Advanced Relevancy Ranking


Published on

Lucene and Solr provide a number of options for query parsing, and these are valuable tools for creating powerful search applications. This presentation given at the 2013 Lucene Revolution will review the role that advanced query parsing can play in building systems, including: Relevancy customization, taking input from user interface variables such as the position on a website or geographical indicators, which sources are to be searched and 3rd party data sources. Query parsing can also enhance data security. Best practices for building and maintaining complex query parsing rules will be discussed and illustrated. Chief Architect Paul Nelson provides this compelling presentation.
Search Technologies provides relevancy tuning services for Solr. For further information, see

Published in: Technology
  • Be the first to comment

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide

Advanced Relevancy Ranking

  1. 1. Advanced Relevancy RankingPaul NelsonChief Architect / Search Technologies
  2. 2. 2Search Technologies Overview• Formed June 2005• Over 100 employees and growing• Over 400 customers worldwide• Presence in US, Latin America, UK & Germany• Deep enterprise search expertise• Consistent revenue growth and profitability• Search Engine Independent
  3. 3. 3Lucene Relevancy: Simple Operators• term(A)  TF(A) * IDF(A)• Implemented with DefaultSimilarity / TermQuery• TF(A) = sqrt(termInDocCount)• IDF(A) = log(totalDocsInCollection/(docsWithTermCount+1)) + 1.0• and(A,B)  A * B• Implemented with BooleanQuery()• or(A, B)  A + B• Implemented with BooleanQuery()• max(A, B)  max(A, B)• Implemented with DisjunctionMaxQuery()3
  4. 4. 4Simple Operators - Exampleandor maxgeorge martha washington custis0.10 0.20 0.60 0.900.1 + 0.2 = 0.30 max(0, 0.9) = 0.900.3 * 0.9 = 0.27
  5. 5. 5Less Used Operators• boost(f, A)  (A * f)• Implemented with Query.setBoost(f)• constant(f, A)  if(A) then f else 0.0• Implemented with ConstantScoreQuery()• boostPlus(A, B)  if(A) then (A + B) else 0.0• Implemented with BooleanQuery()• boostMul(f, A, B)  if(B) then (A * f) else A• Implemented with BoostingQuery()5
  6. 6. 6Problem: Need for More Flexibility• Difficult / impossible to use all operators• Many not available in standard query parsers• Complex expressions = string manipulation• This is messy• Query construction is in the application layer• Your UI programmer is creating query expressions?• Seriously?• Hard to create and use new operators• Requires modifying query parsers - yuck6
  7. 7. 7SolrQuery Processing Language7UserInterfaceQPLEngineSearchQPLScript
  8. 8. 8Introducing: QPL• Query Processing Language• Domain Specific Language for Constructing Queries• Built on Groovy•• Solr Plug-Ins• Query Parser• Search Component• “The 4GL for Text Search Query Expressions”• Server-side Solr Access• Cores, Analyzers, Embedded Search, Results XML8
  9. 9. 9Solr Plug-Ins
  10. 10. 10QPL Configuration – solrconfig.xml<queryParser name="qpl"class="com.searchtechnologies.qpl.solr.QPLSolrQParserPlugin"><str name="scriptFile">parser.qpl</str><str name="defaultField">text</str></queryParser><searchComponent name="qplSearchFirst"class="com.searchtechnologies.qpl.solr.QPLSearchComponent"><str name="scriptFile">search.qpl</str><str name="defaultField">text</str><str name="isProcessScript">false</str></searchComponent>Query Parser Configuration:Search Component Configuration:
  11. 11. 11QPL Example #1myTerms = solr.tokenize(query);phraseQ = phrase(myTerms);andQ = and(myTerms);return phraseQ^3.0 | andQ^2.0 | orQ;Tokenize:Phrase Query:And Query:Put It All Together:orQ = (myTerms.size() <= 2) ? null :orMin( (myTerms.size()+1)/2, myTerms);Or Query:
  12. 12. 12Thesaurus Example #2myTerms = solr.tokenize(query);thes = Thesaurus.load("thesaurus.xml")thesQ = thes.expand(0.8f,solr.tokenizer("text"), myTerms);return and(thesQ);Tokenize:Load Thesaurus: (cached)Thesaurus Expansion:Put It All Together:Original Query: bathroom humor[or(bathroom, loo^0.8, wc^0.8), or(humor, jokes^0.8)]
  13. 13. 13More OperatorsBoolean Query Parser:pQ = parseQuery("(george or martha) near/5 washington")Relevancy Ranking Operators:q1 = boostPlus(query, optionalQ)q2 = boostMul(0.5, query, optionalQ)q3 = constant(0.5, query)Composite Queries:compQ = and(compositeMax(["title":1.5, "body":0.8],"george", "washington"))
  14. 14. 14News Feed Use Case14Order Documents Date1 markets+terms Today2 markets Today3 terms Today4 companies Today5 markets+terms Yesterday6 markets Yesterday7 terms Yesterday8 companies Yesterday9 markets, companies older
  15. 15. 15News Feed Use Case – Step 1markets = split(, "s*;s*")marketsQ = field("markets", or(markets));terms = solr.tokenize(query);termsQ = field("body",or(thesaurus.expand(0.9f, terms)))compIds = split(solr.compIds, "s*;s*")compIdsQ = field("companyIds", or(compIds))Segments:Terms:Companies:
  16. 16. 16News Feed Use Case – Step 2todayDate = sdf.format(c.getTime())todayQ = field("date_s",todayDate)c.add(Calendar.DAY_OF_MONTH, -1)yesterdayDate = sdf.format(c.getTime())yesterdayQ = field("date_s",yesterdayDate)Today:Yesterday:sdf = new SimpleDateFormat("yyyy-MM-dd")cal = Calendar.getInstance()
  17. 17. 17News Feed Use Case17Order Documents Date1 markets+terms Today2 markets Today3 terms Today4 companies Today5 markets+terms Yesterday6 markets Yesterday7 terms Yesterday8 companies Yesterday9 markets, companies older
  18. 18. 18News Feed Use Case – Step 3sq1 = constant(4.0, and(marketsQ, termsQ))sq2 = constant(3.0, marketsQ)sq3 = constant(2.0, termsQ)sq4 = constant(1.0, compIdsQ)subjectQ = max(sq1, sq2, sq3, sq4)tq1 = constant(10.0, todayQ)tq2 = constant(1.0, yesterdayQ)timeQ = max(tq1, tq2)recentQ = and(subjectQ, timeQ)Weighted Subject Queries:Weighted Time Queries:Put it All Together:return max(recentQ, or(marketsQ,compIdsQ)^0.01))
  19. 19. 19Embedded Search Example #1results =, or(qTerms), 50)subjectsQ = or(results*.subjectId)return field("title", and(qTerms)) | subjectsQ^0.9;Execute an Embedded Search:Create a query from the results:Put it all together:qTerms = solr.tokenize(qTerms);
  20. 20. 20Embedded Search Example #2results =, and(qTerms), 10)myList = solr.newList();myList.add("relatedCategories", results*.title);solr.addResponse(myList)Execute an Embedded Search:Create a Solr named list:Add it to the XML response:qTerms = solr.tokenize(qTerms);
  21. 21. 21Other Features• Embedded Grouping Queries• Oh yes they did!• Proximity operators• ADJ, NEAR/#, BEFORE/#• Reverse Lemmatizer• Prefers exact matches over variants• Transformer• Applies transformations recursively to query trees21
  22. 22. 22SolrQuery Processing Language22UserInterfaceQPLEngineSearchData as enteredby user BooleanQuery ExpressionQPLScriptApplicationDev TeamSearch Team
  23. 23. 23SolrQPL: Using External Sources to Build Queries23UserInterfaceQPLEngineSearchQPLScriptRDBMSOtherIndexesThesaurus
  24. 24. CONTACTPaul