Embed Size (px)
Citation preview
Generating Metadata by
MachineBEA 2015Friday, May 29, 11:30-12:20Room 1E10
Presenters
Moderator• Pat Payton, Senior Manager Publisher Relations, Bowker
Speakers• Randi Park, Publishing Officer, The World Bank• Hassan Zaidi, Digital Publishing Officer, International Monetary Fund• Jim Bryant, CEO, Trajectory Inc.
3
Terminology• Automated or Machine Indexing
– Process of assigning index terms against a set vocabulary or taxonomy without human intervention
– Full text or bibliographic records– Multiple vocabularies/rule sets allow for complex text
analysis• Optical Character Recognition (OCR)
– Machine conversion of an image to text– PDF of book content
• Extensible Markup Language (XML)– Set of rules for encoding documents– Both machine readable and human readable
ABOUT THE WORLD BANK
5
• The World Bank Group is the world’s largest source of funding and technical assistance for developing countries.
• Through its five institutions, the Bank Group partners with developing countries to reduce poverty, increase economic growth, and improve the quality of life.
• Comprised of 188 member countries with offices in 120 countries around the world.
around the world.
Our Twin GoalsEnd Extreme Poverty within a Generation &
Boost Shared Prosperity
Like other publishers in some respects but . . .
• Publishing arm of a larger institution, with institutional imperatives
• Open accesso Dissemination trumps revenue
• Research is performed by in-house economists and experts in other fields, by development practitioners working on the ground, and by external contributors.
• Our publishing outputs are meant to enrich the development debate, inform policies, and support the development goals of our client countries.
We are a “Knowledge Bank”The World Bank is the largest source of development
knowledge
Popular Annuals and Flagships
8
Two platforms: The World Bank eLibrary and the Open Knowledge Repository (OKR)
Mobile applications
Topics we cover = 29
• Plus 5 Regions, Countries and Keywords
Metadata strategy
Primary Purpose• Supports user-centered
discovery in WB electronic products
• Semantic fields often exposed and browseable
• Complimented by full text search and filtering
• Book, chapter and article level abstracts, topics, regions, countries, keywords
• Books do not inherit chapter semantics
Secondary Re-purpose• Search and discovery
services• Aggregators• Retail sales channels, both
print and electronic
Our experience with machine generated metadataSet up• Customized our enterprise system as much as was practicalPros• Reasonable solution when
there is a huge corpus• Fast throughput• Inexpensive to run after labor-
intensive set up• PDF source for extraction of
topics, subtopics, countries, regions, keywords
• XML output easily transformed
Cons• Set up effort/cost• Inconsistent use of keyword
terms, depending on how they were used in the text anti-corruption/anticorruption
decision-making/decision makingpolicy-making/policy making
• Abstracts must be written by humans
• False hits due to footnotes, references, names, etc..
Present workflow – human generated
Pros• Book and chapter level
including abstracts• Able to manage keyword
vocabulary using pick-lists with additions as needed
• More accurate, author provides book level draft, EP team does sense check
• New rules and terms can be added any time with little set- up
Cons• Cost per book/chapter• Capacity• Inconsistencies between
legacy (edited machine-generated) and newer content to be addressed
• Single version of keywords may not be ideal for all channels (ie more keywords for discovery services)
Future
• Interested in using technology to improve discovery for direct users and in discovery services
• Full text XML and ePub available for indexing• Institutional need to implement new taxonomy
and full text search for over 200k documents
Introduction: IMF Publications
Objectives: Establish digital publishing program 2010-2011
• New IMF eLibrary• Digital distribution• Digital production• New metadata management system• Create metadata to a granular level (chapters and articles)
***
Digitization and Metadata Challenges2010-2011
Digitization and Metadata Challenges: 2010-2011
New Challenges – New Solutions
Manual vs. Machine
•Metadata quality•Time factor•Cost of labor comparison
Challenge: Cataloging to a granular level (keywords, countries, topics and sub-topics)
New challenges – New solutions
Do the MathIMF example: • 12, 000 titles containing 60,000 chapters/articles (assumes an average of 5 per title),• 15 minutes to catalog each chapter/article with keywords etc,• 15,000 hours/40 (per week) hours =375 weeks • 375 weeks/52 = 7 years of work for one cataloger.
If you pay just $30 per hour to a cataloger, the overall cost would be $450,000. Not to mention new content is being created daily.
Automation allows us to slash the time it takes to catalog our content, saving us time and money.
Machine in Action
Machine in Action
Machine in Action
Results on eLibrary
Super keywords or specific subjects
IMF eLibrary (http://elibrary.imf.org)Browsing the IMF eLibrary
Browsing by Countries
Browse by TopicsBrowsing the IMF eLibrary
Simple Search - Type a word or phrase into the search bar at the top of every page…
Searching on eLibrary
…or Advanced Search allows multiple concepts and filters
Search Results Page
Search within results to search within publications using a single word or phrase.
Select Content Type (Books and Journals/Chapters and Articles), Countries/Region, Topics, Languages, or Date.
Type a word in the Starts with box to go to the first title that begins with the word.
Sort by Title, Date, Source or Author.
Change the number of Items per page.
Keywords
Publication Landing Page
Read on screen in HTML
Read on a variety of devices
Citation tools
Click on a title from the results page to go to the publication landing page.
Publication Landing Page
Related documents
Publication Landing PageRelated documents
Digitization and Reuse of Contents
Digitization and Reuse of Contents
Outcomes• New IMF eLibrary was delivered in March 2011• Digital distribution: Distribute IMF contents to 35
channels in various digital formats• Digital production: Have an established workflow to
generate XML based contents, ePubs, Mobi and PDF ebooks
• New metadata management system. MetaLogic is a full functioning metadata management system
• Create metadata to a granular level (all chapters and articles have individual ) ***
Thank you
™THIS INFORMATION IS PROVIDED IN CONFIDENCE AND MAY NOT BE DISCLOSED TO ANYTHIRD PARTY OR USED FOR ANY OTHER PURPOSE WITHOUT THE EXPRESS WRITTEN PERMISSION OF TRAJECTORY, INC.
Generating Metadata By MachineBEA May 29, 2015 11:30 – 12:20
™THIS INFORMATION IS PROVIDED IN CONFIDENCE AND MAY NOT BE DISCLOSED TO ANYTHIRD PARTY OR USED FOR ANY OTHER PURPOSE WITHOUT THE EXPRESS WRITTEN PERMISSION OF TRAJECTORY, INC.
Natural Language Processing: Processing & Analysis
39
Natural language analysis tools process English language text input, transforming each sentence into data that can be used for search and analysis.
Identify the base forms of words.Identify parts of speech.Identify names of companies, people, places, etc.Describe the structure of sentences in terms of phrases and word dependencies.Indicate which noun phrases refer to the same entities.
™THIS INFORMATION IS PROVIDED IN CONFIDENCE AND MAY NOT BE DISCLOSED TO ANYTHIRD PARTY OR USED FOR ANY OTHER PURPOSE WITHOUT THE EXPRESS WRITTEN PERMISSION OF TRAJECTORY, INC.
Attributes/Entities that Characterize A Book
40
™THIS INFORMATION IS PROVIDED IN CONFIDENCE AND MAY NOT BE DISCLOSED TO ANYTHIRD PARTY OR USED FOR ANY OTHER PURPOSE WITHOUT THE EXPRESS WRITTEN PERMISSION OF TRAJECTORY, INC.
Sentiment: Analyzing the Words Within the Book
“Outstanding” words (5) breathtaking, thrilled, superb
hell, rape, (more unmentionables)
“Catastrophic” words (-5)
torture, fraud, (unmentionables)“Damned” words (-4)
woeful, worsen, kill“Terrible” words (-3)
worthless, travesty, threaten“Upset” words (-2)
numb, provoke, pushy“No” words (-1)
validate, safe, adequate“Yes” words (1):
strengthen, rich, funky“Welcome” words (2)
praise, marvelous, impressive
winning, stunning
“Happy” words (3)
“Wow” words (4)
41
Each word is given a numeric value based on its subjective meaning.
“Positive” words range on a positive scale; “Negative” words range on a negative scale.
Trajectory’s Analytics Engine uses these values to compute the book’s sentiment curve across sentence, paragraph, chapter and entire book.
This sentiment “fingerprint” at an aggregate level yields a unique picture of the book.
™THIS INFORMATION IS PROVIDED IN CONFIDENCE AND MAY NOT BE DISCLOSED TO ANYTHIRD PARTY OR USED FOR ANY OTHER PURPOSE WITHOUT THE EXPRESS WRITTEN PERMISSION OF TRAJECTORY, INC.
Sentiment: Analyzing the Words Within the Book
42
™THIS INFORMATION IS PROVIDED IN CONFIDENCE AND MAY NOT BE DISCLOSED TO ANYTHIRD PARTY OR USED FOR ANY OTHER PURPOSE WITHOUT THE EXPRESS WRITTEN PERMISSION OF TRAJECTORY, INC.
Sentiment: Analyzing the Words Within the Book
43
™THIS INFORMATION IS PROVIDED IN CONFIDENCE AND MAY NOT BE DISCLOSED TO ANYTHIRD PARTY OR USED FOR ANY OTHER PURPOSE WITHOUT THE EXPRESS WRITTEN PERMISSION OF TRAJECTORY, INC.
Trajectory Index
44
™THIS INFORMATION IS PROVIDED IN CONFIDENCE AND MAY NOT BE DISCLOSED TO ANYTHIRD PARTY OR USED FOR ANY OTHER PURPOSE WITHOUT THE EXPRESS WRITTEN PERMISSION OF TRAJECTORY, INC.
Keyword Analysis and Comparison
45
™THIS INFORMATION IS PROVIDED IN CONFIDENCE AND MAY NOT BE DISCLOSED TO ANYTHIRD PARTY OR USED FOR ANY OTHER PURPOSE WITHOUT THE EXPRESS WRITTEN PERMISSION OF TRAJECTORY, INC.
Keyword Translation into Local Languages
46
™THIS INFORMATION IS PROVIDED IN CONFIDENCE AND MAY NOT BE DISCLOSED TO ANYTHIRD PARTY OR USED FOR ANY OTHER PURPOSE WITHOUT THE EXPRESS WRITTEN PERMISSION OF TRAJECTORY, INC.
Recommendations
47
™THIS INFORMATION IS PROVIDED IN CONFIDENCE AND MAY NOT BE DISCLOSED TO ANYTHIRD PARTY OR USED FOR ANY OTHER PURPOSE WITHOUT THE EXPRESS WRITTEN PERMISSION OF TRAJECTORY, INC.
Thank You
48
2015 BEA – BOOTH 1347
United States:50 Doaks LaneMarblehead, Massachusetts01945 United [email protected]
China:No. 3, 8 Chuang Ye RoadHaidan District,Beijing, China 100085
Q & AGenerating Metadata by
MachineBEA 2015Friday, May 29, 11:30-12:20Room 1E10
