magenta-argo by xiangpeng

VIEWS: 8 PAGES: 27

									Argo – Intelligent multi-
agent Meta-Search Engine,
based on domain ontology




                 Samara,
                  2007
        Search: Current Problems
 Advertisement          Common Users              Search Engine Owners
  Companies
  Non-focused and        Search: too many          Hard to be better than
not-integrated         garbage, too little       anyone else
service proposals      value
                                                   Lack of unique and integrated
and advertisements       Not able to adjust      services – how to attract and
  Lack of feedback     search with               keep people
from users             semantics of the
                                                   Lack of profit from
  Hard to attract      user’s knowledge
                                                 advertisements due to their
new users and          domain for individual
                                                 inefficiency and un-targeted
retain old customers   preferences
                                                 marketing
  Absence of             A lot of non-relevant
                                                   How to improve relevancy of
analysis of            results on first places
                                                 current results – they are not
interested/non-        in search engines
                                                 too good but what could be
interested user          Very complicated        better ?
groups and reasons     and unclear rules of
behind it              requests
Keywords Search: Problems

Hard to be sure that certain keyword is related
to the real meaning of the text
Hard to distinguish good results from bad if
both have same keywords
Hard to increase quality of coherent search,
because each keyword is independent
Hard to understand context of the request
A lot of garbage due to extensive usage of
keywords by different sites
            Current Situation - keywords “obsession”
                    Request → Resulting advertisement




1) “Stolen car” →
Get a Free Car!
          Current Situation - keywords “obsession”
                  Request → Resulting advertisement




2) “Family car
pictures” →
Pictures of
your
family
          Current Situation - keywords “obsession”
                  Request → Resulting advertisement




3) “Car air
conditioner” →
Hair Beauty
            Current Situation - keywords “obsession”
                    Request → Resulting advertisement




4) “Where my
wife can repair a
car? →
Save your
Marriage!
Argo – Solution
 Project Task (Argo)

Advertisement companies
 Better targeting of advertisements
 Significantly improve results on search queries
 to attract new users
Common Users
 Give ability to describe user’s problem domain
 and thus increase relevance of the results
 Make support of queries on natural language
             Solution Architecture
Input UI                                  Output UI
                                            Improved                    New marketing                    Visualize
New search        New      New web-                          Relevant                    Trends
                                             search                       campaigns                    market/clients
 requests         ads        sites                             ads                       Analysis
                                             results                       advices                       dynamics


End User                                                                                              Analytic




 Application
  Layer                                                            Analysis and
               Text Understanding     Intelligent matching
                                                                recommendations                      Portal
                     Engine                   module
                                                                     module




 Virtual Market
 Layer
                    Virtual Market        Knowledge Management                                Knowledge
                        Engine                   Toolset                                       Engineer




 Ontology Layer                                                                   DB Layer

   Historical precedents       Decision Making           Problem domain             Advertisements
                                                                                                              Sites DB
     ontological scenes         Logic Ontology               ontology                     DB
  Solution Logic
Create semantic ontology of the problem domain of
interest (in current project - cars)
Initial import of site’s pages from Search Engine on the
basis of keywords
Automatically create semantic descriptors for each site
page with help of Natural Language Understanding
engine
Automatically/Manually create semantic descriptors for
“sponsored” sites
Make requests in either natural language queries or in a
form of semantic descriptor
Sort all pages according to their semantic relevance,
show only relevant advertisements
              Website Ontology Fragment

                                                       Classic
                                                                                                Family
     Seat           Glass
                                                                          Models

                                      Engine
                                                             to
  Tyre                                                   g
                                                      lon                          Sports          Buy         Sell
                                                 Be
            Accessories
                             Part of
                                                                         allow
                                               Car
                                                                                                    Actions       Loan



                                 th
  Dealer                     l wi                                         where
                         a                     Country                                                           Insure
                      De                                                                             Compare
                                                                    Location
                                                                                      On-Line
                  Company
Manufacturer
                                       where

                                                                                 Worldwide
     Services                                                     City
      Text Understanding Process
                                   Syntax stage
 Example phrase: Company will
  provide support for Software
Programs employed by the Client.

 Morphology stage




                                      Semantics stage
             Text Understanding
             Technology
•    How it works ?
        Every word in the sentence is
       assigned an agent                            • Main Advantages
        Each agent has knowledge about                   Ability to understand
       the semantic meaning of the word                 structured and free text
       from the ontology
        Agents negotiate with each other
                                                         Understanding and storing the
       trying to reconstruct the coherent               phrasal context
       meaning of the whole sentence                     Use of knowledge of the
        As a result a semantic network is               problem domain in the text
       created - it’s a concrete scene from             processing
       the ontology                                      Ability to verify and modify the
        Next sentence will be understood                ontology in the course of the
       with the respect to the previously           •   user’s work
       created scene
                                                         Intelligent search and analysis
                                                        on the basis of the problem
    Natural Language Text Understanding algorithm       domain ontology
    is a patented property of Magenta Technology™
          Step 1 – Car Domain ontology

•   Ontology details
       Created on the base of       Ontology
       most popular queries to
       Search Engines
       More than 150 concepts
       Over 30 properties
       Over 15 relations
       Over 50 human experts
       man-hours for ontology
       development
       Synonyms: About 3 words
       or words combination in
       average for each of         Morphology
       concepts, properties,
       relations in the ontology
           Step 2 – Import from Search
           Engine
•   Import details
       Ability to pre-select a
       number of sites with
       keyword search
       Support of many different
       Search Engines
       Import any number of
       sites/request results
       With each site save initial
       query
       Import/Export URL’s
Step 3 – Semantic descriptors for site
pages




Imported site      Corresponding semantic descriptor
         Step 4 – Create semantic, meaningful
         descriptors for “sponsored” sites
                           PoliceAuction – Buy seized cars !
                           http://www.policeauctions.com




INFINITE CAR AUDIO –                                           Consumer Guide – reviews,
any audio for any car!                                         ratings, prices
www.infinitecaraudio.com                                       www.ConsumerGuide.com
         Step 5 – Make requests
Natural Language query:         Corresponding semantic descriptor of request
“Buy used cars online”



                                                                   Results with
                                                                   relevance ratings



“Sponsored link” descriptor –
request’s semantic similarity
          Step 6 – Analyze results
Initial request:         Semantic descriptor of best found web page
“Buy used cars online”




  Semantic descriptor
  of request:
Argo – Analysis
         Argo: Examples of program work

                       Request: “insure family car”
Yahoo Search Results                        Argo Re-Ordered Search Results




                                           Legend:
                                              - Positive (Relevant) result
                                               - Negative (Non-Relevant) result

                                               - Neutral (so-so) result
         Argo: Examples of program work

                       Request: “repair car in Chicago”
Yahoo Search Results                        Argo Re-Ordered Search Results




                                           Legend:
                                              - Positive (Relevant) result
                                               - Negative (Non-Relevant) result

                                               - Neutral (so-so) result
         Argo: Examples of program work

                       Request: “rent used car in London”
Yahoo Search Results                       Argo Re-Ordered Search Results




                                          Legend:
                                             - Positive (Relevant) result
                                              - Negative (Non-Relevant) result

                                              - Neutral (so-so) result
             Results of analysis – comparison with
             keyword search
                                                                 Experiments
         List of requests   Keyword search
                                             Semantic
                                             descriptors         procedure:
                                             search                Search Engines:
# 1 – “Buy a car”                     64%                  74%     Google, Yahoo,
                                                                   Excite, AOL
# 2 – “Rent a car”                                         90%
                                      70%                          25 first pages from
                                                                   each engine
# 3 – “Repair a car”                                       84%
                                      62%                          Human expert
                                                                   evaluation of
# 4 – “Car audio systems”                                  88%
                                      60%
                                                                   returned pages
                                                                   quality
# 5 – “Cars pictures”                                      68%
                                      44%                          Comparison between
                                                                   pages order of
# 6 – “Cars reviews”                                       72%
                                                                   traditional engines
                                      52%                          and ARGO
Total:                              58.67%            79.33%
             Results of analysis – comparison of
             advertisements quality
                                                                            Experiments
 List of requests     Keyword search            Semantic descriptors        procedure:
                      Number       Relevancy    Number        Relevancy       Search Engines:
                                                                              Google, Yahoo,
# 1 – “Buy a car”              8        75%              15           87%     Excite, AOL
# 2 – “Rent a car”                                                            For each request a
                               8        75%              4           100%     number of
# 3 – “Repair a                                                               advertisements
car”                           2        50%              3           100%     was analyzed
# 4 – “Car audio                                                              Human expert
systems”                               100%                          100%     evaluation of
                               5                         3                    suggested
#    5   –    “Cars                                                           advertisements
pictures”                      2       100%              2           100%     The same list of
#    6   –    “Cars                                                           advertisements
reviews”                       0          ---            7            86%     was loaded into
Total:                    4.17       80.00%         5.67            95.5%
                                                                              Argo and semantic
                                                                              descriptors were
                                                                              build
                                                                              Comparison
                                                                              between
                                                                              advertisements
                                                                              quality
   Summary
Implemented a first version of meta-search
engine, which integrate results from all major
search engines
Offered approach allows to significantly improve
quality of search results and relevancy of
advertisements
Easily adjustable to new problem domains – just
changing semantic ontologies

								
To top