1. Unsupervised Extraction of Template Structure in Web Search Queries www 2012 – Session: search
Qingxia Liu

2. Content Motivation Definition of the Problem Generative Model
Alternative Models
Experiments

3. Motivation Web Search Problems: Challenges: fact:
Determine users’ intent
Keyword-search: Small set of keywords
Challenges:
Scalability issues;
Instrumentation issues: user clicks;
Brevity, ambiguity:
E.g “jaguar” the car or the animal?
Sparsity issues:
E.g tail queries such as “jaguar xj engine mount”
fact:
Various queries with the same search intent issue
Usage of intent information:
provide relevant results
to detect whether a query has a commercial intent
to select a useful set of advertisements
to learn from the user’s interaction with the search engine

4. Motivation Goal: Similar works:
extracting the hidden structure behind the observed search queries in a domain;
with no manual intervention;
e.g. “jaguar xj engine mount”
<Brand,Model,Year,Part> pattern
Similar works:
require either direct supervision of the tasks
or use ancillary information such web search click-through data
we seek to solve these issues by enriching search queries with information about the hidden structure underlying them.
analyze queries to obtain segmentations
extract named entities from queries
All these approaches require either direct supervision of the tasks, such as manually labeled seed data, or use ancillary information such web search click-through data, both of which might be expensive or difficult to obtain.
2, sume that the attribute-set as well as the associated vocabularies are given as inputs in form of database relations or entity hierarchies.
detected templates for improving query recommendations

5. Model
we seek to solve these issues by enriching search queries with information about the hidden structure underlying them.
analyze queries to obtain segmentations
extract named entities from queries
All these approaches require either direct supervision of the tasks, such as manually labeled seed data, or use ancillary information such web search click-through data, both of which might be expensive or difficult to obtain.
2, sume that the attribute-set as well as the associated vocabularies are given as inputs in form of database relations or entity hierarchies.
detected templates for improving query recommendations

6. Single-attribute template model
A set of words
template1
attributea W1 W2 W3 W4 ……
attributeb
template2
template3
attributec
attributed
template4

7. multi-attribute template model
A set of words
template1
attra
attrb
attrc W1 W2 W3 W4 ……
template2
attrb
attrc
template3
attra
attrc
e.g.
Brand
Year
Parts
Honda
Toyota
Ford ……

8. Properties
I： constrain the number of distinct templates that can be formed in the model
II： each attribute in the template to generate at least one word
III： Each attribute has a specific word distribution as well as a distinct tendency for the number of words it contributes in a query.

9. PROBLEM DEFINITION
Given a set of queries, extract the underlying schema (templates, attribute, and their vocabularies) and learn the parameters of the generative process in a completely unsupervised manner while respecting the properties mentioned above.

10. Parameter generating process
attributes
Vector
Candidate pool
θt ~ Multinomial(μ)
μ ~ Dirichlet(α)
t q ~ Multinomial(γ)
γ ~ Dirichlet(σ) T q1
tq1 q2
tq2 q3
tq3 q4
tq4 ……
One way to think of this is that
the candidate pool denotes the set of template configurations which are appropriate for the domain. t1 t2 t3 …… tT

11. Generative Model template θ[tq] na ~ Possion(ηa) ηa ~Gamma(g1,g2) zq
wq1
tq1
aq1 q2
wq2
tq2
aq2 q wq tq aq q4 ……
template θ[tq]
attr1
attr2
attr3
attr4 ……
na ~ Possion(ηa)
ηa ~Gamma(g1,g2) zq
zq1
zq2
zq3 ……
W(z[q,i],i) ~ Mutinomial(φz[q,i])
φa~ Dirichlet(β)
Query q w1 w2 w3 ……

12. Generative Model q1 tq1 q2 tq2 q tq q4 …… aq1 aq2 aq t1 t2 t3 …… tT
wq1
tq1
aq1 q2
wq2
tq2
aq2 q wq tq aq q4 …… t1 t2 t3 …… tT

13. Model Learning Gibbs sampling Bayes’ Theorem 分母：Gibbs sampling
分子：根据前面的式子计算出来
Gibbs sampling

14. Model Learning

15. Model Learning

16. Overview random initialization:
iterate over queries and the template set
using the derived conditionals to update the
vectors
compute the likelihood p(φ, μ, γ, |data)
ends with:
query ->template, word -> attribute

17. Alternative approaches
Latent Dirichlet Allocation (LDA)
attribute
Multinomial distribution
topics
a topic
vocabularies
φ distribution
the ith word
a document
query

18. Alternative approaches
Spherical k-Means
nu,v : times wu and wv occur together in queries
co-occurrence behavior
clusters of words -> attributes wu
nu,1
nu,2
nu,3
nu,4 ……
co-occurrence behavior

19. Experiments queries： 100 million Yahoo search queries
43793,83387,15050 queries each domain
ground truth：manually extracted

20. Experiments Automobile domain Travel domain Movies domain
correctly placed：the learnt attribute and the ground truth attribute it belongs to are mapped to each other.
PRECISION(N) is the fraction of words in the first N learnt attributes (in the algorithm’s ordering)that are correctly placed.
CORRECTRECALL(N), is the fraction of words in ground truth attributes mapped to the first N learnt attributes that are correctly placed.
Movies domain

21. Experiments number of iterations number of attributes
φ acts as a prior to the attribute-word multinomial distributions
g1 and g2 are used to generate the Poisson distributions for each attribute : attr - num of words
different values of φ

22. Case Study applications: CTR on sponsored search advertisements
fq1 = ∑i>1qi
tendency to attract ad-clicks from users
for infering its advertisability

23. Thanks for listening ~