Title: Using XML Logical Structure to Retrieve Multimedia Objects
1Using XML Logical Structure to Retrieve
(Multimedia) Objects
- Zhigang Kong Mounia Lalmas
- Information Retrieval Group
- Queen Mary, University of London
2Outline
Motivation
Related Work
XML Logic Structure
Test Collections
Experiments
Conclusions Future Work
3Motivation
The Use of XML Logical Structure
understanding how combining logically structured
document works
logically structured document vs. whole document
4Related Work
Web Image Retrieval
INEX 2005 Multimedia Track
Annotation/ Metadata
My work explores the logical Structure of XML
document
My work exploit existing text content and not
just metadata
My work employs the XML Structure in different
ways.
5XML Logic Structure
article
bdy
abs
sec
sec
subsec
fig
p
p
6Test collections Methodologies
-
- Using an XML text collection to validate XML
multimedia retrieval approach. - Methodology inspired by Dunlop and van
Rijsbergen. - A multimedia element is, actually, just a text
element, which has an attribute value referencing
an external entity (a multimedia object).
7Test Collection
25 topics having more than 10 relevant elements.
INEX
2004 Text Collection
Only consider highly relevant element as
relevant.
8Experiments
Fig. 1. Regions from sibling upwards
The MAP values obtained using regions from
sibling level to 8th ancestor level are 0.1166,
0.1383, 0.1900, 0.1828, 0.0807, 0.0196, 0.0047,
0.0067, and 0.0019.
9Experiments
Fig. 2. Regions from root level down
The MAP values obtained using regions from the
highest level to the 3rd highest level are
0.1294, 0.1858, and 0.1868.
10Experiments
Fig. 3. Combination vs. whole document (1)
The MAP values obtained in the above experiments
are 0.1922, 0.1900, and 0.3114.
11Experiments
Fig. 4. Combination vs. whole document (2)
The MAP values obtained in the above experiments
are 0.2918, 0.2911, and 0.3488.
12Conclusions
logical structure is important in XML multimedia
retrieval
Structure in XML Document
1) All levels 2) Lower level
3) Higher level
Combination 1) improved overall
performance 2) Structured region vs.
whole documents
13Future Work
More Sophisticated Model
Semantic Structure
Large XML Multimedia Collection
14Thank You
- Email cskzg_at_dcs.qmul.ac.uk