XDMR Project Meeting, July 12, 2005

1
Mapping Metadata Across Genres and Languages

• XDMR Project Meeting, July 12, 2005
• Fredric C. Gey, University of California,
  Berkeley
• Colleagues Michael Buckland, Aitao Chen, Ray
  Larson
• Programmers and Students Barbara Norgard,
  Natalia Perelman
• Work performed under
• DARPA research contract Search support for
  unfamiliar metadata vocabularies (1997-2001)
• DARPA TIDES research grant Translingual
  Vocabulary Mappings using Domain Ontologies
  (1999-2004)
• Institute for Museum and Library Services Grant
  Seamless search of textual and numeric databases
  (1999-2001), others (2002-2006)

Fredric C. Gey
2
HETEROGENEOUS DIGITAL INFORMATION SEARCHCurrent
Search Technology (multiple independent searches
without search aids)
Bibliography Full Text
QUERY
Maps and other Geospatial data
Music and other media
3
SEARCHING UNFAMILIAR METADATA PROBLEM STATEMENT

• Numerous databases are indexed by structured
  metadata classifications
• Classification schemes are highly specialized
• As digital libraries multiply in size and
  diversity, there is a need for search engines for
  non-specialists
• Search engines should translate from ordinary
  language to specialized classifications

In U.S. Import-Export Database
computer
No result found
4
Searching Statistical Information Problem
statement

• Numeric statistical information is often thinly
  documented or documented with specialized
  vocabularies
• Census welfare --gt public assistance
• Foreign trade harmonized commodity
  classification computer --gt digital adp machine
• Standardized Industrial Classification (SIC
  codes) automobile --gt motor vehicle
• In search, the users ordinary language term is
  unlikely to match the limited technical
  vocabulary used to document the statistical
  resource
• Can we remedy this situation?

Fredric C. Gey
5
Searching Statistical Information Evidence Poor
Data

• Compared to searching textual databases, numeric
  statistical information and its terminology is
  both evidence-poor and highly technical
• Statistical databases lack the rich set of
  textual clues which can identify items of numeric
  information
• However, if we can find a textual resource
  associated with the numeric metadata, we may be
  able to mine the text to improve numeric search.
• This is possible for some numeric classification
  schemes

Fredric C. Gey
6
SEARCHING UNFAMILIAR METADATA EXAMPLES FROM
FOREIGN TRADE IMPORTS-EXPORTS

• U.S. Foreign Trade Imports-Exports
• Two CD-ROMS/month
• Classified by 16,000 numeric commodity codes in
  12 digit hierarchy
• Search for automobile or computer yields no
  result
• Search for Car yields Railway or Tramway
  Stock, Etc.
• Need to know classifications Passenger Motor
  Vehicle or Digital ADP Machine w/CPU

In U.S. Import-Export Database
Car
Railway or Tramway Stock
7
SEARCHING UNFAMILIAR METADATA U.S. STANDARD
INDUSTRIAL CLASSIFICATION SYSTEM

• U.S. Standard Industrial Classification System
  (SIC)
• Used to classify and aggregate industrial
  activity in the U.S.
• Codes defined by Office of Management and Budget
• County Business Patterns reports annual
  employment, payroll, firm size by county, SIC
  code

In U.S. SIC System
Lobster
Nothing found
8
Searching Statistical Information Economic
Classification Codes

• Standard Industrial Classification (SIC) and
  North American Industrial Classification (NAICS)
  codes have been used to index trade magazines
• This provides a textual resource of hundreds of
  thousands of documents and millions of words
  associated with the numeric data.
• Thus mappings can be made between the words from
  the magazines abstracts and the classifications
• User queries can be matched against words and
  phrases most closely associated with the
  particular numeric data classification
• A ranked list of classifications can be displayed
  to the user in order to improve the search
• Harmonized commodity classifications can be
  searched using SIC codes as a search proxy

Fredric C. Gey
9
SEARCHING UNFAMILIAR METADATASOLUTION -- ENTRY
VOCABULARY TECHNOLOGY

• Maps between ordinary language and specialized
  classifications
• Implemented using text categorization techniques
• Requires training collections which have been
  manually indexed
• Preserves and leverages investment in creation of
  complex classification structures
• Can also be applied to the task of multi-lingual
  information access

In U.S. Patent Database
For Automobile
Try 180/280
10
UNFAMILIAR METADATA PROJECT OVERVIEW

• Provides entry vocabulary modules for searching
  unfamiliar metadata (text, Patents, Standard
  Industrial Classification)
• Fundamental method uses Bayesian inference based
  upon training data (documents) to map between
  ordinary language and complex hierarchical
  metadata structures
• Technique can be applied to searching poorly
  documented statistical information

Unfamiliar Metadata
11
SEARCHING UNFAMILIAR METADATA ENTRY
VOCABULARIES CONSTRUCTED

• Ordinary language to U.S. Patent Classification
• Ordinary language to INSPEC thesaurus terms
• Ordinary language to Library of Congress
  classification codes
• Ordinary language to Standard Industrial
  Classification system
• Experience in entry vocabulary technology
  (training set size, data cleaning, natural
  language)
• Ordinary Language to NAICS with link to 1997
  Economic Census

In U.S. SIC classification
For Lobster
Try Shellfish
12
Translingual Vocabulary Mappings Using Digital
Library Catalogs(Berkeley work with Aitao Chen,
Ray Larson, Michael Buckland)

• Applies digital library multilingual catalogs
  (books, maps, pictures) in 400 languages to
  develop multilingual vocabulary lexicons
• 8,000 Afghan(Pushto) words mapped to English in
  30 minutes
• 20,000 word corpus of Afghan(Pushto) text in 3
  hours
• Records contain one or more sentences from a
  particular language
• Each is categorized by human indexers into
  standardized schemes
• Library of Congress Subject Heading (Islamic
  Fundamentalism)
• International standard (MARC) dataset formatting
• Prototype implemented using 12 million records
  from the University of California electronic
  catalog (http//melvyl.ucop.edu)

13
Search SUBJECT Islamic Fundamentalism and
LANGUAGE Arabic
Yield 119 Arabic language samples on topic
Islamic Fundamentalism
14
BACKGROUND ON DIGITAL LIBRARY CATALOGS

• Library catalogs are being automated at a furious
  pace worldwide
• Library objects (books, maps, pictures) in 400
  languages
• Records contain one or more sentences from a
  particular language
• All categorized by specialists into standardized
  schemes
• Library of Congress Subject Heading (Islamic
  Fundamentalism)
• Library of Congress Classification (BP60, BP63,
  KF27)
• Dewey Decimal Classification (297.2, 306.6,
  320.5)
• International standard (MARC) dataset formatting
• 1000 (est.) remotely searchable catalogs
  worldwide accessible using the international
  search/retrieve protocol Z39.50

15
WHAT CAN DIGITAL LIBRARY CATALOGS PROVIDE?

• Millions of sentence fragments (titles). Many
  languages
• Sentences with precise topical content identified
• From 150,000 Library of Congress Subject Headings
• Subject headings form transfer point
  (interlingua) between English topics and words in
  other languages
• Can be used for
• Query translation in cross-language information
  retrieval
• Creation of bilingual dictionaries

16
BERKELEYS TECHNOLOGY VOCABULARY INDEXES

• Takes large training sets of digital objects
  (book/article titles)
• Software/algorithms perform maximum likelihood
  mappings between words/phrases and subject
  headings
• Search software accepts words and returns subject
  headings and vice versa
• Prototype search software maps from/to English
  and
• Arabic, Chinese, French, German
• Italian, Japanese, Russian, Spanish
• currently uses romanized search for non-roman
  scripts
• currently phrases in English only
• Language-specific and/or domain-specific
• Our prototype for nine languages available at
• http//otlet.sims.berkeley.edu/mulevm2.html

17
BERKELEY TRAINING SET AND PROTOTYPE

• University of California digital library catalog
• Private copy, 10 million records (5 million
  non-English)
• Records in over 100 languages
• Obtained in MARC database standard format
• Foreign language titles use Library of Congress
  transliteration (Romanization) standard
• Prototype search software maps from/to English
  and
• Arabic, Chinese, French, German
• Italian, Japanese, Russian, Spanish

18
Example Library of Congress Subject Heading
Islamic Fundamentalism yields most closely
associated words in multiple languages
19
LIBRARY RESOURCES vs. WEB RESOURCES

• LIBRARY CATALOGS
• Databases in standard formats
• Data tagged with rich metadata
• Database size limited
• Data content identified according to well-defined
  rules

• WEB PAGE MINING
• Data in quite varied formats
• Limited or non-existent metadata
• Database size potentially massive
• Content rarely identified

20
Foreign search words are mapped to English
subject headings
Russian-German bilingual dictionary entries
Perevod Ubersetzung Ubersetzung -- Perevod
21
Translingual Vocabulary Indexes, Place Names,
Stemming CLIR for the TIDES Surprise Language
Exercise Fredric Gey, Aitao Chen, Vivien Petras
(University of California, Berkeley)

• Library catalog records
• Hindi 72,967 records (44,366 Berkeley,
  28,601 UK)
• Urdu 30,206
• Bengali 23,430
• Tamil 20,232
• Cebuano 16 (300 est. worldwide)
• Day 1 English-Hindi vocabulary index
  http//metadata.sims.berkeley.edu/Evms/cdl2/hindi2
  /weblcsh_hindi.html
• Terrorism --atankavada ???????
• Day 5 10,839 English words mapped to
  (transliterated) Hindi (only 492 unique)
• Unable to resolve LOC back transliteration
• Day 2 178 city names from Indian Census, hand
  translated to Hindi by ISI, lat/lon by Beth
  Sundheim of SPAWAR.
• Day 8 5,161 city and village names
• Day 3 Hindi Monolingual UTF-8 Retrieval
• Day 6 added ITRANS transliteration
• Day 18 English-Hindi CLIR system

• Day 24 Berkeleys statistical stemmer clusters
  Hindi words to English stems
• Derived from the IBM/ISI lexicons
• CLIR evaluation participation
• 10 runs submitted, all automatic
• Results very satisfactory
• Also Geographic visualization for stories
  associated with cities
• Lessons learned
• World-wide library retrieval (using Z39.50
  protocol) can increase coverage
• Data limited by language density, timeliness
• Back-transliteration of scripted languages needs
  further research
• Geographic visualization useful

22
HETEROGENEOUS DIGITAL INFORMATION SEARCHEnhanced
Search (augmented with Entry Vocabulary Module
(EVM) Technology)
EVMp
EVMs
EVMt
QUERYplus
EVMg
EVMm
QUERY
23
HETEROGENEOUS DIGITAL INFORMATION SEARCHDirect
Mappings and Search Between Multiple Information
Types
EVMs
EVMp
EVMt
QUERYplus
EVMg
EVMm
QUERY
24
Geographic Visualization of Hindi
Stories(Electronic Cultural Atlas Initiative,
University of California)
25
TWO RELEVANT STANDARDS under development

• SDMX Statistical Data and Metadata Exchange
• EC,
• ISO Standard, Technical Committee 154
• ISO/TS 173692005 SDMX
• www.sdmx.org
• DDI Data Documentation Initiative
• XML standard for documenting statistical dataset
  for archiving
• International committee of statistical data
  archivists
• Coordinated by national archive at University of
  Michigan
• Version 2.0 approved May
• Working on V 3.0 to include detailed geography
• www.icpsr.umich.edu/DDI

26
SEARCHING UNFAMILIAR METADATA Web Site and PI
Contact Information

• www.sims.berkeley.edu/research/metadata
• www.sims.berkeley.edu/research/seamless
• Michael Buckland and Ray Larson (buckland, ray
  _at_sims.berkeley.edu)
• Fredric Gey (gey_at_berkeley.edu)