Providing Expert Advice by Analogy for OnLine Help

1
Providing Expert Advice by Analogy for On-Line
Help

• Henry Lieberman and Ashwani Kumar
• Media Laboratory
• Massachusetts Institute of Technology
• Cambridge, MA, USA
• http//www.media.mit.edu/lieber/

2
What happens when interaction with the Web goes
wrong?

• Increasingly, people use the Web to perform
  procedures (buy things, vote, date, seek jobs) as
  well as just browse
• What happens when things go wrong?
• Debug it yourself
• Get help
• Increasingly, confidence in Web interactions
  depends on effectiveness of help

3
How do you get help?

• Look up things in problem-solution database
• Telephone support
• On-line chat

4
Elicititation Explanation

• Elicitation Helper asks user what is wrong. Get
  enough information from user to determine problem
  and choose solution.
• Explanation Helper tells user how to fix the
  problem. Explain why the solution worked, and
  how to avoid such problems in the future.

5
Mismatch between expert and novice models

• The helper is an expert the user is a novice
• Novice may lack technical vocabulary to
  understand the elicitation questions
• Novice may lack background knowledge to
  understand explanation of soultion
• Expert may be unable to empathize with novice

6
Problem/Solution Database
7
AI Expert Systems

• Expert Systems are the traditional way to encode
  expert behavior
• Knowledge Engineer experts to get their knowledge
  and procedures
• Encode Expert Model
• Deep but narrow knowledge
• If expert and novice dont share vocabulary and
  knowledge, its difficult for the novice to
  interact directly with the Expert Model

8
We need a model of novice knowledge

• How do we model the knowledge of someone who is
  not supposed to know very much?
• Traditionally, novice is modeled as subset of
  expert knowledge
• Intelligent Tutoring Systems, Qualitative
  Physics
• But how do you model novice knowledge in
  general?

9
Open Mind - Push Singh

• Common Sense statements collected from volunteer
  contributors on the Web
• 770,000 English sentences
• Conventional data/knowledge bases Do a lot of
  organizing work up front, so you can get the
  stuff out easily
• OM Get the stuff in any which way, then do a lot
  of work on the back end to interpret it
• Query expansion, Parser, Semantic net miner

10
Open Mind - Push Singh
11
How do human helpers succeed in bridging
Expert/Novice gap?

• We interviewed helpers on the America Online
  (internet service provider) help desk
• Helpers tend to provide cookbook solutions
• We asked for successful interactions where users
  came to understand the problem and solution
• Helpers best explained things by making analogies
  to everyday life situations!

12
Problem/Solution Database

• Problem My browser can see my companys home
  page, but it wont let me access my workgroups
  Web site
• Solution Procedure Check if cipher strength is
  '0. Upgrade Browser to 128 bit Encryption.
• Explanation Generally, Websites require 128-bit
  encryption in order to process information
  securely. If the cipher strength of your browser
  is inadequate, you will not get into secure
  Websites. Upgrading your browser's encryption may
  help it better handle secure Websites.
• NOTE You only need to do this when unable to get
  to secure Websites.

13
Better Explain by Analogy

• Encryption in a browser is like security
  clearance to enter a building. If you don't have
  the proper security clearance, you may be able to
  get into the building, but not into certain
  areas. You must upgrade your security clearance
  status to go further. So without the proper
  encryption, your browser may be able to access a
  website, but not log in.

14
Introducing SuggestDesk

• SuggestDesk is an agent that listens to the
  helper/user chat
• Provides suggestions to helper (unseen to novice
  user)
• Tries to recognize opportunities for making
  analogies between problem/solution DB and Common
  Sense knowledge

15
SuggestDesk
16
Example

• User My browser runs slowly
• Problem/Solution DB
• Browser might run slowly because of network
  congestion
• Browser might run slowly because the browser is
  infected by a virus

17
Elicitation Explanation

• Elicitation
• (Traffic) When did you try to log in?
• (Virus) Have you downloaded any new applications
  lately?
• Solution
• (Traffic) Try again at a later time.
• (Virus) Run an anti-virus program.

18
SuggestDesks help (Traffic)

• What makes things slow down?
• Traffic runs slowly at rush hour.
• Analogy
• The AOL service is like a road.
• The users are like cars.
• If too many users try to use the service at the
  same time, it slows down.
• The solution is to try at a time when things are
  less crowded or find an alternative route.

19
SuggestDesks help (Virus)

• What makes things slow down?
• People slow down when they are tired.
• Being sick can make you tired.
• Analogy
• A computer virus is like a biological virus
• If you have the flu, you cant do things as fast
  as you normally would
• An antivirus program is like medicine.

20
SuggestDesk Architecture

• The Natural Language Processor (NLP),
• The Commonsense Processor (CP),
• The Expert Analyzer (EA),
• The Analogy Mapping Engine (AME), and
• The Elicitation and Explanation Processor (EEP).

21
Natural Language Processing

• Via MontyLingua, Common Sense part-of-speech
  tagger

Result prep_phrases_tagged,
verb_phrases_taggedis/VBZ running/VBG,
verb_arg_structures_concise("run" "browser"
"slow"), noun_phrasesbrowser,
noun_phrases_taggedbrowser/NN, adj_phrases_tagg
edslow/JJ, verb_arg_structuresis/VBZ running
/VBG, browser/NN, slow/JJ, modifiers_taggeds
low/JJ, prep_phrases, verb_phrasesis
running, parameterized_predicatesrun,
past_tense, passive_voice, browser, ,
slow, , modifiersslow,
adj_phrasesslow
22
ExpertNet
(EffectOf 'surf internet' 'download
files') (EffectOf 'surf internet' 'download
applications') (EffectOf 'download files'
'browser cache is large') (EffectOf 'download
applications' 'browser infected by
virus') (EffectOf 'PC infected by virus'
'browser run slow')
23
Analogy Mapping Engine
Analogiescomputer, UsedFor, surf internet,
1.1887218755408673, CapableOfReceivingAction,
run slow, 1.1887218755408673, CapableOfReceiving
Action, crash, 1.1887218755408673, CapableOfRece
ivingAction, start, 1.1887218755408673, 6.188721
8755408675, car, CapableOfReceivingAction,
damage, 1.1887218755408673, CapableOfReceivingAc
tion, crash, 1.1887218755408673,
CapableOfReceivingAction, start, 1.18872187554086
73, 5.930167946706389, software, CapableOfRe
ceivingAction, run slow, 1.1887218755408673, Cap
ableOfReceivingAction, crash, 1.1887218755408673,
CapableOfReceivingAction, install,
1.1887218755408673, CapableOfReceivingAction,
install, 1.1887218755408673, 5.855516191543203
24
Weighting

• log(f0.5i4),
• where foutgoing edges and i incoming edges

25
SuggestDesk Architecture

• The Natural Language Processor (NLP),
• The Commonsense Processor (CP),
• The Expert Analyzer (EA),
• The Analogy Mapping Engine (AME), and
• The Elicitation and Explanation Processor (EEP).

26
User Testing
27
Woodstein (with Earl Wagner)

• Debugger for Web procedures
• By analogy to debugger for programming
• Provides visualization, explanation, tools for
  incremental exploration
• Self-help debugging
• Co-operative debugging between expert and novice

28
Woodstein
29
Conclusion

• The effectiveness of online help is key to the
  success of Web interactions
• Help needs to bridge the gap between expert and
  novice knowledge
• Common Sense Reasoning can help find analogies
  that allow expert and novice to communicate
• We need debuggers that help us systematically
  explore the causes of problems
• Dont worry, help is on the way!