SlideShare uma empresa Scribd logo
1 de 48
Introduction to Search Engine-
     Building with Lucene
             Kai Chan
    SoCal Code Camp, June 2012
How to Search
• One (common) approach to searching all your
  documents:

for each document d {
  if (query is a substring of d’s content) {
    add d to the list of results
  }
}
sort the results

                                                1
How to Search
• Problems
  – Slow: Reads the whole database for each search
  – Not scalable: If your database grows by 10x, your
    search slows down by 10x
  – How to show the most relevant documents first?




                                                        2
Inverted Index
• (term -> document list) map
Documents:   T0 = "it is what it is"
             T1 = "what is it"
             T2 = "it is a banana"

Inverted     "a":        {2}
index:       "banana":   {2}
             "is":       {0, 1, 2}
             "it":       {0, 1, 2}
             "what":     {0, 1}
 E                                     3
Inverted Index
• (term -> <document, position> list) map

 T0 = "it is what it is”
       0 1 2      3 4

 T1 = "what is it”
       0    1 2

 T2 = "it is a banana”
       0 1 2 3

 E                                          4
Inverted Index
• (term -> <document, position> list) map

 T0 = "it is what it is"
 T1 = "what is it"
 T2 = "it is a banana"

 "a":        {(2,   2)}
 "banana":   {(2,   3)}
 "is":       {(0,   1), (0, 4), (1, 1), (2, 1)}
 "it":       {(0,   0), (0, 3), (1, 2), (2, 0)}
 "what":     {(0,   2), (1, 0)}

 E                                                5
Inverted Index
• Speed
  – Term list
     • Very small compared to documents’ content
     • Tends to grow at a slower speed than documents
       (after a certain level)
  – Term lookup: O(1) to O(log of number of terms)
  – Document lists are very small
  – Document + position lists still small


                                                        6
Inverted Index
• Relevance
  – Extra information in the index
     • Stored in a easily accessible way
     • Determine relevance of each document to the query
  – Enables sorting by (decreasing) relevance




                                                           7
Determining Relevancy
• Two models used in the searching process
  – Boolean model
     • AND, OR, NOT, etc.
     • Either a document matches a query, or not
  – Vector space model
     • How often a query term appears in a document vs.
       how often the term appears in all documents
     • Scoring and sorting by relevancy possible



                                                          8
Determining Relevancy
Lucene uses both models

     all documents


             filtering (Boolean Model)

   some documents
      (unsorted)

             scoring (Vector Space Model)

   some documents
   (sorted by score)
                                            9
Vector Space Model
f(frequency of term B)

                         document 1
                                        query


                                           document 2




                                f(frequency of term A)   10
Scoring
• Term frequency (TF)
  – How many times does this term (t) appear in this
    document (d)?
  – Score proportional to TF
• Document frequency (DF)
  – How many documents have this term (t)?
  – Score proportional to the inverse of DF (IDF)



                                                       11
Scoring
• Coordination factor (coord)
  – Documents that contains all or most query terms
    get higher scores
• Normalizing factor (norm)
  – Adjust for field length and query complexity




                                                      12
Scoring
• Boost
  – “Manual override”: ask Lucene to give a higher
    score to some particular thing
  – Index-time
     • Document
     • Field (of a particular document)
  – Search-time
     • Query



                                                     13
Scoring
                     coordination factor           query normalizing factor



       score(q, d) = coord(q, d) . queryNorm(q) .
     Σ t in q (tf (t in d) . idf(t)2 . boost(t) . norm(t, d))

           term            inverse
        frequency        document
                         frequency
                                         term boost          document boost,
                                                                field boost,
                                                         length normalizing factor
http://lucene.apache.org/core/3_6_0/scoring.html                               14
Work Flow
• Indexing
  – Index: storage of inverted index + documents
  – Add fields to a document
  – Add the document to the index
  – Repeat for every document
• Searching
  – Generate a query
  – Search with this query
  – Get back a sorted document list (top N docs)

                                                   15
Adding Field to Document
• Store?
• Index?
  – Analyzed (split text into multiple terms)
  – Not analyzed (treat the whole text as ONE term)
  – Not indexed (this field will not be searchable)
  – Store norms?




                                                      16
Analyzed vs. Not Analyzed
             Text: “the quick brown fox”


Analyzed: 4 terms                Not analyzed: 1 term
1. the                           1. the quick brown fox
2. quick
3. brown
4. fox




                                                      17
Index-time Analysis
• Analyzer
  – Determine which TokenStream classes to use
• TokenStream
  – Does the actual hard work
  – Tokenizer: text to tokens
  – Token filter: tokens to tokens




                                                 18
Text:
San Franciso, the Bay Area’s city-county
http://www.ci.sf.ca.us controller@sfgov.org

WhitespaceAnalyzer:
[San] [Francisco,] [the] [Bay] [Area’s]
[city-county] [http://www.ci.sf.ca.us/]
[controller@sfgov.org]

StopAnalyzer:
[san] [francisco] [bay] [area] [s] [city] [county]
[http] [www] [ci] [sf] [ca] [us] [controller]
[sfgov] [org]

StandardAnalyzer:
[san] [francisco] [bay] [area's] [city] [county]
[http] [www.ci.fs.ca.us] [controller] [sfgov.org]
                                                     19
Notable TokenStream Classes
• ASCIIFoldingFilter
  – Converts alphabetic characters into basic forms
• PorterStemFilter
  – Reduces tokens into their stems
• SynonymTokenFilter
  – Converts words to their synonyms
• ShingleFilter
  – Creates shingles (n-grams)

                                                      20
Tokens
• Information about a token
  – Field
  – Text
  – Start offset, end offset
  – Position increment




                               21
Attributes
• Past versions of Lucene: Token object
• Recent version of Lucene: attributes
  – Efficiency, flexibility
  – Ask for attributes you want
  – Receive attribute objects
  – Use these object for information about tokens




                                                    22
create token stream
TokenStream tokenStream =
analyzer.reusableTokenStream(fieldName, reader);
tokenStream.reset();

CharTermAttribute term =                                   obtain each
stream.addAttribute(CharTermAttribute.class);              attribute you
                                                           want to know
OffsetAttribute offset =
stream.addAttribute(OffsetAttribute.class);

PositionIncrementAttribute posInc =
stream.addAttribute(PositionIncrementAttribute.class);

while (tokenStream.incrementToken()) {           go to the next token
  doSomething(term.toString(),
              offset.startOffset(),      use information about
              offset.endOffset(),        the current token
              posInc.getPositionIncrement());
}

tokenStream.end();           close token stream
tokenStream.close();                                                 23
Query-time Analysis
• Text in a query is analyzed like fields
• Use the same analyzer that analyzed the
  particular field

 +field1:“quick brown fox” +(field2:“lazy dog” field2:“cozy cat”)



    quick    brown      fox        lazy   dog        cozy    cat


                                                                    24
Query Formation
• Query parsing
  – A query parser in core code
  – Additional query parsers in contributed code
• Or build query from the Lucene query classes




                                                   25
Term Query
• Matches documents with a particular term
  – Field
  – Text




                                             26
Term Range Query
• Matches documents with any of the terms in a
  particular range
  – Field
  – Lowest term text
  – Highest term text
  – Include lowest term text?
  – Include highest term text?



                                             27
Prefix Query
• Matches documents with any of the terms
  with a particular prefix
  – Field
  – Prefix




                                            28
Wildcard/Regex Query
• Matches documents with any of the terms
  that match a particular pattern
  – Field
  – Pattern
     • Wildcard: * for 0+ characters, ? for 0-1 character
     • Regular expression
     • Pattern matching on individual terms only




                                                            29
Fuzzy Query
• Matches documents with any of the terms
  that are “similar” to a particular term
  – Levenshtein distance (“edit distance”):
    Number of character insertions, deletions or
    substitutions needed to transform one string into
    another
     • e.g. kitten -> sitten -> sittin -> sitting (3 edits)
  – Field
  – Text
  – Minimum similarity score

                                                              30
Phrase Query
• Matches documents with all the given words
  present and being “near” each other
  – Field
  – Terms
  – Slop
     • Number of “moves of words” permitted
     • Slop = 0 means exact phrase match required




                                                    31
Boolean Query
• Conceptually similar to boolean operators
  (“AND”, “OR”, “NOT”), but not identical
• Why Not AND, OR, And NOT?
  – http://www.lucidimagination.com/blog/2011/12/
    28/why-not-and-or-and-not/
  – In short, boolean operators do not handle > 2
    clauses well



                                                32
Boolean Query
• Three types of clauses
  – Must
  – Should
  – Must not
• For a boolean query to match a document
  – All “must” clauses must match
  – All “must not” clauses must not match
  – At least one “must” or “should” clause must
    match

                                                  33
Span Query
• Similar to other queries, but matches spans
• Span
  – particular place/part of a particular document
  – <document ID, start position, end position> tuple




                                                        34
T0 = "it is what it is”
      0 1 2      3 4

T1 = "what is it”
      0    1 2

T2 = "it is a banana”
      0 1 2 3


           <doc ID, start pos., end pos.>
“it is”:   <0,      0,          2>
           <0,      3,          5>
           <2,      0,          2>
                                        35
Span Query
• SpanTermQuery
  – Same as TermQuery, except your can build other
    span queries with it
• SpanOrQuery
  – Matches spans that are matched by any of some
    span queries
• SpanNotQuery
  – Matches spans that are matched by one span
    query but not the other span query

                                                     36
spanTerm(apple)                 spanOr([apple, orange])


apple                  orange   apple                  orange




        spanTerm(orange)            spanNot(apple, orange)



                                                         37
Span Query
• SpanNearQuery
  – Matches spans that are within a certain “slop” of
    each other
  – Slop: max number of positions between spans
  – Can specify whether order matters




                                                        38
the                quick           brown       fox

                     2                 1             0

1. spanNear([brown, fox, the, quick], slop = 4, inOrder = false)        ✔

2. spanNear([brown, fox, the, quick], slop = 3, inOrder = false)        ✔

3. spanNear([brown, fox, the, quick], slop = 2, inOrder = false)        ✖

4. spanNear([brown, fox, the, quick], slop = 3, inOrder = true)         ✖

5. spanNear([the, quick, brown, fox], slop = 3, inOrder = true)         ✔


                                                                   39
Filtering
• A Filter narrows down the search result
  – Creates a set of document IDs
  – Decides what documents get processed further
  – Does not affect scoring, i.e. does not score/rank
    documents that pass the filter
  – Can be cached easily
  – Useful for access control, presets, etc.



                                                        40
Notable Filter classes
• TermsFilter
   – Allows documents with any of the given terms
• TermRangeFilter
   – Filter version of TermRangeQuery
• PrefixFilter
   – Filter version of PrefixQuery
• QueryWrapperFilter
   – “Adapts” a query into a filter
• CachingWrapperFilter
   – Cache the result of the wrapped filter

                                                    41
Sorting
• Score (default)
• Index order
• Field
  – Requires the field be indexed & not analyzed
  – Specify type (string, int, etc.)
  – Normal or reverse order
  – Single or multiple fields


                                                   42
Interfacing Lucene with “Outside”
• Embedding directly
• Language bridge
  – E.g. PHP/Java Bridge
• Web service
  – E.g. Jetty + your own request handler
• Solr
  – Lucene + Jetty + lots of useful functionality


                                                    43
Books
• Lucene in Action, 2nd Edition
  – Written by 3 committers and PMC members
  – http://www.manning.com/hatcher3/
• Introduction to Information Retrieval
  – Not specific to Lucene, but about IR concepts
  – Free e-book
  – http://nlp.stanford.edu/IR-book/


                                                    44
Web Resources
• Official Website
   – http://lucene.apache.org/
• StackOverflow
   – http://stackoverflow.com/questions/tagged/lucene
• Mailing lists
   – http://lucene.apache.org/core/discussion.html
• Blogs
   – http://www.lucidimagination.com/blog/
   – http://blog.mikemccandless.com/
   – http://lucene.grantingersoll.com/

                                                        45
Getting Started
• Getting started
  – Download lucene-3.6.0.zip (or .tgz)
  – Add lucene-core-3.6.0.jar to your classpath
  – Consider using an IDE (e.g. Eclipse)
  – Luke (Lucene Index Toolbox)
    http://code.google.com/p/luke/




                                                  46
47

Mais conteúdo relacionado

Mais procurados

Vu Semantic Web Meeting 20091123
Vu Semantic Web Meeting 20091123Vu Semantic Web Meeting 20091123
Vu Semantic Web Meeting 20091123
Rinke Hoekstra
 
Perl%20Tutorial.!Picking%20Up%20Perl
Perl%20Tutorial.!Picking%20Up%20PerlPerl%20Tutorial.!Picking%20Up%20Perl
Perl%20Tutorial.!Picking%20Up%20Perl
tutorialsruby
 

Mais procurados (10)

An Introduction to NLP4L (Scala by the Bay / Big Data Scala 2015)
An Introduction to NLP4L (Scala by the Bay / Big Data Scala 2015)An Introduction to NLP4L (Scala by the Bay / Big Data Scala 2015)
An Introduction to NLP4L (Scala by the Bay / Big Data Scala 2015)
 
Text Mining Infrastructure in R
Text Mining Infrastructure in RText Mining Infrastructure in R
Text Mining Infrastructure in R
 
Vu Semantic Web Meeting 20091123
Vu Semantic Web Meeting 20091123Vu Semantic Web Meeting 20091123
Vu Semantic Web Meeting 20091123
 
Learning Multilingual Semantic Parsers for Question Answering over Linked Dat...
Learning Multilingual Semantic Parsers for Question Answering over Linked Dat...Learning Multilingual Semantic Parsers for Question Answering over Linked Dat...
Learning Multilingual Semantic Parsers for Question Answering over Linked Dat...
 
NLP - Prédictions de tags sur les questions Stackoverflow
NLP - Prédictions de tags sur les questions StackoverflowNLP - Prédictions de tags sur les questions Stackoverflow
NLP - Prédictions de tags sur les questions Stackoverflow
 
Perl%20Tutorial.!Picking%20Up%20Perl
Perl%20Tutorial.!Picking%20Up%20PerlPerl%20Tutorial.!Picking%20Up%20Perl
Perl%20Tutorial.!Picking%20Up%20Perl
 
엘라스틱서치 적합성 이해하기 20160630
엘라스틱서치 적합성 이해하기 20160630엘라스틱서치 적합성 이해하기 20160630
엘라스틱서치 적합성 이해하기 20160630
 
Webinar: OpenNLP and Solr for Superior Relevance
Webinar: OpenNLP and Solr for Superior RelevanceWebinar: OpenNLP and Solr for Superior Relevance
Webinar: OpenNLP and Solr for Superior Relevance
 
Phpconf2008 Sphinx En
Phpconf2008 Sphinx EnPhpconf2008 Sphinx En
Phpconf2008 Sphinx En
 
The vector space model
The vector space modelThe vector space model
The vector space model
 

Semelhante a Introduction to search engine-building with Lucene

SDEC2011 NoSQL Data modelling
SDEC2011 NoSQL Data modellingSDEC2011 NoSQL Data modelling
SDEC2011 NoSQL Data modelling
Korea Sdec
 
SDEC2011 NoSQL concepts and models
SDEC2011 NoSQL concepts and modelsSDEC2011 NoSQL concepts and models
SDEC2011 NoSQL concepts and models
Korea Sdec
 

Semelhante a Introduction to search engine-building with Lucene (20)

Tutorial 1 (information retrieval basics)
Tutorial 1 (information retrieval basics)Tutorial 1 (information retrieval basics)
Tutorial 1 (information retrieval basics)
 
Intent Algorithms: The Data Science of Smart Information Retrieval Systems
Intent Algorithms: The Data Science of Smart Information Retrieval SystemsIntent Algorithms: The Data Science of Smart Information Retrieval Systems
Intent Algorithms: The Data Science of Smart Information Retrieval Systems
 
Chapter 6 Query Language .pdf
Chapter 6 Query Language .pdfChapter 6 Query Language .pdf
Chapter 6 Query Language .pdf
 
Web search engines
Web search enginesWeb search engines
Web search engines
 
SDEC2011 NoSQL Data modelling
SDEC2011 NoSQL Data modellingSDEC2011 NoSQL Data modelling
SDEC2011 NoSQL Data modelling
 
Architecture of a search engine
Architecture of a search engineArchitecture of a search engine
Architecture of a search engine
 
Enterprise Search Solution: Apache SOLR. What's available and why it's so cool
Enterprise Search Solution: Apache SOLR. What's available and why it's so coolEnterprise Search Solution: Apache SOLR. What's available and why it's so cool
Enterprise Search Solution: Apache SOLR. What's available and why it's so cool
 
Intro to Elasticsearch
Intro to ElasticsearchIntro to Elasticsearch
Intro to Elasticsearch
 
The Intent Algorithms of Search & Recommendation Engines
The Intent Algorithms of Search & Recommendation EnginesThe Intent Algorithms of Search & Recommendation Engines
The Intent Algorithms of Search & Recommendation Engines
 
search engine
search enginesearch engine
search engine
 
Relevance in the Wild - Daniel Gomez Vilanueva, Findwise
Relevance in the Wild - Daniel Gomez Vilanueva, FindwiseRelevance in the Wild - Daniel Gomez Vilanueva, Findwise
Relevance in the Wild - Daniel Gomez Vilanueva, Findwise
 
3_Indexing.ppt
3_Indexing.ppt3_Indexing.ppt
3_Indexing.ppt
 
SDEC2011 NoSQL concepts and models
SDEC2011 NoSQL concepts and modelsSDEC2011 NoSQL concepts and models
SDEC2011 NoSQL concepts and models
 
Search pitb
Search pitbSearch pitb
Search pitb
 
Text features
Text featuresText features
Text features
 
Some Information Retrieval Models and Our Experiments for TREC KBA
Some Information Retrieval Models and Our Experiments for TREC KBASome Information Retrieval Models and Our Experiments for TREC KBA
Some Information Retrieval Models and Our Experiments for TREC KBA
 
What is in a Lucene index?
What is in a Lucene index?What is in a Lucene index?
What is in a Lucene index?
 
Lucene for Solr Developers
Lucene for Solr DevelopersLucene for Solr Developers
Lucene for Solr Developers
 
Building Search & Recommendation Engines
Building Search & Recommendation EnginesBuilding Search & Recommendation Engines
Building Search & Recommendation Engines
 
Semi-automated Exploration and Extraction of Data in Scientific Tables
Semi-automated Exploration and Extraction of Data in Scientific TablesSemi-automated Exploration and Extraction of Data in Scientific Tables
Semi-automated Exploration and Extraction of Data in Scientific Tables
 

Último

EIS-Webinar-Prompt-Knowledge-Eng-2024-04-08.pptx
EIS-Webinar-Prompt-Knowledge-Eng-2024-04-08.pptxEIS-Webinar-Prompt-Knowledge-Eng-2024-04-08.pptx
EIS-Webinar-Prompt-Knowledge-Eng-2024-04-08.pptx
Earley Information Science
 
Artificial Intelligence: Facts and Myths
Artificial Intelligence: Facts and MythsArtificial Intelligence: Facts and Myths
Artificial Intelligence: Facts and Myths
Joaquim Jorge
 

Último (20)

08448380779 Call Girls In Diplomatic Enclave Women Seeking Men
08448380779 Call Girls In Diplomatic Enclave Women Seeking Men08448380779 Call Girls In Diplomatic Enclave Women Seeking Men
08448380779 Call Girls In Diplomatic Enclave Women Seeking Men
 
Raspberry Pi 5: Challenges and Solutions in Bringing up an OpenGL/Vulkan Driv...
Raspberry Pi 5: Challenges and Solutions in Bringing up an OpenGL/Vulkan Driv...Raspberry Pi 5: Challenges and Solutions in Bringing up an OpenGL/Vulkan Driv...
Raspberry Pi 5: Challenges and Solutions in Bringing up an OpenGL/Vulkan Driv...
 
Axa Assurance Maroc - Insurer Innovation Award 2024
Axa Assurance Maroc - Insurer Innovation Award 2024Axa Assurance Maroc - Insurer Innovation Award 2024
Axa Assurance Maroc - Insurer Innovation Award 2024
 
🐬 The future of MySQL is Postgres 🐘
🐬  The future of MySQL is Postgres   🐘🐬  The future of MySQL is Postgres   🐘
🐬 The future of MySQL is Postgres 🐘
 
Workshop - Best of Both Worlds_ Combine KG and Vector search for enhanced R...
Workshop - Best of Both Worlds_ Combine  KG and Vector search for  enhanced R...Workshop - Best of Both Worlds_ Combine  KG and Vector search for  enhanced R...
Workshop - Best of Both Worlds_ Combine KG and Vector search for enhanced R...
 
Boost Fertility New Invention Ups Success Rates.pdf
Boost Fertility New Invention Ups Success Rates.pdfBoost Fertility New Invention Ups Success Rates.pdf
Boost Fertility New Invention Ups Success Rates.pdf
 
2024: Domino Containers - The Next Step. News from the Domino Container commu...
2024: Domino Containers - The Next Step. News from the Domino Container commu...2024: Domino Containers - The Next Step. News from the Domino Container commu...
2024: Domino Containers - The Next Step. News from the Domino Container commu...
 
Handwritten Text Recognition for manuscripts and early printed texts
Handwritten Text Recognition for manuscripts and early printed textsHandwritten Text Recognition for manuscripts and early printed texts
Handwritten Text Recognition for manuscripts and early printed texts
 
Exploring the Future Potential of AI-Enabled Smartphone Processors
Exploring the Future Potential of AI-Enabled Smartphone ProcessorsExploring the Future Potential of AI-Enabled Smartphone Processors
Exploring the Future Potential of AI-Enabled Smartphone Processors
 
Strategize a Smooth Tenant-to-tenant Migration and Copilot Takeoff
Strategize a Smooth Tenant-to-tenant Migration and Copilot TakeoffStrategize a Smooth Tenant-to-tenant Migration and Copilot Takeoff
Strategize a Smooth Tenant-to-tenant Migration and Copilot Takeoff
 
How to Troubleshoot Apps for the Modern Connected Worker
How to Troubleshoot Apps for the Modern Connected WorkerHow to Troubleshoot Apps for the Modern Connected Worker
How to Troubleshoot Apps for the Modern Connected Worker
 
[2024]Digital Global Overview Report 2024 Meltwater.pdf
[2024]Digital Global Overview Report 2024 Meltwater.pdf[2024]Digital Global Overview Report 2024 Meltwater.pdf
[2024]Digital Global Overview Report 2024 Meltwater.pdf
 
How to Troubleshoot Apps for the Modern Connected Worker
How to Troubleshoot Apps for the Modern Connected WorkerHow to Troubleshoot Apps for the Modern Connected Worker
How to Troubleshoot Apps for the Modern Connected Worker
 
EIS-Webinar-Prompt-Knowledge-Eng-2024-04-08.pptx
EIS-Webinar-Prompt-Knowledge-Eng-2024-04-08.pptxEIS-Webinar-Prompt-Knowledge-Eng-2024-04-08.pptx
EIS-Webinar-Prompt-Knowledge-Eng-2024-04-08.pptx
 
GenAI Risks & Security Meetup 01052024.pdf
GenAI Risks & Security Meetup 01052024.pdfGenAI Risks & Security Meetup 01052024.pdf
GenAI Risks & Security Meetup 01052024.pdf
 
04-2024-HHUG-Sales-and-Marketing-Alignment.pptx
04-2024-HHUG-Sales-and-Marketing-Alignment.pptx04-2024-HHUG-Sales-and-Marketing-Alignment.pptx
04-2024-HHUG-Sales-and-Marketing-Alignment.pptx
 
Boost PC performance: How more available memory can improve productivity
Boost PC performance: How more available memory can improve productivityBoost PC performance: How more available memory can improve productivity
Boost PC performance: How more available memory can improve productivity
 
Artificial Intelligence: Facts and Myths
Artificial Intelligence: Facts and MythsArtificial Intelligence: Facts and Myths
Artificial Intelligence: Facts and Myths
 
Scaling API-first – The story of a global engineering organization
Scaling API-first – The story of a global engineering organizationScaling API-first – The story of a global engineering organization
Scaling API-first – The story of a global engineering organization
 
08448380779 Call Girls In Civil Lines Women Seeking Men
08448380779 Call Girls In Civil Lines Women Seeking Men08448380779 Call Girls In Civil Lines Women Seeking Men
08448380779 Call Girls In Civil Lines Women Seeking Men
 

Introduction to search engine-building with Lucene

  • 1. Introduction to Search Engine- Building with Lucene Kai Chan SoCal Code Camp, June 2012
  • 2. How to Search • One (common) approach to searching all your documents: for each document d { if (query is a substring of d’s content) { add d to the list of results } } sort the results 1
  • 3. How to Search • Problems – Slow: Reads the whole database for each search – Not scalable: If your database grows by 10x, your search slows down by 10x – How to show the most relevant documents first? 2
  • 4. Inverted Index • (term -> document list) map Documents: T0 = "it is what it is" T1 = "what is it" T2 = "it is a banana" Inverted "a": {2} index: "banana": {2} "is": {0, 1, 2} "it": {0, 1, 2} "what": {0, 1} E 3
  • 5. Inverted Index • (term -> <document, position> list) map T0 = "it is what it is” 0 1 2 3 4 T1 = "what is it” 0 1 2 T2 = "it is a banana” 0 1 2 3 E 4
  • 6. Inverted Index • (term -> <document, position> list) map T0 = "it is what it is" T1 = "what is it" T2 = "it is a banana" "a": {(2, 2)} "banana": {(2, 3)} "is": {(0, 1), (0, 4), (1, 1), (2, 1)} "it": {(0, 0), (0, 3), (1, 2), (2, 0)} "what": {(0, 2), (1, 0)} E 5
  • 7. Inverted Index • Speed – Term list • Very small compared to documents’ content • Tends to grow at a slower speed than documents (after a certain level) – Term lookup: O(1) to O(log of number of terms) – Document lists are very small – Document + position lists still small 6
  • 8. Inverted Index • Relevance – Extra information in the index • Stored in a easily accessible way • Determine relevance of each document to the query – Enables sorting by (decreasing) relevance 7
  • 9. Determining Relevancy • Two models used in the searching process – Boolean model • AND, OR, NOT, etc. • Either a document matches a query, or not – Vector space model • How often a query term appears in a document vs. how often the term appears in all documents • Scoring and sorting by relevancy possible 8
  • 10. Determining Relevancy Lucene uses both models all documents filtering (Boolean Model) some documents (unsorted) scoring (Vector Space Model) some documents (sorted by score) 9
  • 11. Vector Space Model f(frequency of term B) document 1 query document 2 f(frequency of term A) 10
  • 12. Scoring • Term frequency (TF) – How many times does this term (t) appear in this document (d)? – Score proportional to TF • Document frequency (DF) – How many documents have this term (t)? – Score proportional to the inverse of DF (IDF) 11
  • 13. Scoring • Coordination factor (coord) – Documents that contains all or most query terms get higher scores • Normalizing factor (norm) – Adjust for field length and query complexity 12
  • 14. Scoring • Boost – “Manual override”: ask Lucene to give a higher score to some particular thing – Index-time • Document • Field (of a particular document) – Search-time • Query 13
  • 15. Scoring coordination factor query normalizing factor score(q, d) = coord(q, d) . queryNorm(q) . Σ t in q (tf (t in d) . idf(t)2 . boost(t) . norm(t, d)) term inverse frequency document frequency term boost document boost, field boost, length normalizing factor http://lucene.apache.org/core/3_6_0/scoring.html 14
  • 16. Work Flow • Indexing – Index: storage of inverted index + documents – Add fields to a document – Add the document to the index – Repeat for every document • Searching – Generate a query – Search with this query – Get back a sorted document list (top N docs) 15
  • 17. Adding Field to Document • Store? • Index? – Analyzed (split text into multiple terms) – Not analyzed (treat the whole text as ONE term) – Not indexed (this field will not be searchable) – Store norms? 16
  • 18. Analyzed vs. Not Analyzed Text: “the quick brown fox” Analyzed: 4 terms Not analyzed: 1 term 1. the 1. the quick brown fox 2. quick 3. brown 4. fox 17
  • 19. Index-time Analysis • Analyzer – Determine which TokenStream classes to use • TokenStream – Does the actual hard work – Tokenizer: text to tokens – Token filter: tokens to tokens 18
  • 20. Text: San Franciso, the Bay Area’s city-county http://www.ci.sf.ca.us controller@sfgov.org WhitespaceAnalyzer: [San] [Francisco,] [the] [Bay] [Area’s] [city-county] [http://www.ci.sf.ca.us/] [controller@sfgov.org] StopAnalyzer: [san] [francisco] [bay] [area] [s] [city] [county] [http] [www] [ci] [sf] [ca] [us] [controller] [sfgov] [org] StandardAnalyzer: [san] [francisco] [bay] [area's] [city] [county] [http] [www.ci.fs.ca.us] [controller] [sfgov.org] 19
  • 21. Notable TokenStream Classes • ASCIIFoldingFilter – Converts alphabetic characters into basic forms • PorterStemFilter – Reduces tokens into their stems • SynonymTokenFilter – Converts words to their synonyms • ShingleFilter – Creates shingles (n-grams) 20
  • 22. Tokens • Information about a token – Field – Text – Start offset, end offset – Position increment 21
  • 23. Attributes • Past versions of Lucene: Token object • Recent version of Lucene: attributes – Efficiency, flexibility – Ask for attributes you want – Receive attribute objects – Use these object for information about tokens 22
  • 24. create token stream TokenStream tokenStream = analyzer.reusableTokenStream(fieldName, reader); tokenStream.reset(); CharTermAttribute term = obtain each stream.addAttribute(CharTermAttribute.class); attribute you want to know OffsetAttribute offset = stream.addAttribute(OffsetAttribute.class); PositionIncrementAttribute posInc = stream.addAttribute(PositionIncrementAttribute.class); while (tokenStream.incrementToken()) { go to the next token doSomething(term.toString(), offset.startOffset(), use information about offset.endOffset(), the current token posInc.getPositionIncrement()); } tokenStream.end(); close token stream tokenStream.close(); 23
  • 25. Query-time Analysis • Text in a query is analyzed like fields • Use the same analyzer that analyzed the particular field +field1:“quick brown fox” +(field2:“lazy dog” field2:“cozy cat”) quick brown fox lazy dog cozy cat 24
  • 26. Query Formation • Query parsing – A query parser in core code – Additional query parsers in contributed code • Or build query from the Lucene query classes 25
  • 27. Term Query • Matches documents with a particular term – Field – Text 26
  • 28. Term Range Query • Matches documents with any of the terms in a particular range – Field – Lowest term text – Highest term text – Include lowest term text? – Include highest term text? 27
  • 29. Prefix Query • Matches documents with any of the terms with a particular prefix – Field – Prefix 28
  • 30. Wildcard/Regex Query • Matches documents with any of the terms that match a particular pattern – Field – Pattern • Wildcard: * for 0+ characters, ? for 0-1 character • Regular expression • Pattern matching on individual terms only 29
  • 31. Fuzzy Query • Matches documents with any of the terms that are “similar” to a particular term – Levenshtein distance (“edit distance”): Number of character insertions, deletions or substitutions needed to transform one string into another • e.g. kitten -> sitten -> sittin -> sitting (3 edits) – Field – Text – Minimum similarity score 30
  • 32. Phrase Query • Matches documents with all the given words present and being “near” each other – Field – Terms – Slop • Number of “moves of words” permitted • Slop = 0 means exact phrase match required 31
  • 33. Boolean Query • Conceptually similar to boolean operators (“AND”, “OR”, “NOT”), but not identical • Why Not AND, OR, And NOT? – http://www.lucidimagination.com/blog/2011/12/ 28/why-not-and-or-and-not/ – In short, boolean operators do not handle > 2 clauses well 32
  • 34. Boolean Query • Three types of clauses – Must – Should – Must not • For a boolean query to match a document – All “must” clauses must match – All “must not” clauses must not match – At least one “must” or “should” clause must match 33
  • 35. Span Query • Similar to other queries, but matches spans • Span – particular place/part of a particular document – <document ID, start position, end position> tuple 34
  • 36. T0 = "it is what it is” 0 1 2 3 4 T1 = "what is it” 0 1 2 T2 = "it is a banana” 0 1 2 3 <doc ID, start pos., end pos.> “it is”: <0, 0, 2> <0, 3, 5> <2, 0, 2> 35
  • 37. Span Query • SpanTermQuery – Same as TermQuery, except your can build other span queries with it • SpanOrQuery – Matches spans that are matched by any of some span queries • SpanNotQuery – Matches spans that are matched by one span query but not the other span query 36
  • 38. spanTerm(apple) spanOr([apple, orange]) apple orange apple orange spanTerm(orange) spanNot(apple, orange) 37
  • 39. Span Query • SpanNearQuery – Matches spans that are within a certain “slop” of each other – Slop: max number of positions between spans – Can specify whether order matters 38
  • 40. the quick brown fox 2 1 0 1. spanNear([brown, fox, the, quick], slop = 4, inOrder = false) ✔ 2. spanNear([brown, fox, the, quick], slop = 3, inOrder = false) ✔ 3. spanNear([brown, fox, the, quick], slop = 2, inOrder = false) ✖ 4. spanNear([brown, fox, the, quick], slop = 3, inOrder = true) ✖ 5. spanNear([the, quick, brown, fox], slop = 3, inOrder = true) ✔ 39
  • 41. Filtering • A Filter narrows down the search result – Creates a set of document IDs – Decides what documents get processed further – Does not affect scoring, i.e. does not score/rank documents that pass the filter – Can be cached easily – Useful for access control, presets, etc. 40
  • 42. Notable Filter classes • TermsFilter – Allows documents with any of the given terms • TermRangeFilter – Filter version of TermRangeQuery • PrefixFilter – Filter version of PrefixQuery • QueryWrapperFilter – “Adapts” a query into a filter • CachingWrapperFilter – Cache the result of the wrapped filter 41
  • 43. Sorting • Score (default) • Index order • Field – Requires the field be indexed & not analyzed – Specify type (string, int, etc.) – Normal or reverse order – Single or multiple fields 42
  • 44. Interfacing Lucene with “Outside” • Embedding directly • Language bridge – E.g. PHP/Java Bridge • Web service – E.g. Jetty + your own request handler • Solr – Lucene + Jetty + lots of useful functionality 43
  • 45. Books • Lucene in Action, 2nd Edition – Written by 3 committers and PMC members – http://www.manning.com/hatcher3/ • Introduction to Information Retrieval – Not specific to Lucene, but about IR concepts – Free e-book – http://nlp.stanford.edu/IR-book/ 44
  • 46. Web Resources • Official Website – http://lucene.apache.org/ • StackOverflow – http://stackoverflow.com/questions/tagged/lucene • Mailing lists – http://lucene.apache.org/core/discussion.html • Blogs – http://www.lucidimagination.com/blog/ – http://blog.mikemccandless.com/ – http://lucene.grantingersoll.com/ 45
  • 47. Getting Started • Getting started – Download lucene-3.6.0.zip (or .tgz) – Add lucene-core-3.6.0.jar to your classpath – Consider using an IDE (e.g. Eclipse) – Luke (Lucene Index Toolbox) http://code.google.com/p/luke/ 46
  • 48. 47

Notas do Editor

  1. I bet this is exactly how many systems are handling search right now.Perhaps many systems do not think about how to sort the result and just throws back the result list to the user, without considering what should go first.
  2. Image the slowdown if your website goes from &quot;nobody besides our employees and friends use it&quot; to being &quot;the next FaceBook”.People loose interest in your application easily,if the first few things your search result present do not look exactly like what they are trying to find.
  3. Expand the inverted index we just saw.Positions start with zero.
  4. There are only so many words that people commonly use.You can hash the terms, organize them as a prefix tree, sort them and use binary search, and so on.For the purpose of deciding which documents match, you only need to store document IDs (integers).
  5. Extra info: determine how good of a match a document is to a query.Put the best matches near the topof the search result list.
  6. The highest-scored (most relevant) document is the first in the result list.
  7. In VSM, documents and queries are presented as vectors in an n-dimensional space, where n is the total number of unique terms in the document collection, and each dimension corresponds to a separate term. A vector&apos;s value in a particular dimension is not zero if the document or the query contains that term.Document vector closer to query vector = document more relevant to the query
  8. The term might be a common word that appears everywhere.
  9. Storing the field means that the original text is stored in the index; can retrieve it at search time.Indexing the fieldmeans that the field is made searchable.
  10. Token = term, at index time, with start/end position information, and not tied to a document already in the index.
  11. WhitespaceAnalyzer:whitespaces as separators;punctuations are a part of tokens. StopAnalyzer: non-letters as separators; makes everything lowercase; removes common stop-words like &quot;the”.StandardAnalyzer:sophisticated rules to handle punctuations, hyphens, etc.; recognizes (and avoids breaking up) e-mail addresses and internet hostnames.
  12. Character folding: turns the &quot;a&quot; with an accent mark above into an &quot;a&quot; without the accent markStemming: the words &quot;consistent&quot; and &quot;consistency&quot; have the same stem, which is &quot;consist”Synonyms: like &quot;country&quot; and &quot;nation”Shingles: “the quick”, “the brown”, “brown fox”; useful for searching text in Asian languages like Chinese and Japanese; reduces the number of unique terms in an index and reduces overhead.
  13. Offsets: character offsets of this token from the beginning of the field&apos;s textPosition increment: position of this token relative to the previous token; usually 1
  14. This query have clauses about 3 fields. So you analyze 3 pieces of text and get back 3 sets of tokens.A good practice is to use the same analyzer that analyzed the particular field that you are searching.
  15. Examples of range:January 1st to December 31st of 2012 (inclusive)1 to 10 (excluding 10)
  16. Your pattern describe a term, not a document, so you cannot put a phrase or a sentence in a pattern and expect the query to match that phrase or sentence.
  17. Minimum similarity score isbased on the editing distance.
  18. It takes two moves to swap two words in a phrase.
  19. Lucene does not have the standard boolean operators.
  20. Lucene has these instead (of the “standard” boolean operators).
  21. End position is actually one plus the position of the last term in the span
  22. This &quot;slop&quot; is different from the &quot;slop&quot; in Phrase Query.
  23. total number of positions between spans = 2 + 1 + 0 = 3The first two queries match this document because the slops are at least 3. The third query does not match, because the slope is less than 3. The fourth query does not match because even though the required slop is large enough, the query require all the spans to be in the given order, and the spans in this document are not. The fifth query matches because the given order matches the order of the spans in the document.
  24. CachingWrapperFilter good for filters that don’t change a lot, e.g. access restriction.
  25. Index order = order in which docs are added to the indexIndex and not analyzed = whole field as one token/term
  26. Embedding directly: good when the rest of your application is also in Java.In most uses cases, you would be dealing with Solr rather than Lucene directly. But you would still be indirectly using Lucene, and you can still benefit from understanding many of the things discussed in this session.
  27. Eclipse has many useful features such as setting up the classpath and compiling your code for you.
  28. It shows you what your index looks like and what fields and terms it has. You can look at individual documents, run queries, try out different analyzers.