SEARCHING WEB 3.0 CONTENT

A Semantic Search Approach

Kinga Schumacher

German Research Center for Artiﬁcial Intelligence, IUI Group, Alt-Moabit 91c, Berlin, Germany

Michael Sintek

German Research Center for Artiﬁcial Intelligence, KM Group, Trippstadter Str. 122, Kaiserslautern, Germany

Keywords:

Web 3.0, Semantic web, Semantic search, Fact retrieval, Spreading activation.

Abstract:

The Web 3.0 is commonly understood as the combination of the Semantic Web and Web 2.0, where conven-

tional and social websites and data sources (e. g., DBs, XML, HTML, plain text) should be integrated and

linked as well. Thus, there is a plethora of information in various representation forms which can be mapped

to an information pool composed of a knowledge base (in RDF/S) and a text index. In doing so, structured

data (e. g., DBs) is usually mapped to the knowledge base while unstructured data (e. g., plain text) to the text

index and semi-structured (e. g., XHTML) data to both. Therefore, a search method is required which is able

to explore both the knowledge base and the text index exploiting the cross-linking of data. For this purpose,

we propose a search approach which combines fact retrieval and semantic document retrieval. It is able to

answer queries with facts and documents as well as documents together with facts, and it supports free text

and formal queries as well as queries composed of free text and formal parts.

1 MOTIVATION

Web 3.0 applications, e. g., semantic wikis, semantic

blogs, social semantic networks and social semantic

information spaces make available a huge amount of

linked content in various representation forms. To

search such content, two main tasks have to be per-

formed: First, heterogeneous data sources have to be

integrated, and, second, a search approach is required

which is able to explore linked content with various

degrees of formality. Fig. 1 shows the architecture of

such a search engine sketching the two major phases:

ofﬂine and online part. The ofﬂine part includes con-

nection of the data sources by processing the contents

in order to map/convert them to the knowledge base

(KB), which contains the ontologies and instances,

and to the text index. Depending on the formality of

data, i. e., structured, semi-structured or unstructured,

different processing steps like annotation, mapping,

natural language processing (NLP) are required to ﬁll

the KB. Statistical analysis is typically applied to cre-

ate the text index. The information pool, composed

of the KB and the text index, integrates and enables

access to information from various data sources, po-

tentially covering the whole spectrum from simple

text documents over linked data to formally described

knowledge and, as the result, combining Semantic

Web, Web 2.0, Web 1.0 and legacy data. However,

the important beneﬁt is that it enables to recognize

coherence between information elements from differ-

ent data sources and to reﬂect it in the KB by link-

ing them to each other. Therefore, this kind of data

source connection is applied in several Web 3.0 ap-

plications, e. g., in NEPOMUK (Sauermann et al.,

2007), Aletheia (Stieger and Aleksy, 2009) and In-

foSleuth (Nodine et al., 2000).

In this paper, we propose a hybrid semantic search

approach developed to explore such an information

pool (Fig. 1, online part). Our approach is not con-

stricted to simply retrieve plain facts or documents,

it deals with free-text, structured and also mixed

queries, and it is able to answer them with a combi-

nation of elements, i. e., documents and facts, speciﬁc

to the query. The rest of this paper is organized as fol-

lows: Sect. 2 introduces our approach in detail, Sect. 3

deals with the evaluation, Sect. 4 covers related work,

and Sect. 5 concludes the paper.

308

Schumacher K. and Sintek M..

SEARCHING WEB 3.0 CONTENT - A Semantic Search Approach.

DOI: 10.5220/0003340703080313

In Proceedings of the 7th International Conference on Web Information Systems and Technologies (WEBIST-2011), pages 308-313

ISBN: 978-989-8425-51-5

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

OFFLINE

ONLINE

core

Know-

ledge

Base

formal

Semi-

structured

(XML

Structured

(DB, RDF)

Annotation

Mapping

query and res

Facts

Documents

Formal

(SPARQL)

OFFLINE

ONLINE

formal

and

informal

engine

Text-

index

informal

Un-

structured

(text,

multimedia)

(XML

XHTML)

Statistical

Analysis,

NLP

lt processing

Documents

and facts

Documents

INFORMATION

RESULT

Informal

(free

text)

informal

elements

DATA

SOURCES

PROCESSING

INFORMATION

POOL

ENGINE

QUER

RESULT

Figure 1: Semantic Search Engine Architecture.

2 SEARCH APPROACH

Since the information pool integrates linked content

from different data sources as facts or as documents

with metadata, the search results are elements of the

KB and documents where one result can be a com-

bination of both, e. g., documents with facts. For

comprehensive search support, all kinds of queries,

i. e., free text queries, formal queries, and queries

composed of free text and also formal parts should

be allowed. This facilitates both purposeful formal

and vague free text queries and thus supports begin-

ner as well as advanced users and also users with

more and users with less knowledge about the back-

ground KB. Allowing all kinds of queries combined

with query construction assistance (e. g., by KB-based

auto-completion while the user is typing the query

terms) supports the user to create queries with as

many formal parts as possible. Given that formal parts

are precise unlike free text, they mitigate the problem

of resolving syntactic and structural ambiguity (ambi-

guity of the underlying structure of complex expres-

sions) by pre-query disambiguation. An example for a

mixed query is “which nations competed in free swim-

ming”. By typing the query terms, the system looks

for labels with the same beginning in the KB and rec-

ommends associated resources, in this case the class

Nation and the properties competesIn and hasCom-

petitor. By selection of a recommendation, the appro-

priate part of the internal query representation is re-

placed by the URI of the chosen resource. “free swim-

ming” cannot be syntactically matched to the KB, it

is interpreted as free text. A mixed result is a docu-

ment which contains the result table (person, stand-

ing) of the free swimming competition, together with

the facts that state the nationalities of the people.

For a comprehensive search on an information

pool as described above, the core search engine

should explore the KB and the text index using the

cross-linking between them. Since searching in a for-

mal KB and semantic document retrieval require dif-

ferent methods, a fact retrieval approach is needed

to ﬁnd triples from the KB (incl. metadata of docu-

ments), and a semantic document retrieval approach

is needed which is able to accomplish enhanced docu-

ment retrieval embracing matched facts and their on-

tological context. The approaches have to work to-

gether in a way that the engine is able to exploit and

combine both facts and documents meaningfully con-

sidering the links and available additional (derived)

knowledge. Further requirements are the ability to re-

solve ambiguities and the feasibility to apply an ad-

equate ranking function which is suitable for mixed

result lists where a single result can be a fact or facts,

a document as well as a document with facts.

For our prototype, we selected a triple-based fact

retrieval and a graph-traversal-based semantic docu-

ment retrieval algorithm for the following reasons:

triple-based search provides the resolving of ambi-

guities (see Sect. 2.1), and the graph traversal algo-

rithm Spreading Activation enables an effective com-

bination of fact retrieval and document retrieval which

will later be shown in Sect. 2.3. In our hybrid ap-

proach, at ﬁrst, the query is processed by the fact re-

trieval component which determines matching facts

and also the expanded query for document retrieval.

The matching facts together with the retrieved docu-

ments are used to set up the activation network and to

accomplish spreading. The approach is able to handle

all kind of queries, and it returns documents, facts and

also a combination of them speciﬁc to the query.

In terms of the KB, in addition to the availability

of the ontologies plus an instance base, this approach

requires either the instantiation of documents, i. e.,

each document is associated with an instance in the

KB, or the metadata of documents should explicitly

refer to instances of the KB. Furthermore, for a good

syntactic matching and improved retrieval, synonyms

of the ontological elements should be available.

2.1 Fact Retrieval

The triple-based fact retrieval approach maps query

terms to literals in the KB resulting in URIs of match-

ing and thus potentially intended properties (p

) and

non-properties (n

). These resources are used as sub-

ject, predicate or object to generate RDF queries, e. g.,

, p

?, p

. In (Goldschmidt and Krish-

namoorthy, 2005), the process of creating and apply-

ing queries is repeated for each new resource detected

but it is limited to only two hops in the KB in order to

avoid irrelevant inferences. We adopted and extended

this basic idea in order to enhance capabilities by pro-

viding more hops in the KB guided by the query.

In the following, G

denotes our KB as graph in-

SEARCHING WEB 3.0 CONTENT - A Semantic Search Approach

309

cluding the ontology and instances, q the query as

an ordered list of terms q = (t

,. ..,t

),n ∈ N. To

explain the syntactic matching, we deﬁne the tex-

tual content L

of G

as L

= {l| ∃

r, p,l

∈

, l is a Literal}. Furthermore, we deﬁne the set

of resources without literals and statements (RDFS

deﬁnes rdfs:Literal and rdf:Statement as Resources)

and divide it in the set of properties P

and

the set of non-properties N

: R

= {r| ∃

r, p,o

∈

∨ ∃

s,r, o

∈ G

∨ ∃

s, p, r

∈ G

, r /∈ L

= {p| ∃

p,rdf:type, rdf:Property

∈ G

}

and N

= R

where

s, p, o

is an RDF triple

that consists of subject s, predicate p and object o.

Our KB supports the search engine with synonyms

of the elements (classes, properties, and instances)

which are used for the syntactic matching step. If

the user does not choose one of the recommendations

in the query construction phase (see Sect. 2) we ap-

ply the n-gram syntactic matching method to match

a query term against the KB. In the ﬁrst case, the

similarity value is 1.0, in the second case it is the

computed n-gram value. Thus, the syntactic match-

ing returns for one query term t

∈ q a set of 3-tuples

,r, w

): M(t

) = {(t

,r, w

)| r ∈ R

, ∃l

∈

= ngram(t

)}. Only resources r were

included in the result set which are matched with a

> H where H ∈ [0,1] ⊂ R is a predeﬁned simi-

larity threshold. To match phrases, we compare the

weights of possible phrases (processed as one term)

with the weights of their single terms. If the weight

of the phrase is higher than the average weight of its

terms then we propose that the phrase is intended. The

total result M(q, G

) is the union of the results per

query term which is the starting basis for semantic

matching: M(q,G

) = {

i=1

M(t

)}.

For semantic matching, we do not consider the

whole KB but only the proper instances and desig-

nate this graph with G

. Furthermore, we deﬁne the

set of resources (again without literals and statements)

, properties P

and non-properties N

in G

anal-

ogously to the deﬁnitions for G

. We partition N

the pairwise disjunctive set of classes C

, the set of

things T

, i. e., resources, which are neither classes

nor literals, and the set of literals L

. When the query

is composed of only one term the semantic match-

ing step returns resources and triples of the graph

, dependent on what kind of resources the term

has matched. It returns the matched things, the

instances of matched classes, statements containing

the matched literals and statements with the matched

properties as predicate. When the query is composed

of more than one term, we iterate over every two ad-

jacent terms t

i+1

and consider their results from the

syntactic matching (M(t

),M(t

i+1

)). We create

and apply possible SPARQL queries with the matched

resources in order to ﬁnd suited triples in G

. In case

of two properties, the possible queries ask for either a

subject with both properties or two triples where the

object of the ﬁrst triple is the subject of the second

one. If there are no results for two query terms and

one or both of the terms match against a class, we

search for triples with the instances of the matched

class. In the next iteration, we consider the resulting

triples and the query terms which produced no triples

so far. We create query templates with these terms and

the subjects/objects of the triples which were found

based on the adjacent query terms. The ordered pro-

cessing of unmatched terms enables to handle also

enumerations of instances, properties or classes. The

process is iterated and stops when either all query

terms are matched or it is not possible to include all

terms since some terms do not match existing triples

in G

. We also make as many hops in the KB as pos-

sible guided by the query. The result consists of a set

of instances and a set of triples. At least, we identify

triple sets in the result which build a coherent sub-

graph of G

. To avoid the merging of instances to a

big subgraph in order to deliver well arranged results,

each triple of a subgraph is connected to another by

the same subject or the same object but we exclude

the connection by the same class. For this, we pick

a triple

, p

from the result and add all other

triples



, p



to the subgraph where s

= s

= o

or s

= o

or o

= o

, o

/∈ C

. One subgraph

is also a set of connected triples where the connec-

tion by class is excluded or it is one triple if there

are no connected triples in the result set. The rank-

ing is based on the w

which are computed in the

syntactic matching step by exact (recommended re-

source) or n-gram match. The weight of a matched

triple is the sum of the w

of participating ontolog-

ical elements r

. Each expansion with new triples in-

creases the weight of the partial result by the appro-

priate w

value. Finally, the rank of a result in S

is the sum of participating elements’ weights divided

by the number of query terms. Resolving ambigui-

ties is supported by triple-based processing. Since we

do not directly transform the user query to an RDF

query, the triples found step by step lead to possible

interpretations based on the existing triples in G

2.2 Semantic Document Retrieval

The pure semantic document retrieval comprises doc-

ument retrieval and Spreading Activation and it is car-

ried out if no facts have been matched.

The idea of using SA for information retrieval is

to ﬁnd more relevant information based on retrieved

WEBIST 2011 - 7th International Conference on Web Information Systems and Technologies

310

information elements by exploiting associations rep-

resented by semantic networks (as graphs). The on-

tological concepts are the nodes, the properties the

edges of the network, usually directed and weighted.

SA starts with the initial incoming activation of nodes

which propagate the activation along the edges acti-

vating the connected nodes. This process is iterated,

i. e., the activation spreads through the network, until

the stop condition is fulﬁlled. Result is the activa-

tion level of each node at termination time (Crestani,

1997). For semantic document retrieval, our network

model includes all instances, their classes and rela-

tions which connect these nodes, i. e., all properties

which link instances but not instances with literal val-

ues. To each edge a default weight w

is assigned. We

insert an inverse edge also if the property has no in-

verse in order to make sure that properties without an

inverse are involved if one of the connected instances

is activated. And we insert edges from classes to their

instances (but not the other way around) in order to

avoid noise in results since it stops spreading from

one instance to all other instances of a class but en-

ables to spread from a class to its instances if a class

is initially activated. We use a Lucene index for key-

word matching on the document corpus. The result is

a set of weighted documents where the associated in-

stances in the network are our initial activation points.

We apply the activation function I

∑

i j

(1 −α)

where I

is the incoming activation of nodes, O

the

outgoing activation (determined by an output func-

tion), w

i j

∈ [0,1] ⊂ R is the weight of the edge from n

to n

, and α is an attenuation factor which decreases

the activation strength with each propagation. Since

both α and the edge weights already decrease the acti-

vation level we use the simple output function O

= I

We apply an activation constraint which stops spread-

ing at a node when its activation level does not exceed

a deﬁned threshold. Furthermore, a fan-out constraint

averts the danger of a too wide spreading through

nodes with high connectivity, thus to become noise

in results. It is important to assure that each directed

edge is processed only once in order to avoid endless

spreading in cycles. In each iteration, the node with

the highest activation level and pending edges is pro-

cessed. SA stops when no more nodes have an acti-

vation level above the deﬁned threshold or the nodes

above the threshold have no pending edges. The re-

sult is the set of document instances weighted by their

activation level.

2.3 Hybrid Approach

In our hybrid search approach, at ﬁrst, fact retrieval

is carried out and deliver a set of matched resources.

We use the synonyms (in KB) of matched resources

to perform query expansion before querying the text

index. The result of the document retrieval with the

expanded query is a set of weighted documents as

described in Sect. 2.2. To be able to create the ac-

tivation network, the set of matched resources of the

fact retrieval result set is to extract where we differ-

entiate between instances, properties and classes. For

the extraction of the classes and the properties only

the results of the syntactic matching are relevant. To

get the matched instances, we extract not only the re-

sources found by the syntactic matching, but also sub-

jects and objects from the triples found by the seman-

tic matching. Now, all information to set up the acti-

vation network is available. We design our semantic

network as described in Sect. 2.2 and apply the same

rules to ﬁll the matrix. The difference is: we assign to

the properties which are matched by the fact retrieval

their computed weight instead of the low initial de-

fault weight w

. The set of activation nodes contains

the instances of found documents and the instances

and classes from the fact retrieval results where their

weights are applied as initial activation weight. The

spreading function, constraints and stop condition are

the same as described in Sect. 2.2. The results of the

activation process are a set of weighted nodes includ-

ing the documents. The last step of the hybrid ap-

proach is to merge the results of the spreading acti-

vation and the subgraphs. For this, we start by the

subgraphs and add the connected documents from the

results of the SA process. Note that the subgraphs

also contains the properties which are not spread since

they have a literal value. For all other found docu-

ments, the results of SA deﬁne a set of facts which de-

scribes the answer and which is also part of the result.

We only collect the high ranked resources as facts to

avoid confusing results. The rank of one object from

the result set is the average rank of the contained in-

formation elements.

3 EVALUATION

We evaluated our approach based on a manually an-

notated test bed (Grothkast et al., 2008). It contains

an ontology about Olympic Games, an instance base

which describes the Olympic Games 2004 and 122

news articles about it. The news have been manually

annotated, their metadata directly refers to the KB,

e. g., news about the 200m Swimming for Men are

annotated with the competition and people from the

instance base which are mentioned in the text. Fur-

thermore, they are associated with 8 queries, i. e., it

is known which articles are relevant for a particular

SEARCHING WEB 3.0 CONTENT - A Semantic Search Approach

311

query. The queries are: q1: standings of Australians;

q2: disciplines with gold for Australians; q3: teams

of South Korea, q4: when did Chinese win gold; q5:

places of competitions in cycling; q6: in which disci-

plines did British sportsmen compete; q7: who com-

peted in swimming; q8: which nations have a women

football team. We applied simple keyword search (1),

semantic document retrieval (2), and the hybrid ap-

proach (3) on the test set and computed the precision,

recall and F-measure per query.

Table 1: Evaluation Results - Keyword Search.

Query Precision Recall FMeasure

q1 0.2752 1.0 0.4316

q2 0.0354 1.0 0.0683

q3 0.3421 0.8125 0.4814

q4 0.1563 1.0 0.2703

q5 0.2858 1.0 0.4444

q6 0.125 0.2857 0.1739

q7 0.0338 0.6667 0.0645

q8 0.0138 1.0 0.0274

average 0.1584 0.8456 0.2452

Table 2: Evaluation Results - Semantic Document Re-

trieval.

Query Precision Recall FMeasure

q1 0.2632 1.0 0.4167

q2 0.0345 1.0 0.0667

q3 0.2192 1.0 0.36

q4 0.0408 1.0 0.0784

q5 0.0317 1.0 0.0615

q6 0.5 1.0 0.6667

q7 0.0256 1.0 0.05

q8 0.0435 1.0 0.0833

average 0.1448 1.0 0.223

Table 3: Evaluation Results - Hybrid Approach.

Query Precision Recall FMeasure

q1 0.6042 1.0 0.7533

q2 0.069 1.0 0.129

q3 0.4688 1.0 0.6383

q4 0.2134 1.0 0.3517

q5 0.4 1.0 0.5714

q6 0.4615 1.0 0.6315

q7 0.4 1.0 0.5714

q8 0.25 1.0 0.4

average 0.3584 1.0 0.5058

The precision of the semantic document retrieval

in comparison with the keyword search decreases in

some cases but, at the same time, the recall increases.

It is caused by Spreading Activation where also news

articles are activated which are relevant for only a part

of the query, i. e., annotated with only some instances

which constitute the metadata of the documents found

by the keyword search. Recall increases for the same

reason, since documents which do not contain one of

the query terms but, e. g., a synonym of it, were acti-

vated via the shared metadata, ontological context.

The hybrid approach performs very well for both

precision and recall. It tackles the problem of too

wide spreading since it applies: 1. more precise ac-

tivation points by involving matching facts and their

ranks; 2. a query speciﬁc network setup, especially

for weights of involved properties. The best improve-

ments in comparison to keyword search and seman-

tic document retrieval can be achieved if the query

matches some properties (q2, q6 and q7), since they

guide the ﬂooding during SA. Fewer improvements

are observable if only classes and commonly used in-

stances are matched since classes spread to all of their

instances and commonly used instances are not spe-

ciﬁc enough and cause spreading from nodes which

are far apart. E.g., in q4 where “when” matches a

class and “Chinese” is a commonly used instance, and

in q5 where both “places” and “competitions” match

classes. The fact retrieval delivers in some cases (q3,

q5, q7) the answer to the query as facts. The hy-

brid approach does not perform well if no facts are

matched because it only performs pure semantic doc-

ument retrieval in this case.

4 RELATED WORK

In (Schumacher et al., 2008), we informally presented

an early stage of our hybrid semantic search approach

that features a combination of semantic document re-

trieval and fact retrieval, but the approach described

in this paper is able to process mixed queries and the

algorithms (esp. the hybrid approach) have been dras-

tically improved with respect to efﬁciency.

Existing hybrid search systems address mainly the

problem of processing free text or mixed queries on

formal KBs (fact retrieval) where parts/metadata of

documents are transformed into formal knowledge

or they are considered as facts in the KB. Semplore

(Wang et al., 2009) supports faceted search and com-

plex hybrid queries on structured data by transform-

ing Web of Data into a text index. The ﬁelds represent

predeﬁned hierarchical relations and they also hold

textual properties of an instance. (Tran et al., 2007)

describes an approach translating keyword queries

to DL conjunctive queries using background knowl-

edge and so supports semantic based declarative ques-

tion answering on formal KBs. An advanced version

of this approach ﬁrst computes conjunctive formal

queries from keywords by exploration of top-k match-

ing subgraphs allowing the user to choose one (Tran

et al., 2009). This system supports free text queries in-

deed, but the users have to examine complex conjunc-

tive queries before continuing with the search. (Lad-

wig and Tran, 2010) introduces a tight integration of

the approach with query answering, i. e., graph pattern

matching. There are several further methods for graph

WEBIST 2011 - 7th International Conference on Web Information Systems and Technologies

312

based query processing in oder to search over graph-

structured data, e. g., BLINKS (He et al., 2007). Some

approaches apply a predeﬁned set of query templates,

e. g., (Sacaleanu et al., 2008), where the latter real-

izes a multilingual entailment-based question answer-

ing approach. Other methods for query interpretation

are using deep NLP, e. g., Powerset (Converse et al.,

2008) or large background knowledge created with

high effort, e. g., WolframAlpha. Document search

by fact retrieval is supported by DBPedia, Semantic

Wikis and documents enriched with RDFa or micro-

formats since the document search is processed by re-

trieving the ‘included’ facts.

Our approach is essentially different to these

search methods as it provides fact and document re-

trieval, formal, free text and mixed queries, and also

mixed results, i. e., documents with facts, while all

other existing approaches support only a subset of

this.

5 CONCLUSIONS AND FUTURE

WORK

The evaluation shows the power of our hybrid ap-

proach. It performs best if properties are involved

in the query since they alleviate the weakness of SA,

i. e., noise results caused by uncontrolled spreading.

Also in other cases, the combination performs quite

well due to the more precise fact retrieval results. If

no facts are available for a query, the search approach

performs semantic document retrieval using the meta-

data of the documents. Furthermore, our approach

delivers facts if they are available, and so the user

doesn’t need to browse the documents of the result

list. In future versions, to improve the precision if no

facts are matched, we are going to setup our seman-

tic network with edge weights which express the im-

portance of relations and we plan further evaluations

(e. g., with DBPedia). We also foresee to integrate this

approach in the digital library assistant DiLiA (Seifert

and Kruppa, 2010), by extending it to support com-

plex queries for expert users.

ACKNOWLEDGEMENTS

This research has been supported in part by the THE-

SEUS Program CTC, which is funded by the BMWi

(gn 01MQ07016), and the research project DiLiA

which is co-funded by the ERDF (gn 10140159). The

responsibility for this publication lies with the au-

thors.

REFERENCES

Converse, T., Kaplan, R. M., Pell, B., Prevost, S., Thione,

L., and Walters, C. (2008). Powerset’s natural lan-

guage wikipedia search engine. Report, Powerset inc.,

California.

Crestani, F. (1997). Application of spreading activation

techniques in information retrieval. Artiﬁcial Intelli-

gence Review, 11(6):453–482.

Goldschmidt, D. E. and Krishnamoorthy, M. (2005). Archi-

tecting a search engine for the semantic web. In AAAI

Workshop on Contexts and Ontologies.

Grothkast, A., Adrian, B., Schumacher, K., and Dengel, A.

(2008). OCAS: Ontology-based corpus and annota-

tion scheme. In Proc. of the High-level Information

Extraction Workshop 2008, pages 25–35.

He, H., Wang, H., Yang, J., and Yu, P. S. (2007). BLINKS:

Ranked keyword searches on graphs. In Proc. of

the International Conference on Management of Data,

pages 305–316.

Ladwig, G. and Tran, D. T. (2010). Combining keyword

translation with structured query answering for efﬁ-

cient keyword search. In Proc. of the 7th Extended

Semantic Web Conference.

Nodine, M., Fowler, J., Ksiezyk, T., Perry, B., Taylor, M.,

and Unruh, A. (2000). Active information gathering

in InfoSleuth. International Journal of Cooperative

Information Systems, 9(1/2):3–28.

Sacaleanu, B., Orasan, C., Spurk, C., Ou, S., Ferrandez, O.,

Kouylekov, M., and Negri, M. (2008). Entailment-

based question answering for structured data. In

Posters and Demonstrations, 22nd Int. Conference on

Computational Linguistics, pages 29–32.

Sauermann, L., van Elst, L., and Dengel, A. (2007). PIMO

– a framework for representing personal information

models. In Proc. of I-SEMANTICS 2007, pages 270–

277.

Schumacher, K., Sintek, M., and Sauermann, L. (2008).

Combining fact and document retrieval with spread-

ing activation for semantic desktop search. In Proc.

of the 5th European Semantic Web Conference, pages

569–583.

Seifert, I. and Kruppa, M. (2010). A pool of topics: In-

teractive relational topic visualization for information

discovery. In Proc. of the Conference on Visual Infor-

mation Communication, pages 195–207.

Stieger, B. and Aleksy, M. (2009). Utilization of knowl-

edge management for service business processes im-

provement. In Proc. of the Int. Multiconference on

Computer Science and Information Technology, pages

171–175.

Tran, T., Cimiano, P., Rudolph, S., and Studer, R. (2007).

Ontolgy-based interpretation of keywords for seman-

tic search. In Proc. of the 6th Int. Semantic Web Con-

ference, pages 523–536.

Tran, T., Wang, H., Rudolph, S., and Cimiano, P. (2009).

Top-k exploration of query candidates for efﬁcient

keyword search on graph-shaped (RDF) data. In

ICDE, pages 405–416.

Wang, H., Liu, Q., Penin, T., Fu, L., Zhang, L., Tran, T.,

Yu, Y., and Pan, Y. (2009). Semplore: A scalable IR

approach to search the web of data. Journal of Web

Semantics, 7(3):177–188.

SEARCHING WEB 3.0 CONTENT - A Semantic Search Approach

313