Keynote talk at ICWE 2015, Rotterdam

1. Making the Web Searchable
P R E S E N T E D B Y P e t e r M i k a , Y a h o o L a b s ⎪ J u n e 2 5 , 2 0 1 5

2. Agenda
2
 Web Search
› How it works… and where it fails
 Semantic Web
› The promise and the reality
 Semantic Search
› Research and Applications at Yahoo
 What’s next?
› More intelligence!

3. Search is really fast, but not particularly intelligent

4. What it’s like to be a machine?
Roi Blanco

5. What it’s like to be a machine?
↵⏏☐ģ
✜Θ♬♬ţğ√∞§®ÇĤĪ✜★♬☐✓✓
ţğ★✜
✪✚✜ΔΤΟŨŸÏĞÊϖυτρ℠≠⅛⌫
≠=⅚©§★✓♪ΒΓΕ℠
✖Γ♫⅜±⏎↵⏏☐ģğğğμλκσςτ
⏎⌥°¶§ΥΦΦΦ✗✕☐
⏎↵⏏☐ģğğğ

6. Real problem

7. Search needs (more) intelligence
 Current paradigm: ad-hoc document retrieval
› Set of documents that each may fulfill the information need
› Relevance of those documents is clearly established by
• Textual similarity
• Authority
 Queries outside this paradigm are hard
› Semantic gap between the query and the content
• Queries that require a deeper understanding of the world at large
› Complex needs where no document contains the answer
 Analysis of aggregate behavior is a challenge
› We may answer two queries perfectly, without knowing how they are related

8.  Semantic gap
› Ambiguity
• jaguar
• paris hilton
› Secondary meaning
• george bush (and I mean the beer brewer
in Arizona)
› Subjectivity
• reliable digital camera
• paris hilton sexy
› Imprecise or overly precise searches
• jim hendler
 Complex needs
› Missing information
• brad pitt zombie
• florida man with 115 guns
• 35 year old computer scientist living in
barcelona
› Category queries
• countries in africa
• barcelona nightlife
› Transactional or computational queries
• 120 dollars in euros
• digital camera under 300 dollars
• world temperature in 2020
Examples of hard queries
Are there even
any true
keyword
queries?
Users may
have stopped
asking them

9. The Semantic Web

10. Enter the Semantic Web
10
 A social-technical ecosystem where the
meaning of content is explicit and shared
among agents (humans and machines)
› Shared identifiers for real-world entities
› Standards for exchanging structured data
• Data modeled as a graph
› Shared formal, schema languages
• Names of entity and relationship types
• Constraints on the entities, relationships and attributes

11.  “At the doctor's office, Lucy instructed her
Semantic Web agent through her
handheld Web browser. The agent
promptly retrieved information about
Mom's prescribed treatment from the
doctor's agent, looked up several lists of
providers, and checked for the ones in-
plan for Mom's insurance within a 20-mile
radius of her home and with a rating of
excellent or very good on trusted rating
services. It then began trying to find a
match between available appointment
times (supplied by the agents of individual
providers through their Web sites) and
Pete's and Lucy's busy schedules.”
› The Semantic Web. Tim Berners-Lee, James
Hendler, Ora Lassila. Appeared in: Scientific
American 284(5):34-43 (May 2001)
6/25/201511
Huge expectations

12. In a perfect world, the Semantic Web is the end-game for IR
#ROI_BLANCO
#ROI_BLANCO
#ROI_BLANCO

13. The view from IR: skepticism
13
 Mismatch to the IR problem
› End-users
• Do not know the identifiers of things
• Not aware of the schema of the data
• Can’t be expected to learn complex query languages
› Focus on document retrieval
• Limited recognition of the value of direct answers (initially)
 “Who is going to do the annotation?”
› Focus still largely on text
› Automated annotation/extraction tools produce poor results
 Early experiments using external knowledge are unsuccessful
› Query/document expansion using thesauri etc.

14. The Semantic Web off to a slow start
14
 Complex technology
› Set of standards sold as a stack
• URI, RDF, RDF/XML, RDFa, JSON-LD,
OWL, RIF, SPARQL, OWL-S, POWDER …
› Not very developer friendly
 Ideals of knowledge representation
hard to enforce on the Web
 No clear value proposition
› Chicken and egg problem
• No users/use cases, hence no data
• No data, because no users/use cases

15. … and became practical
15
 Simplified technology
› Fewer, more developer friendly representations
› Focusing on the lower layers of the stack
› Data first, schemas/logic second
 Giving up ideals about knowledge representation
› Shared identifiers
› Logical consistency
› Distinction between real-world entities and web resources
 Motivation for adoption
› Search engines, Facebook, Pinterest, Twitter etc. investing in information extraction

16. Two important achievements
16
 Linked Open Data
› Social movement to (re)publish existing datasets
• 100B+ triples of data
• Encyclopedic, governmental, geo, scientific datasets
• Impact: background knowledge
– Basis for knowledge graphs in search engines
 Metadata inside HTML pages
› Facebook’s OGP and schema.org
• Over 15% of all pages have schema.org markup (2013)
• Personal information, images, videos, reviews, recipes etc.
• Impact: remove the need for automated extraction

17. Example
17

18. View source
18

19. Caveat: not your perfect Semantic Web
19
 Outdated, incorrect or incomplete data
› Lack of write access or feedback mechanisms
 Mistakes made by tools
› Noisy information extraction
› Entity linking (reconciliation)
 Limited or no reuse of identifiers
 Metadata not always representative of content

20. Semantic Search at Yahoo
20

21. Semantic Search research (2007-)
 Emergence of the Semantic Search field
› Intersection of IR, NLP, DB and SemWeb
• ESAIR at SIGIR
• SemSearch at ESWC/WWW
• EOS and JIWES at SIGIR
• Semantic Search at VLDB
 Exploiting semantic understanding in the retrieval process
› User intent and resources are represented using semantic models
• Semantic models typically differ across NLP, DB and Semantic Web
› Semantic models are exploited in the matching and ranking of resources

22. Semantic Search – a process view
Query
Constructi
on
•Keywords
•Forms
•NL
•Formal language
Query
Processin
g
•IR-style matching & ranking
•DB-style precise matching
•KB-style matching & inferences
Result
Presentation
•Query visualization
•Document and data presentation
•Summarization
Query
Refinement
•Implicit feedback
•Explicit feedback
•Incentives
Document Representation
Knowledge Representation
Semantic Models
Resources
Documents

23. Result presentation using metadata
Personal and
private
homepage
of the same
person
(clear from the
snippet but it
could be also
automatically
de-duplicated)
Conferences
he plans to attend
and his vacations
from homepage
plus bio events
from LinkedIn
Geolocation
“Microsearch”
internal prototype
(2007)

24. Yahoo SearchMonkey (2008)
1. Extract structured data
› Semantic Web markup
• Example:
<span property=“vcard:city”>Santa Clara</span>
<span property=“vcard:region”>CA</span>
› Information Extraction
2. Presentation
› Fixed presentation templates
• One template per object type
› Applications
• Third-party modules to display data (SearchMonkey)

25. Effectiveness of enhanced results
 Explicit user feedback
› Side-by-side editorial evaluation (A/B testing)
• Editors are shown a traditional search result and enhanced result for the same page
• Users prefer enhanced results in 84% of the cases and traditional results in 3% (N=384)
 Implicit user feedback
› Click-through rate analysis
• Long dwell time limit of 100s (Ciemiewicz et al. 2010)
• 15% increase in ‘good’ clicks
› User interaction model
• Enhanced results lead users to relevant documents (IV) even though less likely to clicked than textual (III)
• Enhanced results effectively reduce bad clicks!
 See
› Kevin Haas, Peter Mika, Paul Tarjan, Roi Blanco: Enhanced results for web search. SIGIR 2011:
725-734

26. Adoption among consumers of web content
 Google announces Rich Snippets - June, 2009
› Faceted search for recipes - Feb, 2011
 Bing tiles – Feb, 2011
 Facebook’s Like button and the Open Graph Protocol (2010)
› Shows up in profiles and news feed
› Site owners can later reach users who have liked an object

27. schema.org
 Collaborative effort sponsored by large consumers of Web data
› Bing, Google, and Yahoo! as initial founders (June, 2011)
› Yandex joins schema.org in Nov, 2011
 Agreement on a shared set of schemas for the Web
› Available at schema.org in HTML and machine readable formats
› Free to use under W3C Royalty Free terms

28. Yahoo’s Knowledge Graph
Chicago Cubs
Chicago
Barack Obama
Carlos Zambrano
10% off tickets
for
plays for
plays in
lives in
Brad Pitt
Angelina Jolie
Steven Soderbergh
George Clooney
Ocean’s Twelve
partner
directs
casts in
E/R
casts
in
takes place in
Fight Club
casts in
Dust Brothers
casts
in
music by
Nicolas Torzec: Making knowledge reusable at Yahoo!: a Look at the Yahoo! Knowledge Base
(SemTech 2013)

29. Building the Knowledge Graph
 Information extraction
› Automated information extraction
• e.g. wrapper induction
› Metadata from HTML pages
• Focused crawler
› Public datasets (e.g. Dbpedia)
› Proprietary data
 Data fusion
› Manual mapping from the source
schemas to the ontology
› Supervised entity reconciliation
 Ontology management
› Editorially maintained OWL ontology
with 300+ classes
› Covering the domains of interest of
Yahoo
 Curation and quality assessment
› Editors and user feedback still play a
large role
Bellare et al: WOO: A Scalable and Multi-tenant Platform for Continuous Knowledge Base Synthesis. PVLDB 2013
Welch et al.: Fast and accurate incremental entity resolution relative to an entity knowledge base. CIKM 2012

33.  Entity linking/entity retrieval
› Identifying the most relevant entity to
the query
 Entity recommendation
› Given that the user is interested in one
entity, which entity to recommend next?
Roi Blanco, Berkant Barla Cambazoglu,
Peter Mika, Nicolas Torzec: Entity
Recommendations in Web Search. ISWC
2013
Entity displays in web search

34. The importance of entities
34
 Entity mention query = <entity> {+ <intent>}
› ~70% of queries contain a named entity (entity mention queries)
• brad pitt height
› ~50% of queries have an entity focus (entity seeking queries)
• brad pitt attacked by fans
› ~10% of queries are looking for a class of entities
• brad pitt movies
› Jeffrey Pound, Peter Mika, Hugo Zaragoza:
Ad-hoc object retrieval in the web of data. WWW 2010: 771-780
 Intent is typically an additional word or phrase
• Disambiguate, most often by type e.g. brad pitt actor
• Specify action or aspect e.g. brad pitt net worth, toy story trailer
brad pitt height
how tall is
tall
…

35.  Inverted index
› Inspired by text retrieval
• Match individual keywords
• Score and aggregate
 Parsing
› Inspired by text parsing
• Find potential mentions of entities (spots)
in query
• Score candidates for each spot
Two broad approaches to entity retrieval
brad
(actor) (boxer) (city)
(actor) (boxer) (lake)
pitt
brad pitt
(actor) (boxer)

36. Retrieval-based approach
 Experimented with different index structures
› Horizontal: one field for text and one for property name
› Vertical: One field per property
› Combination: one field per property weight (best performance in both AND/OR mode)
Horizontal
Vertical
R-Vertical

37. Retrieval-based approach
37
 Ranking based on BM25F
› R. Blanco, P. Mika, S. Vigna: Effective and Efficient Entity Search in RDF Data. ISWC 2011
› 42% improvement in MAP over best method in SemSearch 2010
› <100ms time for simple conjunctive queries
 Open source implementation and demo using WebDataCommons data
› glimmer.research.yahoo.com
› https://github.com/yahoo/Glimmer/
Doc
map
map
reduce
reduce
map reduce
Index

38. Entity linking approach
38
 Large-scale entity/alias dictionaries
› Alias mining from usage data, Wikipedia etc.
 Dynamic segmentation
 Novel method for scoring alias matches
› Completely unsupervised
› Combination of
• Keyphraseness: how likely is a segment to be an entity mention?
• Commonness: How likely that a linked segment refers to a particular entity?
• Context-model based on word2vec representation
› Roi Blanco, Giuseppe Ottaviano and Edgar Meij.
Fast and space-efficient entity linking in queries. WSDM 2015

39. Results: effectiveness
39
 Significant improvement over external baselines and internal system
› Measured on public Webscope dataset Yahoo Search Query Log to Entities
Search over Bing, top
Wikipedia result
State-of-the-art in literature
A trivial search engine
over Wikipedia
Our method:
Fast Entity Linker (FEL)
FEL + context

40.  Two orders of magnitude faster
than state-of-the-art
› Simplifying assumptions at scoring time
› Adding context independently
› Dynamic pruning
 Small memory footprint
› Compression techniques, e.g. 10x
reduction in word2vec storage
40
Results: efficiency

41. Related entity recommendations
 Some users are short on time
› Need for direct answers
› Query expansion, question-answering, information boxes, rich results…
 Other users have time at their hand
› Long term interests such as sports, celebrities, movies and music
› Long running tasks such as travel planning

42. Example user sessions

43. Spark system for related entity recommendations
Entity
graph
Data
preprocessing
Feature
extraction
Model
learning
Feature
sources
Editorial
judgements
Datapack
Ranking
model
Ranking and
disambiguation
Entity
data
Features

44. Machine learned ranking
 Features from the Knowledge Graph and large-scale text sources
› Unary
• Popularity features from text: probability, entropy, wiki id popularity …
• Graph features: PageRank on the entity graph, wikipedia, web graph
• Type features: entity type
› Binary
• Co-occurrence features from text: conditional probability, joint probability …
• Graph features: common neighbors …
• Type features: relation type
 Regression model using Gradient Boosted Decision Trees (GBDT)
› Trained on editorial data (cf. clicks)

45. Evaluation
45
1. 10-fold cross-validation
2. Side-by-side testing
› More appropriate for judging sets of results
• “Blondie and Mickey Gilley are 70’s performers and do not belong on a list of 60’s musicians.”
3. Online evaluation (bucket testing)
› Small % of search traffic redirected to test system, another small % to the baseline system
› Data collection over at least a week, looking for stat. significant differences that are also
stable over time
› Metrics
• Coverage and Click-through Rate (CTR)
• Searches per browser-cookie (SPBC)
• Other key metrics should not impacted negatively, e.g. Abandonment and retry rate, Daily Active Users
(DAU), Revenue Per Search (RPS), etc.

46. Click-through rate (CTR) before and after the new system
Before release:
Gradually
degrading performance
due to lack of fresh data
After release:
Learning effect:
users are starting to
use the tool again
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80
Days
CTR
CTR before Spark
Trend before Spark
CTR after Spark
Trend after Spark
Spark is deployed
in production

47. What’s next?

48. Summary
 Information Retrieval
› Reached the limits of the ad-hoc text retrieval paradigm
› Needs to go beyond syntactic representations
 Semantic Web
› Provides means for knowledge representation and reasoning across the Web
› Adoption has been slow, but picking up steadily
 Applications in Web Search
› Entity-based experiences
• Rich results, information boxes and related entities
› Question-answering

49. Search needs even more intelligence
49
 Representation
› Modeling the World, not just what is on the Web
› Modeling personal information and preferences
› Modeling of intents (actions that can be taken on the World)
 Understanding
› Need better understanding of context
› User profile, history and current state
 Retrieval
› (Guided) interaction
› Predictive search

50. Q&A
 Many thanks to members of the Semantic Search team
at Yahoo Labs London and to Yahoos around the world
 Contact me
› pmika@yahoo-inc.com
› @pmika
› http://www.slideshare.net/pmika/