Empowering Customers to Self Solve - A Findability Journey - Manikandan Sivanesan & Rutvij Vyas, Red Hat
1. Empowering Customers to
Self Solve - A Findability
Journey
Manikandan Sivanesan
Senior Software Engineer, Red Hat
@manisnesan
#Activate18 #ActivateSearch
9. Why self-solve matters for Business?
“Self-service offers companies a tantalizing opportunity to reduce
spending, often drastically. The cost of a do-it-yourself
transaction is measured in pennies, while the average cost of a
live service interaction (phone, e-mail, or webchat) is more
than $7 for a B2C company and more than $13 for a B2B
company.”
- HBR Business Review 2017
14. Red Hat Customer Portal
• Value of Red Hat Subscription
• FAQ, Docs, Knowledge Base
• Certifications
• Product Security
• Award Winning Support 24x7
• Onsite Search and Apps
18. Categories of queries
• Low intent High Frequency
• rhel, ansible, openshift - Product based
• sosreport, openjdk - technology/component specific
• High intent Low Frequency
• open files hard limit
• multipath configuration
• Error messages
• org.jboss.ejb3.annotation.TransactionTimeout
• NFS: nfs4_discover_server_trunking unhandled error -22. Exiting
with error EIO
19. Product Entity Recognition
• Customer Product Vocabulary vs Official
• rhel 7, red hat linux , linux, red hat enterprise linux
• SynonymGraphFilterFactory
• Build our own synonyms list
• Manually Maintained
• Acronyms
• Hard to identify components or technology specific words
20. Word2Vec for Product Inference
• Query Preprocessing
• Word2Vec - Learning representation
• “relatedness” between words
• q = kubernetes
• "bq": “boostProduct:openshift^4"
• q = karaf
• “bq”: "boostProduct:fuse^4"
• Source the Product Detector Synonyms
• neutron, packstack, director => OpenStack
21. Tier Model Approach
• Fields from Content mapped to Tiers
• Search across multiple tiers
• Different boosts for different tiers
• Common parlance with Content Writers
• qf = tierA^10.0 tierB^5.5 tierc^4.5 tierD^3.5
Content/Tier Tier A Tier B Tier C Tier D
Knowledge Base title issue resolution resolution_stripPunc
Documentation chapterTitle sectionTitles content content_stripPunc
Vulnerability title content content_stripPunc
24. Problem Query
● Product, Version
● Summary
● Description
Challenges
• Vague
• Important tokens
• Diagnostics Info
Don’t Force the customer
25. Only the important tokens
• edismax query parser
○ stopword removal in phrase query
•Missing bigram/trigram with leading/following stopword
• Stopwords handling before sending to Solr
• Improved phrase matching
((capsule does not have installation media)^5.0) => (title:"instal
media 5.0"~2 | body:"instal media 5.0"~2)~0.1
(title:"media 5.0 instal"~2 | body:"media 5.0 instal"~2)~0.1
26. Relevance Signals
• Customer Feedback
• User Clicks
• Fusion - Signals and Aggregation
• Associate Feedback
• Case resolution
• Popularity based on linked knowledge content
• Linked Solutions from similar cases
"boost": "if(exists(caseCount_365),sum(1,div(log(map(caseCount_365,0,0,1)),log(8.0))),1)"
27. Evaluation framework
• Leverages past closed cases & linked knowledge content
• Exact answer to a particular problem
• Search effectiveness
• % of linked content in Top 1, 3, 6 ranks of results
• Tuning
• Best parameters value chosen based on maximized score
• Hard to tune in practice
28. Product Specific Tuning
• Boost factors determined by Major Products
• Link Popularity not relevant for emerging products
• Individual Tuning for Top products
29. Journey Ahead - Product Specific LTR
• RankNet
• Training Dataset
• Neural Network Model
• Available in Solr 7.3
30. Results
• Self-Solve based Home Page
• Increased traffic
• Customers are motivated to self-solve
• A/B testing
• 7% decrease in case creation rate for experiment group
31. Faster Case Resolution
• Speciality Case Routing
• Rule Engine -> Mahout Classifier -> Deep Learning
• Accuracy ~ 85%
• Language Detection
• Case Routing based on Customer Locale
• Case text in multiple languages
• FastText based Language Classifier
• Accuracy ~ 99%
32. Credits
• Charles Sanders
• Tom Butt
• Michael Alcorn
• Scot Floess
• Diego Fernandez
• JP Sherman
• Search Team
• AIR Team