Web and applications to use that data
• TAP Scrapping retrieves information from HTML
  sites and convert it into machine understandable
  data
• TAPache was used to publish data onto their
  Semantic Web, ABS
• TAP has a query interface called GetData

8
GetData

• A lightweight query language
• Unpredictable users
• Good for aggregating data from multiple sites
• It is not an expressive query language
• Built on SOAP
• Designed to query directed labeled graphs

9
GetData

• Queries have two arguments the resource of the
  properties to be accessed (nodes) and the
  properties to be accessed (arcs)
• GetData(, ),
• Returned from the query is a graph with resources
  as URIs

10
GetData

• Example
• GetData (, livesIn),
• http//tap.stanford.edu/data/CityDublin,_Ohio

11
Semantic Search

• ABS contains data from many sites, such as
  AllMusic, Ebay, Amazon, AOL Shopping,
  TicketMaster, Weather.com and Mapquest
• There are millions of triples in the ABS Semantic
  Web
• TAP knowledge base has a broad range of domains
  including people, places, organizations, and
  products
• Resources have a rdftype and rdfslabel

12
Semantic Search

• W3C Semantic Search has five different data
  sources People, Activities, Working Groups,
  Documents, and News
• Both the ABS and W3C Semantic Search have a basic
  ontology about people, places, events,
  organizations, vocabulary terms, etc.
• The plan is to augment a traditional search with
  data from the Semantic Web

13
Semantic Search
14
Semantic Search

• A secondary objective is to use an understanding
  of the denotation to improve traditional search,
  this will be covered under Text Search
• Semantic Search is built as a client of TAP
• Search terms are mapped to nodes on the Semantic
  Web

15
Semantic Search

• Three problems facing Semantic Search
• Denotation
• What to Show
• Presentation

16
Denotation

• Need to determine the concept denoted by the
  search query
• To overcome ambiguity
• Popularity of the search term
• A user profile
• Search context
• Offer other denotations to the user

17
Denotation

• Complex Search Terms
• eric miller rdf
• eric miller rdf
• eric miller rdf
• eric miller rdf
• Restricted to two denotations
• Nothing will be contributed to the traditional
  search if terms do not denote anything known to
  the Semantic Web

18
What to Show

• What data should be included in the results and
  in what order?
• The selected node or nodes, which corresponds to
  the denotation of the term, is the anchor node
  (for combination of terms there are two anchor
  nodes)
• A sub graph must then be selected to be returned

19
What to Show
Approach 1

• The graph is selected by walking through the
  graph in breadth first order starting at the
  anchor node
• The first N triples are collected (N is some
  predefined limit)
• Included are at most R triples with the same
  source and arc label (R is computed based on the
  average branching factor)

20
What to Show
Approach 1 cont.

• Included in the sub graph are at most M triples
  with the same source (M is computed based on the
  bushiness or can be a function of the distance
  from the anchor node)
• Adv no hand coding by the user, incorporates
  information about the anchor and its neighbors
• Disadv sensitive to representation on the
  Semantic Web, ignores the search context

21
What to Show
Approach 2

• Manually specify the set of properties that
  should be gathered
• Walk through the graph breadth first starting at
  the anchor node, collecting values for the
  properties until N triples have been collected (N
  is some predefined number)
• Adv able to specify certain properties, able to
  factor in search context
• Disadv takes time to setup, does not incorporate
  information about the anchor node and its
  neighbors

22
What to Show

• A hybrid approach would have the best results
• The second approach is applied first, then if N
  triples were not collected, the second approach
  is applied looking at properties not previously
  examined
• The triples are ordered based first on their
  distance from the anchor and second the arch label

23
Formatting

• Results are displayed using a set of templates
• Each class of object has an associated set of
  templates
• The templates specify the class and the
  properties and a HTML template
• A template is identified for each node in the
  ordered list and the HTML is generated
• The HTML is included in the results page

24
Formatting
25
Text Search

• Most searches are about a relatively small number
  of categories, such as person, place,
  organization, document, etc.
• Knowing the category greatly increase accuracy
• Applying type to the category narrows the search
  further
• This is in need of further research, however the
  current proposed solution is to offer an example
  from each category and have the user select the
  context which they want

26
Evaluation

• Testing is being conducted on the W3C website
• Individuals searching on the website are
  anonymously chosen at random
• Their results are augmented with the W3C Semantic
  Web Search
• User activity is recorded, when a user clicks on
  a link supplied by the Semantic Search as opposed
  to a link from the traditional search

27
Conclusion

• Research searches are suitable to exploit the
  information on the Semantic Web
• More work is needed on Text Search
• More work is needed on denotation

28
My Thoughts

• It may be an unachievable goal to have a computer
  know what we are looking for when what we are
  looking for is not easily expressed and
  categorized.
• An achievable goal may be to have searches filter
  out the junk, leaving us with a minimal number
  of resources to sift through manually
• The authors were thorough and moderately easy to
  follow

29
References

• Semantic Search, R. Guha, Rob McCool, Eric
  Miller 2003
• http//www2003.org/cdrom/papers/refereed/p779/ess
  .html
• TAP
• http//tap.stanford.edu/