Toward the SelfAnnotating web

About This Presentation

Title:

Toward the SelfAnnotating web

Description:

the Hilton hotel ... the hotel Hilton... DEFINITE1: the INSTANCE CONCEPT ... Hilton, a hotel in ... Hilton is a hotel... APPOSITION: INSTANCE , a CONCEPT ... – PowerPoint PPT presentation

Number of Views:43

Avg rating:3.0/5.0

Slides: 26

Provided by: sea59

Category:

more less

Transcript and Presenter's Notes

Title: Toward the SelfAnnotating web

1
Toward the Self-Annotating web

Presented byHaibo zhaoThe University of Georgia

2
Outline

Introduction
The process of PANKOW
Pattern-based categorization
Evaluation
Conclusion
Possible ways to improve

3
1 Introduction

The success of semantic web depends on the
proliferation of web pages annotated with
metadata
How to generate these annotated web pages
Manually, Semi-automatically, Or automatically?

4
Problems with current large-scale information
exaction

Manual definition of an information extraction
system is a laborious task requiring a lot of
time
Learning of extraction rules requires too many
examples for learning the rules
Vicious circle there is no semantic web because
of lack of metadata, and there is no metadata
because of there is no semantic web that one can
learn from

5
PANKOW

Pattern-based annotation through Knowledge On the
Web
An unsupervised, pattern-based approach to
categorize instances with regard to an ontology
Use globally available web data and structures to
semantically annotate (or at least facilitate
annotation of) web pages

6
Linguistic Patterns (1)

Hearst patternsProfessors such as Dr. Sheth
such Professors as Dr. Sheth Professors,
especially Dr. Sheth Dr. Sheth and other
Professors

H1 s such as H2 such
s as H3 s,
(especiallyincluding) H4
(andor) other s
7
Linguistic Patterns (2)

Definites patternsthe Hilton hotel the hotel
Hilton

DEFINITE1 the DEFINITE2
the
8
Linguistic Patterns (3)

Apposition and CopulaHilton, a hotel in
Hilton is a hotel

APPOSITION , a COPULA
is a
9
2 The Process of PANKOW
10
Process details of PANKOW

(1) Identify candidate proper nouns
(2) Derive hypothesis phrases
(3) Query Google DB to get the number of
hits for each hypothesis phrase
(4) Sum up the query results to a total
for each instance-concept pair. Then the
system categorizes the instances into
their highest ranked concepts

11
(1) Identify candidate proper nouns

Input A web page we want to annotate
Output Set of candidate proper nouns
How POS tagger (Part-Of-Speech tagger)Pos
tagger is an approach which can assigns the
correct syntactic category, like adjective,
common noun, proper noun to words
Example

12
(2) Derive hypothesis phrases

Input Set of candidate proper nouns
Output Set of hypothesis phrase
How Introduce all candidate proper nouns and
candidate ontology concepts into linguistic
patterns to derive hypothesis phrases
Example

13
(3) Query Google DB to get the number of hits for
each hypothesis phrase

Input Set of hypothesis phrases
Output the number of hits for each hypothesis
phrase
How Google DB is queried for the hypothesis
phrases via google web service API
More about google web apihttp//www.google.com/a
pis/

14
(4) Categorization

Input The number of hits for each hypothesis
phrase
Output Annotated Web Page
How The system sumps up the query results to a
total for each instance-concept pair, then the
system categorizes the candidate proper nouns
into the highest ranked concepts
Example

15
2 Determine the best categorization

Baseline
Linear weighting
Interactive SelectionReturn the best n matches
for each proper noun

16
3 Evaluation

30 web pages needed to annotate, 277 proper
nouns, 59 concepts, and 10 patterns mentioned
earlier.
2 human subjects (A and B) are asked to annotate
manually.
Compare the system results with manually
annotation results

17
Results got by the system (top 60
Instance-concept relations)
18
Evaluation measures

Reference standards (Results produced by Human A
and Human B)

19
Evaluation measures

Define Precision and Accuracy
y belongs to A, B
Compute seperately for A and B, then average the
results to be the final precision and accuracy

20
(No Transcript)
21
Interactive Selection

Instead of returning a unique concept for each
instance, it returns the best n concepts to end
users.
The results are very good, the best accuracy is
49.56
This means that for almost half of the instances
in a web page, the system can provide the user
the correct answer (concept)

22
(No Transcript)
23
4 Conclusion

A simple but smart idea to do categorization
Relatively effective approach to do
auto-annotation
Heavily rely on GOOGLE
It seems we can do more than what this paper did

24
My Opinions

It only takes advantage of the number of hits,
but google web api can actually return an
abstract for the phrase, which is a context. How
to make these contexts useful ?
It only consider is-a relationship. Is it
possible to figure out other patterns for other
relationships
Any way to generate these patterns automatically ?

25
Questions ?

Write a Comment

User Comments (0)