1. Structure and Content Scoring for XML
Amélie Marian (Rutgers University)
Joint work with:
Sihem Amer-Yahia (AT&T Research Labs)
Nick Koudas (University of Toronto)
Divesh Srivastava (AT&T Research Labs)
David Toman (University of Waterloo)

2. Motivations: XML Data Heterogeneity
book
book
Data
book
title
(Great Expectations)
edition
(paperback)
info
author
(Dickens)
info
info
author
(Dickens)
title
(Great Expectations)
edition
(paperback)
title
(Great Expectations)
author
(Dickens)
Heterogeneous XML Data about books
Query:
book[./info[./title=“Great Expectations” and
./author=“Dickens”] and ./edition=“paperback”]
book
title
(Great Expectations)
edition
(paperback)
info
author
(Dickens)
Query root node:
Distinguished node
1/2/2019
Amélie Marian - Rutgers University

3. Amélie Marian - Rutgers University
XML Query Relaxation
Query
[Amer-Yahia et al. EDBT’02]
book
title
(Great Expectations)
edition
(paperback)
info
author
(Dickens)
Tree pattern relaxations:
Leaf node deletion
Edge generalization
Subtree promotion
book
book
book
title
(Great Expectations)
edition
(paperback)
info
author
(Dickens)
Data
edition?
info
info
author
(Dickens)
title
(Great Expectations)
edition
(paperback)
title
(Great Expectations)
author
(Dickens)
1/2/2019
Amélie Marian - Rutgers University

4. Amélie Marian - Rutgers University
Motivations
Top-k query processing suitable for relaxed XML queries over heterogeneous collections
Return k XML nodes that are closest to query structure
Opportunity for more efficient query processing
Need scoring mechanism to identify best k answers
1/2/2019
Amélie Marian - Rutgers University

5. Amélie Marian - Rutgers University
Contributions
Scoring mechanism for XML queries
Data structures for top-k query processing
Experimental evaluation
1/2/2019
Amélie Marian - Rutgers University

6. Scoring Functions Critical for Top-k Query Processing
Top-k answer quality depends on scoring function
Efficient top-k query processing requires scoring function:
Monotonic
Fast to compute
Little attention given to scoring functions for structured and semi-structured data
Extensively studied over text data (e.g., tf.idf)
Proposed scoring function inspired by tf.idf for XML data
1/2/2019
Amélie Marian - Rutgers University

7. Adaptation of tf.idf to XML Queries
Document Collection (Information Retrieval)
XML Document
Document
XML Node (result is a subtree rooted at a distinguished node, i.e., a node with a given label and structural properties)
Keyword(s)
Query Pattern
idf (inverse document frequency) is a function of the fraction of documents that contain the keyword(s)
idf is a function of the fraction of distinguished nodes that match the query pattern
tf (term frequency) is a function of the number of occurrences of the keyword in the document
tf is a function of the number of ways the query pattern matches the distinguished node
1/2/2019
Amélie Marian - Rutgers University

8. Scoring Function for XML Approximate Matches
book
title
(Great Expectations)
edition
(paperback)
info
book
title
(Great Expectations)
edition
(paperback)
info
author
(Dickens)
book
book
Required properties:
Exact matches should be scored higher than relaxed matches (idf)
Distinguished nodes with several matches should be ranked higher than those with fewer matches (tf)
How to combine tf and idf?
tf.idf, as used by IR, violates above properties
Ranking based on idf, then breaking ties using tf satisfies the properties
info
info
edition
(paperback)
edition
(paperback)
author
(Dickens)
title
(Great Expectations)
(a)
(b)
score(a) <= score(b)
score(a) >= score(b)
1/2/2019
Amélie Marian - Rutgers University

9. A Family of Scoring Methods for XML Path Queries
book
title
(Great Expectations)
edition
(paperback)
info
author
(Dickens)
Query
Twig predicate
High quality
Expensive computation
Path predicates
Binary predicates
Low quality
Fast computation
book
title
(Great Expectations)
edition
(paperback)
info
author
(Dickens)
book
title
(Great Expectations)
edition
(paperback)
author
(Dickens)
info +
book
title
(Great Expectations)
edition
(paperback)
author
(Dickens)
info +
1/2/2019
Amélie Marian - Rutgers University

10. Amélie Marian - Rutgers University
Contributions
Scoring mechanism for XML queries
Data structures for top-k query processing
Experimental evaluation
1/2/2019
Amélie Marian - Rutgers University

11. Matrix Representation of Twigs
Twigs (queries and tuples) can be represented by matrices that capture all relationships in the query: a b c d e
Partial Tuple:
Query: a1
(not joined with e yet)
(no matches for e)
(e1 matches) b1 d1 c1 e1 a b c d e = / // X a b c d e = // X / ?
// X X / =
X X X X X
Matrix subsumption used to compare tuple and queries
1/2/2019
Amélie Marian - Rutgers University

12. Representing Relaxed Query Patterns: DAG Structureb c
Each child is more relaxed (has more matches) than its parent
idf of a child is no higher than the idf of its parent
idf scores are accessible in constant time for any match (complete or partial) using hash function a b c a b c a b c a b c a b c a b a a
Exhaustive algorithm to build the DAG c b a
1/2/2019
Amélie Marian - Rutgers University

13. Information stored in the DAGb c
idf score information:
idf=(1+|a|)/(1+|ap|), where |ap| is the number of a nodes that satisfy the query predicate
For query processing:
Best possible score from here
Best possible score after each remaining join operations
Number of matches (useful for tf)
1.228 a b c a b c
1.2
1.195 a b c a b c
1.167
1.195 a b c a
1.167
1.156 b a a
1.049
1.156 c b a 1
1/2/2019
Amélie Marian - Rutgers University

14. Query Processing using the DAG
Benefits:
Score computation done in a preprocessing phase (using exact or approximate information)
Score access during query processing done in constant time
Additional information needed for query processing precomputed and accessed in constant time (e.g., score upper bound)
tf estimated at runtime based on available information
1/2/2019
Amélie Marian - Rutgers University

15. Quality/Space/Time tradeoff
Binary Predicates
Smaller DAG (O(4q))
Faster pre-processing (and processing)
Lower Quality (fewer possible scores)
Path Predicates and Twig
DAG is O(4q^2/2)) in space (still reasonable in practice)
More pre-processing
Higher Quality (more differences between scores)
1/2/2019
Amélie Marian - Rutgers University

16. Amélie Marian - Rutgers University
Contributions
Scoring mechanism for XML queries
Data structures for top-k query processing
Experimental evaluation
1/2/2019
Amélie Marian - Rutgers University

17. Amélie Marian - Rutgers University
Experimental Setup
Data:
Synthetic heterogeneous document collections generated with Toxgene
Real dataset: Wall Street Journal Treebank corpora
Pregenerated queries exhibiting different sizes, query structures and predicates
Measures:
DAG size
DAG preprocessing time
Query processing time
Precision (percentage of top-k answers that are actual top-k answers, as given by Twig)
1/2/2019
Amélie Marian - Rutgers University

18. Amélie Marian - Rutgers University
XML Scoring Precision
1/2/2019
Amélie Marian - Rutgers University

19. XML Scoring Preprocessing Time
1/2/2019
Amélie Marian - Rutgers University

20. Amélie Marian - Rutgers University
XML Scoring Real data
1/2/2019
Amélie Marian - Rutgers University

21. Amélie Marian - Rutgers University
Conclusions
Scoring method for XML queries
Inspired from tf.idf
Accounts for structure and content
Accounts for structural relaxations
Efficient data structures to compute and access scores during top-k query processing
DAG
Matrix representation of queries and tuples
Evaluation of the scoring methods tradeoffs
Answer quality vs. preprocessing time
1/2/2019
Amélie Marian - Rutgers University

22. Amélie Marian - Rutgers University
Related Work
IR Scoring
Content only
XML Scoring
Content with structure
XIRQL [XML&IR’00], JuruXML [SIGIR’03], IR-CADG [WebDB’04]
None of these techniques account for structural relaxations (with the exception of our previous work [ICDE’05])
XML Structural Relaxation
FleXPath [SIGMOD’04], Kanza and Sagiv [PODS’01], Schlieder [EDBT’02], Delobel and Rousset [FMII’01]
1/2/2019
Amélie Marian - Rutgers University

23. Amélie Marian - Rutgers University
Future Work
Streaming scenarios
Incremental updates on DAG
Approximate scoring
Integration with approximate text scoring
Extend proposed XML scoring function to handle text content approximation (e.g., misspellings)
Unify structure and content score
Quality evaluation (INEX)
1/2/2019
Amélie Marian - Rutgers University