Semantic Annotation of Images Extracted from the Web

using RDF Patterns and a Domain Ontology

Rim Teyeb Jaouachi

, Mouna Torjmen Khemakhem

, Nathalie Hernandez

, Ollivier Haemmerl

and Maher Ben Jemaa

ReDCAD Laboratory, National School of Engineers of Sfax, University of Sfax, Sfax, Tunisia

IRIT, Universit

e Toulouse - Jean Jaur

es, Toulouse, France

Keywords:

Semantic Annotation, RDF Patterns, Domain Ontology, Information Retrieval.

Abstract:

Semantic annotation of web resources presents a point of interest for several research communities. The use

of this technique improves the retrieval process because it allows one to pass from the traditional web to the

semantic web. In this paper, we propose a new method for semantically annotating web images. The main

originality of our approach lies in the use of RDF (Resource Description Framework) patterns in order to guide

the annotation process with contextual factors of web images. Each pattern presents a group of information to

instantiate from contextual factors related to the image to annotate. We compared the generated annotations

with annotations made manually. The results we obtain are encouraging.

1 INTRODUCTION

The concept of the semantic web is the brainchild

of Tim Berners-Lee (Cunningham et al., 2002), the

original creator of the World Wide Web. The main

idea behind this kind of web is to weave a web that

not only links documents to each other but also that

recognises the meaning of the information in those

documents. The aim of Tim Berners-Lee was to trans-

form the current web from a set of interconnected

data by simples links semantically isolated into a huge

mass of information linked in a semantic manner.

In other words, the semantic web consists in

adding formal semantics to the web content in order to

allow a more efﬁcient access and management. This

is possible thanks to the improvement of the capabil-

ity of computers to manipulate data meaningfully by

providing meaning into web resources

. In doing so,

external software agents have to carry out complex

tasks on behalf of a human user and to improve the

degree of cooperation between humans and comput-

ers.

However, the transformation from traditional web

to semantic web depends on the presence of a critical

mass of metadata (Krestel et al., 2010) corresponding

A web resource is an entity that can be described on the

web. Each resource is identiﬁed by a unique URI (Uniform

Resource Identiﬁer).

to web resources. The acquisition of these metadata

is a major challenge for the semantic web commu-

nity. As a solution, many manual tools ((Kahan and

Koivunen, 2001), (McDowell et al., 2003), (Hand-

schuh et al., 2001), (Bechhofer and Goble, 2001)),

semi-automatic tools ((Cunningham et al., 2002),

(Laclavik et al., 2009), (Vargas-Vera et al., 2002))

and automatic tools ((Popov et al., 2004), (Kogut and

Holmes, 2001)) for semantic annotation have been de-

veloped.

In our case, we are interested in the semantic an-

notation of web images. With the presence of a huge

number of web images, many approaches were devel-

oped. There are many approaches based on the image

content (color, texture, etc.) in order to produce an-

notations and only a few works using the contextual

factors of the image to annotate it without human in-

tervention.

Our goal is to obtain a fully automatic approach

for web images annotation based on contextual fac-

tors such as image caption, document title and sur-

rounding text. The main idea is to generate an RDF

graph from each contextual factor. The elementary

RDF graph (concerning one contextual factor) is com-

posed of concepts and instances of concepts linked

between them by semantic relations. After the genera-

tion of all elementary RDF graphs from all the contex-

http://www.w3.org/tr/2004/rec-rdf-primer-20040210/

137

Teyeb Jaouachi R., Torjmen Khemakhem M., Hernandez N., Haemmerle O. and Ben Jemaa M..

Semantic Annotation of Images Extracted from the Web using RDF Patterns and a Domain Ontology.

DOI: 10.5220/0005374301370144

In Proceedings of the 17th International Conference on Enterprise Information Systems (ICEIS-2015), pages 137-144

ISBN: 978-989-758-097-0

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

tual factors, the next step consists in combining them

into a global RDF graph which is considered as an

image annotation.

The originality of our work is the use of a set of

RDF patterns and a domain ontology in order to guide

the step of annotation.

This paper is organised as follows: Section 2 re-

views the related works. Section 3 details the ap-

proach we propose for the image annotation. In sec-

tion 4, we present the domain ontology used in our

work. Section 5 ﬁnally demonstrates the obtained re-

sults.

2 RELATED WORKS

2.1 Annotation Approaches for Images

We can classify the existing approaches into two cat-

egories: (1) Content-Based Image Annotation; and

(2) Context-based Image Annotation.

Several works have studied Content-Based Image

Annotation. Among these approaches, we ﬁnd (Li

and Wang, 2003) (Cusano et al., 2004), (Halaschek-

wiener et al., 2006), (Bellini et al., 2011), (Arndt

et al., 2007). The main idea of this type of approaches

is to associate a semantic description with the image

in totality or to describe a speciﬁc region of the im-

age. (Wang et al., 2007) shows a good classiﬁcation

of these different approaches. Using the content of the

image for annotation purpose resolves, partially, the

Semantic Gap Problem deﬁned by (Smeulders et al.,

2000).

Only a few works use contextual factors of the im-

age in order to annotate it. Among these approaches,

we ﬁnd (Declerck et al., 2004). This approach was

proposed in the project Esperonto. It was based

on natural language techniques, ontologies and other

knowledge bases.

The approach proposed in (Nguyen, 2007) used

two contextual factors associated with web images. It

exploited the caption and keywords associated with

the image in order to construct a graph representing

its semantics.

2.2 SPARQL Query Generation using

Patterns

Our work is inspired by the SWIP system presented

in (Pradel et al., 2013). This system allows the trans-

lation of natural language queris into formal ones, ex-

pressed in SPARQL. The translation process is done

thanks to the use of query patterns, each pattern repre-

senting a family of typical queries. After the selection

of the pattern which is the best match to the natural

language query, that pattern is modiﬁed in order to

build the SPARQL query corresponding to the natural

query. The formal deﬁnition of a pattern is given in

(Pradel et al., 2012). An extension of that work has

been presented in (Gillet, 2013). It allows the genera-

tion of SPARQL queries based on different ontologies

of a same domain, thanks to ontology alignments.

Even if our goal is to annotate documents instead

of querying them, we propose to use patterns in our

work. These patterns are deeply inspired by the pat-

terns deﬁned in (Pradel et al., 2012), but they are used

for multimedia document annotation purpose. The

patterns are adapted in order to take into account the

notions related to images. The details of our approach

is presented in Section 3.

3 OUR APPROACH FOR IMAGE

ANNOTATION

In this section, we present our approach aiming at an-

notating web images. Our idea is to use RDF patterns

in order to guide the extraction of relevant informa-

tion from contextual factors surrounding the image to

be annotated.

Among these factors, we cite (1) the image cap-

tion; (2) the paragraph title; (3) the text around the

image; (4) the hyperlinks between documents con-

taining images (if they exist); (5) the image name (if

it is signiﬁcant) and (6) the table content if the images

are grouped in this structure.

After the choice of the factors to be used, the next

step is to apply, for each factor, a set of processings.

The goal of this step is to instantiate the RDF pat-

terns and to select the best instantiated pattern. More

details about our approach can be found in (Jaouachi

et al., 2013). Figure 1 shows an overview of the pro-

posed approach.

In this paper, we will focus on the step of the def-

inition, the instantiation and the use of RDF patterns.

A pattern is an RDF graph which represents a pro-

totype regrouping information considered as impor-

tant by the domain experts (the cinema in our case).

Each pattern is built by focusing on a group of pieces

of information related to the cinema ﬁeld. Different

relationships between various concepts forming the

patterns are inspired from the “Movie ontology”.

In addition to these relations, we use two proper-

ties belonging to the foaf ontology (foaf:depiction and

foaf:img) which allows us to express the relationship

ICEIS2015-17thInternationalConferenceonEnterpriseInformationSystems

138

Figure 1: General approach overview.

between an element of the pattern (object or person

respectively) and the related image.

Deﬁnition

Our deﬁnition of a pattern is inspired from (Pradel

et al., 2013).

A pattern p is composed of 5 elements

(G, Q, SP, Img, S):

• G is a connected RDF graph which describes the

general structure of the pattern to be instantiated.

Such a graph is composed of triples according to

the structure presented in Figure 2. This structure

is formed by a subject (which can be a concept

or an instance of a concept), a predicate and an

object (which can be a concept or an instance of a

concept or an image);

• Q is a subset of elements of G, these elements

are considered to be characteristics of the pattern.

Such an element can be a class or an object prop-

erty or an image of G;

• SP is the set of sub-patterns sp of p;

• Img is the set of distinct images called qualifying

images present in the pattern. An image can illus-

trate an element of the pattern;

• S is a description of the meaning of the pattern in

natural language.

Example

Figure 3 shows an example of an RDF pat-

tern used for the step of annotation. It is

composed of three sub-patterns which are

[Movie,releaseDate,Date], [Movie,AsTitle,String]

Figure 2: Triples constituting the graph patterns.

Figure 3: Example of a pattern used for the annotation.

and [Movie,foaf:depection,Image]. All of them are

optional because they can remain uninstantiated.

Sub-patterns [Movie,releaseDate,Date] and

[Movie,AsTitle,String] are not repeatable and have

as cardinalities ReleaseDate*0..1 and Title*0..1

respectively with 0 is the minimal cardinality and 1

is the maximum cardinality. Indeed, a movie has one

title and one release date.

However, the sub-pattern

[Movie,foaf:depiction,Image] is repeatable and

has as cardinality image*0..n, n being the maximum

cardinality.

3.1 RDF Pattern Deﬁnition

In order to annotate web images, we deﬁne some RDF

patterns. Each pattern represents a prototype of a

group of information related to the domain studied.

In our case, we used six patterns considered as im-

portant by the experts of the domain. Each pattern is

centered around a vertex which is the Movie concept

and formed by a set of sub-patterns.

A sub-pattern can be deﬁned as a simple triple

with the subject which is an instance of the Movie

concept (eg. triple [Movie:*,hasSoundmix,Sound-

Mix:*] with Movie:* is a vertex of the pattern)

or a set of triples whith one of them which is

linked at the vertex of the pattern (eg. triples

[Movie:*,hasComposer,Musical Artist:*] and [Musi-

cal Artist:*,foaf:img,image:*]) . The ﬁgure 4 rein-

forces the two deﬁnitions.

In order to deﬁne formally a pattern, we need three

type of vocabularies:

SemanticAnnotationofImagesExtractedfromtheWebusingRDFPatternsandaDomainOntology

139

Figure 4: Examples of sub-patterns.

• The ﬁrst vocabulary allows the description of the

patterns in general. Indeed, we use the patters on-

tology

proposed in (Pradel et al., 2013). This

ontology deﬁnes the grammar allowing the repre-

sentation of patterns.

• The second vocabulary is the vocabulary of the

domain addressed in the context of the image.

It is necessary to associate each element or re-

lationship of a pattern with a concept or a rela-

tionship belonging to the domain ontology used

for the evaluation. In our case, we use the do-

main ontology named Movie ontology which will

be presented in section 4. We use properties of

this ontology to link the various properties used

in the construction of the RDF patterns. For ex-

ample, we have exploited the properties release-

Date from the Movie ontology to specify the pred-

icate between Movie concept and Date concept

and consequently to construct a sub-pattern of the

pattern shown in ﬁgure 3.

• The third vocabulary is necessary to describe all

the elements of the RDF patterns in relation to im-

ages. Indeed, by this vocabulary, we link an image

to the appropriate element of the pattern. In our

case, we use two properties from the foaf project:

foaf:depiction and foaf:img.The ﬁrst property rep-

resents a relationship between a thing and an im-

age that depicts it, and the second property relates

a Person to an image that represents him/her.

To conclude, this step allows us to obtain several

patterns that have to be instantiated based on any

document related to the domain (cinema in our

case).

Example

Figure 5 shows an example of RDF pattern according

to the formal deﬁnition and uses the three vocabular-

ies mentioned above.

http://swip.univ-tlse2.fr/SwipWebClient/welcome.

html

Figure 5: Examples of pattern presented in formal deﬁni-

tion.

This pattern is composed of two sub-patterns: lo-

cation and country. The three numbers after the name

of each sub-pattern are respectively : the minimal car-

dinality, the maximal cardinality and the identifying

of the element in order to instantiate the concerned

pattern.

3.2 Instanciation of RDF Patterns

The purpose of deﬁning and using RDF patterns is to

guide the annotation procedure. We are not willing

to extract all the pieces of information contained in a

document but only to extract the pieces of information

which allow the instantiation of the patterns.

For example, by using the pattern shown in ﬁgure

3, we want to extract triples from a document that can

be instance of one of the three sub-patterns. In other

words, we look for the date of the Movie, its title and

its images.

This is possible by the following steps that will be

repeated as many times as the number of sub-patterns:

- The ﬁrst step consists in identifying the predi-

cate of the sub-pattern. This step is based on the ap-

plication of extraction rules that will be described in

section 5.1.

- Once found, the second step consists in identify-

ing the subject and the predicate related by the iden-

tiﬁed relationship. This step is based on the identiﬁ-

cation of instances forming two approximates of the

identiﬁed relationship.

- The third step consists in linking the web image

to the appropriate element of the pattern. This step

is based on the analysis of the image caption, if there

is any indication in the text describing the instance of

ICEIS2015-17thInternationalConferenceonEnterpriseInformationSystems

140

the concept concerned by the image.

We present in ﬁgure 7 the instantiation of the pat-

tern shown in ﬁgure 3 by using the text and the image.

Figure 6: Instantiation of the pattern presented in ﬁgure 3.

3.3 Selecting the Best Instantiated

Patterns to Represent the Text

The purpose of using patterns is to generate an RDF

annotation for the web document containing the im-

age. Our strategy is to instantiate patterns and then to

choose the best instantiated pattern.

In order to rank the instantiated patterns, we focus

on a set of criteria such as recall, precision, number

of correct extracted triples, number of automatic ex-

tracted triples. It is possible to use the three ﬁrst crite-

ria only if we have a manual annotation as reference.

However the last criteria is independent of manual an-

notation.

In this paper, we present the results based on the

number of automatic extracted triples. The results us-

ing this criteria are presented in section 5. As a ref-

erence, we will compare the results obtained by using

our approach with manual ranking.

4 THE MOVIE ONTOLOGY

Like in most research domains, there are ontologies

which are used in order to represent, share and reuse

knowledge. Ontologies contain an effective structure

of the domain knowledge which improves the efﬁ-

ciency of the retrieval system. The semantic web

techniques and technologies provide a manner to con-

struct and use web resources by attaching semantic

information to them.

In our work, we are interested in improving the

management of multimedia information by means

of knowledge representation, indexing and retrieval.

Among the multimedia entertainment, cinema stands

a good position so that we are interested, especially,

in the cinema domain.

In the literature, there is an ontology of the cin-

ema ﬁeld which describes movie scenes. It is called

“Movie ontology”

and it is developed by the De-

partment of Informatics at the University of Zurich.

This ontology contains concept hierarchies for movie

categorisation, instances of concepts and relations be-

tween concepts.

It contains distinct concepts like Award, Certiﬁca-

tion, Film, Person (Actor, Actress, Writer, Producer,

etc.).

5 EXPERIMENTATIONS AND

EVALUATION

In order to validate our proposition, we used a corpus

composed of 10 Wikipedia pages related to the cin-

ema domain, written in English language and contain-

ing images. The choice of these pages was arbitrary.

In our experiments, we have used only the surround-

ing text of images in order to instantiate six patterns.

As a reference, we annotated the corpus manually

and we generated the RDF graph corresponding to the

semantic annotation of each image. Figure 7 shows an

example of a web image and its surrounding text.

Figure 7: Example of image used for the evaluation.

5.1 Extraction Rules for Semantic

Relation Detection

In order to instantiate the RDF patterns, we use the

NLP (Natural Language Processing) tools, GATE

platform (Cunningham et al., 2002), Porter stemmer

(Porter, 1997) and our own extensions dedicated to

determinate the semantic relations for different pat-

terns.

http://www.movieontology.org/

SemanticAnnotationofImagesExtractedfromtheWebusingRDFPatternsandaDomainOntology

141

In this step, we used JAPE language (Cunningham

et al., 2002). It is a language based on regular expres-

sions. Using extraction rules, we try to detect an in-

stance of the Movie ontology relations (adapted as a

predicate for RDF pattern) and to detect instances of

concepts linked by this relationship.

Such a rule represents a set of phases, each of

which consists of a set of pattern/action rules. It has

always two sides: Left (LHS:Left-Hand Side) and

Right(RHS:Rigth-Hand Side). The LHS of the rule

contains the identiﬁed annotation pattern that may

contain regular expressions and the RHS outlines the

action to be taken on the detected pattern and con-

sists of annotation manipulation statements. The ex-

ample below (ﬁgure 8) shows a grammar which al-

lows the detection of instances of the semantic rela-

tion ”isAwardedWith”.

Figure 8: Rule example of JAPE grammar.

In the ﬁgure 8, we ﬁnd an example of a rule

that is labeled isAwardedWith. ”Token.stem” cor-

responds to the lemmatised form of the word, ”To-

ken.category” corresponds to the grammatical cate-

gory of the word, ”Token.Kind” design the kind of

word. It means the word represents a number or a

simple word or a punctuation. ”Lookup.majortype”

means that the word is considered as the default con-

cept. ”Lookup.minortype” corresponds to the speciﬁc

categories of the word. "|" means that there are many

alternatives. "-->" is the boundary of the LHS rule.

Our relation will be part of the annotation prop-

erties that can be seen in GATE. This is possi-

ble using :is Awarded with.RelationShip = kind =

”is Awarded with”, rule=isAwardedwith.

5.2 Results

In order to choose the best pattern to be considered

as the basis for the annotation of the image, we did a

ranking according to the number of extracted triples.

Table 1 and table 2 present the results for the im-

age shown in Figure 7 using our approach ﬁrstly, and

the result of the manual classiﬁcation secondly.

Table 1: Automatic ranking of patterns corresponding to

image 7.

Rank Patterns Ranking Number of automatic triple

1 pattern 2 7

2 pattern 4 6

3 pattern 1,pattern 6 4

4 pattern 3,pattern 5 0

Table 2: Manual RDF patterns classiﬁcation.

Rank Patterns Ranking Number of manual triple

1 pattern 2 6

2 pattern 6 4

3 pattern1 3

4 pattern 3,pattern 4 1

We note that for this document, we obtained the

same best rated pattern (pattern 2 is the highest rated

pattern obtained by our approach and by a manual ap-

proach).

We note that the number of triples in a pattern

can inﬂuence the ﬁnal ranking. Indeed, the pattern

2 (the best ranked pattern) contains the highest num-

ber of triples. In addition, the presence of the generic

triples having a maximum cardinality equal to n, can

affect scheduling. For example, we can instantiate the

generic triple [Movie: * hasActor, Actor *] repeatedly

since a ﬁlm can have the participation of several ac-

tors.

We repeated the same work for the ten web docu-

ments used for the evaluation.

In order to evaluate the quality of the annotation

for the test collection, we used the following measure:

QA =

Nr.o f wellannotateddocuments

Totalnumbero f documents

(1)

with QA is the quality of the annotation.

We consider a well-annotated document if it is an-

notated by P

(i ∈ [1..6] with 6 is the number of pattern

used) having the ﬁrst place automatically and the ﬁrst

place manually (the same P

We also calculate QA where P

obtains the second

rank automatically.

Table 3 presents the results on the entire collec-

tion.

Obtained results are encouraging. Indeed, having

high rates for the two best patterns (compared to man-

ual annotation) shows the importance of our work.

ICEIS2015-17thInternationalConferenceonEnterpriseInformationSystems

142

Table 3: Quality of annotation.

Automatic Rank of P

1 0.44

2 0.55

We succeeded to associate eighteen images to dif-

ferent elements of patterns that can be illustrated by a

picture.

The success of these association shows the interest

of our approach to annotate text and image at a time.

Choosing the number of extracted triples as cri-

terion of classiﬁcation is not arbitrary. In fact, it is

impossible to use precision and recall as criterion of

selection because it is not possible to obtain manual

annotation for every test.

However, we note that the use of this criterion

(number of extracted triples) has the disadvantage of

promoting the pattern with the greatest number of

sub-patterns or with a maximum cardinality greater

than 1 (repeatable sub-patterns).

To overcome this problem, we plan to propose a

new ranking function in our future work.

6 CONCLUSION

The potential of the semantic web to resolve informa-

tion retrieval problems is tremendous. Based on se-

mantic annotation techniques, adding formal seman-

tics to the web content is vital to improve information

indexing and retrieval.

In our case, our goal is to improve web images re-

trieval. To achieve this aim, we propose an automatic

approach to semantically annotate images through

their context. Indeed, we use contextual factors such

as the caption of the image, the surrounding text, etc.

in order to instantiate RDF patterns.

In this paper, we focused on the process of instan-

tiation of RDF patterns using a domain ontology and

patterns of extraction written in Jape language and the

exploitation of instantiated patterns in order to choose

the suitable annotation.

Preliminary results are encouraged to automate

and use a bigger cinema corpus.

Working with all image contextual factors repre-

sents our next step. The aim of this step is to instan-

tiate RDF patterns from each factor and to generate a

global annotation.

REFERENCES

Arndt, R., Troncy, R., Staab, S., Hardman, L., and Vacura,

M. (2007). Comm: Designing a well-founded multi-

media ontology for the web. In Proceedings of the 6th

International Semantic Web Conference (ISWC2007),

pages 11–15.

Bechhofer, S. and Goble, C. (2001). Towards annotation

using daml+oil. In Workshop Knowledge Markup &

Semantic Annotation, Victoria B.C., Canada.

Bellini, P., Bruno, I., and Nesi, P. (2011). Exploiting intelli-

gent content via axmedis/mpeg-21 for modelling and

distributing news. International Journal of Software

Engineering and Knowledge Engineering, 21(1):3–

32.

Cunningham, H., Maynard, D., Bontcheva, K., and Tablan,

V. (2002). GATE: A Framework and Graphical Devel-

opment Environment for Robust NLP Tools and Ap-

plications. In ACL’02,the 40th Anniversary Meeting

of the Association for Computational Linguistics.

Cusano, C., Ciocca, G., and Schettini, R. (2004). Image an-

notation using SVM. Proceedings of SPIE, 5304:330–

338.

Declerck, T., Crispi, C., Contreras, J., and ?scar Cor-

cho (2004). Text-based semantic annotation service

for multimedia content in the esperonto project. In

Knowledge-Based Media Analysis for Self-Adaptive

and Agile Multi-Media, Proceedings of the European

Workshop for the Integration of Knwoledge, Seman-

tics and Digital Media Technology, EWIMT 2004,

November 25-26, 2004, London, UK.

Gillet, P. (2013). Gnration de patrons de requtes par-

tir dalignements dontologies: Application un systme

dinterrogation du web smantique fond sur les patrons.

Halaschek-wiener, C., Golbeck, J., Schain, A., Parsia,

B., and Hendler, J. (2006). Annotation and prove-

nance tracking in semantic web photo libraries. In In

International provenance and annotation workshop.

Chicago, pages 82–89.

Handschuh, S., Staab, S., and Maedche, A. (2001). Cream:

creating relational metadata with a component-based,

ontology-driven annotation framework. In K-CAP,

pages 76–83.

Jaouachi, R. T., Khemakhem, M. T., Jemaa, M. B., Hernan-

dez, N., and Haemmerle, O. (2013). Multi-factor rdf

graph based image annotation -application in cinema

domain-. In ICWIT’13, Fifth International Conference

on Web and Information Technologies.

Kahan, J. and Koivunen, M.-R. (2001). Annotea: An

open rdf infrastructure for shared web annotations.

In Proceedings of the 10th International Conference

on World Wide Web, pages 623–632, New York, NY,

USA. ACM.

Kogut, P. and Holmes, W. (2001). Aerodaml: Applying

information extraction to generate daml annotations

from web pages. In First International Conference

on Knowledge Capture (K-CAP 2001). Workshop on

Knowledge Markup and Semantic Annotation.

Krestel, R., Witte, R., and Bergler, S. (2010). Predicate-

Argument EXtractor (PAX). In New Challenges for

NLP Frameworks.

Laclavik, M., Hluch?, L., Seleng, M., and Ciglan, M.

(2009). Ontea: Platform for pattern based automated

semantic annotation. Computing and Informatics,

pages 555–579.

SemanticAnnotationofImagesExtractedfromtheWebusingRDFPatternsandaDomainOntology

143

Li, J. and Wang, J. Z. (2003). Automatic linguistic indexing

of pictures by a statistical modeling approach. IEEE

Trans. Pattern Anal. Mach. Intell., 25(9):1075–1088.

McDowell, L., Etzioni, O., Gribble, S. D., Halevy, A. Y.,

Levy, H. M., Pentney, W., Verma, D., and Vlasseva, S.

(2003). Mangrove: Enticing ordinary people onto the

semantic web via instant gratiﬁcation. In The Seman-

tic Web - ISWC 2003, Second International Semantic

Web Conference, Sanibel Island, FL, USA, October

20-23, 2003, Proceedings, pages 754–770.

Nguyen, M.-T. (2007). Vers une plate-forme d’annotations

smantiques automatiques partir de documents multi-

mdias.

Popov, B., Kiryakov, A., Kirilov, A., Manov, D.,

Ognyanoff, D., and Goranov, M. (2004). Kim seman-

tic annotation platform. Journal of Natural Language

Engineering, pages 375–392.

Porter, M. F. (1997). Readings in information retrieval.

chapter An algorithm for sufﬁx stripping. Morgan

Kaufmann Publishers Inc., San Francisco, CA, USA.

Pradel, C., Haemmerl

e, O., and Hernandez, N. (2012). Des

patrons modulaires de requtes sparql dans le systme

swip. In 23eme Journes Francophones dIngnierie des

Connaissances.

Pradel, C., Haemmerl

e, O., and Hernandez, N. (2013). Nat-

ural language query translation into sparql using pat-

terns. In COLD.

Smeulders, A. W. M., Worring, M., Santini, S., Gupta, A.,

and Jain, R. (2000). Content-based image retrieval at

the end of the early years. IEEE Trans. Pattern Anal.

Mach. Intell., 22(12):1349–1380.

Vargas-Vera, M., Motta, E., Domingue, J., Lanzoni, M.,

Stutt, A., and Ciravegna, F. (2002). Mnm: Ontology

driven semi-automatic and automatic support for se-

mantic markup. pages 379–391.

Wang, C., Jing, F., Zhang, L., and Zhang, H.-J. (2007).

Content-Based Image Annotation Reﬁnement. Com-

puter Vision and Pattern Recognition, 2007. CVPR

’07. IEEE Conference on, pages 1–8.

ICEIS2015-17thInternationalConferenceonEnterpriseInformationSystems

144