Controlling Overlap in

Content-Oriented XML Retrieval

Charles L. A. Clarke

School of Computer Science

University of Waterloo

Waterloo, Canada

Content-Oriented XML Retrieval

Documents not Data

Natural Language Queries “text index compression”

Ranked Results

Goal

In response to a user query, select and return an appropriately ranked mix of document components… paragraphs, sections, subsections, bibliographic entries, articles, etc.

In response to a user query, select and return an appropriately ranked

IEEE Journal Article in XML

<article> <fm> <atl>Text Compression for Dynamic Document Databases</atl> <au>Alistair Moffat</au> <au>Justin Zobel</au> <au>Neil Sharman</au> <abs> <p> <b>Abstract</b> For compression of text...</p> </abs> </fm> <bdy> <sec> <st>INTRODUCTION</st> <ip1>Modern document databases contain vast quantities of text...</ip1> <p>There are good reasons to compress the text stored in a...</p> </sec> <sec> <st>REDUCING MEMORY REQUIREMENTS</st>... <ss1><st>2.1 Method A</st> <ip1>The first method... </sec> ... </bdy></article>

XML Tree

article

fm bdy

atl au au auabs

p

b

sec

st ip1 p

sec

stss1

st

Ranking Document Components

1) Split into components

2) Treat each as a “document”

3) Rank using standard IR methods

#1 #6

#2

#103

#8

#4

Problem: Top ranks dominated by related components

eg. “text index compression”

32.000923 /co/2000/ry037.xml /article[1]/bdy[1]31.861366 /co/2000/ry037.xml /article[1]31.083460 /co/2000/ry037.xml /article[1]/bdy[1]/sec[2]30.174324 /co/2000/ry037.xml /article[1]/bdy[1]/sec[5]29.420393 /tk/1997/k0302.xml /article[1]29.250019 /tk/1997/k0302.xml /article[1]/bdy[1]/sec[1]29.118382 /tk/1997/k0302.xml /article[1]/bdy[1]29.075621 /co/2000/ry037.xml /article[1]/bdy[1]/sec[3]28.417294 /tk/1997/k0302.xml /article[1]/bdy[1]/sec[6]28.106693 /tp/2000/i0385.xml /article[1]27.761749 /co/2000/ry037.xml /article[1]/bdy[1]/sec[7]27.686905 /tk/1997/k0302.xml /article[1]/bdy[1]/sec[3]27.584927 /tp/2000/i0385.xml /article[1]/bdy[1]27.273247 /co/2000/ry037.xml /article[1]/bdy[1]/sec[4]27.186977 /tp/2000/i0385.xml /article[1]/bdy[1]/sec[1]27.181036 /tk/1997/k0302.xml /article[1]/bm[1]27.072521 /tk/1997/k0302.xml /article[1]/bdy[1]/sec[3]/ss1[1]26.992224 /co/2000/ry037.xml /article[1]/bdy[1]/sec[5]/ss1[1]

Is overlap always bad?

General overview

Topic-focused components

article

fm bdy

atl au au auabs

p

b

sec

st ip1 p

sec

stss1

st

Overview

• Baseline retrieval method

• Evaluation methodology

• Re-ranking algorithm

Baseline Retrieval Method

• Treat each component as a separate document.

• Rank using Okapi BM25.

• Tune retrieval parameters to task and collection.

Okapi BM25

Given a query Q, a component x is scored:

where:D = number of documents in the collectionDt = number of documents containing tqt = frequency that t occurs in the queryxt = frequency that t occurs in xK = k1((1 – b) + b·lx/lavg)lx = length of xlavg = average document length

Σ qt(k1 + 1) xt

K + xt

D – Dt + 0.5Dt + 0.5( )log

t є Q

k1 = 10.0b = 0.80

Overview

• Baseline retrieval method

• Evaluation methodology

• Re-ranking algorithm

XML Evaluation - INEX 2004

• Third year of the “INitiative for the Evaluation of XML retrieval” (INEX).

• Encourages research into XML information retrieval technology (similar to TREC).

• Over 50 participating groups.

INEX 2004 Content-Only Retrieval Task

• Test collection of articles taken from IEEE journals between 1995 and 2002.

• 40 “adhoc” queries:–text index compression– software quality– new fortran 90 compiler

• Each group could submit up to three runs consisting of the top 1500 components for each query.

Relevance Assessment

• INEX uses two dimensions for relevance– exhaustivity: the degree to which a

component covers a topic– specificity: the degree to which a

component is focused on a topic

• A four-point scale is used in both dimensions (eg. a (3,3) component is highly exhaustive and highly specific)

INEX 2004 Evaluation Metrics

• Mean average precision (MAP)

• XML cumulated gain (XCG)

• Various quantization functions:– strict quantization - (3,3) elements only

– generalized quantization

– specificity-oriented generalization

Overview

• Baseline retrieval method

• Evaluation methodology

• Re-ranking algorithm

Controlling Overlap

Starting with a component ranking generated by the baseline method, elements are re-ranked to control overlap.

Scores of those components containing or contained within higher ranking components are iteratively adjusted.

Controlling Overlap

#1#2

#4

#6

#10

#87#203

#21

#863

Re-ranking Algorithm

1. Report the highest ranking component.

2. Adjust the scores of the unreported components.

3. Repeat steps 1 and 2 until the top m components have been reported.

Adjusting Scores

xt = frequency that t occurs in component x

Σ qt(k1 + 1) xt

K + xt

D – Dt + 0.5Dt + 0.5( )log

t є Q

ft = frequency that t occurs in xgt = frequency that t has been reported in

subcomponents of x

xt = ft – α·gt

Adjusting Scores (α = 0.5)

ft = 6gt = 0xt = 6

ft = 13gt = 0xt = 13

#1

ft = 13gt = 6xt = 10

Re-ranking Algorithm

1. Maintain tree nodes in a priority queue ordered by score.

2. Report component at front of queue.

3. Propagate adjustments up and down the tree re-ordering nodes in the queue.

4. Repeat steps 2 and 3 until the top m components have been reported.

Mean Average Precision vs. XCG

Extended Re-Ranking Algorithm

• One parameter (α) not sufficient.

• Reported in descendant (α) vs. ancestor (β).

• Number of times term reported in ancestor: 1 = β0 ≥ β1 ≥ β2 ≥ … ≥ βn = 0

xt = βi·( ft – α·gt )

Concluding Comments

• Overlap isn’t always bad

• Overlap should be controlled not eliminated

• Current and future work:– Modifying k1 to reflect term frequency adjustments

– More appropriate evaluation methodology

– Learning α, βi

Questions?

Controlling Overlap in

Content-Oriented XML Retrieval

Charles L. A. ClarkeSchool of Computer Science

University of WaterlooWaterloo, Canada