The well tempered search application
Download as PPTX, PDF7 likes1,516 views
This document discusses the challenges of improving search applications by addressing the disconnect between how humans think semantically and how search engines operate syntactically. It emphasizes the need for better algorithms and the importance of continually refining search engines with semantic intelligence to enhance user experiences. The author also outlines various techniques and solutions, such as autophrasing and query filtering, to overcome limitations and improve search effectiveness.
![Variation II: The Red Sofa Problem
{
"response":{"numFound":3,"start":0,"docs":[
{
"color":"red",
”text":"This is the red sofa example. Please find with 'red sofa' query.",
{
"color":"red",
”text":"This is a red beach ball. It is red in color but is not something that you
should not sit on because you would tend to roll off.",
{
"color":"blue",
”text":"This is a blue sofa, it should only hit on sofas that are blue in color."]
}}
OOTB – q=red sofa is interpreted as text:red text:sofa (default OR)
http://localhost:8983/solr/collection1/select?q=red+sofa&wt=json](https://image.slidesharecdn.com/thewelltemperedsearchapplication-150219065138-conversion-gate01/85/The-well-tempered-search-application-20-320.jpg)
![Query Autofiltering
{
"response":{"numFound":1,"start":0,"docs":[
{
"id":"1",
"color":"red",
"description":"This is the red sofa example. Please find with 'red sofa' query."]
}
http://localhost:8983/solr/collection1/infer?q=red+sofa&wt=json
Now search for “red sofa” only returns ….. red sofas!
But – is this too “brute force”? The takeaway is that using
the search index AS a knowledge store can be very powerful!](https://image.slidesharecdn.com/thewelltemperedsearchapplication-150219065138-conversion-gate01/85/The-well-tempered-search-application-24-320.jpg)
The well tempered search application
- 1. The Well-Tempered Search Application Variations on a Theme: Why does my search app suck, and what can I do about it? Ted Sullivan – (old Phuddy Duddy) Senior (very much so I’m afraid) Solutions (I hope) Architect (and sometime plumber) Lucidworks Technical Services
- 2. Our Basic Premises (Premisi?) • Lemma 1: Search Applications use algorithms that make finding chunks of text within large datasets possible in HTT (human-tolerable time). • Lemma 2: These algorithms work by breaking text into primitive components and building up a search “experience” from that. • Lemma 3: Lemma 2 is not sufficient to achieve Lemma 1.
- 3. The Basic Disconnect • Text can be analyzed at the level of tokens (syntax) and at the level of meaning (semantics). • We think one way (semantics), search engines think another (syntax – i.e. token order). • How do we bridge the gap? … More clever algorithms!
- 4. Art and Science • We need to be intelligent curators of these algorithms. Craftsmen (craftswomen?) that think of these as tools with a specific purpose. • Like any good craftsperson – we need a wide array of tools to get the job done (well almost).
- 5. When is my search app done? • Quick answer: NEVER (ain’t consultin’ great?) • Long answer – As long as it is continues to improve, like fine wine or bourbon, you are on the path to enlightenment. • How do you get there grasshopper? Add semantic intelligence to the engine!
- 6. Search cannot be shrink-wrapped!! What have we got for Donny behind Curtain #1 Jay? Well Monty - Heeeeeeeeeeeerrrrrrrreeeeesssss the Google … SEARCH Appliance!!!!
- 7. Search cannot be shrink-wrapped!! What have we got for Donny behind Curtain #1 Jay? Well Monty - Heeeeeeeeeeeerrrrrrrreeeeesssss the Google … SEARCH Appliance!!!!* Sorry Donny – It’s a ZONK! * but Google Web Search has some Serious Mojo!
- 8. Prelude part 1– The basic problem The inverted index and “bag-of-words” search: The red fox jumped over the fence. Time flies like an arrow. Fruit flies like a banana. the 1,6 red 2 fox 3 jumped 4 over 5 fence 7 flies 2,7 like 3,8
- 9. Prelude part B – The Tried and True • Phrase and Proximity boosting and “Slop” • Synonyms and stop words • Stemming or Lemmatization • Autocomplete • Best Bets / Landing Pages – the sledgehammer • Spell check – spell suggest – aka the warm fuzzies.
- 10. Fugue - Subject or Exposition Search engines need more ‘semantic awareness’ or at least the illusion of this. There is a heavy duty solution called Artificial Intelligence – which except in the fertile imagination of Hollywood screenwriters, is not there yet. So we need to fake it just a bit.
- 11. Theme and Variations I autophrasing and the red sofa Theme: When multiple words mean just one thing. Fuzzy way: Boosting phrases (proximity and phrase slop) - pushes false positives down – i.e. out of the limelight - i.e. - shoves ‘em under the rug This encounters a problem with faceted search Like the eye of Sauron in LOTR or Santa Claus, the faceting engine SEES ALL (sins)! Brake Pads example: hit on things that have ‘brake’ (like children’s stroller brakes) and ‘pads’ – like mattress pads.
- 12. Variation I: Autophrasing AutophrasingTokenFilter tells Lucene not to tokenize when a noun phrase represents a single thing - by providing a flat list of phrases. Creates one-to-one token mapping that Lucene prefers because it avoids the “sausagization” problem. https://github.com/LucidWorks/auto-phrase-tokenfilter
- 13. income tax refund income tax tax refund “income tax” is not income. A “tax refund” is not a tax. Solution: Autophrasing + synonym mapping income tax => tax tax refund => refund Autophrasing Example
- 14. Autophrasing Setup autophrases.txt: income tax tax refund tax rebate sales tax property tax synonyms.txt income_tax,property_tax,sales_tax,tax tax_refund,refund,rebate,tax_rebate <fieldType name="text_autophrase" class="solr.TextField" positionIncrementGap="100"> <analyzer type="index"> <tokenizer class="solr.StandardTokenizerFactory"/> <filter class="solr.LowerCaseFilterFactory"/> <filter class="com.lucidworks.analysis.AutoPhrasingTokenFilterFactory" phrases="autophrases.txt" includeTokens="true” replaceWhitespaceWith="_" /> <filter class="solr.StopFilterFactory" ignoreCase="true" words="stopwords.txt" enablePositionIncrements="true" /> <filter class="solr.PorterStemFilterFactory"/> </analyzer> </fieldType>
- 15. Multi-term synonym problem • New York, New York – it’s a HELLOVA town! Subject was inspired by an old JIRA ticket: Lucene-1622 “if multi-word synonyms are indexed together with the original token stream (at overlapping positions), then a query for a partial synonym sequence (e.g., “big” in the synonym “big apple” for “new york city”) causes the document to match” (or “apple” which will hit on my blog post if you crawl lucidworks.com !)
- 16. This means certain phrase queries should match but don't (e.g.: "hotspot is down"), and other phrase queries shouldn't match but do (e.g.: "fast hotspot fi"). Other cases do work correctly (e.g.: "fast hotspot"). We refer to this "lossy serialization" as sausagization, because the incoming graph is unexpectedly turned from a correct word lattice into an incorrect sausage. This limitation is challenging to fix: it requires changing the index format (and Codec APIs) to store an additional int position length per position, and then fixing positional queries to respect this value. Sausagization: from Mike McCandless blog Changing Bits: Lucene's TokenStreams are actually graphs! http://blog.mikemccandless.com/2012/04/lucenes-tokenstreams-are-actually.html
- 17. Multi-term synonym demo new york new york state empire state new york city new york new york big apple ny ny city of new york state of new york ny state autophrases.txt new_york => new_york_state,new_york_city,big_apple, new_york_new_york,ny_ny,nyc,empire_state,ny_state, state_of_new_york new_york_state,empire_state,ny_state, state_of_new_york new_york_city,big_apple,new_york_new_york,ny_ny,nyc, city_of_new_york synonyms.txt
- 18. This document is about new york state. This document is about new york city. There is a lot going on in NYC. I heart the big apple. The empire state is a great state. New York, New York is a hellova town. I am a native of the great state of New York. new york new york city new york state Multi-term synonym demo /select /autophrase
- 19. This document is about new york state. This document is about new york city. There is a lot going on in NYC. I heart the big apple. The empire state is a great state. New York, New York is a hellova town. I am a native of the great state of New York. empire state Multi-term synonym demo /select /autophrase (Even a blind squirrel finds a nut once in a while)
- 20. Variation II: The Red Sofa Problem { "response":{"numFound":3,"start":0,"docs":[ { "color":"red", ”text":"This is the red sofa example. Please find with 'red sofa' query.", { "color":"red", ”text":"This is a red beach ball. It is red in color but is not something that you should not sit on because you would tend to roll off.", { "color":"blue", ”text":"This is a blue sofa, it should only hit on sofas that are blue in color."] }} OOTB – q=red sofa is interpreted as text:red text:sofa (default OR) http://localhost:8983/solr/collection1/select?q=red+sofa&wt=json
- 21. Closing the Loop: Content Tagging and Intelligent Query Filtering Using the search index itself as the knowledge source:
- 22. Solution for the Red Sofa problem Query Autofiltering: Search Index driven query introspection / query rewriting:
- 23. Lucene FieldCache Magic Lucene FieldCache (to be renamed UninvertedIndex in Lucene 5.0) Inverted Index: Show all documents that have this term value in this field. Uninverted or Forward Index: Show all term values that have been indexed in this field. SolrIndexSearcher searcher = rb.req.getSearcher(); SortedDocValues fieldValues = FieldCache.DEFAULT.getTermsIndex( searcher.getAtomicReader( ), categoryField ); … StringTokenizer strtok = new StringTokenizer ( query, " .,:;"'" ); while (strtok.hasMoreTokens( ) ) { String tok = strtok.nextToken( ).toLowerCase( ); BytesRef key = new BytesRef( tok.getBytes() ); if (fieldValues.lookupTerm( key ) >= 0) {
- 24. Query Autofiltering { "response":{"numFound":1,"start":0,"docs":[ { "id":"1", "color":"red", "description":"This is the red sofa example. Please find with 'red sofa' query."] } http://localhost:8983/solr/collection1/infer?q=red+sofa&wt=json Now search for “red sofa” only returns ….. red sofas! But – is this too “brute force”? The takeaway is that using the search index AS a knowledge store can be very powerful!
- 25. Architecture: its all about Plumbing • Pipelines for every occasion. Indexing Pipelines – good ‘ole ETL - Content enrichment, tagging - Metadata cleanup Query Pipelines – identification, query preprocessing - introspection One is the “hand” the other, the “glove”
- 26. Index Pipelines Lots of choices here: • Internal to Solr – DIH, UpdateRequestProcessor Pros and cons • External – Morphlines, Open Pipeline, Flume, Spark, Hadoop, Custom SolrJ • Lucidworks Fusion
- 27. Entity and Fact Extraction Entities: Things, Locations, Dates, People, Organizations, Concepts Entity Relationships Company was acquired by Company Drug cures Disease Person likes Pizza Annotation Pipelines (UIMA, Lucidworks Fusion): Entity Extraction followed by Fact Extraction Pattern method: $Drug is used to treat $Condition Parts of Speech (POS) analysis Subject Predicate Object
- 28. Theme and Variations II The Classification Wars • Machine Learning or Taxonomy – is it a Floor Wax or a Dessert Topping? Answer: It’s a floor wax AND a dessert topping! Its delicious and just look at that shine!
- 29. Machine Learning Use mathematical vector-crunching algorithms like Latent Dirichlet Allocation (LDA), Bayesian Inference, Maximum Entropy, log likelihood, Support Vector Machines (SVM) etc., to find patterns and to associate those patterns with concepts. Can be supervised (i.e. given a training set) or unsupervised (the algorithm just finds clusters). Supervised learning are called semi-automatic classifiers. Check out Taming Text by Ingersoll, Morton and Farris (Manning)
- 30. Machine Learning In LucidworksFusion Training Data NLP Trainer Stage NLP Model Test Data NLP Classifier Stage Classified Documents
- 31. Taxonomy or Ontology “Knowledge graphs” that relate things and concepts to each other either hierarchically or associatively. Pros: Works without large amounts of content to analyze Encapsulates the knowledge of human subject matter experts Cons: Often not well designed for search (mixes semantic relationship types / organizational logic) Requires curation by subject matter experts whose time is costly
- 32. Taxonomies Designed for Search Category nodes and Evidence nodes Category Node: A ‘parent’ node Can have child nodes that are: Sub Categories Evidence Nodes Evidence Node: Tends to be a leaf node (no children) Contains keyterms (synonyms) May contain “rules” e.g. (if contains term a and term b but not term c) Evidence Nodes can have more than one category node parent Hits on Evidence Nodes add to the cumulative score of a Category Node. Scores can be diluted as the accumulate up the hierarchy – so that the nearest category gets the strongest ‘vote’.
- 33. US Corporations Foreign Corporations British Chinese French German Japanese Russian etc. Fortune 100 Companies Energy Financial Services Investment Banks Commercial Banks Health Care Health Insurance HMO Medical Devices Pharmaceuticals Hospitality Manufacturing Aircraft Automobiles Electrical Equipment
- 34. Ford, GM, Chrysler Fortune 100 Companies Energy Financial Services Investment Banks Commercial Banks Health Care Health Insurance HMO Medical Devices Pharmaceuticals Hospitality Manufacturing Aircraft Automobiles Electrical Equipment US Corporations Foreign Corporations British Chinese French German Japanese Russian etc.
- 35. Ford, GM, Chrysler,Toyota,BMW Fortune 100 Companies Energy Financial Services Investment Banks Commercial Banks Health Care Health Insurance HMO Medical Devices Pharmaceuticals Hospitality Manufacturing Aircraft Automobiles Electrical Equipment US Corporations Foreign Corporations British Chinese French German Japanese Russian etc.
- 36. Fortune 100 Companies Energy Financial Services Investment Banks Commercial Banks Health Care Health Insurance HMO Medical Devices Pharmaceuticals Hospitality Manufacturing Aircraft Automobiles Electrical Equipment Ford, GM, Chrysler,Toyota,BMW GE, Boeing US Corporations Foreign Corporations British Chinese French German Japanese Russian etc.
- 37. Fortune 100 Companies Energy Financial Services Investment Banks Commercial Banks Health Care Health Insurance HMO Medical Devices Pharmaceuticals Hospitality Manufacturing Aircraft Automobiles Electrical Equipment Ford, GM, Chrysler,Toyota,BMW GE, Boeing Bank of America, Hyatt US Corporations Foreign Corporations British Chinese French German Japanese Russian etc.
- 38. Query Pipelines The ‘Wh’ Words: Who, What, When, Where Who are they (authentication)? What can they see (security - authorization)? When can they see it (entitlement)? What are they interested in (personalization / recommendation)? Where are they now (location)?
- 39. Query Pipelines Inferential Search Query introspection -> Query modification. Query Autofiltering Are you feeling lucky today? Topic boosting / spotlighting Use ML to detect the topic, then boost and/or spotlight results tagged this way. Use a specialized collection to store ‘facet knowledge’
- 40. The Art of the Fugue: Inferential Search • Infer what the user is looking for and give them that • Clever software infers meaning aka query “intent” • When we do this right, it appears to be magic!
- 41. Machine Learning Drives Query Introspection Training Data NLP Trainer Stage NLP Model Test Data NLP Classifier Stage Classified Documents
- 42. Machine Learning models can drive Query Introspection NL Query NLP Query Stage NLP Model Tagged Query Landing Page Boost Documents
- 43. Da Capo al Coda • Killer search apps are crafted from fine ingredients and like fine whiskey will get better with age - if you are paying attention to ‘what’ your users are looking for. • Putting the pieces together requires an understanding of ‘what’ things, independent of what words they use to describe it.
- 44. Thanks for your attention! Ted Sullivan Lucidworks, Technical Services ted.sullivan@lucidworks.com Skype: ted.sullivan5 LinkedIn Metuchen, New Jersey (You gotta problem with that?)
Editor's Notes
- #7: ( * but the Google web app has some serious Mojo!)
- #8: ( * but the Google web app has some serious Mojo!)
- #10: Suga “Plague” story here – on Spell Check
- #14: Sales tax is a type of tax, not a type of sale.
- #29: Cue – Utube Video here
- #40: Who are they? – Security – what can they see? Personalization – what do they like?