stop stupid fuzzy searches Table of contents 01 Fuzzy search 02 - - PowerPoint PPT Presentation

stop stupid fuzzy searches

01 Fuzzy search 02 Smart Query Rewriting 03 Conclusion 04 Surprise

Table of contents

01 Fuzzy search

37% 23% 9% 5% 15% 11% 26% 20% 10% 12% 18% 14% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Edit distance 0 Edit distance 1 Edit distance 2 Edit distance >2 Singluar & Plural Decomposition Frequency Value/Search

Why we need it / Distribution of spelling errors

10% 25% 27% 10% 17% 11% 6% 23% 34% 7% 16% 14% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Insert Delete Replace Transpose Singular & Plural Decomposition Desktop Mobile

Why we need it / Distribution of spelling errors by device type

Causes of spelling errors

format phonetic typo decomposition

• spannbettlaken

spannbettlacken spanbettlaken spannbettllaken spammbettlaken Spann bettlaken Bettlaken zum spannen

1% 3% 13% 9% 7% 1% 4% 1%

spann-bettlaken

spannbettlaken

61% 4%

22% 8% 5%

…42 additional spellings

83 56 50 47 43

Query-Intent Error-type query Result size

generates

835

candidates

How it works

generates

~650k

candidates GET catalog/products/_search { “query”: { “fuzzy”: { “title”: { “value”: “spannbettlacken”, “fuzziness”: 1 } } } }

EditDistance 1

GET catalog/products/_search { “query”: { “fuzzy”: { “title”: { “value”: “spannbettlacken”, “fuzziness”: 2 } } } }

EditDistance 2

Resulting in / high recall but low precision

0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1 0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0

Precision (PREC) Recall (TPR)

Resulting in / low search throughput ~0.1 seconds for spelling a short word

5000 10000 15000 20000 25000 30000 35000 1 2 3 4 5 6

Searches per Second Query Terms

• r - term

and - term

• r - fuzzy 2

and - fuzzy 2

• +

Searches for all possible candidates inside a given edit- distance Natively implemented in Elasticsearch and Lucene Increased CPU usage and query response time Inconsistent and not always relevant results Skewed search analytics

Observations

02 Smart Query rewriting

MAKE FUZZY SEARCH AS FAST, EASY AND RELEVANT AS EXACT SEARCH

spannbettlaken spann-bettlaken spannbettlacken schpanbettlaken spannbettllaken spammbettlaken spanmbettlaken spannbettlaken

Cluster similar Queries Test & Select MasterQuery

spannbettlaken

MasterQuery Search Engine

Our Solution / smart query rewrites

Cluster similar Queries

Based on deep learning & crafted algorithms we clean and cluster queries with the same meaning We use the concept of controlled precision reduction

Exact Match Fingerprint Lemmatization & Phonems Fuzzy Match

Our Solution / smart query rewrites

spannbettlaken spann-bettlaken spannbettlacken schpanbettlaken spannbettllaken spammbettlaken spanmbettlaken spannbettlaken

Test & Select MasterQuery Based on tracking KPIs and deep learning and global parameter

• ptimization we

test & select the query which maximises the balance between the search result interaction probability and the economic outcome

Our Solution / smart query rewrites

spannbettlaken spann-bettlaken spannbettlacken schpanbettlaken spannbettllaken spammbettlaken spanmbettlaken spannbettlaken

CXP search|hub / Query Intelligence Platform

Frontend Search Endpoint

High performance Caching & Logging Da Data|hub Semantic Query Parsing Site Search Analytics Guided Selling Personalization … Solr Elasticsearch FACT-Finder Fredhopper Celebros Algolia ACS

Search Engine

Sm Smart|Quer uery Query Segmentation Query Scoping

03 Conclusion

Impact – top-10 ecom player A

90% 100% 110% 120% 130% 140% Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec w/o smart|query w smart|query

Uses an already a highly optimized state-of-the-art eCommerce Search solution

Impact – top-50 ecom player B

90% 100% 110% 120% 130% 140% Apr May Jun Jul Aug Sep Oct Nov Dec Jan Feb Mar w/o smart|query w smart|query

Uses an optimized SolR implementation

Resulting in / High recall & high precision

0,97 0,98 0,99 1 0,5 0,55 0,6 0,65 0,7 0,75 0,8 0,85 0,9 0,95 1 500.000 1.000.000 1.500.000 2.000.000 2.500.000 3.000.000 3.500.000 4.000.000 4.500.000 5.000.000

Precision (PREC) Recall (TPR) Queries

Recall (TPR) Precision (PREC)

Resulting in / insane query performance

~0.00005 seconds for spelling a short word – 80 ops/ms

5000 10000 15000 20000 25000 30000 35000 1 2 3 4 5 6

Searches per Seconds search|hub & Elastic Query Terms

• r - term

and - term

• r - fuzzy 2

and - fuzzy 2

more relevant results consistent results reduced manual effort for curated search results save CPU usage improved query response time consistent site search analytics additional complexity

Observations

+

04 Surprise

CXP smart|query- PreDictLib

fast & accurate spell correction at scale

search|hub -PreDictLib

fast & accurate spell correction at scale Qui Quick Highl hlight hts:

§ extremely fast & constant index access § truly language independent edit distance § ability to add records to the index at runtime without performance decrease based on one of the most efficient spell correction implementations out there called symspell by Wolf Grabe

Symspell/ some Benchmarks

88,7% 45,8% 69,2% 88,3% 88,7% 1,0% 1,0% 1,7% 2,2% 100,0% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Lucene WordCorrect ElasticSearch WordCorrect No.2 eCommerce Search No.1 in eCommerce Search SymSpell

Throughput vs Accuracy

Accuracy Searches/sec

• modified edit distance to a

weighted edit distance

• changed Damerau Levenshtein

distance with a weighted Damerau Levenshtein distance – taking into account keyboard distance

• re-rank the candidate list by

applying additional similarity algorithms

search|hub -PreDictLib

fast & accurate spell correction at scale

89% 46% 69% 88% 89% 89% 99% 1% 1% 1% 2% 86% 100% 98% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

Lucene WordCorrect ElasticSearch WordCorrect No.2 eCommerce Search No.1 in eCommerce Search Symspell CXP PreDict (CE) CXP Searchhub

Throughput vs. Accuracy

Accuracy Searches/sec

Search|hub– PreDice(CE) & PreDict(EE) / some Benchmarks

what you‘ll get

CXP SmartQuery – PreDictLib (CE)

fast & accurate spell correction at scale

§ the Lib as Java source § accuracy and benchmark tests § real-life test data

ht https://gi github. b.com/se searchhub/pr preDict

Questions