kevin c. chang. about the collaboration -- cazoodle 2 coming next week: vacation rental search

Post on 21-Jan-2016

217 Views

Category:

Documents

0 Downloads

Preview:

Click to see full reader

TRANSCRIPT

Kevin C. Chang

About the collaboration -- Cazoodle

2

Coming next week: Vacation Rental Search

3

How do you greet people in your culture?

What have you been searching

lately?

4

What have you been searching lately? The university and areas of Kevin Chang? The email of Marc Snir? Customer service phone number of Amazon? What profs are doing databases at UIUC? The papers and presentations of SIGMOD 2007? Due date of SIGMOD 2008? Sale price of “Canon PowerShot A400”? “Hamlet” books available at bookstores?

5

The Web is a Big Library.Huge Supermarket!

6

Queries can be any things, too!

Search Engine

7

Are there certain “regularities” to

exploit?

Let’s try out…

8

Survey 1: How likely does a query follow a pattern?

9

9 out of 10 samples share a pattern with others!

Survey 2: How likely do queries in a domain follow patterns w.r.t. pre-specified attributes?

10

Over 28,000 manually labeled queries:

Some domains have as high as 90+% patterned queries.

Survey 3: How many patterns are there?

11

Hundreds of patterns needed to cover 80% queries.

Simple concept: What is Query Template?

(this paper) Sequence of keywords and attributes #celebrity affairs #category jobs in #location #movie showtimes in #zipcode …

(In general) Patterns that can be induced from queries e.g., regular expressions.

12

13

How would such templates be

useful?

We advocate Rich Query Interpretation. t = “#category jobs in #location” for Job

q = “accounting jobs in chicago”

By matching query q to template t:

1) Intent Classifier: recognize intended domain. q Job

2) Query Parser: recognize associated attributes. #category = “accounting”, #location = “chicago”

14

Rich query interpretation is useful.

Tailored responses by query patterns:

Finding results directly No longer 10 blue links.

Ranking results Relevant to attributes desired.

Dispatching verticals Bring verticals into search.

Matching ads More likely to click.

15

Query: Finding flights

16

Query: Finding movie showtimes

17

Query: Finding weather

18

But much more patterns can be leveraged!

19

20

Now, how to systematically discover such

templates?

Problem: Query Template Discovery Given:

Query log L e.g., we use MSN query log 2006.

Domain schema D e.g., (#category, #location, #title) with vocabulary. Incomplete schema can be handled, too.

Seed knowledge (queries, sites, templates, or mix) E.g., 5 queries; or 2 sites; or 2 templates.

Output: “Good” templates T* = {t1, t2, …}

t1 = #location jobs t2 = #location #category positions ……..

21

22

Step 1:Define quality metrics.

How to measure quality of templates?

Some templates are more “popular.” “#city1 #city2”, “#make #model”

Some templates are more “accurate.” “#city1 #city2 flights”, “#location #make used cars”

23

Precision:

Recall:

24

Step 2:From seeds, infer

templates with good quality.

1) Can P and R be “inferred”? (or, estimated.)

Probabilistic Recall:

Probabilistic Precision:

25

Sites Ss1: monster.coms2: motorola.coms3: us401k.com

Queries Qq1: jobs in chicagoq2: jobs in bostonq3: jobs in microsoftq4: jobs in motorolaq5: marketing jobs in motorolaq6: 401k plansq7: illinois employment statistics

Templates Tt1: jobs in #locationt2: jobs in #companyt3: #category jobs in #company t4: #location employment statistics

t1

t2

t3

t4

q1

q2

q3

q4

q5

q6

q7

s1

s2

s3

1025

124

4

24

1

1

11

1

1

1

2) What relationships can we use to infer? Log QST “Quest” Graph

3) How to infer on this graph?

Duality of Random Walk:

When we walk back and forth, we are inferring precision and recall, respectively.

R(t) is forward random walk from seeds.

P(t) is backward random walk to seeds.

27

Recall is forward random walk from seeds.

28

tq

xIq Iqt

DR0(x)

Recall is just like (personalized) PageRank.

Precision is backward random walk to seeds.

29

Precision is harmonic energy minimization.

tq

xItIqt

DP0(x)

Experimental results

Quest is effective in finding templates by inferred P and R, achieving 90% on actual F-measures.

Top results:

30

31

Thank You!

And they did the real work…

Ganesh Agarwal Govind Kabra

top related