webinar: simpler semantic search with solr
TRANSCRIPT
Ted Sullivan
Simpler Semantic Search
lucidworks.com
Senior Solutions Architect
Building Search Applications
Search is about Technology & Language
• These are difficult but also different problems
• Solving the “language problem” requires that we understand how language is used in search
• We understand language at the semantic level - where “meaning” or intent lives
• Search Engines deal with language at the syntactic level
• Most problems relating to search quality stem from this basic “disconnect” – the “what” vs “what words” dichotomy
Better^
Technology – Horizontal Concerns
Search applications share these requirements with other information retrieval systems
• Performance – returning results in HTT (Human Tolerable Time)
• Scalability – being able to search “billions and billions”of documents serving thousands or tens of thousands of users at a time.
• Reliability – fault tolerance, fail-over, redundancy
• Maintainability – easy to upgrade, search index can be kept current in the face of rapidly changing content.
• Usability – User Experience is critical to success. UI and UX Mobile Technology Is a Game Changer here!!!
Language – Vertical Concerns
These requirements are more specific to search systems.
• Accuracy – returning the “correct” results.
• Precision – few false positives
• Recall – few false negatives
• Relevance – returning the “best” results at the top
Returning the wrong results very fast is notnecessarily a good thing. Returning too manyresults can affect performance.
Time flies like an arrow
Fruit flies like a banana
Our mental image for the second sentence depends on how we “parse” it. It depends on what the subject noun or noun phrase is.
The subject can be “fruit” or “fruit flies”. This decision changes the verb which is either “flies” or “like” respectively.
Fruit flies like a banana
Fruit flies like a banana
We can do this because we know that both “fruit” and “fruit flies” represent single concepts – even though “fruit flies” is two words – i.e. a “noun phrase”.
Fruit flies like a banana
Fruit flies like a banana
Search algorithms and semantics
Tokenization plus vector mathematics(TF/IDF or one of its cousins) – “bag-of-words” Algorithmic tweaks – enhanced bag-of-words:
1. Some fields are more relevant than others
2. Hitting on more terms in the query is better than hitting on fewer (token scores are summed)
3. The nearer the query terms are to each other in the document the better – same order as query is best
4. Getting 0 results provides no feedback – OR is safer than AND (we already have “fuzzy” & with bullet (2)
Problem: Search engines don’t understand semantics
Better Search: Detecting Noun Phrases
Can algorithms be used to detect noun phrases?
Yes, but not perfectly and may need too much CPU at query-time
Another way is to use knowledge bases – a lot of extra work, but in some cases – we already have one - the search index itself!
Better Search: Detecting Noun Phrases
The basic technique is called “autophrasing” – recognizing when more than one word represents just one thing.
Autophrasing – uses an extra knowledge-base file “autophrases.txt”
Query Autofiltering – uses the phrases that are stored as metadata values in the index.
Multi-term Synonym Problem
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 !)
Sausagization
From Mike McCandless blog: Changing Bits: Lucene's TokenStreams are actually graphs!
• 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.
http://blog.mikemccandless.com/2012/04/lucenes-tokenstreams-are-actually.html
Multi-term Synonym Demo
autophrases.txt
new yorknew york state empire state new york citynew york new yorkbig apple ny nycity of new yorkstate of new yorkny state
synonyms.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
Multi-term Synonym Demo
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
/select /autophrase
Multi-term Synonym Demo
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
/select /autophrase
Query Autofiltering
Content Tagging and Intelligent Query Filtering. Using the search index itself as the knowledge source:
Search Index
Content ContentTagging
Auto FilteringQuery The Answer
Lucene FieldCache “In Action”
Standard “Inverted Index” (Lucene itself): • Show all documents that have this term value in this field
• Used to get initial set of search result IDs
Uninverted or Forward Index (FieldCache): • Show all term values that have been indexed in this field
• Can lookup term value for a doc ID
• Used to facet and get display values for doc IDs.
Query Autofiltering Implementation
Use Lucene FieldCache to build a map of field values to field names (of string fields)
Add synonym mappings from synonyms.txt and stemming to this value(s) -> field(s) map
Use this map to discover noun phrases in the query that correspond to field values in the index – longest contiguous phrase wins
Build filter or boost queries based on these discovered mappings
QueryAutoFilteringComponent
Solr SearchComponent
github: https://github.com/LucidWorks/query-autofiltering-component
JIRA: SOLR-7539
<requestHandler name="/autofilter" class="solr.SearchHandler"> <lst name="defaults"> <str name="echoParams">explicit</str> <int name="rows">10</int> <str name="df">text</str> </lst> <arr name="first-components"> <str>queryAutofiltering</str> </arr> </requestHandler>
<searchComponent name=“queryAutofiltering" class="org.apache.solr.handler.component.QueryAutoFilteringComponent" />
QueryAutofiltering DemoHypothetical eCommerce App for a Fictional department store
• Metadata has Noun Phrases!
<doc> <field name="id">95</field> <field name="product_type">sweat shirt</field> <field name="product_category">shirt</field> <field name="style">V neck</field> <field name="style">short sleeve</field> <field name="brand">J Crew</field> <field name="color">grey</field> <field name="material">cotton</field> <field name="consumer_type">womens</field></doc> <doc> <field name="id">154</field> <field name="product_type">crew socks</field> <field name="product_category">socks</field> <field name="color">white</field> <field name="brand">Joe Boxer</field> <field name="consumer_type">mens</field></doc><doc> <field name="id">17</field> <field name="product_type">boxer shorts</field> <field name="product_category">underwear</field> <field name="color">white</field> <field name="brand">Fruit of the Loom</field> <field name="consumer_type">mens</field></doc>
Query Autofiltering – Basic Behaviorq = red socks -> fq=color:red&fq=product_type:socks
or bq=(color:red AND product_type:socks)^20
q = Red Lion socks -> fq=brand:”Red Lion”&fq=product_type:socks
q = scarlet Chaise Lounge -> color:red AND product_type:”Lounge Chair”
q = white dress shirts -> color:white AND product_type:”dress shirt”
Dealing With “Unstructured” Text
This term ITSELF is evidence that we think of language as unstructured when we know that it actually is not - It HAS to have structure or we couldn’t communicate very well.
“The Lady Is A Tramp” vs “Lady And The Tramp”
Dealing with unstructured text means better handling of phrases.
Little words – like “if” can have big meaning!
Classification Technologies
Machine Learning • Automated vs Semi-Automated
Natural Language Processing (NLP) • Parts Of Speech
Taxonomy / Ontology • Relationships
• Handles Phrases naturally
• Knows what is what and what is related to what!
Ontologies Designed for Search
Category Nodes – ‘parent’ nodes that can have child nodes, including:
• Sub Categories
• Evidence Nodes
Evidence Node – tend to be a leaf nodes (with no children) and contain 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 they traverse the graph – so that the nearest category gets the strongest ‘vote’.
Fortune 100 Companies
Energy • Financial Services
• Investment Banks
• Commercial Banks
Health Care • Health Insurance
• HMO
• Medical Devices
• Pharmaceuticals
Hospitality
Manufacturing • Aircraft
• Automobiles
• Electrical Equipment
Corporations • US
• British
• Chinese
• French
• German
• Japanese
• Russian
• +
Fortune 100 Companies
Energy • Financial Services
• Investment Banks
• Commercial Banks
Health Care • Health Insurance
• HMO
• Medical Devices
• Pharmaceuticals
Hospitality
Manufacturing • Aircraft
• Automobiles
• Electrical Equipment
Corporations • US
• British
• Chinese
• French
• German
• Japanese
• Russian
• +
The Basic Search “Use Case”
Traditional - Brief display – snippeting,hyperlinks and paging
• Faceted Navigation
• Highlighting
• Need To RETHINK for Mobile!!!
Query Formulation
–> Result Inspection
–> Query Refinement
Shortening The Loop
Query Suggestion (aka autocomplete, typeahead)
• “Predictive” search
• Single field restriction
Recommendation • Query – result – click – store – aggregate
• Boosting results or Suggesting queries
Best Bets (Query Elevation) – i.e. Punting • Spotlighting
• Making it dynamic
Faceting • Takes advantage of classification tagging
• Can be used to generate multi-field phrases for suggestion
Inferential Search • “I’m Feeling Lucky”
• Query Autofiltering
Enhanced Search: Pipelines
Document and Query Pre-Processing
Internal to Solr:
• Update Request Processor
• Data Import Handler (DIH)
• Search Component Chain
Big Data = Big Problem or just a Big Opportunity:
• Hadoop – Solr
• Spark – Solr
• Morphlines
External to Solr: • Custom ETL + SolrJ Integration
• Apache UIMA *
• DIH Client (SOLR-7188)
• Lucidworks Fusion
• Modular Informatic Designs framework (coming soon to Open Source?)
Index Pipelines – Good Ole ETL + ______
Annotations!Subject - Verb - Object
Entity Extractors – Identify Subject and Object (noun phrases)
Annotations – mark locations of entities in document
Discover Facts from Semantic Patterns • $Person joined $Company
• $Drug is used to treat $Disease
• $Company acquired $Company
• $Person wrote $Song
Watson used IBM’s (now Apache’s) UIMA (+40,000 PC’s)
Jeopardy is a “guess subject given object and verb - posed as a question” – game
Who Needs Query Pipelines?
Who, What, Where, When: • Security Filtering - Entitlements
• Dynamic Boost Block based on Preferences, Search History
• Geo Filtering – IP to geolocation
• Content Spotlighting based on time, place and search history
• Query Introspection – Infer User Intent
Lucidworks Fusion: Pipelines Proliferate
Documents and Queries are dynamic Metadata Objects • PipelineDocument QueryRequestAndResponse respectively
Lots of Stages – more coming with every release • Metadata -> metadata – lookup, clone, map, join
• Content -> metadata – extract, transform, classify
Index Pipelines: One-Way Query Pipelines: Round-Trip • Both pre- and post-Query filtering opportunities
Connectoror Query Stage Stage Stage Stage Solr Cloud