Dr. Susan Gauch

1
Intelligent Access to Time-Sensitive Information

• Dr. Susan Gauch

2
Motivation

• Soldiers in the field need effective access to
  time-sensitive information
• Need to be able to make new information available
• Quickly
• Effectively
• Accurately
• Need to be able to validate quality of
  informally-collected information
• Need to tailor information to soldiers mission
  and expertise

3
Phase I Goals

• Goal 1 Make informally collected information
  available quickly
• Develop automatic techniques to harvest informal
  information sources
• Develop automatic techniques to identify
  content-bearing chunks
• Develop automatic techniques to segment and tag
  chunks to produce learning objects
• Develop a Web accessible search site for the
  learning objects

4
Phase I Goals

• Goal 2 Improve accuracy of informal information
• Develop manual editing system to improve/extend
  tagging of learning objects
• Develop manual annotation system to allow
  end-users to add comments, corrections, ratings
• Develop new search algorithms that preferentially
  select highly-rated items

5
Phase II Goals

• Goal 1 Tailor the information to the individual
• Develop automatic techniques to implicitly create
  profiles for the users of the system
• Incorporate explicit information about the users
• Develop automatic techniques to select most
  relevant information for the user
• Develop automatic techniques to tailor
  information presentation to the user

6
Phase II Goals

• Goal 2 Improve the usefulness of the
  information
• Create a system to analyze the annotations to
  identify objects requiring editing
• Create an alert system to automatically notify
  interested parties about urgent new information,
  changes to information they contributed

7
(No Transcript)
8
Related Projects

• IKME
• Intelligent Knowledge Management Archive
• Goal Given XML-tagged content
• Provide automatic data cleanup
• Track schema changes
• Migrate objects to new schema
• Provide Web-accessible searching
• XML database
• Added full-text relevance ranking

9
IKME (contd)

• Provide Web-form to reuse learning objects
• Manuals
• Lesson objects
• Identify similar learning objects
• Automatically cluster related objects together

10
ChatTrack

• Goal Intelligent Access to Chat data
• http//www.ittc.ku.edu/chattrack
• 3 components
• Archiving
• Profiling
• Searching
• XML-markup of chat messages
• Relational database for fast access

11
Architecture
Client
Internet
Chat DataArchive (XML/SQL)
ConceptDatabase
IRCClient
IRCClient
Chat Server(with ChatLog)
Classifier
Indexer
Administrator /Intelligence Agent
ChatProfile
ChatRetrieve
12
Chat Archiving (2)
ltjoingt ltdategt2004-04-17lt/dategt
lttimegt085750lt/timegt ltusernamegtjmblt/username
gt ltchannelgtjayhawklt/channelgt lt/joingt
ltmsggt ltdategt2004-04-17lt/dategt
lttimegt085814lt/timegt ltsendergtJasongtlt/sendergt
ltreceivergtjayhawklt/receivergt
ltdatagtThere is going to be weather, whether or
not. Uh oh, Ill be RIGHT back!lt/datagt lt/msggt
ltquitgt ltdategt2004-04-17lt/dategt
lttimegt085819lt/timegt ltusernamegtjmblt/username
gt lt/quitgt ltmsggt ltdategt2004-04-17lt/dategt
lttimegt085946lt/timegt ltsendergtAlicelt/send
ergt ltreceivergtjayhawklt/receivergt
ltdatagtPoof..Left in a hurry! Must be a tornado
outside his door or something. lollt/datagt lt/msggt

• XML chat data ? SQL database

• ChatLog Library XML schema can be used for
  almost any client/server-based system

13
Chat Profile

• User Profile focuses on one chat participant
• Session Profile filters by chat room name only
• Analyst selects criteria of interest

14
Chat Profile

• Chat data collected from archive (stopword
  removal Porter stemming)
• Classification performed once chat utterances
  collected
• Classifier creates vector of keywords from chat
  data
• Similarity measure
• Determines similarity between chat data vector
  vectors for each trained concept
• Concepts sorted top matches returned
• Represented in hierarchical fashion
• Asterisks represent relative concept weights at
  each level

15
Classification in american-politics(two hours
in January 2004 Undernet) SESSION PROFILE
16
Classification in american-politics(two hours
in January 2004 Undernet) USER PROFILE
selected one public chat participant
17
Chat Retrieve

• Some profiles warrant further analysis
• Agent/admin needs ability to explore chat
  session linked to profiles in question
• Incremental Indexing system keeps data current

18
Chat Retrieve (3)

• Queries based on
• speaker name
• keywords
• date/time range
• listener name
• (combo of above)
• Keyword retrieval based on tf idf

19
Chat Retrieve (4)

• Selecting chat room name replays chat session
• Includes utterances spoken by all participants
• Tracks all chat room participants
• Even if they do not contribute to sessions

20
Public Demo http//www.ittc.ku.edu/chattrack

• ChatTrack Provides agents with new tools for
  vigilance against crime
• ChatProfile generates conceptual profiles from
  chat data
• Reduces manual efforts to classify chat sessions
• ChatRetrieve facilitates manual analysis for
  session retrieval by agents, administrators,
  parents
• Filter based on listener/sender, date/time, chat
  room session name

21
Beta Version ChatTrend
22
Future (ChatTrend 2)
23
Search Engines Today

• Common problems of search engines
• ambiguity ( e.g., rock, salsa )
• retrieved results are based on web popularity
  rather than user's interests
• Goal
• Improve search accuracy by retrieving by concept
  (e.g., music, dance)
• Improve search accuracy by matching user
  interests

24
Search Engine Personalization

• Ongoing research to investigate ways
• to implicitly collect information about the user
• to represent information about the user.
• Use user profiles
• to re-rank the results returned from the initial
  retrieval process
• to filter results that better fits user's
  interests

25
Sources of User Information

• User explicit information
• - users too lazy
• - information becomes out of date/inaccurate
• User browsing histories
• - must collect information via desktop robot or
  have user connect to Internet via a proxy
• User desktops
• - contextual retrieval
• User search histories
• - information available to search engine itself

26
User Profile Creation

• Collect information about the users interests
  (search history)
• Categorize representative texts into concept
  hierarchy
• - use Open Directory Project for concepts
• - train classifier on representative pages
• - compare representative texts to training
  texts to identify the concepts discussed
• Concept weights represent user interests

27
Personalizing Search Results

• Submit query to Internet search engine (e.g.,
  Google)
• Categorize each result into same concept
  hierarchy (e.g., ODP) to create result profiles
• Conceptual match is calculated based on
  similarity between result profiles and user
  profile
• Rerank results based on conceptual match
• - rank order produced called conceptual rank

28
Summary

• Built user profiles based on queries submitted
  and snippets of user-selected results.
• This information was sufficient to build user
  profiles that were able to significantly improve
  personalized rankings.
• A query-based profile produced an improvement of
  33.
• A snippet-based profile produced an equivalent
  improvement of 34.
• http//www.ittc.ku.edu/mirco/demo/

29
Conclusions

• Search engines can capture information submitted
  to their site in order to create personalized
  search.
• Users need not install proxy servers or desktop
  bots.
• Privacy issues arise with any personalized
  service.
• Need to look at combination of short-term,
  long-term user interests with current task focus.
• http//www.ittc.ku.edu/mirco/demo/search.php