Semantic%20Search%20Facilitator:%20Concept%20and%20Current%20State%20of%20Development

1
Semantic Search Facilitator Concept and Current
State of Development
InBCT Tekes PROJECT Chapter 3.1.3 Industrial
Ontologies and Semantic Web (year 2004)
2
Industrial Ontologies Group

• Researchers
• Vagan Terziyan
• Oleksandr Kononenko
• Andriy Zharko
• Oleksiy Khriyenko
• Olena Kaykova
• Olga Klochko
• Andriy Taranov

• Contact
• e-mail vagan_at_it.jyu.fi
• Phone [phone]
• URL http//www.cs.jyu.fi/ai/OntoGroup

3
Resources
Resources used from InBCT Project in 2004

• 12 000  EURO salaries for 5 months

4
Semantic-based Enhancement of the Information
RetrievalMotivation from Industrial Ontologies
GroupWhile recently there is luck of
annotated resources in the Web, which makes
metadata-based search useless, we should develop
enhanced Web search tool based on Google and
WordNet ontology and provide semantic search user
interface
5
Semantic Web and Information Retrieval

• Semantic Web promises many advantages and
  benefits, but
• We are only in transition towards the Semantic
  Web
• Resources are not yet annotated semantically
• Not enough metadata available in the Web for more
  smart search
• Semantic search of non-semantic data ???
• Yes, why not? We need a Semantic Facilitator !

6
Semantic Facilitator Concept

• What is it?
• Search service that uses other services
• Utilizes other search engines as Web services
  and
• makes their performance better due to smart
  query generation algorithms
• Supports search within heterogeneous resources
  (Web pages, Web databases, local file system,
  etc.)
• Filters returned results based on user
  preferences
• Intelligent semantic query-based tool that
  really understands what users want to find
• What it is not?
• Search engine, indexing tool, registry, etc.
• Data storage, database browser, etc.

7
Web search - Whats the Problem?

• Search in the web is not always convenient
• Polysemy of words gives irrelevant results
• Synonymy does not supported by search engines gt
  loss of relevant results
• There is a need to capture semantics from search
  query

8
Semantic Search Assistantlight version of
Semantic Facilitator

• Semantic Search Assistant (SSA) is a software
  that
• helps user to obtain more relevant results while
  using standard search engine (Google) by
  interaction with WordNet ontology
• finds possible contexts for words in search query
• can broaden or constrict search query with other
  relevant words and phrases for result improvement
• works with not annotated documents
• is not restricted to any concrete domain

9
Sense Determination

• WordNet is an open source ontology, which
  contains information about different meanings of
  a term, synonyms, antonyms and other lexical and
  semantic relations
• Having several words in search query we can
  determine in which context (sense) each of them
  is used with the help of WordNet
• by comparing words synsets
• by comparing words textual descriptions and
  examples
• by finding common roots going up in WordNet
  hierarchy tree for each word
• by asking a user

10
How does it work?

1. Gets keyword query
2. Translates original query into series of queries
   to Google taking into account the semantics of
   keywords
3. Combines returned results

11
Basic idea
SemanticSearchFacilitator is Semantic Interface
for search mechanism, which use combination and
integration of the set of other existing search
engines. This interface provides additional
semantic enhancement and filtering of a complex
information retrieval.
Local search
? search
DB search
Web search
SemanticSearchFacilitator
12
Ontology Personalization
Ontology Personalization is mechanism, which
allows users to have own conceptual view and be
able to use it for semantic querying of search
facilities.
Search
13
Semantic Search Enhancement
Semantic Search Enhancement
Semantic Search Facilitator transforms
concept-based queries on the intermediate stage
of query execution into a set of simple
(keyword-based) queries to the used search
services of various types (Internet
search-engine, database, intranet, etc.). Such
semantic facilitator uses knowledge of how to
translate semantic queries into the query formats
of several different search services and can
perform routine actions that most of users do in
order to achieve better performance and get more
relevant results
Common ontology
Personal ontology
(
)
Query
14
Semantic Search Enhancement
Semantic Search Enhancement
Common (linguistic) ontology
Semantic Search Facilitator uses ontologically
(WordNet) defined knowledge about words and
embedded support of advanced Google-search query
features in order to construct more efficient
queries from formal textual description of
searched information. Semantic Search Facilitator
hides from users the complexity of query language
of concrete search engine and performs routine
actions that most of users do in order to achieve
better performance and get more relevant results.
Domain ontology
(
)
Query
15
Capturing Semantics from Search
PhrasesMotivation according to our Ukrainian
colleague Vadim ErmolayevGoogle query should
be transformed based on domain ontology
16
(No Transcript)
17
(No Transcript)
18
(No Transcript)
19
(No Transcript)
20
(No Transcript)
21
(No Transcript)
22
Semantic Search Assistant
23
Semantic Search Assistant
24
Semantic Search Assistant
Sense Processor

• Defines the sense of words in search query if
  polysemantic interpretation is possible
• This determination is used for narrowing the
  choice from plenty of sense combinations
• Combination choice algorithm is based on
  analyzing the intersected items for different
  words in search query
• Intersections can be found, for example, in
  definition of words, in their synsets, or in
  definitions of some pointer type for words (e.g.
  in upper level hypernym)
• Is written in Python

25
Semantic Search Assistant
Google Search Query Generator Builds the new
search queries based on determined semantic sense
of query and user's measure of results relevance.
Search Preferences Processor Interprets user's
preferences for the query.
26
Semantic Search Assistant

• User Interface
• Allows sending search queries to search engine,
  getting search results and customization of query
• Customization
• selecting measure of importance for each word in
  query
• prioritizing certain sense for words in query
• Irrelevant results are excluded by choosing words
  that give negative characteristics to the
  resource

27
Algorithm for the New Query Generation
Rij
- relevance of the words sense
j 1, p
Sense (i1)
-1
1

k 1, mij

i 1, n
Syn (ij1)
Word(i)
Sense (ij)
1
-1

Nijk
Syn (ijk)

0
1

- number of the synonyms senses
Syn (ijmij )
Ri
- significance of the word in query
Sense (ip)
-1
1

n number of the words from query
p number of the words senses
mij number of the words synonyms in senses
28
Algorithm for the New Query Generation
Synonym Quality
j1
Rij
1
, if Synijk is a member of the synsetj
Qijk
p

L
Nijk
L number of the synsets which contain Synijk
Word(i)

Reduction of the synonym quality absolute value
if Qijk gt 0, then synonym will used via OR in
a query if Qijk lt 0, then will used via AND NOT
29
Algorithm for the New Query Generation
Algorithm 1
Word(1)
Syn
Syn
Syn

Syn
Word(i)
Syn
Syn
Syn

Syn

Word(n)
Syn
Syn
Syn

Syn
AND
AND
Query
Syn
Syn
Syn
Syn
Syn
Syn
Syn
Syn
Syn
OR (AND NOT)
OR (AND NOT)
OR (AND NOT)
30
Algorithm for the New Query Generation
Algorithm 2
gt
Q
Word(1)
Syn
Syn
Syn

Syn
Word(i)
Syn
Syn
Syn

Syn

Word(n)
Syn
Syn

Syn
Syn
Ri
Filtering based on a significance of the word

AND
AND
Query
Syn
Syn
Syn
Syn
Syn
Syn
Syn
Syn
Syn
OR (AND NOT)
OR (AND NOT)
OR (AND NOT)
31
Google APIAdaptation to search engine
32
We use Google because..

• Developers write software that connects remotely
  to the Google Web APIs service and access
  Google's index of more than 4 billion web pages
• Google Web APIs support the same search syntax as
  the Google.com site
• Communication is performed via the Simple Object
  Access Protocol (SOAP), an XML-based mechanism
  for exchanging typed information

..but that could be virtually any of existing
search engine
33
Requests

• Search query and a set of parameters is submitted
  to the Google Web APIs service and a set of
  search results is returned
• Cache requests submit a URL to the Google Web
  APIs service and receive the contents of the URL
  when Google's crawlers last visited the page (if
  available)
• Spelling requests submit a query to the Google
  Web APIs service and receive a suggested spell
  correction for the query (if available)

34
Requests Additional Commands

• intitle restricts your search to the titles of
  web pages
• inurl restricts your search to the titles of web
  pages
• intext searches only body text
• inanchor searches for text in pages link
  anchors
• site allows you to narrow your search by site or
  top-level domain
• link returns a list of pages linking to the
  specified URL
• cache finds a copy of the page that Google
  indexed from Googles cache

35
Requests Additional Commands

• daterange limits your search to the particular
  date or range of dates that a page was indexed
• filetype searches the suffixes or filename
  extensions
• related finds pages, which are related to the
  specified page
• info provides a page of links to more
  information about specified URL
• phonebook looks up phone numbers

36
Search Results Format Search Response

• Search response you get each time after search
  request
• It contains
• ltdocumentFilteringgt - A Boolean value indicating
  whether filtering was performed on the search
  results. This will be "true" only if (a) you
  requested filtering and (b) filtering actually
  occurred
• ltsearchCommentsgt - A text string intended for
  display to an end user. One of the most common
  messages found here is a note that "stop words"
  were removed from the search automatically. (This
  happens for very common words such as "and" and
  "as.")

37
Search Results Format Search Response (2)

• ltestimatedTotalResultsCountgt - The estimated
  total number of results that exist for the query
• ltestimateIsExactgt - A Boolean value indicating
  that the estimate is actually the exact value
• ltresultElementsgt - An array of ltresultElementgt
  items. This corresponds to the actual list of
  search results
• ltsearchQuerygt - This is the value of ltqgt for the
  search request
• ltstartIndexgt - Indicates the index (1-based) of
  the first search result in ltresultElementsgt

38
Search Results Format Search Response (3)

• ltendIndexgt - Indicates the index (1-based) of
  the last search result in ltresultElementsgt
• ltsearchTipsgt - A text string intended for
  display to the end user. It provides instructive
  suggestions on how to use Google
• ltdirectoryCategoriesgt - An array of
  ltdirectoryCategorygt items. This corresponds to
  the ODP directory matches for this search
• ltsearchTimegt - Text, floating-point number
  indicating the total server time to return the
  search results, measured in seconds

39
Search Results Format Result Element

• ltsummarygt - If the search result has a listing
  in the ODP directory, the ODP summary appears
  here as a text string
• ltURLgt - The URL of the search result, returned
  as text, with an absolute URL path
• ltsnippetgt - A snippet which shows the query in
  context on the URL where it appears. This is
  formatted HTML and usually includes ltBgt tags
  within it. Note that the query term does not
  always appear in the snippet
• lttitlegt - The title of the search result,
  returned as HTML

40
Search Results Format Result Element (2)

• ltcachedSizegt - Text (Integer "k"). Indicates
  that a cached version of the ltURLgt is available
  size is indicated in kilobytes
• ltrelatedInformationPresentgt - Boolean indicating
  that the "related" query term is supported for
  this URL
• lthostNamegt - When filtering occurs, a maximum
  of two results from any given host is returned.
  When this occurs, the second resultElement that
  comes from that host contains the host name in
  this parameter

41
Limitations of Google APIs

• Search request length 2048 bytes
• Maximum number of words in the query 10
• Maximum number of site terms in the query 1 (per
  search request)
• Maximum number of results per query 10
• Maximum value of ltstartgt ltmaxResultsgt 1000

42
WordNet( online access http//www.cogsci.prince
ton.edu/cgi-bin/webwn )
43
WordNet 2.0 Search Example

• Search word "driver ? The noun "driver" has 5
  senses in WordNet.1. driver -- (the operator of
  a motor vehicle)2. driver -- (someone who drives
  animals that pull a vehicle)3. driver -- (a
  golfer who hits the golf ball with a driver)4.
  driver, device driver -- ((computer science) a
  program that determines how a computer will
  communicate with a peripheral device)5. driver,
  number one wood -- (a golf club (a wood) with a
  near vertical face that is used for hitting long
  shots from the tee)
• Sense 1driver -- (the operator of a motor vehicle
  )       gt busman, bus driver -- (someone who dri
  ves a bus)       gt chauffeur -- (a man paid to d
  rive a privately owned car)       gt designated d
  river --(the member of a party who
  is designated to refrain from alcohol
  and so is sober when it is time to drive home)   
      gt honker -- (a driver who causes his car's ho
  rn to make a loud honking sound
  "the honker was fined for disturbing the peace") 
        gt motorist, automobilist -- (someone who dr
  ives (or travels in) an automobile)       gt owne
  r-driver -- (a motorist who owns the car that he/s
  he drives)       gt racer, race driver, automobil
  e driver -- (someone who drives racing cars at
  high speeds)       

44
WordNet Basic Terminology

• Syntactic category part of speech noun, verb,
  adjective, adverb
• Synonymic set (synset) list of synonymic words
  or collocations
• Every word can have several senses
• Every sense of a word is associated with synonyms
  (synset) of the word in that specific sense
• Synsets are organized in hierarchies interlinked
  with semanticrelations

45
WordNet Relations

• hypernymy / hyponymy
• ( is kind of / has kind )
• - hypernym, smth hiponym
• gull is a kind of bird, gull has specific kind
  called seagull
• meronymy / holonymy
• (part of / has a part)
• eye is a part of head / head has eyes as a part
• domain
• category
• usage

46
WordNet Organization

• Building Blocks
• Word forms common word orthography
• Word meanings by synsets
• Relations
• Lexical between word forms
• Semantic between word meanings
• gt Pointers
• Lexical pertain only to specific word
• Semantic pertain to all of the words in
  semantic set.

47
WordNet Organization
Synset1
Synset2
Semantic Relation
Lexical Relation
WordForm1
WordForm2
48
WordNet Organization

• Nouns and verbs
• organized into hierarchies based on the
  hypernymy/hyponymy relation between synsets
• additional pointers are be used to indicate other
  relations
• Adjectives
• arranged in clusters containing head synsets and
  satellite synsets
• Adverbs
• often derived from adjectives
• sometimes have antonyms
• therefore the synset for an adverb usually
  contains a lexical pointer to the adjective from
  which it is derived

49
Simple Example of WordNet usage
Syntactic Categories
Synsets
50
Simple Example of PyWordNet usage
Syntactic Categories
Synsets
51
Semantic Search Assistantprototype
52
Features of SSA

• Platform independent (written in Java)
• Works in 2 modes
• common mode, implements almost all of Google
  functionality
• extended mode, extends common mode, makes several
  requests with the same semantic sense, returns
  compound results.
• Keeps results in XML format

53
Common mode

• SSA has clear and simple interface, which helps
  user makes advanced Google search without special
  knowledge

• SSA transforms values of fields into Google
  request according to special format, which Google
  provides for advanced search

54
Extended mode

• More powerful mode than the common one
• SSA takes user request, makes a try to choose
  more convenient sense with users help
• Makes a set of requests, which extend users
  request by synonyms and exclude unsuitable words

55
Extended mode (2)

• SSA makes generated and base requests in series
• It compounds taken results into one according
  developed rules
• Provides possibility to highlight results for
  specific request for result analysing

56
Generating of requests set

• WordNet API and dictionaries are used for
  generating the set of requests

• When user enters original request, SSA switches
  to the panel, where different senses of typed
  word are presented

57
Generating of requests set (2)

• For every presented sense on this panel a user
  can see some description (even example) extracted
  from WordNet dictionary
• Also he/she can set rate of correspondence for
  every sense in range -1, 1

58
Making compound result

• SSA sends generated requests to Google one by one
• It keeps obtained results for each request
  separately
• User finally will get an integrated result, which
  was generated according special rules

59
Integrated resultsgenerating rules

• Unique identifier for each result is its URL
• SSA counts amount of URL appearances in returned
  results and sets this amount as index for every
  URL
• Results with bigger index are showed first
• If indexes are equal, results are shown according
  the order as Google returned them

60
Results analysis

• After making all requests, SSA shows final
  results
• All results are keeping also in files in XML
  format for further analysing
• User can highlight results for specific request,
  if there were more than one request

61
Python - WordNet

• PyWN("pin"), a python interface to WordNet,
  developed by WordNet employees John Asmuth and
  Jesse Fischer
• PyWordNet is a Python interface to WordNet
  developed by Oliver Steele

62
Java-Python Integration Jython
Python
pyClass.py
jythonc
Jython
pyClass.java
javac
pyClass.class (inner classes)
Java
63
Results

• Methods for automatic sense determination using
  WordNet Lexical Database were studied and
  correspondent algorithms were implemented
• Algorithm for new query generating were
  implemented and embedded to the programming
  complex
• User Interface for advanced search (with Google
  integration) was developed with Semantic Search
  Assistant functionality

64
Example

• Initial query
• hotel reservation agency
• (1, 7 and 5 senses correspondingly)
• From first 5 results only 3 are relevant
• (results with whole sequence of query words even
  does not appear in first three pages)

• Generated query
• ("hotel") ("booking" OR "reserve")
  (-"qualification") ("bureau" OR "agency")
  (-"means")
• From first 5 results all are relevant
• (using synonym booking along with
  reservation was helpful)

65
Example

• Results of initial query

Results of generated query
66
More Examples
67
Test 1
Initial query cork mousepad
68
Test 1
Initial query cork mousepad
Enhanced query ("phellem" OR "bobfloat" OR
"bobber" OR "cork" OR "bob") ("mousepad" OR
"mouse mat")
69
Test 2
Initial query flowers present shop
70
Test 2
Initial query flowers present shop
Enhanced query ("flower") (-"heyday" -"prime"
-"efflorescence") ("present") (-"nowadays"
-"present tense") ("store" OR "shop")
(-"workshop")
71
Test 3
Initial query hotel reservation agency
72
Test 3
Initial query hotel reservation agency
Enhanced query ("hotel") ("booking" OR
"reserve") (-"qualification") ("bureau" OR
"agency") (-"means")
73
Test 4
Initial query zodiac fish
74
Test 4
Initial query zodiac fish
Enhanced query ("zodiac") ("pisces" OR "fish" OR
"pisces the fishes")
75
Drawbacks

• Lack highly specialized terminology for narrow
  domains in WordNet gt difficult to get better
  results with SSA in such cases
• Frequent absence of sense relation between words
  in whole phrases gt difficulty of context
  determination by used algorithms
• Presence of several very close senses for many
  terms in WordNet gt no clear belonging of the
  word to some sense
• Possible wrong determination of part of speech
  for word in query gt using improper synonyms and
  antonyms for making query

76
Possible Improvements and further work

• Additional Adaptive Learning (for personalized
  context definition)
• Creating Global Sense Ontology on the basis of
  WordNet Database
• Improving algorithms for automatic computing of
  relevance indexes
• Adding algorithms for smart cutting off for
  generated queries
• Using fuzzy logic for determination of query
  context
• Adding other lexical databases for supporting
  search in specific domains (like programming,
  medicine)
• Multilingual support

77
Current status

• During Jan-May 2004 main efforts for the InBCT
  Semantic Search Facilitator project were put
  into the research and design of the basic
  features of SSA and implementation of
  ontology-based search method.
• The development of the prototype Semantic Search
  Assistant software has been started and pilot
  version is ready.
• Starting 1.06.2004 kernel part of the Industrial
  Ontologies Group start working on TEKES project
  SmartResource Proactive Self-Maintained
  Resources in Semantic Web
• at Agora Center, University of Jyväskylä
• Further development (from the point of stability
  and usability) of SSA will be continued during
  Jul-Sep 2004