ACL/LaTeCH-Portland, June 24th 2011
Enrichment and Structuring of Archival Description
Metadata
Kalliopi Zervanou*, Ioannis Korkontzelos**,
Antal van den Bosch* & Sophia Ananiadou** * Tilburg Centre for Cognition &
CommunicationThe University of Tilburg, NL
[email protected] [email protected]
** National Centre for Text MiningThe University of Manchester, UK
ACL/LaTeCH-Portland, June 24th 2011
Research on Metadata• Developing standards:
– collection specific (e.g. EAD, MARC21)– cross-collection (e.g. Dublin Core)
• Provide mappings: – across schemas– ontologies (ad hoc or standard CDOC-CRM)
• Discard metadata for IR (Koolen et al., 2007)
• Exploit metadata for IR (Zhang&Kamps, 2009)
ACL/LaTeCH-Portland, June 24th 2011
The IISH EAD dataset• EAD: XML standard for encoding archival
descriptions
• Challenges: – Variety of languages used– Varying type and amount of information– Style: enumerations, lists, incomplete
sentences
ACL/LaTeCH-Portland, June 24th 2011
Motivation & Objectives• Improved search and retrieval
– content-based metadata document clustering
– content-based/semantic search– support exploratory search– link across collections, metadata formats &
institutions– create unified metadata knowledge
resources
ACL/LaTeCH-Portland, June 24th 2011
Method overview
ACL/LaTeCH-Portland, June 24th 2011
Method overview
ACL/LaTeCH-Portland, June 24th 2011
Pre-processing• EAD/XML element selection & extraction
– EAD elements containing free-text & archive content information
• Language identification (n-gram method)– Identifier trained on Europarl corpus
• Text snippets length: ~20 tokens
ACL/LaTeCH-Portland, June 24th 2011
Snippet length based on language
ACL/LaTeCH-Portland, June 24th 2011
Method overview
ACL/LaTeCH-Portland, June 24th 2011
Method overview
ACL/LaTeCH-Portland, June 24th 2011
Enrichment & Structuring• Topic detection: Automatic term
recognition using C-value method
• Agglomerative hierarchical term clustering:– complete, single & average linkage criteria– document co-occurence & lexical similarity
measures
ACL/LaTeCH-Portland, June 24th 2011
Method overview
ACL/LaTeCH-Portland, June 24th 2011
Method overview
ACL/LaTeCH-Portland, June 24th 2011
Term results (auto eval)
ACL/LaTeCH-Portland, June 24th 2011
Results• C-value best performance: candidates that
occur as non-nested at least once
• Average linkage criterion & Doc Co-occurence: provide broader and richer hierarchies
Questions?Check-out our poster!
