Interconnection Semantics for Keyword Search in XML

1
Interconnection Semantics for Keyword Search in
XML

• Sara Cohen
• Technion Israel

Yaron Kanza University of Toronto
Benny Kimelfeld Hebrew University
Yehoshua Sagiv Hebrew University
CIKM 2005 Bremen, Germany
2
Goal

• We present a general framework to efficiently
  determine when different parts of an XML document
  are semantically related

Who cares?

• A step towards bridging the gap between keyword
  search and database querying
• For example
• Simplification and enhanced flexibility of query
  languages for XML (e.g., schema free)
• Taking into account structural properties for
  ranking/filtering results of keyword search

3
Schema-Free Queries
Can be formulated in the tools proposed by Cohen
et al., ICDT03 and Li et al., VLDB04
SELECT article/title FROM interconnected
//article, //author, //publisher
WHERE author/nameA. Cohen and
publisher/nameAm. Publishing
Find the titles of articles that were written by
A. Cohen and were published by Am. Publishing
4
Schema-Free Queries
SELECT article/title FROM interconnected
//article, //author, //publisher
WHERE author/nameA. Hunt and
publisher/nameAm. Publishing
5
XML Structure 1
author is nested inside article
6
XML Structure 2
article is nested inside author
7
XML Structure 3
article and author are connected through ID
references
8
Keyword Search over XML

• Typically, ranking of XML fragments is determined
  by the frequencies of the keywords in the
  fragments
• We propose to take into account the answer to the
  following test
• Do the keywords appear in semantically related
  parts of the fragment?

9
Keyword-Search Example
Cohen , IR
10
A Result
Cohen , IR
Cohen and IR are in the same department
11
Another Result
Cohen , IR
Cohen and IR are in the same department and
Cohen wrote an article about IR
This fragment should have a higher rank
Identifying meaningful relationships can improve
ranking in keyword search
12
Existing Approaches

• Each of the existing approaches proposes a
  specific method for deriving meaningful
  relationships
• ID references are ignored
• That is, documents are always trees
• The schema is ignored
• Therefore, missing information is not taken into
  account

13
Our Contribution

• We propose a framework that enables a large
  variety of approaches
• A single method is not likely to be appropriate
  in all cases
• Both ID references and the schema can be taken
  into account
• We propose specific semantics that are
  automatically derived from schemas
• We present efficient algorithms for applying
  interconnection semantics

14
Contents

• Introduction
• Formal Framework
• Derived Semantics
• Computing Interconnectivity
• Undirected and Universal Semantics
• Conclusion and Future Work

15
Contents

• Introduction
• Formal Framework
• Derived Semantics
• Computing Interconnectivity
• Undirected and Universal Semantics
• Conclusion and Future Work

16
An XML Document
A rooted directed graph
A node is an object that has a label and
(possibly) a value
Edges represent element nesting and ID references
17
A Schema
A rooted directed graph
A node is a label
18
A Schema Defines the Document Structure
The root of the schema is the label of the root
of the document
19
A Schema Defines the Document Structure
An edge in the document is allowed only if an
edge between the corresponding labels appears in
the schema
20
Relationships and Trees

• In our framework, relationships among objects are
  represented by trees
• Trees represent atomic relationships, i.e.,
  relationships that are indivisible
• Our trees are directed
• That is, represent hierarchies
• Relaxed later

21
Patterns

• In the formal framework, patterns are the basic
  building blocks

Formally, a pattern is a pair (L,C)
C is a tree of labels
L is a set of labels
(
)
,
title,publication,author

• C contains L
• C has no redundant edges

22
Interconnection by Patterns

• A pattern defines when a set O of objects with a
  given set L of labels is interconnected
• O is interconnected if the objects are in a tree
  that is isomorphic to the pattern

(
)
,
title,publication,author
23
Interconnection by Patterns
24
Interconnection by Patterns
Interconnected
25
Interconnection Semantics

• An interconnection semantics P is a set of
  patterns
• A set of objects is interconnected by P if it is
  interconnected by a pattern of P

(title,name , )
(title,name , )
26
A Framework of Semantics

• Various approaches for deriving semantic
  relationships can be represented by means of
  interconnection semantics
• Interconnection semantics can be given explicitly
  (e.g., by manually defining the patters)
• Alternatively, interconnection semantics can be
  given implicitly (i.e., be derived) by defining
  conditions that characterize patterns
• Derived semantics are discussed next

27
Contents

• Introduction
• Formal Framework
• Derived Semantics
• Computing Interconnectivity
• Undirected and Universal Semantics
• Conclusion and Future Work

28
Derived Semantics

• In principle, an interconnection semantics can be
  given explicitly by authors (or users) of XML
  documents
• However, generating a semantics may be a
  cumbersome task
• Moreover, the semantics may be very large

We need semantics that are automatically derived
29
Derived Semantics (cont.)

• A derived semantics is essentially a rule that
  states which subtrees of the schema appear in
  patterns
• A default schema can be obtained from the
  document
• One semantics that always captures our intuition
  is not likely to exist, therefore it is of
  interest to explore several approaches for
  deriving semantics
• We explore a wide spectrum of specific derived
  semantics
• Our derived semantics demonstrate the strengths
  and weaknesses of different approaches for
  obtaining such semantics

30
A Simple Example
title,publication
31
The Interconnection Semantics Pall

• The semantics Pall generalizes the approach of
  Cohen et al., ICDT03 to graph (rather than
  tree) documents
• Given a schema S, the semantics Pall(S) is the
  set of all patterns (L,C), where C is a subtree
  of S
• Note that a set of objects is Pall(S)-interconnect
  ed if it is contained in a uniquely labeled
  subtree of the document

32
The Subtrees of title,publication
33
The Subtrees of title,author
34
The Subtrees of title,author
?
Is this what we mean?
35
The Interconnection Semantics Pmin

• The semantics Pall(S) may contain patterns that
  imply rather weak relationships
• We follow the convention that small trees imply
  strong relationships
• The semantics Pmin(S) contains, for each set L of
  labels, only the patterns (L,C) of Pall(S), such
  that C is of minimal size

36
The Subtrees of title,author
This subtree of the schema is minimal w.r.t.
title,author
37
Pmin(S)-Interconnected title and author
This subtree of the document shows
Pmin(S)-interconnection!
38
The Subtrees of title,author
This subtree of the schema is not minimal w.r.t.
title,author
39
Pmin(S)-Interconnected title and author
This subtree does not show Pmin(S)-interconnectio
n!
40
What about these title and author?
41
What about these title and author?
This subtree does not show Pmin(S)-interconnectio
n
42
What about these title and author?
This subtree does not show Pmin(S)-interconnectio
n
because this subtree is smaller
43
The Interconnection Semantics Puca

• We consider a different notion of minimality that
  is based on structure
• Consider an acyclic schema S and a pattern
  p(L,C) in Pall(S)
• Intuitively, p is structurally minimal if
  internal nodes in C cannot be roots of trees
  containing L
• Formally, p is structurally minimal if only the
  root of C is a common ancestor of L in S
• The semantics Puca(S) is the set of all
  structurally minimal patterns in Pall(S)

44
Some Examples
45
One Structurally Minimal Pattern
(title,author , )
In the schema, article is the only
common ancestor of title,author
46
Another Structurally Minimal Pattern
(title,author , )
In the schema, inproc. is the only
common ancestor of title,author
47
A Third Structurally Minimal Pattern
(title,author , )
In the schema, inproc. is the only
common ancestor of title,author
48
Not a Structurally Minimal Pattern
(title,author , )
In the schema, department, publications and
incproc. are all common ancestors of
title,author
49
Back to the Document
50
Puca(S)-Interconnected title and author
This subtree shows Puca(S)-interconnection!
51
Puca(S)-Interconnected title and author
This subtree shows Puca(S)-interconnection!
52
Not Puca(S)-Interconnected
This subtree does not show Puca(S)-interconnectio
n!
53
Contents

• Introduction
• Formal Framework
• Derived Semantics
• Computing Interconnectivity
• Undirected and Universal Semantics
• Conclusion and Future Work

54
Interconnectivity Problems

• Given an interconnection semantics, we are
  interested in solving two problems
• Determine whether a given set of objects is
  interconnected
• Arises in keyword search
• Generate all interconnected sets of objects
  having a given set of labels
• Arises in evaluation of queries that incorporate
  interconnectivity

55
Computing Explicit Semantics

• When the interconnection semantics is given
  explicitly, there are highly efficient algorithms
  for solving the interconnectivity problems
• I.e., polynomial under query-and-data complexity
• In query evaluation, the first k results can be
  enumerated quickly
• I.e., evaluation in incremental polynomial time
• We can solve the interconnectivity problems by
  translating patterns into queries of standard
  languages, e.g., XQuery
• Thus, existing engines can be used

56
Computing Derived Semantics

• There are two approaches for solving
  interconnectivity problems when the semantic is
  derived (implicit)
• Pattern Extraction
• Direct Computation (w/o extracting patterns)

57
1st Approach Pattern-Extraction

• This approach is basically a reduction to the
  case where interconnection semantics are given
  explicitly
• In particular, given an implicit semantics P and
  a set L of labels, we extract from the schema all
  patterns (L,C) of P
• Thus, we create an explicit representation of the
  implicit semantics and use the corresponding
  explicit-semantics algorithms

58
Puca(S) Patterns for title,author
59
Extracting the Patters

• Given a set L of labels, we can extract the
  relevant patterns of our derived semantics very
  efficiently, when measuring the complexity in
  terms of the combined size of input and the
  output
• I.e., with polynomial delay
• For Pall(S) and Puca(S), we can use the
  keyword-search algorithms given in Kimelfeld and
  Sagiv, DBPL05
• An algorithm for Pmin(S) is given in the
  proceedings

60
2nd Approach Direct Computation

• There are highly efficient algorithms for
  evaluating the extracted patterns
• However, there may be a large (i.e., exponential)
  number of relevant patterns
• When the number of patterns is large, we are
  interested in solutions that compute
  interconnectivity directly, i.e., without
  extracting the patterns first

61
Algorithms for Direct Computation

• There are efficient algorithms for direct
  computation of our derived semantics
• An exception is Pall(S) in cyclic schemas, which
  is intractable
• We use data complexity as a yardstick of
  efficiency, which is conventional in DB theory
• In some common cases, interconnectivity problems
  are more tractable
• I.e., there are efficient algorithms under
  query-and-data complexity (like tree join vs.
  general join)
• Usually, when either the schema is a tree or the
  document has no ID references
• Details are in the proceedings

62
Contents

• Introduction
• Formal Framework
• Derived Semantics
• Computing Interconnectivity
• Undirected and Universal Semantics
• Conclusion and Future Work

63
Undirected Relationships
An article was written by authors from two
different departments
64
Undirected Relationships
65
Undirected Relationships
66
Undirected Patterns and Semantics
Undirected Patterns

• (L,C)
• L is a set of labels
• C is an undirected tree of labels

An undirected interconnection semantics consists
of undirected patterns
67
Universal Interconnectivity

• Until now, an interconnection semantics was
  applied in an existential manner
• That is, a set of nodes is interconnected if
  there is an evidence for interconnectivity (e.g.,
  a uniquely labeled subtree that contains the
  given objects)
• When applying interconnection semantics in a
  universal manner, we consider all subtrees of the
  document (contexts) that contain the given
  objects
• A set of objects is universally interconnected if
  every subtree contains an evidence for
  interconnectivity

68
An Example
69
An Example
This subtree shows interconnectivity
70
An Example
The two objects are NOT universally
interconnected!
This subtree does not show interconnectivity
71
Derived Semantics

• We considered two undirected derived semantics
• Pu-all(S) all undirected patterns of a schema S
• Pu-min(S) all minimal patterns in Pu-all(S)
• We considered the universal versions of the
  semantics Pall(S) and Pu-all(S)
• Usually, these derived semantics can be computed
  efficiently
• ?Pall(S) is harder than Pall(S) and ?Pu-all(S) is
  easier
• However, completely different proof techniques
  are required
• More details, including algorithms and complexity
  results, are in the proceedings

72
Contents

• Introduction
• Formal Framework
• Derived Semantics
• Computing Interconnectivity
• Undirected and Universal Semantics
• Conclusion and Future Work

73
Interconnection Semantics

• A framework for determining meaningful
  relationships among XML objects
• Interconnection semantics can be either
  explicitly constructed or automatically
  (implicitly) derived
• Users can tweak implicit semantics by adding
  specific patterns
• The framework enables many different types of
  semantics, with different strengths

74
Containments among our Derived Semantics
75
Complexity

• Explicit semantics can be computed very
  efficiently
• Under reasonable conditions, our derived
  semantics can also be efficiently computed
• Different semantics require totally different
  techniques
• Two computation approaches
• Pattern extraction
• Direct computation

76
Future Work

• Experimentation
• Which semantics works better in practical
  scenarios?
• As a result of the experimentation, discover new
  (better) semantics
• Information Retrieval
• Combine our approach with IR techniques for
  ranking XML fragments
• Richer Data Model
• That is, richer schema and document models