Answering Structured Queries on Unstructured Data

1
Answering Structured Queries on Unstructured Data

• Jing Liu, Xin (Luna) Dong, Alon Halevy
• Univ. of Washington
• _at_WebDB 2006

2
Seamless Querying on the Structured and
Unstructured Data
3
Seamless Querying on the Structured and
Unstructured Data
4
Dataspaces and DSSPs

• Dataspace ? Collection of both structured and
  unstructured data PODS06 Keynote talk
• Dataspace Support Platforms (DSSPs) ? Services
  over dataspaces (e.g., search, query and source
  discovery, etc)

5
SEMEX- Personal Information Management
Title
Year
Paper
Author
PublishedIn
CitedBy
6
SEMEX- Personal Information Management
Time
Paper
FromFile
7
SEMEX- Personal Information Management
Article (Title dataspace) (Author Alon
Halevy)
8
SEMEX- Personal Information Management
Mentioned In Article alon keynote pods06
9
SEMEX- Personal Information Management
Web search results by Google Alon Halevys
Home Page DBLP David Maier
10
Current Approach

• Information-extraction approach
• Use supervised learning
• Hard to scale to data in a large number of
  domains
• Hard to apply to the case where the query schema
  is unknown beforehand

11
Solution in SEMEX

• Semex solution
• Transform a structured query into keyword search
• Keyword search on unstructured data.
• Advantages
• Apply to different domains
• Handle different queries

12
SEMEX Transform a Structured Query into a
Keyword Search
Article (Title dataspace) (Author alon
halevy) ? Article dataspace alon halevy
13
Challenges

• Example
• SELECT title
• FROM paper
• WHERE title LIKE Dataspaces AND year 2005

select title from paper where title LIKE
dataspaces and year 2005
Top-10 Precision 0
14
Challenges

• Example
• SELECT title
• FROM paper
• WHERE title LIKE Dataspaces AND year 2005

title paper title dataspaces and year 2005
Top-10 Precision 0
15
Challenges

• Example
• SELECT title
• FROM paper
• WHERE title LIKE Dataspaces AND year 2005

dataspaces 2005
Top-10 Precision 0.2
16
Challenges

• Example
• SELECT title
• FROM paper
• WHERE title LIKE Dataspaces AND year 2005

dataspaces 2005 paper title
Top-10 Precision 0.2
17
Challenges

• Example
• SELECT title
• FROM paper
• WHERE title LIKE Dataspaces AND year 2005

dataspaces 2005 paper
Top-10 Precision 0.6
18
Outline

• Motivation
• Problem Definition
• Our Algorithm
• Construct Query Graph
• Extract Keywords
• Experimental Results
• Conclusions and Future Work

19
Problem Definition

• Keyword extraction (Query transformation)
• Input a structured query
• Output a set of keywords
• Measure the quality of the extraction using top-k
  precision of keyword-search answers

20
Queries Considered

• Only consider basic SPJ (selection, projection,
  simple joining) queries in our first step
• Do not consider
• Disjunctions
• Comparison predicates (e.g., ?, lt, gt)
• Aggregations

21
How to Select Keywords?

• Example
• SELECT title
• FROM paper
• WHERE title LIKE Dataspaces AND year
• 2005

22
How to Select Keywords?

• Example
• SELECT title
• FROM paper
• WHERE title LIKE Dataspaces AND year
• 2005

23
How to Select Keywords?

• Example
• SELECT title
• FROM paper
• WHERE title LIKE Dataspaces AND year
• 2005

24
How to Select Keywords?

• Example
• SELECT title
• FROM paper
• WHERE title LIKE Dataspaces AND year
• 2005

25
Outline

• Motivation
• Problem Definition
• Our Algorithm
• Construct Query Graph
• Extract Keywords
• Experimental Results
• Conclusions and Future Work

26
Architecture Overview
SQL Queries
XML Queries
Triple Queries
Query-graph
Construction
Query Graph
Keyword
Extraction
Keyword Set
27
Construct Query Graph

• SELECT title
• FROM Paper, Person
• WHERE title LIKE Dataspaces
• AND Paper.author Person.id
• AND Person.name LIKE Halevy

title
?paper
?
author
title
person
Dataspaces
name
Halevy
28
Construct Query Graph

• SELECT title
• FROM Paper, Person
• WHERE title LIKE Dataspaces
• AND Paper.author Person.id
• AND Person.name LIKE Halevy

title
?paper
?
author
title
person
Dataspaces
name
Halevy
29
Construct Query Graph

• SELECT title
• FROM Paper, Person
• WHERE title LIKE Dataspaces
• AND Paper.author Person.id
• AND Person.name LIKE Halevy

title
?paper
?
author
title
person
Dataspaces
name
Halevy
30
Construct Query Graph

• SELECT title
• FROM Paper, Person
• WHERE title LIKE Dataspaces
• AND Paper.author Person.id
• AND Person.name LIKE Halevy

title
?paper
?
author
title
person
Dataspaces
name
Halevy
31
Construct Query Graph

• SELECT title
• FROM Paper, Person
• WHERE title LIKE Dataspaces
• AND Paper.author Person.id
• AND Person.name LIKE Halevy

title
?paper
?
author
title
person
Dataspaces
name
Halevy
32
Informativeness and Representativeness

• Example
• A paper authored by a person with name Halevy
• Informativeness ? Measure the amount of
  information provided by a label term (i-score)
• Representativeness ? Roughly correspond to the
  probability that searching the given term returns
  documents or webpages in the queried domain
  (r-score 1 - distractiveness)
• Informativeness gt distractivenessi.e., i-score
  r-score gt 1

? Halevys paper
33
Informativeness of a Label Depends on the Already
Selected Labels
Paper
34
Informativeness of a Label Depends on the Already
Selected Labels
Dataspace Paper in 2005
35
Informativeness of a Label Depends on the Already
Selected Labels
Dataspace Paper of Halevy and Franklin in 2005
36
Effect of a Selected Label on the i-scores of
Other Labels
title (0.8,0.2)
?paper (0.8,0.6)
? (0.8,0)
author (0.8,0.4)
title (0.8,0.2)
person (0.8,0.6)
Dataspaces (1,0.8)
name (0.8,0.2)
Halevy (1,0.8)
37
Effect of a Selected Label on the i-scores of
Other Labels
title (0.8,0.2)
?paper (0.8,0.6)
? (0.8,0)
author (0.8,0.4)
title (0.8,0.2)
-0.8
person (0.8,0.6)
Dataspaces (1,0.8)
name (0.8,0.2)
Halevy (1,0.8)
38
Effect of a Selected Label on the i-scores of
Other Labels
title (0.8,0.2)
?paper (0.8,0.6)
? (0.8,0)
author (0.8,0.4)
title (0,0.2)
-0.8
person (0.8,0.6)
Dataspaces (1,0.8)
name (0.8,0.2)
Halevy (1,0.8)
39
Effect of a Selected Label on the i-scores of
Other Labels
title (0.8,0.2)
?paper (0.8,0.6)
? (0.8,0)
-0.4
author (0.8,0.4)
title (0,0.2)
-0.8
person (0.8,0.6)
Dataspaces (1,0.8)
name (0.8,0.2)
Halevy (1,0.8)
40
Effect of a Selected Label on the i-scores of
Other Labels
title (0.8,0.2)
?paper (0.4,0.6)
? (0.8,0)
-0.4
author (0.8,0.4)
title (0,0.2)
-0.8
person (0.8,0.6)
Dataspaces (1,0.8)
name (0.8,0.2)
Halevy (1,0.8)
41
Effect of a Selected Label on the i-scores of
Other Labels
-0.1
title (0.8,0.2)
?paper (0.4,0.6)
? (0.8,0)
-0.4
-0.1
author (0.8,0.4)
title (0,0.2)
-0.8
person (0.8,0.6)
Dataspaces (1,0.8)
name (0.8,0.2)
Halevy (1,0.8)
42
Effect of a Selected Label on the i-scores of
Other Labels
-0.1
title (0.7,0.2)
?paper (0.4,0.6)
? (0.8,0)
-0.4
-0.1
author (0.7,0.4)
title (0,0.2)
-0.8
person (0.8,0.6)
Dataspaces (1,0.8)
name (0.8,0.2)
Halevy (1,0.8)
43
Extract Keywords Using Greedy Algorithm
title (0.8,0.2)
?paper (0.8,0.6)
? (0.8,0)
author (0.8,0.4)
title (0.8,0.2)
person (0.8,0.6)
Dataspaces (1,0.8)
name (0.8,0.2)
Halevy (1,0.8)
44
Extract Keywords Using Greedy Algorithm
title (0.8,0.2)
?paper (0.8,0.6)
? (0.8,0)
author (0.8,0.4)
title (0.8,0.2)
person (0.8,0.6)
Dataspaces (1,0.8)
name (0.8,0.2)
Halevy (1,0.8)
Step 1 Choose all labels of value nodes,
update i-scores of the rest labels
45
Extract Keywords Using Greedy Algorithm
title (0.6688,0.2)
?paper (0.5375,0.6)
? (0.8,0)
author (0.575,0.4)
title (0.3688,0.2)
person (0.5,0.6)
Dataspaces (1,0.8)
name (0.05,0.2)
Halevy (1,0.8)
Step 1 Choose all labels of value nodes,
update i-scores of the rest labels
46
Extract Keywords Using Greedy Algorithm
title
?paper (0.5375,0.6)
?
author
title
person (0.5,0.6)
Dataspaces (1,0.8)
name
Halevy (1,0.8)
Step 2 Choose the label with highest i r if i
r gt 1, update i-scores of the rest labels
47
Extract Keywords Using Greedy Algorithm
title
?paper (0.5375,0.6)
?
author
title
person (0.5,0.6)
Dataspaces (1,0.8)
name
Halevy (1,0.8)
Step 2 Choose the label with highest i r if i
r gt 1, update i-scores of the rest labels
48
Extract Keywords Using Greedy Algorithm
title
?paper (0.5375,0.6)
?
author
title
person (0.2,0.6)
Dataspaces (1,0.8)
name
Halevy (1,0.8)
Step 3 Iterate step 2 until no more labels can
be added
49
Extract Keywords Using Greedy Algorithm
title
?paper (0.5375,0.6)
?
author
title
person
Dataspaces (1,0.8)
name
Halevy (1,0.8)
Final keyword set Dataspaces, Halevy, paper
50
Extract Keywords Using Greedy AlgorithmAnother
Example
title
?paper
?
author
title
author
person
person
Dataspaces (1,0.8)
name
name
Franklin (1,0.8)
Halevy (1,0.8)
Final keyword set Dataspaces, Halevy, Franklin
51
Outline

• Motivation
• Problem Definition
• Our Algorithm
• Construct Query Graph
• Extract Keywords
• Experimental Results
• Conclusions and Future Work

52
Experiment Setup

• Six different domains movie, geography, company
  profiles, bibliography, DBLP, and car profiles
• Randomly select text values
• Vary two parameters in the selected queries
• values ? Number of attribute values in the query
  (information given)
• Length ? Longest path from a queried instance to
  other instances (complexity of structure
  information)
• Measure the quality of extracted keywords with
  top-k precision

53
Initialize i-scores and r-scores ? Without domain
knowledge

• i-scores
• Text-value labels 1
• Labels of queried instances 1
• Other labels 0.8
• r-scores
• Text-value node labels 0.8
• Labels of association edges between instances of
  the same type 0.8
• Instance node labels 0.6
• Association edge labels 0.4
• Attribute edge labels 0.2
• Number-value node labels 0

54
High Precisions in All Data Domains w/o Domain
Knowledge

• Average top-2 precision was 0.68
• Average top-10 precision was 0.59

55
Initialize i-scores and r-scores ? With domain
knowledge

• Can obtain more meaningful r-scores
• How
• Do keyword search on the labels
• Calculate the percentage of top-k answers that
  are related to the queried domain

56
Applying Domain Knowledge Increases Performance

• The top-10 precisions increased 39 on average.

57
Increasing Value Increases Precision
Movie
Geography
Movie
Geography
58
Increasing Length Decreases Precision
Movie
Geography
Movie
Geography
59
Conclusions

• Dataspace Support Platforms require answering
  structured queries on unstructured data
• Solution Transform a structured query into
  keyword search by keyword extraction
• Our algorithm obtains good results in various
  domains

60
Future Work

• Refine the extracted keyword set by considering
  the schema or a corpus of schemas
• Use existing structured data to supplement the
  selected keyword set
• Perform linguistic analysis of the words in the
  structured query
• Develop methods for ranking answers from
  structured and unstructured data sources

61
Precise Data Integration
Cost
Benefit
Heterogeneity
62
Approximate Data Integration
Benefit
Cost
Heterogeneity
63
Answering Structured Queries on Unstructured Data

• Jing Liu, Xin (Luna) Dong, Alon Halevy
• Univ. of Washington
• http//data.cs.washington.edu/semex/semex.html

64
Related Work

• SCORE CIKM, 2005
• Extract keywords from query results on structured
  data.
• Not generic.

65
Algorithm of Updating i-scores

• Effect of selected label on other labels
• Source node (or edge) has flow volume rs.
• The flow value is divided among the neighbors.
• The flow value decreases exponentially with the
  number of hops.
• Update i-scores
• inew iold Effect of selected label