Multimedia Retrieval based on Geometric Distance in Semi-structured

Document

Sana Fakhfakh, Mohamed Tmar and Walid Mahdi

Laboratory MIRACL, Institute of Computer Science and Multimedia of Sfax, Sfax University, Sfax, Tunisia

Keywords:

Geometric Distance, Image, Multimedia Retrieval, XML Element, Structure.

Abstract:

This paper is included on multimedia retrieval in XML document, whose goal is to ﬁnd relevant multimedia

element. In this article, we are particularly interested by studing the impact of various structural factors for

image retrieval by our proposed method using a new evidence of source applied to media ”image”. This

method consist to deﬁne a geometric distance between XML nodes. Experiments are undertaken into two data

sets ”INEX 2007” and ”ImageCLEF 2010”. The obtained results showed effectiveness of our approach.

1 INTRODUCTION

This paper falls under the context of multimedia re-

trieval in XML documents. The need with this kind

of information is justiﬁed by quick change of scopes

of application which use structural documents (for-

mat HTML or XML) what imposes new challenges in

the ﬁeld of search for information. Indeed, nowadays

XML document passed a simple tool for exchanging

data to a new storage medium. XML document in-

cludes textual element and multimedia element such

as image, audio and video. These elements are orga-

nized according to structure which includes informa-

tion notably although there is not only one manner to

organize contents. However, the choice of structure

depends greatly on the context of use of the textual

contents.

Mainly in the literature, there are two main classes

of approaches in the ﬁeld of multimedia retrieval: re-

trieval methods based on multimedia content (MR-

content) and multimedia methods to retrieval based on

context (MR-Context). The approaches of the multi-

media retrieval based on content use speciﬁc features

of low level according to type of media (Lew, 2006).

We can cite for example image retrieval that exploits

visual features (the color, texture, forms ···). These

methods have proven effective with media ”image” in

well deﬁned ﬁelds such as medical ﬁeld this is due

to requirement for thorough knowledge of distinctive

media. This type of research can be applied to only

one type of media in system due to lack of semantic

representation in media content.

The approaches of the multimedia retrieval based

on context do not depend on type of media in question

(Elghazel et al., 2005) (Tjondronegoro et al., 2005).

Indeed, these methods rely on information surround-

ing the multimedia element representing its seman-

tic description. Multimedia retrieval based on tex-

tual context is most used, although the structural con-

text remains an obvious source which plays a part

paramount in understanding of structured documents.

In this article, we focus on techniques for multi-

media retrieval based on textual and structural con-

text in XML documents. This type of document in-

cludes textual information and structural constraints.

So, XML document cannot be effectively exploited

by classical techniques of IR, which regard document

as a plane source of information.

The implicit incorporation of multimedia ele-

ments in XML documents requires the exploitation of

textual context for multimedia retrieval. However, the

textual context remains insufﬁcient in most of time.

The idea is to calculate the relevancy score of me-

dia element based on information from the textual

and structural context to answer a speciﬁc informa-

tion needs of user, expressed as query composed of

set of keywords.

Let us take for example a image media. If we ex-

ploit the image context which is composed by descrip-

tion of its contents such as its title, name, descriptive

texts which surround it, title of XML document ... In

following ﬁgure, we present document extracted from

”WIKIPEDIA” encyclopedia describing lion. We no-

tice the existing simultaneous textual and multime-

dia information. For image retrieval from time after

”Pleistocene”, we extract information from the tex-

220

Fakhfakh S., Tmar M. and Mahdi W..

Multimedia Retrieval based on Geometric Distance in Semi-structured Document.

DOI: 10.5220/0004964102200225

In Proceedings of the 10th International Conference on Web Information Systems and Technologies (WEBIST-2014), pages 220-225

ISBN: 978-989-758-023-9

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

XML Document

Male

Female (lioness)

The lion (Panthera leo) is one of

the four big cats in the genus

Panthera and a member of the

family Felidae. With some males

exceeding 250 kg (550 lb) in

weight,[4] it is the second-largest

living cat after the tiger.....

Pleistocene

Lion

Cave lions in the

Chauvet Cave,

France

Several additional subspecies of

lion existed in prehistoric times:...

Figure 1: Example of a multimedia object context.

tual description, not from title (ﬁgure 1).

In our work, we will be interested by media ”im-

age”. Most existing work in this area uses the in-

formation from textual description of image. There

are other sources of evidence that were used as vi-

sual descriptors, information from link around the

image (Hliaoutakis et al., 2006), structure of XML

document. To resolve difﬁculties in mutlimedia re-

trieval ﬁeld, you must deﬁne adequate source of ev-

idence for representation a multimedia element and

deﬁning appropriate indexing model. In this context,

we present our structural indexing system combin-

ing conceptual information for semi-structured doc-

uments dedicated to approximate retrieval data. We

begin with an overview of existing work in multi-

media retrieval. Then we turn to the presentation of

our approach while detailing the preprocessing, ex-

traction of textual and structural and phase calcula-

tion relevance of multimedia element in information

a better response to needs expressed by user. Finally,

we present the results of applying our method on two

data sets ”INEX 2007” and ”ImageCLEF 2010”.

2 STATE OF THE ART

The advent of structured documents has caused new

problems in information retrieval world, and more

speciﬁcally in multimedia elements retrieval. These

problems are strongly related to nature of these doc-

uments that provide the structure as a new source

of evidence. Thus, nowadays, XML documents in-

clude multimedia elements of different types (audio,

video and image)implicitly embedded in the textual

elements. These multimedia elements (such as phys-

ical objects) do not contain enough information to be

able to answer a given query. Therefore, the calcula-

tion of relevance score of multimedia element must be

linked to textual and structural information provided

by other nodes XML (Hliaoutakis et al., 2006).

Several works deal XML documentas a ﬂat source

of information and ignore the structure of XML doc-

uments. In this context, (Schlieder and Holger, 2002)

say: ”Ignore the document structure is to ignore its

semantics”. Indeed, XML document is used to de-

scribe a set of data by a structure that provides a se-

mantic lexicon. Thus, it facilitates the presentation

of information in terms of interpretation and exploita-

tion. Replying to this need, new works appear in the

ﬁeld of multimedia retrieval that takes in account the

structure as source of relevant information. Existing

work in structured retrieval of multimedia elements is

decomposed in two classes. The ﬁrst class includes

some works which proceed to adopt some traditional

technical of retrieval information as language model.

In this context, the team CWI/UTwente performs a

step of ﬁltering results to keep the fragments contain-

ing at least one multimedia element (Tsikrika et al.,

2008)(Westerveld et al., 2007). The second class in-

cludes the speciﬁc work to be structured multimedia

retrieval. This class uses the structure as a source of

evidence in the process of selection of multimedia el-

ements. As ﬁrst step, (Kong and Lalmas, 2005) pro-

posed a method which combines structure of XML

document (XPath) with the use of links (XLink). This

method consist to divide XML documentinto regions.

Each region represent a area of ancestors of the multi-

media element. His score is calculated in function of

the scores of each region. This method exploits verti-

cal structure only. In a second time, (Torjmen et al.,

2010) have used the addition of horizontal structure

to the notion of hierarchy. (Torjmen et al., 2010) use

a method called ”CBA” (Children, Brothers, Ances-

tors), which takes into consideration the information

carried by the children , brothers and fathers nodes

for calculate the relevance of multimedia elements.

The authors propose an alternative method ”Ontolo-

gyLike” which is based on the identiﬁcation of XML

document to ontology. To calculate the similarity be-

tween nodes the authors use similarity measures that

are mainly based on the number of edges to calcu-

late the distance between nodes. There are other

approaches to multimedia retrieval are based on ex-

ploitation of links in XML document (Awadi and Tor-

jmen, 2010). This work was improved by proposing

a hybrid approach that combines structure with using

of links who is consider as semantic links (Aouadi

et al., 2012). This method above to divide the docu-

MultimediaRetrievalbasedonGeometricDistanceinSemi-structuredDocument

221

ment into regions according the hierarchical structure

and the location of image in document. This factor

plays a role in the weighting of links for compute the

score of image.

In this paper, we propose a new metric for mul-

timedia retrieval in XML documents which involves

the use of geometric distances to calculate the rele-

vance of each node from the multimedia node. This

method consists of placing the nodes of XML doc-

ument in Euclidean space and deﬁne each node by

a vector of coordinates to calculate then the distance

between each pair of nodes. This distance will play a

beneﬁcial role in to calculate the score of multimedia

element.

3 PROPOSED APPROACH

We focus on techniques for multimedia retrieval

based on textual and structural context in XML doc-

uments. XML documents cannot be effectively ex-

ploited by classical techniques of IR, which regard

document as a bog of words. Therefore, the calcula-

tion of relevance score of multimedia element must be

linked to textual and structural information provided

by other nodes XML (Hliaoutakis et al., 2006). Thus,

it facilitates the presentation of information in terms

of interpretation and exploitation. Replying to this

need, we propose a new method in the ﬁeld of mul-

timedia retrieval that takes into account the structure

as a source of evidence and its impact on search per-

formance. We present a new source of evidence ded-

icated to multimedia retrieval based on the intuition

that each textual node contains information that de-

scribes semantically a multimedia element. And the

participation of each text node in the score of a mul-

timedia element varies with its position in there XML

document. To compute the geometric distance, we

initially place the nodes of each XML document in an

Euclidean space to calculate the coordinates of each

node by algorithm 1. Then, we compute the score of

a multimedia element depending on the distance be-

tween each textual node (Fakhfakh et al., 2013).

<name>Coronation of Louis VIII and Blanche...</name>

<description>Coronation of Louis VIII ...</description>

<comment>Coronation of Louis VIII ...</comment>

<caption article="text/en/1/308531">a miniature from ...</caption>

<caption article="text/en/2/314411"> ... circa 1450.</caption>

</text>

<description>Couronnement de Louis VIII le Lion fol...</description>

<comment>Couronnement de [[Louis VIII of Fra)</comment>

<caption article="text/fr/1/501190">Couronnement de Louis VIII ...</caption>

<caption article="text/fr/5/540615">Couronnement de Louis VIII ...</caption>

</text>

<comment>({{en|Coronation of Louis VIII and Blanche...</comment>

<license>Public Domain</license>

</image>

image

&2 &4

name

text

comment

license

&14&13&12&11

&10

description

comment

caption

comment

description

&14

&10

&12

&11

&13

Geometric representation of XML document XML treeXML document

Figure 2: The steps of passing an XML document to geo-

metric representation.

Figure 2 shows the steps of passing an XML doc-

ument to a geometric representation of the XML ele-

ments in a Euclidean space. The ﬁrst step consist to

present a XML document as XML tree to take into

account XML document properties.

An XML tree is described by a set of relationships

between nodes. Formally an XML tree is a pair A =

(E, R) where E is a set of XML elements and R ⊂ E

((p, q) ∈ R if p is the parent of q) is a set of relations

satisfying:

∃!r ∈ E, ∀q ∈ E − {r}, (r, q) ∈ R (1)

With r is the root of the tree.

∀p ∈ E − {r}, ∃!q ∈ E, (p, q) ∈ R (2)

Each node has a parent except the root r.

In second step, we will spend to presentation of

XML tree in a geometric representation. This step

is mainly based on equalities extraction in XML tree

according to our proposed hypotheses.

The XML tree representation allowed us to unveil

certain relationships of neighboring, brotherhood and

offspring. Indeed, the distance d which separate two

or more brothers with their common ancestors itera-

tively is the same. And brothers of the same hierar-

chical level are equidistant.

These distances are deﬁned according to the re-

lationship of contiguity and semantic similarity be-

tween nodes. These distances are not quantized but

will be extracted in function of the position of each

textual node in XML tree.

All these properties result in: For all q

, x

···x

) and q

= (x

, x

···x

) where Q is

a set of vectors in R

• In the same hierarchy, if there are more than two

brothers then their adjacent nodes are equidistant:

property 1

∀q

, q

∈ Q, if A

) = A

)

d(q

, q

) = d(q

, q

)

• The distance between any node and its descen-

dants is the same:

property 2

∀q

, q

, q ∈ Q, n ∈ N, A

) = A

) = q

d(q

, q) = d(q

, q)

With ∀n ∈ N

∗

, we deﬁne function A

by:

∀q ∈ E,

(q) =











{q} if n = 0

n−1

(p) if ∃ p ∈ E, (p, q) ∈ R and n > 0

∅ else

WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies

222

From these relationships, we can generate system

of equations taking into account for kinship relation-

ships nodes based on hierarchy and adjacency. These

relationships are decried by equalities in this order

(these equations are only examples):

d(n

, n

) = d(n

, n

)

d(n

, n

) = d(n

, n

)

d(n

, n

) = d(n

, n

)

d(n

, n

) = d(n

, n

)

These distances are deﬁned according to the re-

lationship of contiguity and semantic similarity be-

tween nodes. They are not quantized but will be ex-

tracted in function of the position of each textual node

in the XML tree. The resulting system is nonlinear,its

resolution requires the use of an approximate resolu-

tion iteratively method where we used iterative solu-

tion method (see Algorithm 1).

Algorithm 1: Resolution algorithm approximate nonlinear

system of equations.

Require: (Q = (q

, q

...q

|Q|

), R) :an XML tree as q

=(q

...q

) ∀i ∈ [1, |Q|]

m:dimension

for (i, j) ∈ [1, |Q|]

← random value

end for

← (q

, q

...q

|Q|

)

repeat

P ← Q

for (i, j) ∈ [1, |Q|]

← (q

, q

...q

i−1

, q

+ d

(1), q

i+1

·· ·q

|Q|

)

← (q

, q

...q

i−1

, q

+ d

(ε), q

i+1

·· ·q

|Q|

)

← (q

, q

...q

i−1

, q

+ d

(1− ε), q

i+1

·· ·q

|Q|

)

t ← 0

while error(Q

) > error(Q

) do

= (q

, q

...q

i−1

, q

+ 2

(1), q

i+1

·· ·q

|Q|

)

t=t+1

end while

t ← 0

while error(Q

) < error(Q

) do

= (q

, q

...q

i−1

, q

− 2

(1), q

i+1

·· ·q

|Q|

)

t=t+1

end while

while |error(Q

) − error(Q

)| > ε do

←

+ Q

let Q

= (p

, p

... p

|Q|

)

if error(p

, p

... p

i−1

, p

− d

(ε), p

i+1

·· · p

|Q|

) >

error(p

, p

... p

i−1

, p

+ d

(ε), p

i+1

·· · p

|Q|

) then

← Q

else

← Q

end if

end while

end for

until P = Q

The process begins by assigning to each XML

node a random vector it. Tries to improve the coor-

dinate values of each node according to an error value

(the sum of the squared deviations). At each iteration,

the coordinates are improved together with the mini-

mization of this error. The algorithm stops when the

error reaches its minimum value (no improvement is

possible). Let Q the set of vectors obtained at a given

iteration during the running of the algorithm, the error

is deﬁned by:

error(Q) =

∑

∈Q,

)=A

)

(d(q

, q

) − d(q

, q

))

∑

,q∈Q,n∈N,

)=A

)=q

(d(q

, q)− d(q

, q))

Where m is the dimension of the Euclidean space

and ∀v ∈ R, D

(v) = (d

, d

···d

) is such as:

= {

0 if k 6= j

v otherwise

3.1 Indexing System

We propose a indexing system MXS − index com-

posed by two party: party of textual indexing and

party of structural indexing. In ﬁrst party, our

approach uses NLP (Natural Language Processing)

techniques to extract the candidate XML nodes of

the resulting indexing. The weight of these nodes is

depending on the frequency of each of these terms

and the number of elements in the corpus according

to the number of elements containing the term. In

Second party, we built structural index using infor-

mation extract from XML tree and geometric metric.

Each XML node will presented by characteristic vec-

tor. We start by extract geometric proprieties. And

we compute coordinates of each XML nodes. This

party is accompanied by generating XML data model

which processes ancestor, descendant and proximity

relationships (ﬁgure 3).

Figure 4 schematize the process of textual and

structural indexing XML documents with our index-

ing system. Well as the transition of XML document

as a tree presentation to geometric presentation in Eu-

clidean space.

3.2 Multimedia Element

A multimedia element (eg image) does not contain

textual content. Its score is based on textual nodes in

its neighborhood. The transition from the XML tree

structure representation of elements in an Euclidean

space, where we exploit the dissimilarity distances

separating a multimedia node and other textual nodes,

is performed by extracting the equations satisfying the

properties deﬁned earlier and the application of algo-

rithm 1. To calculate the distance between a node

n and multimedia element H, we calculate the Eu-

clidean distance between their respective feature vec-

MultimediaRetrievalbasedonGeometricDistanceinSemi-structuredDocument

223

Textual Indexing

Word splitting

Stop words

removing

Lemmatisation

Tokenization

Structural Indexing

Extracting geometrical properties

Computing coordinates of XML node

XML

data model

XML

DataBase

NPL

Tools

Text Index Structural Index

Text Index

Structural Index

XML Node

Term weighting

General Index

Term weighting

Coordinates of XML node

Figure 3: Architecture of our indexing model MXS−index.

<name>Coronation of Louis VIII and Blanche...</name>

<description>Coronation of Louis VIII ...</description>

<comment>Coronation of Louis VIII ...</comment>

<caption article="text/en/1/308531">a miniature from ...</caption>

<caption article="text/en/2/314411"> ... circa 1450.</caption>

</text>

<description>Couronnement de Louis VIII le Lion fol...</description>

<comment>Couronnement de [[Louis VIII of Fra)</comment>

<caption article="text/fr/1/501190">Couronnement de Louis VIII ...</caption>

<caption article="text/fr/5/540615">Couronnement de Louis VIII ...</caption>

</text>

<comment>({{en|Coronation of Louis VIII and Blanche...</comment>

<license>Public Domain</license>

</image>

image

&2 &4

name

text

comment

license

&14&13&12&11

&10

description

comment

caption

comment

description

&14

&10

&12

&11

&13

Geometric representation of XML document

text

comment

caption

license

comment

description

comment

caption

id parent tag

&11

&10

&14

&13

&12

imageNULL

name

text

Structural indexation of XML documentTextual indexation of XML document

term

frequence

in element

frequence

in documen

Coronation

Louis

2 6

Inverse index

Figure 4: Treatment process of XML document.

tors q

and q

dist(n, H) =

∑

i=1

− q

)

(3)

With m is the dimension of the Euclidean space. q

is deﬁned by: q

=(xn

, xn

... xn

) with xn are the

vector characteristics of node n. And q

is deﬁned by:

=(xH

, xH

... xH

with xH represent the coordi-

nates compose the vector characteristics of a node H.

We calculate the score for each textual node depend-

ing on the frequencyof each term (t f) and the number

of elements in the corpus according to the number of

elements containing the term (id f ). A textual node

is presented by: n = (n

, n

···n

|v|

) where n

is the

weight of the term t

, v is the set of indexing terms:

= t f(t

, n) × id f(t

) (4)

With

id f(t

) = log(

) (5)

Where N is the total number of XML elements in the

corpus, N

is the number of elements that contain the

term t

and t f(t

, n) is the frequency of the term t

node n. The score of textual node depends on the

weight of each indexing term. A query is made by

the list v = (v

, v

···v

|v|

) where v

∈ {0, 1} (0:not ex-

ist, 1:exist) according membership t

at the query. The

score of textual node n for the query q is deﬁned by:

rsv(q, n) = q × n

|V|

∑

i=1

× n

(6)

Where µ is the set of textual elements. The score of

multimedia node H is deﬁned by:

rsv(q, H) =

∑

n∈µ

rsv(q, n)

dist(n, H)

(7)

With dist(n, H) is the distance between the feature

vectors corresponding to the nodes n and H. This

equation leads to assign the importance of contribu-

tion of all nodes in computing the score of multimedia

element that shows its beneﬁcial impact in multime-

dia retrieval.

4 EVALUATION AND RESULTS

We evaluate our system into two databases extracted

from two collections : INEX 2007 (Initiative for the

Evaluation of XML Retrieval) Ad Hoc task (Fuhr

et al., 2007) and ImageCLEF 2010 Wikipedia image

retrieval task (Popescu et al., 2010). These databases

are composed by XML documents extracted from

Wikipedia. The evaluation results show that this

method provides a MAP which is equal to 0.2102

as MAP with using ”ImageCLEF 2010” collection.

The result has been improved signiﬁcantly with the

”INEX 2007” collection to 0.3102 as MAP. This in-

crease is due to nature of ”INEX 2007” collection

who includes XML documents with heterogeneous

structure. So in ”INEX 2007” collection we ﬁnd doc-

uments with high depth. This factor highlights struc-

tural information and ampliﬁes effect textual informa-

tion based on computed distances . For against, our

system is more stable with ”ImageCLEF 2010” col-

lection, this is due to rapid convergence of results (al-

gorithm 1). With our measure, we have shown that

combined use of textual and structural context can

properly determine the relevance of multimedia ele-

ment, and the structure plays a primordial role in mul-

timedia retrieval.

Table 1: Results of the impact our approach on INEX 2007

and ImageCLEF 2010 based in MAP(Mean Average Preci-

sion).

Company INEX 2007 ImageCLEF 2010

Task Collection XML Ad Hoc Wikipedia Retrieval

Number of XML document 659,388 237,434

Number of image 246,730 237,434

Topics 19 70

MAP 0.3102 0.2572

WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies

224

5 CONCLUSIONS

In this paper, we propose a novel approach for multi-

media retrieval in XML documents. This method con-

sist to calculate the score of element mulltimedia ac-

cording the textual context provided by nodes in prox-

imity and structural context from distance between

nodes and multimedia element. Experiments shows

the interest of our method on INEX 2007 and Image-

CLEF 2010 collections. In the future, we want to ex-

ploit another factor to calculate the relevance of mul-

timedia element such as the title of image, the weight-

ing of the links in XML document ... As well as an-

other source of evidence as visual descriptors and the

study parameters combination of using of structural,

textual and visual context.

REFERENCES

Aouadi, H., Torjmen-Khemakhem, M., and Jemaa, M. B.

(2012). Combination of document structure and links

for multimedia object retrieval. Journal of Informa-

tion Science, 38(5):442–458.

Awadi, H. and Torjmen, M. (2010). Exploitation des liens

pour la recherche d’images dans des documents xml.

Elghazel, H., Idrissi, K., Baskurt, A., and Ben Amar, C.

(2005). Approche textuelle pour la recherche d’image.

In 3rd International Conference on Sciences of Elec-

tronic, Technologies of Information and Telecommu-

nications SETIT 2005.

Fakhfakh, S., Tmar, M., and Mahdi, W. (2013). A new met-

ric for multimedia retrieval in structured documents.

In ICEIS (2), pages 240–247.

Fuhr, N., Kamps, J., Lalmas, M., Malik, S., and Trotman,

A. (2007). Overview of the inex 2007 ad hoc track. In

INEX, pages 1–23.

Hliaoutakis, A., Varelas, G., Voutsakis, E., Petrakis, E.

G. M., and Milios, E. (2006). Information retrieval

by semantic similarity. In Intern. Journal on Semantic

Web and Information Systems (IJSWIS).Special Issue

of Multimedia Semantics, pages 55–73.

Kong, Z. and Lalmas, M. (2005). Xml multimedia retrieval.

In SPIRE, pages 218–223.

Lew, M. S. (2006). Content-based multimedia information

retrieval: State of the art and challenges. ACM Trans.

Multimedia Comput. Commun. Appl, 2:1–19.

Popescu, A., Tsikrika, T., and Kludas, J. (2010). Overview

of the wikipedia retrieval task at imageclef 2010. In

CLEF (Notebook Papers/LABs/Workshops).

Schlieder, T. and Holger, M. (2002). Querying and ranking

xml documents. Journal of the American Society for

Information Science and Technology, 53:489–503.

Tjondronegoro, D., Zhang, J., Gu, J., Nguyen, A., and Geva,

S. (2005). Integrating text retrieval and image retrieval

in xml document searching. In INEX, pages 511–524.

Torjmen, M., Pinel-Sauvagnat, K., and Boughanem, M.

(2010). Using textual and structural context for

searching multimedia elements. IJBIDM, 5(4):323–

352.

Tsikrika, T., Serdyukov, P., Rode, H., Westerveld, T., Aly,

R., Hiemstra, D., and de, A. P. V. (2008). Structured

document retrieval, multimedia retrieval, and entity

ranking using pf/tijah. In 6th Initiative on the Eval-

uation of XML Retrieval, INEX 2007, volume 4862 of

Lecture Notes in Computer Science, pages 306–320,

London. Springer Verlag.

Westerveld, T., Rode, H., van, R. O., Hiemstra, D.,

Ramirez, G., Mihajlovic, V., and de, A. V. (2007).

Evaluating structured information retrieval and multi-

media retrieval using pf/tijah. In Fuhr, N., Lalmas, M.,

and Trotman, A., editors, Comparative Evaluation of

XML Information Retrieval Systems, volume 4518 of

Lecture Notes in Computer Science, pages 104–114,

Berlin, Germany. Springer Verlag.

MultimediaRetrievalbasedonGeometricDistanceinSemi-structuredDocument

225