patrick pantel - akbcwekex2012.files.wordpress.com · 1x listen to music (on a creativework or...
TRANSCRIPT
Patrick Pantel Joint work with: Tom Lin (UW), Michael Gamon (MSR), Anitha Kannan (MSR), Ariel Fuxman (MSR)
June 2012
Structured Data
Entity Detection
Play trailer
Structured Recommendations
Price prediction
Task completion
Task completion
Aggregate ratings
Active Objects 4
Direct Answer Structured Data
Current Experience
Active Objects 21
Better experience
Active Objects 23
Rec
ogn
ize
en
tity
in q
uer
y A
ctio
ns
easi
ly
acce
ssib
le
Better experience
Active Objects 24
Rec
ogn
ize
en
tity
in q
uer
y A
ctio
ns
easi
ly
acce
ssib
le
Better experience
Active Objects 25
Rec
ogn
ize
en
tity
in q
uer
y A
ctio
ns
easi
ly
acce
ssib
le
Politicians
Active Objects 29
Rec
ogn
ize
en
tity
in q
uer
y A
ctio
ns
easi
ly
acce
ssib
le
Films
Active Objects 30
Rec
ogn
ize
en
tity
in q
uer
y A
ctio
ns
easi
ly
acce
ssib
le
32
tt
tt
n22
TT
aae
wwKK
css s
f f
KK
y y
TT
KK
Active Objects
Actions vs Intents
Active Objects 36
User Intents and Goals
Query Intent
plan vacation
Query hilton orlando reviews sea world location
Finer-grained Intents
Actions on Entities
read reviews(hotel)
get address(landmark) add to Netflix queue(film)
buy(camera)
Navigational Informational Transactional [Broder, 2002]
Advice Locate … Download Obtain Interact …
[Rose and Levinson, 2004]
get in shape
how to lose weight
Do web queries contain entities?
Active Objects 38
43% entity (e.g., “GoldenEye”, “Horne Auto”)
14% entity category (e.g., “golf cart battery”, “global sim card”)
15% no entity (e.g., “xxx”, “good reading quotes”)
28% website (e.g., “yahoo mail”, “girlybox.com”)
category 4%
category + refiner
10%
entity 29%
entity + refiner 14%
website 28%
other 15%
Entity Distribution in Web Search Queries
* From a query traffic-weighted sample
creativework 40%
organization 37%
product 9%
person 8%
event 3% other
3%
Schema.org types for entity-bearing queries
Ontology of Actions
Active Objects 43
Navigational 10x Login Action (on a Website entity) 4x Search Action (on a Website entity) Informational (need satisfied by reading content, or could be satisfied by written transcript of content) 1x Find Location(s) (on an Organization entity) 1x Find Lyrics (on a CreativeWork / MusicalTrack entity) 2x Find Recipe For (on a food) 1x Find Where to Buy (on a Product entity) 2x Get Contact Information (on an Organization entity) 1x Get Directions To (on an Organization / Location entity) 2x Get Domain Information (on a Website entity) 1x Get Event Details (on an Event entity) 2x Get Event Results (on an Event entity) 4x Product Detail (on a Product entity) 29x Learn (on any entity) 6x Learn / Educational (on a Person / Product / Organization entity) 1x Learn / Trivia (on any entity) 1x Operating Hours (on an Organization entity) 3x Read Articles (on a News / Magazine entity) 1x Read Guide (on a Product entity) 1x Read Help (on a Product entity) 8x Read News About (on any entity) 3x Read Reviews (Shopping on a CreativeWork / Product / Service entity) 1x Read Spoilers (on a CreativeWork) 8x Research (focused information gathering, on any entity) 8x Search Database of (e.g., obituaries, on an Organization / Website) See Menu (on a Restaurant) 3x See Pictures (on a Person / Product / Organization entity) Side Effects / Safety (on a Product entity) Stock Price (on an Organization entity)
Transactional (navigating to a web-mediated action)
1x Apply for Job (on a LocalBusiness / Organization entity)
Buy (Shopping on a Product entity)
Buy Tickets (on an Event / Product / Person entity)
3x Content Creation (on a Website entity)
Discuss Online (on any entity)
5x Download (on a CreativeWork or Software entity)
1x Listen to Music (on a CreativeWork or Website entity)
Manage Account (on a Local Business / Website / Org entity)
Pay Bill (on a Website / Organization entity)
14x Play Game (on a Game entity)
Rent (on a CreativeWork / Product entity)
2x Reservation (on a Hotel entity)
Schedule Appointment (on a LocalBusiness entity)
Sell (Shopping on a Product entity)
1x Use Service On (e.g., translate, on a Website)
6x Watch Video About (on any entity)
1x Web Chat
Other
13x Shopping (category of actions including reviews and buying)
19x Various/Unknown
Actions are tied to entity types 47 actions in current list
Note: No existing Actions equivalent for Schema.org
How many Actions should there be?
Active Objects 44
0
1
2
3
4
5
6
7
8
9
10
0 50 100 150 200 250
Ne
w A
ctio
ns
pe
r 1
0 A
nn
ota
tio
ns
(ave
rage
)
Annotations
Discovery Rate of New Actions
Rapidly decreasing
discovery rate
45
tt
tt
n22
TT
aae
wwKK
css s
f f
KK
y y
TT
KK
Active Objects
Learning Actions from Web Usage Logs
Active Objects 49
21 types
58,123 hosts
2,164,579 (query, host) pairs
over 3 months 129,088 contexts
235,385 entities • Three months of us-en
web logs
• Annotate with Freebase entities
• Keep queries with an entity in set of 21 types
• Filter out navigational queries
• Filter out clicked hosts that weren’t clicked at least 100 times
Orlando hotel reviews
Does Hope Solo have a boyfriend?
Free Winzip download
watch family guy online
get reviews
read biography
download software
watch shows online
←
Model 1
Active Objects 52
Action Context
Goal: Define a theory for how actionable queries are generated.
The story for p(actionable query), or more formally The story for p(f,q,a,n | a,b) For each action a
fa ~ Dirichlet(b) For each query q
q ~ Dirichlet(a) For each context position in q (pre or post)
action a ~ Multinomial(q) ngram n ~ Multinomial(fa)
Model 1.01Model 1.01
bb
f f
KK
a a
q q
a1a1
n1
a2a2
n2
Star Wars “ebert review"
read reviews action
“______ ______”
read reviews action
(action→contexts die)
Active Objects 53
Model 2
Action Context
The story for p(f,q,a,n | a,b) For each action a
fa ~ Dirichlet(b) For each query q
q ~ Dirichlet(a) For each context position in q (pre or post)
action a ~ Multinomial(q) ngram n ~ Multinomial(fa)
action a ~ Multinomial(q) ngram n1 ~ Multinomial(fa) ngram n2 ~ Multinomial(fa)
a a
q q
n1 n2
bb
f f
KK
Model 1.02Model 1.02
aa
Star Wars “ebert review"
read reviews action
Active Objects 54
Model 3
Clicked hosts matter…
Action Context
buy action read reviews action
amazon.com
ebay.com
walmart.com
rottentomatoes.com
metacritic.com
efilmcritic.com
“______ ______”
read reviews action
Active Objects - 55 -
Model 1.13Model 1.13
Model 3
The story for P(f,q,w,a,n,c | a,b,) For each action a
fa ~ Dirichlet(b)
For each query q q ~ Dirichlet(a) action a ~ Multinomial(q) ngram n1 ~ Multinomial(fa) ngram n2 ~ Multinomial(fa)
“ebert review”
Action Host Context
bb
f f
KK
ww
KK
a a
q q
n22
c
aa
click c ~ Multinomial(wa)
wa ~ Dirichlet(i)
www.rottentomatoes.com
w, ,c ,
(to contexts) (to clicks)
Star Wars
Active Objects 57
Model 3
The type matters…
Action Host Context
Active Objects 58
Model 1.14Model 1.14
Model 4
Action Host Context Type
Model 1.03Model 1.03
bb
f f
KKn
22
aa
ww
KKc
g g
t t
TT
a a
q q
tt
The story for P(f,q,t,w,t,a,n,c | a,b,g,) For each action a
fa ~ Dirichlet(b) wa ~ Dirichlet(i)
For each query q
q ~ Dirichlet(a) action a ~ Multinomial(q) ngram n1 ~ Multinomial(fa) ngram n2 ~ Multinomial(fa) click c ~ Multinomial(wa)
a a
q q
For each type t tt ~ Dirichlet(g)
type t ~ Multinomial(q) action a ~ Multinomial(q) action a ~ Multinomial(tt)
“ebert review” “______ ______”
film type
read reviews action
www.rottentomatoes.com
t, t, g,
(to contexts) (to clicks)
(to actions)
Star Wars
Active Objects 60
Model 3
We also have entity data…
Action Host Context Type
Active Objects - 61 -
a a
q q
bb
f f
KK
Model 1.05bModel 1.05b
tt
n22
g g
t t
TT
aa
ww
KKc
Model 5
Action Host Context Type Entity
h h
y y
TTe
The story for P(f,q,t,y,w,t,a,e,n,c | a,b,g,h,) For each action a
fa ~ Dirichlet(b) wa ~ Dirichlet(i)
For each type t tt ~ Dirichlet(g)
For each query q q ~ Dirichlet(a) type t ~ Multinomial(q) action a ~ Multinomial(tt) ngram n1 ~ Multinomial(fa) ngram n2 ~ Multinomial(fa) click c ~ Multinomial(wa)
yt ~ Dirichlet(h)
entity e ~ Multinomial(yt)
y, e, h,
Active Objects - 63 -
Model 6
Action Host Context Type Entity Empty
The story for P(f,q,t,y,w,s,t,a,e,s,n,c | a,b,g,h,,e) For each action/type pair, {a, t}
fa ~ Dirichlet(b) wa ~ Dirichlet() sa,t ~ Beta(e)
For each type t tt ~ Dirichlet(g) yt ~ Dirichlet(h)
For each query q q ~ Dirichlet(a) type t ~ Multinomial(q) action a ~ Multinomial(tt) entity e ~ Multinomial(yt) switch s1 ~ Bernou (sa) switch s2 ~ Bernoulli(sa,t) if (s1) ngram n1 ~ Multinomial(fa) if (s2) ngram n2 ~ Multinomial(fa) click c ~ Multinomial(wa)
The story for P(f,q,t,y,w,s,t,a,e,s,n,c | a,b,g,h,,e) For each action/type pair, {a, t}
fa ~ Dirichlet(b) wa ~ Dirichlet() sa ~ Beta(e)
switch s1 ~ Bernoulli(sa) switch s2 ~ Bernoulli(sa) if (s1) if (s2)
s,
a a
q q
Model 1.06bModel 1.06b
tt
n22
g g
t t
TT
aa
e
ww
KKc
bb
f f KK
h h
y y
TT
e e
s s
KK
s
Apply Generative Model
Active Objects 66
66
Action (hidden)
Clicked Host (observed)
Query Context (observed)
Type (observed)
P(Action|Type)
P(Host|Action) P(Context|Action)
A new query comes in: (e.g., “New York City hotels”) Entity Recognition(query) → entity (“New York City”) entity → types (“city”, “employer”, “travel destination”) (query, entity) → context (“Ø”, “hotels”) Historical Data(query) → distribution over hosts • EM Posterior Probabilities to give us likelihood of
each action cluster. • action cluster → action phrase (“book hotel in”) • (action cluster, action phrase, historical data) →
best hosts (“travel.bing.com”)
query: jetbeam rrt-0
Active Objects 69
Action words from Web Trigrams
Active Objects 70
• Patterns (similar to Hearst patterns) on a Web Trigram corpus to get actions.
Web Trigrams
Pattern Match “want to (x)” “have to (x)” “you can (x)” “I can (x)”
Filter Adverbs (e.g., “honestly”, “quickly”)
Filter noise (the 25% with
lowest 𝑓𝑟𝑒𝑞𝑢𝑒𝑛𝑐𝑦
𝑢𝑛𝑖𝑔𝑟𝑎𝑚 𝑐𝑜𝑢𝑛𝑡
e.g., “a”, “boy”)
13,417 action words make download find torrent say eBay Pay login Buy podcast help … …
Method scales to longer actions, e.g., 4-grams for 2-word actions (“read review”)
Finds modern/web actions that older annotated corpora might miss.
↑ ↑
Web Action words from Trigrams
Active Objects 71
• Not all actions can be recommended over the Web (e.g., “shock” or “kill”). How do we find the ones that can?
Web Action ngram pattern “(x) at (y)” where y has the form of a web site URL “buy at Amazon.com” “download at cnet.com”
13,417 action words make download find torrent say eBay Pay login Buy podcast help … …
1,279 web actions buy review shop unsubscribe book download …
Human Annotation of Action Phrases
Active Objects 73
19, Read biography of 19, See pictures of 19, Read blog of 19, Contact 19, Read interview with 19, Watch video of
12, Download 12, Find reviews of 12, Update 12, Get help for
24, Apply for jobs at 24, View career options at 24, View map of 24, Read news about 24, Find locations of 24, Find address of 24, Get stock quote of
From each model, we first automatically generate:
Use as training data
Evaluation Setup
Active Objects 79
a a
q q
n1 n2
bb
f f
KK
Model 1.02Model 1.02
aa
Model 2 Model 3 Model 4 Model 5 Model 6
+clicked host +entity types +entity +empty switch action → context
a a
q q
bb
f f
KK
Model 1.13Model 1.13
n22
aa
ww
KKc
a a
q q
bb
f f
KK
Model 1.14Model 1.14
tt
n22
g g
t t
TT
aa
ww
KKc
a a
q q
bb
f f
KK
Model 1.05bModel 1.05b
tt
n22
g g
t t
TT
aah h
y y
TTe
ww
KKc
a a
q q
Model 1.06bModel 1.06b
tt
n22
g g
t t
TT
aa
e
ww
KKcs
e e
s s
KK
bb
f f KK
h h
y y
TT
21 types
58,123 hosts
2,164,579 (query, host) pairs
over 3 months 129,088 contexts
235,385 entities
Data
+ Filter out Navigational
Queries
50 clusters, 2-step learning over 100 total EM iterations, 2 folds per model
Evaluation Framework
Active Objects 82
Co
mb
ined
set
of
acti
on
s o
ver
all m
od
els
x 7
UHRS Fair κ
agreement
Performance on HEAD vs. TAIL vs. Type-Balanced queries
Active Objects 87
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
HEAD TAIL Type-Balanced
nD
CG
nDCG vs. Query Sets (with 95% confidence bounds)
Model 2
Model 3
Model 4
Model 5
Model 6
Tail is dominated by “People” type
+ click
+ type
+ entity + switch
Action Discovery, Diversity
Active Objects 89
0%
100%
200%
300%
400%
500%
600%
700%
800%
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Tota
l P(a
ctio
n |
typ
e)
Cluster Rank
Model 4
0%
100%
200%
300%
400%
500%
600%
700%
800%
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Tota
l P(a
ctio
n |
typ
e)
Cluster Rank
Model 4
Model 5
0%
100%
200%
300%
400%
500%
600%
700%
800%
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Tota
l P(a
ctio
n |
typ
e)
Cluster Rank
Model 4
Model 5
Model 6
Examples
Active Objects 92
Query: Webster University
Entity: Webster University
Context: (“”, “”)
Types: /business/employer, education/university, /location/location
Model 2 (context) Model 3 (+click) Model 4 (+type) Model 5 (+entity) Model 6 (+switch)
1. Torrent 2. Read biography 3. Find adult pictures of 4. Watch videos 5. See picture of 6. Get quotes from 7. Apply for jobs at
1. Torrent 2. Read biography 3. Read news about 4. See pictures of 5. Apply for jobs at 6. Get quotes from 7. See videos with
1. Read reviews of 2. See map of 3. Follow sports teams of 4. Get weather in 5. Apply for jobs at 6. Find address of 7. See rankings of
1. Read reviews of 2. See map of 3. Follow sports teams of 4. Get weather in 5. Apply for jobs at 6. Find address of 7. See tuition of
1. Find address 2. See pictures of 3. Find map of 4. Read news about 5. Apply for jobs at 6. See cost of 7. See ranking of
Models 4, 5, 6 automatically generate reasonable actions for this query
Model 2 (context) Model 3 (+click) Model 4 (+type) Model 5 (+entity) Model 6 (+switch)
1. Torrent 2. Read biography 3. Find adult pictures of 4. Watch videos 5. See picture of 6. Get quotes from 7. Apply for jobs at
1. Torrent 2. Read biography 3. Read news about 4. See pictures of 5. Apply for jobs at 6. Get quotes from 7. See videos with
1. Read reviews of 2. See map of 3. Follow sports teams of 4. Get weather in 5. Apply for jobs at 6. Find address of 7. See rankings of
1. Read reviews of 2. See map of 3. Follow sports teams of 4. Get weather in 5. Apply for jobs at 6. Find address of 7. See tuition of
1. Find address 2. See pictures of 3. Find map of 4. Read news about 5. Apply for jobs at 6. See cost of 7. See ranking of
+ User Model?
Examples
Active Objects 93
Query: download Skype
Entity: Skype
Context: (“download”, “”)
Types: /computer/software,/business/employer,/business/business_operation
Model 2 (context) Model 3 (+click) Model 4 (+type) Model 5 (+entity) Model 6 (+switch)
1. Download 2. Login to 3. Read celebrity gossip about 4. Watch videos with 5. Hack 6. Find games with 7. Watch movies with
1. Download 2. Play games 3. Hack 4. Chat at 5. Create account at 6. Torrent 7. Read biography of
1. Download 2. Login to 3. Hack
1. Find on social networks 2. Download 3. Hack
1. Download 2. Find reviews of 3. Update 4. Get help for
Again, we can now automatically generate reasonable actions for queries!
144
tt
tt
n22
TT
aae
wwKK
css s
f f
KK
y y
TT
KK
Active Objects
Active Objects (Model IM)Active Objects (Model IM)
QQn
22
qqTT
aae
wwKK
css s
f f KK
y y TT
KK
Inferring Type Distributions
• Entity types are modeled as latent variables, jointly with the intended actions.
• Extensions: – Given admissible types from a
KB such as Freebase, learn their priors and contextual disambiguation
– Given a new term, induce the types of the term
– Automatically induce type list and admissible types for arbitrary entities
Active Objects 145
tt
tt
Prior Art (Guo, Xu, Cheng, and Li, SIGIR’09)
Active Objects 146
Generative process for entity bearing queries. For each query q
entity e ~ Multinomial(y) type t ~ Multinomial(te) ngram n1 ~ Multinomial(ft) ngram n2 ~ Multinomial(ft)
Guo’09Guo’09y y
e
ttttEE
n22
f f TT
Active Objects 147
y y
tt
n22
e
f f TT
ttEE
Guo’09Guo’09
+ Empty Switch
y y
tt
n22
e
f f TTtt
EE
Model M0Model M0
ss s TT
Active Objects 148
y y
tt
n22
e
f f TT
ttEE
Guo’09Guo’09
+ Empty Switch + Click
wwTT
css s TT
y y
tt
n22
e
f f TTtt
EE
Model M1Model M1
Active Objects 149
y y
tt
n22
e
f f TT
ttEE
Guo’09Guo’09
+ Empty Switch + Click + Action
Model IMModel IMtt
tt
n22
e f f KK
y y TT
qqTT
aa
wwKK
css s KK
Experimental Setting
• Training – Queries from 3 months of US Bing search usage logs
– Entities from 73 Freebase types, accounting for 50% query traffic in US market
– Model parameters trained using 2-step learning over 100 EM iterations, 2 folds per model
• Testing – Query-weighted random sample of 500 HEAD and 500 TAIL entity-
bearing queries
– 7 paid independent annotators identified all applicable Freebase types to the entities in the queries • Two annotators per query
• Fleiss’ was 0.445, moderate agreement
Active Objects 150
Performance Analysis
Active Objects 151
• † indicates statistical significance over BFB, and ‡ over both BFB and Guo’09. • Bold indicates statistical significance over all non-bold models in the
column. • M1 (empty context + click signal) significantly outperforms baseline and
Guo’09, on HEAD. • IM significantly better over all models and across all metrics
– Biggest gains in first position of its ranking (Prec@1 metric).
Switch Parameter Analysis
Active Objects 153
• Switch improves performance across all models
• More expressive models benefit more from switch
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
M0 M1 IM
Re
lati
ve g
ain
of
swit
ch v
s. n
o s
wit
ch
Effect of Empty Switch Parameter (s) on HEAD
No switch
nDCG
MAP
MAPW
Prec@1
Discussion
• Why performance on TAIL lower than expected? – TAIL is skewed towards the PEOPLE types
– Latent actions are over-expressive and they do not help in differentiating PEOPLE types • Inspection of latent Action parameter in IM shows that most PEOPLE
types have all their mass distributed to three generic and common intents (see pictures of, find biographical information about, and see video of)
• Success case in the TAIL – “ymca” -> {song, place, educational_institution}
• Marginalizing out the context words gives the following intent priors: 0.63, 0.29, 0.08
– q1 = “jamestown ymca ny” IM correctly classified “ymca” as a place
– q2 = “ymca palomar” IM correctly classified “ymca” as a educational_institution
Active Objects 155
Wrap-Up
Active Objects 157
Active Objects 159
• A hodgepodge of related strings
• Only actionable through search
Active Objects 160
2009
2010
Active Objects 161
Big Wins
• Typed relations
2009
2010
Active Objects 162
Big Wins
• Typed relations
• User Interface accesses structured data
2009
2010
Active Objects 163
Big Wins
• Typed relations
• User Interface accesses structured data
• Click through experience can now leverage strongly-typed identifier
2009
2010
Active Objects 164
Big Wins
• Typed relations
• User Interface accesses structured data
• Click through experience can now leverage strongly-typed identifier
• Brokered Actions (one click conversions)
2009
2010
Add to Amazon cart
Connect on Twitter
Play track
Active Objects 165
2009
2010
To Do
• Fixed actions
• Model tasks
• Annotate URLs/Apps with actions
• Active Objects as an entity-centric UI