EECS 395495: Web Information Retrieval and Extraction

1
EECS 395/495 Web Information Retrieval and
Extraction

• Spring 2008

2
Outline

• Introductions
• Course goals and logistics
• Why take this class?
• What is Web Search?
• What will it be in five years? Ten years?

3
Introductions

• Professor Doug Downey
• TA Junsong Yuan

4
Goals

• How does Web search work?
• First half of course
• Whats the future of Web search?
• Second half of course

5
Logistics

• First half the class lectures
• Second half presentations of research papers
• Groups lead discussions
• Each individual presents a handful of slides
• See Web page and sign up (e-mail TA) this week

6
Logistics

• Grading
• Participation (30)
• During lectures/discussions (10)
• Leading a discussion (15)
• Search feature pitch (5)
• Projects (70)
• Project grade, based on individual contribution
  (60)
• Review of another teams project (10)

7
Projects

• Groups of 2-3 (not necessarily discussion
  groups)
• Examples
• Read 3-5 recent research papers and summarize the
  state of the art in some area of IR/IE
• Implement an IR/IE system and report on the
  results
• Answer theoretical questions
• Etc.

8
Specific Examples

• Systems
• Automatically place blogs on liberal/conservative
  continuum
• Create a search knob for specifying reading
  level
• Execute relevant background searches as I write
  an e-mail
• Theoretical questions
• Read a paper on search ad auctions or PageRank
  and attempt to extend the results
• Summary of Research
• Read 5 papers on automated question answering
  summarize the field and suggest future directions

9
Project Milestones

• April 8 Proposal (1 page)
• Meetings with me to finalize April 9/10
• lots of progress
• May 5 Report of preliminary results (3 pages)
• May 7 Review group provides feedback (1 page)
• June 2 Final Report (4 pages)
• June 4/June 8 (finals week) Final Presentations
  (10 mins 5 min for QA)

10
Search Feature pitch

• This Thursday (!)
• Individually deliver a 2.5 minute pitch of a
  new search engine feature
• Plausible, not necessarily possible(see example
  later)
• Send me your ppt/pdf slides by Thursday at 9AM
• Class discusses for one minute while next
  presenter sets up

11
Outline

• Introductions
• Course goals and logistics
• Why take this class?
• What is Web Search?
• What will it be in five years? Ten years?

12
Who cares about search?

• "The most important application for the
  foreseeable future...is search.
• Steve Ballmer, CEO of MicrosoftFinancial Times
  http//news.cnet.com/8301-10784_3-9973650-7.html
• Why?
• Searchs utility scales as the Web scales
• People use it all the time
• Control
• Profit

13
Why should you care about search?

• Opportunity
• Fascinating important enabling technologies
• Scaling
• Machine Learning/Data Mining
• Graph-based algorithms
• Language Understanding
• Auction theory
• User Interfaces

14
Graph-based Algorithms

• Example PageRank
• Googles original claim to fame
• Idea Quality of p is proportional to the
  aggregate quality of the pages linking to p
• PageRank(p) probability that a random surfer
  lands on p
• Pick a starting page at random
• Follow links uniformly at random
• Every now and then, jump to a random page
• PageRank(p) proportion of visits to p

15
PageRank example
15 probability of a random jump
C
B
A
F
D
E
16

• Wikipedia.com

17
How to compute PageRank

• Simulate a random surfer?
• On 20 billion pages and 400 billion hyperlinks
• It can be done
• Youll learn how
• What about
• Personalized PageRank? Link spam?

18
What do you need to take this class?

• EECS 311
• basic understanding of algorithms and data
  structures
• Basics of linear algebra and probability theory
• Tolerance for non-linearity
• Willingness to participate

19
Outline

• Introductions
• Course goals and logistics
• Why take this class?
• What is Web Search?
• What will it be in five years? Ten years?

20
Web Search today

• Performs an easy task
• Extremely quickly
• At massive scale
• Relatively well

21
Easy?

• Belief Web Search engines have to understand my
  query and find a needle in a haystack of 15
  billion documents!
• Reality Most search queries are
• short (avg. 2.5 words 2005)
• satisfied by pages from a small subset of the Web
• Millions rather than billions Mei et al., 2008
• gt More like finding a pencil in a haystack!

22
Extremely Quickly

• Results returned in lt 1 sec
• For almost any query
• For any engine
• It was not always thus!
• 3-4 seconds in early days (Chu Rosenthal,
  1996 Garratt et al. 2001)
• How? Inverted indices, enormous data centers,
  clever algorithms

23
Users are quick too
Downey, Dumais, Horvitz 2007
24
At Massive Scale

• (in millions)
• How? Inverted indices, enormous data centers,
  clever algorithms

http//blog.searchenginewatch.com/comscoresearchsh
arefeb2009_0309.jpg
25
Relatively Well
26

• Note, for images missing from this online
  version see links at bottom of page

http//www.seoresearcher.com/distribution-of-click
s-on-googles-serps-and-eye-tracking-analysis.htm
27
Web Search Engines today

• For the most part
• Perform an easy task
• Extremely quickly
• At massive scale
• Relatively well
• The future of Web search is in more difficult
  tasks

28
Trend toward more difficult queries

• Search queries are getting longer.

http//www.readwriteweb.com/archives/hitwise_searc
h_queries_are_getting_longer.php
29
Things you cant do with search today

• Query by description rather than content
• Humorous anecdotes about Perry Farrell
• Extracting and synthesizing over multiple pages
• Nanotechnology companies hiring on the West Coast
• Substances the FDA has banned
• Organizing bodies of documents
• Show me the most compelling cases for and against
  the recent stimulus package
• Who says drinking from aluminum soda cans
  increases my Alzheimer's risk? Should I believe
  them?

30
Image Search for famous pink building
31
Search Feature Pitch

• Guidelines
• Introduce yourself (name, major, degree, year)
• Whats the problem
• Whats your solution
• What makes it feasible
• 2 1/2 minutes isnt a lot of time!!

32
Mechanism for over-specified queries

• Florida gators football depth chart 2002
• 0 results
• Florida gators football depth chart 2002
• Many results -- none correct in top 10
• After protracted manual reformulation process
• 2002 gator football media guide
• 8 results, top several correct

33
Proposed Solution

• User submits over-specified query
• one with zero hits, e.g.Florida gators football
  depth chart 2002
• Search engine
• Tries moving quotes, substituting phrases, etc.
• Florida gators football depth chart 2002
• Florida gators depth chart 2002 pigskin
• After a few minutes, returns summary of results

34
Why plausible?

• Substitutions can be obtained from thesauri or
  statistical co-occurrence, e.g.
• P(depth chart football media guide) is
  largegt try substituting football media guide
  for depth chart in query
• Given a few minutes, we can scalably execute
  several hundred candidate queries

35
Reminder

• Your search engine feature pitch
• Presented this Thursday in class
• Send ppt or pdf to Junsong I by 9AM Thurs
• See course Web page
• Also
• Start forming project groups
• Sign up for team mtg with me next Thurs/Fri
• Look for discussion papers dates later today

36
(No Transcript)