CONTEXTUAL SEMANTIC SEARCH

Capturing and using the User’s Context to Direct Semantic Search

Caio Stein D’Agostini, Renato Fileto, Mário Antônio Ribeiro Dantas and

Fernando Alvaro Ostuni Gauthier

Depto. de Informática e Estatística, Universidade Federal de Santa Catarina, Caixa Postal 476

88.040-900, Florianópolis - SC, Brazil

Keywords:

Semantic web, context, personalization, user proﬁle, semantic search.

Abstract:

One orthodox perspective of the semantic Web depends on establishing a uniﬁed representation of ontology

for an universe of discourse. However this expectation is unrealistic, because different people have different

views of the world, making it impossible to devise a uniﬁed knowledge view that satisﬁes everyone in certain

cases. One solution for this problem is to allow different users’ views of consensual knowledge formalized

in an ontology, and keep track of the mappings between these views and the underlying ontology. This work

proposes to collect context information from the users interactions with a semantic search system, in order to

gradually build individual users’ views mapped to an ontology. This approach allows the user to pose queries

based on keywords or his personal knowledge view. In addition, each personalized knowledge view captures

the preferences of a single, speciﬁc user, enabling the system to provide better search results, based on its

previous experience with that user.

1 INTRODUCTION

Semantic search approaches (Mangold, 2007) try to

augment and improve searches on a set of resources

that are initially unknownto the user, by using ontolo-

gies and semantic annotations of the resources. Even

though there are several problems to be addressed

when dealing with semantic technologies, such as the

ontology creation and the semantic annotation pro-

cess, we chose to center the scope of this paper on

only one issue. The focus of our work is on how

to capture user’s contextual information and how to

employ it, along with a given ontology and resources

annotated according to its terms, to improve the pre-

cision, the coverage and the ranking of search results

for individual users.

The ontologies used to describe resources in the

semantic Web are potentially conﬂicting with partic-

ular users’ views of an universe of discourse. These

conﬂicts create a tension between individual users and

formal ontologies. The former wants to see and in-

teract with semantic Web systems according to their

particular views, while the latter tries to describe a do-

main in a standardized way. In order to alleviate this

tension, the systems’ appearance and behavior can

be customized according to particular users’ views

or, at least, the system must avoid forcing the user

to strictly comply to cumbersome ontologies, some-

times built by forcing consensus even when they are

developed for small groups of people. Nevertheless,

allowing different users’ views characterizes just the

stage of Semantic Coexistence in the Web (Naeve,

2005). In order to reach the stage of Semantic Col-

laboration (Naeve, 2005) it is also necessary to estab-

lish bridges between individual users’ views and stan-

dardized ontological descriptions (Park and Cheyer,

2006).

This paper reports the progress of an ongoing

study to create a personalized knowledge view for

each user mapped to a consensual ontology. The goal

is to provide individual tailored views of the knowl-

edge base, so that he can beneﬁt from the systems’

use of ontologies and semantic annotations, without

the need to cope with their full extent and details.

These knowledge views are automatically created and

updated by capturing contextual information during

user interactions with the system, without additional

user’s effort. It alleviates the tension between individ-

ual users’ views and the consensual ontology. In addi-

tion, as the individual views capture the user’s context

and preferences, they enable the estimation of user in-

tentions in order to drive the disambiguation and se-

mantic extension of keyword-based searches for re-

sources annotated to a domain ontology.

154

Stein D’Agostini C., Fileto R., Antônio Ribeiro Dantas M. and Alvaro Ostuni Gauthier F. (2008).

CONTEXTUAL SEMANTIC SEARCH - Capturing and using the User’s Context to Direct Semantic Search.

In Proceedings of the Tenth International Conference on Enterprise Information Systems - SAIC, pages 154-159

DOI: 10.5220/0001706801540159

 SciTePress

This paper is organized as follows. Section 2

presents a deﬁnition of context relevant for the current

work and some proposals on how to represent and use

contexts. Section 3 presents our proposal to capture

the context corresponding to the users’ views and use

it to improve the search process. Section 4 describes

our representation of particular users’ views and their

mappings to an underlying ontology. Section 5 pro-

vides algorithms capture the users’ views and to use

it to drive the search process and potentially improve

the search results by infering the probable user inten-

tions. Finally, section 6 presents closing comments

and work left to be done or currently under develop-

ment.

2 RELATED WORK

Context has a broad set of deﬁnitions according to

how it is used. An appropriate deﬁnition for this work

is the one from (Mani and Sundaram, 2007), which

says that context is a dynamic subset of knowledge

that affects the communication between entities. For

information retrieval, context is a representation of a

user’s view of a domain. The context describes his

preferences according to this particular view of a do-

main. It affects how the user’s queries are interpreted

and processed.

(Challam et al., 2007) and (Sieg et al., 2007) use

ontological proﬁles to represent individual user con-

texts. In these proﬁles, the degree of interest of a user

for a subject is represented by a weight in the on-

tology term referring to that subject. These weights

are used to direct the searching algorithm (Sieg et al.,

2007) and to rank the retrieved results (Challam et al.,

2007). On both solutions the user’s behavior is mon-

itored. Information such as the websites visited by

the user can be used to update his proﬁle. However,

this type of ontological proﬁle only captures the in-

terest for terms, failing to capture relevant relations

between them.

(Michlmayr et al., 2007) uses publicly available

bookmarks, from sites such as del.icio.us

, to analyze

how users correlate words. Since bookmarks are cre-

ated at an speciﬁc time and stored for years, it is pos-

sible to create contextual proﬁles that reﬂects the user

context at different time periods. A weighted graph

is created by parsing all tags employed by the user to

describe his bookmarks. Each tag is represented by a

node in the graph, and every co-occurrence between

two tags in the annotation of the same bookmark con-

tributes to increase the weight of the edge between the

www.del.icio.ous

corresponding nodes. The edges with higher weights

are selected an user proﬁle. This proﬁle can be used to

expand search expressions with other words that are

probably also relevant to the user.

A similar graph structure is suggest by (Park and

Cheyer, 2006). However, this work separates the

users’ views, represented as topic maps, from the on-

tological description of a domain. The topic maps

are graphs where the nodes - called topics - repre-

sent subjects relevant to the users and the edges rep-

resent correlations between topics. The topics are

named according to the users’ views and are mapped

to the terms with corresponding names in the on-

tology. However, (Park and Cheyer, 2006) fails do

provide a detailed description of how to use these

topic maps. Several questions are left unanswered,

such as details of the topic maps structure (e.g., how

homonym topics are stored and differentiated), their

mapping to the ontology (e.g., how to know that dif-

ferent topics are synonyms), the criteria used to deﬁne

browsing options in a topic map (i.e, which correlated

topic is more important when there are several edges

leaving a topic) and the creation and maintenence of

topic maps.

The analysis of these works shows that there

are several alternatives to represent contextual in-

formation. The solution proposed in (Michlmayr

et al., 2007) individually represents the context from

different users, but with no relation with a com-

monly accepted ontology. However, the ideias from

(Michlmayr et al., 2007) might help to solve the is-

sues left in (Park and Cheyer, 2006), by providinge

suitable mappings between users’ views and ontolo-

gies.

3 CAPTURING THE USER

CONTEXT

This paper proposes to explicitly represent particular

users’ views of an universe of discourse and establish

mappings between each view and a consensual ontol-

ogy. It helps to alleviate the tension between partic-

ular users’ views and formal ontologies. In addition,

it enables the alignment of the search results with the

user’s context.

Example 1 (Keyword-search for São Paulo).

Consider a user who is travelling to Brazil by air-

plane. Suppose that he wants to learn about the

airport he is going to arrive. He is going to São

Paulo, which is a city, but this composite word

is also the name of a state and a soccer team,

among other entities. However, this foreign user does

CONTEXTUAL SEMANTIC SEARCH - Capturing and using the User’s Context to Direct Semantic Search

155

not have all this knowledge. Assume that the user sub-

mits a search for the keyword São Paulo and waits

for the results.

The ﬁrst time the user poses a search, his con-

text is empty. Since the system cannot infer his in-

tentions, it checks the occurrences of São Paulo

in the ontology. The system returns pointers to re-

sources annotated with the keyword São Paulo,

grouped according to different concepts associated to

São Paulo, such as city, state and soccer

team, as shown in Figure 1.

Resources related to city(São Paulo)

Possible meanings for the keyword São Paulo

city(São Paulo)

state(São Paulo)

soccer team(São Paulo)

city(São Paulo), city

city(São Paulo), Guarulhos

city(São Paulo), airport, Guarulhos

Doc

Annotated to:

Figure 1: Results for a search for the keyword

São Paulo

The user selects only the results that, besides

city of São Paulo city(São Paulo), are

also related to airport or city(Guarulhos).

The system then updates the weighted topic map

representing the user’s context. In the next time

the user searches for São Paulo, the system em-

ploys the contextual information to elevate the rank

of resources annotated to city(São Paulo),

city(Guarulhos) and airport.

Note that the association of a keyword with a

speciﬁc concept and the respective instance from the

ontology (e.g.,

city(São Paulo)

provides a more

speciﬁc connotation for the keyword. The relevance

of a connotation in the user context enables the

disambiguation of future queries, and the relevance

of the associations of the keyword with the annotation

of the resources selected by the user enables means to

drive semantic expansion of future queries. To make

this context-driven search possible, there are some

problems that need to be addressed: How to formally

represent the user’s context aligned to the ontological

description of a universe of discourse; how to develop

algorithms to keep the context information up-to-date

with the continuously changing user’s view; how to

employ this information to improve search results.

The proposed solutions for these problems are

presented next.

4 REPRESENTING PERSONAL

CONTEXTS ALIGNED TO AN

ONTOLOGY

Our contextual semantic search system follows the

three-layer architecture proposed by (D’Agostini

et al., 2007). Figure 2 illustrates this architecture.

Its layers are organized as follows:

Figure 2: Mapping between the user’s context and domain

ontology.

• Users’ Contextual Views (Weighted Topic

Maps): It maintains the users’ knowledge views.

These views are updated with information col-

lected from previous user’s interactions with the

system. Each user has his individual view. The

subjects of the views are indexed by their names

in order to speed up the retrieval algorithms.

• Deﬁnitions (Ontology): It mantains the ontol-

ogy used to formally describe the universe of dis-

course and the non-intrusive annotations of re-

sources based on that ontology. The terms of the

ontology and the semantic annotation are also in-

dexed to speed up their retrieval. The resources

are considered to be already annotated

• Content (Annotated Resources): This layer

refers to the stored content. It includes a repos-

itory of resources associated with ordered seman-

tic annotations to support their retrieval. A re-

source can be a data set, a corpus (e.g., a doc-

ument in a digital library) or a service (e.g., a re-

source in a computational grid), etc, depending on

the type of application.

We represent each user’s context by a weighted

topic map. A topic can refer to any subject and

has a name, properties and relations with other top-

ics (Garshol, 2004). Topic maps have been chosen to

ICEIS 2008 - International Conference on Enterprise Information Systems

156

describe the users’ contexts because they are human

readable and organize knowledge in a similar way as

people do (Novak and Gowin, 1984), while they can

still provide a formal and machine processable knowl-

edge representation. For the purpose of this work, a

topic map is represented as a graph TM(T, A), where

T is the set of topics representing subjects of interest

to the user and A is the set of associations which rep-

resent how the topics are correlated according to the

user’s view.

Each topic t ∈ T is tagged with the word employed

by the user to name the respective subject. Each topic

corresponds to only one term (concept or instance of

concept) formally described in the ontology. It is pos-

sible for two topics to have the same name, but they

must refer to different terms with the same name.

In Example 1, the keyword

São Paulo

renders

an ambiguous search, since

São Paulo

is the name

of different subjects. To disambiguate searches, each

topic t ∈ T has a weight w ∈ [0, 1]. The sum of the

weights from all topics with the same name equals

1. The topic weights are assigned based on how

frequently the user referred to each topic (i.e., each

connotation of the same word) in the past. Associa-

tions also have weights. Each association a ∈ A has a

weight w ∈ [0, 1]. The sum of the weights from all as-

sociations departing from the same topic also equals

1. The weights of the associations express the like-

lihood that a user searching for one subject is (also

or instead) interested in another subject, as in Exam-

ple 1, where the user searches for

São Paulo

, when

he is in the airport located in the neighboring city of

Guarulhos

The process of creating new topics and associa-

tions in the topic map, as well as keeping them aligned

to the underlying ontology is described next.

5 CONTEXT MANAGEMENT

The topic map represents a user’s knowledge view.

It is created and updated based on the user’s feedback

over the returned results for each search he poses. The

system considers that the more frequently used topics

are more relevant to the user than the less frequently

used topics with the same name. This assumption al-

lows the system to disambiguate the searched key-

word. Since the searches are directed by the user’s

topic map, the topic map evolution and the search pro-

cess are dependent on each other. The same context-

search-context dependency happens at (Sieg et al.,

2007), where it is called the cold-start problem. This

dependency is broken by allowing searches to be per-

formed directly on the ontology.

That type of search happens when there is no in-

formation available in the topic map to infer the user’s

particular intentions for a keyword, the search for that

keyword is performed directly in the ontology. If

the system ﬁnds different connotations for that key-

word in the ontology, it presents them to the user,

and asks him to choose results related to particular

connotations. On the other hand, when the keyword

is already in the topic map, the system can use the

weights of topics referring to different connotations

of the keyword to make a ranked list of results. The

user’s choices over the returned results are used, in

both cases, to update his context. This process can be

repeated for each search or until enough contextual

information has been gathered.

5.1 Context Maintenance

The topic map evolves continuously at each search,

by adjusting the topic and association weights accord-

ing to the user choices over the search results. A

topic weight is increased when the term it refers to

is used to annotate a resource considered relevant by

the user. The weights of the associations are updated

according to the correlation between relevant topics.

The associations correlating the topics corresponding

to the terms used to annotate the selected results and

the topic corresponding to the relevant meaning of the

search keyword have their weights increased. The

higher the number of times a term is used to anno-

tate the chosen resources, the higher the weight of its

corresponding association with the keyword topic. If

the association does not exists, a new one is created.

This process is described in Algorithm 1.

Algorithm 1 (Context Maintenance Algorithm).

1. The user checks the resources he considers relevant

2. For each topic checked by the user{

3. Increment the weight from the topic matching;}

4. For each resource checked by the user{

5. Increment the weights from the associations

between the topic corresponding to the keyword

and the topics corresponding to the terms used to

annotate the checked resources;}

6. Normalize the weights from topics and associations

to the interval [0,1];

The weights of topics and associations which are

not considered relevant in a search can be decayed, so

that feedbacks from recent searches have more inﬂu-

ence in the weights than older ones. This procedure

can be used to reﬂect the temporal aspects of the con-

text.

Consider the results from the search in Exam-

ple 1, in Figure 1. The user selected the resources

, which are annotated to

city(São Paulo)

. But

so is resource

, which was not checked by the

user. Also, both resources

and

are annotated

CONTEXTUAL SEMANTIC SEARCH - Capturing and using the User’s Context to Direct Semantic Search

157

Figure 3: Possible relevant result based on the user’s topic

map.

Guarulhos

and

is also annotated to the term

airport

, to which none of the rejected resources are.

With this information the system can infer that, when

the user searched for

São Paulo

, he was interested

in the

city

, not in the

state

soccer team

. And

since the

Guarulhos

and

airport

annotation are the

ones which differs resources

and

from the re-

jected ones, it can be inferred that, by searching for

city(São Paulo)

, the user was actually searching

for

Guarulhos

airport

. Figure 3 shows the corre-

sponding created topic map and its mapping to the on-

tology. The correlation between

São Paulo

(the key-

word searched for) and

Garulhos

is stored by apply-

ing a weight 1.0 to a topic corresponding to

city(São

Paulo)

and creating an association between that topic

and the topic

Guarulhos

. This association has as-

signed a weight of 0.66, since it correspond to 2 of

the 3 annotations characteristic to the selected re-

sources (2 x

Guarulhos

and 1 x

airport

). The topic

airport

receives a lower weight, since

airport

used less times than

Guarulhos

to annotate the re-

sults checked by the user.

5.2 Contextual Semantic Search

During the search process, summarized in Algorithm

2, the weighted topic maps are used to direct the

search in the ontology. This directioning is done by

disambiguating keywords and expanding the search

to include other subjects the user relates to the subject

he is searching for, allowing the search to consider

the user’s particular view of how different elements

from a domain relate. Depending on the keywords

searched, three different situations might occur: the

searched keyword has a correspondence in the topic

map, meaning that there is context information avail-

able in the topic map to direct the search process; the

keyword has no correspondence in the topic map, but

there are terms in the ontology which are referred by

the keyword (the ’cold-start’ problem); the keyword

has no correspondence either on the topic map or at

the ontology and no results are returned. For all these

situations the search process described here considers

that the keyword received as parameter does not need

any type of processing. This means that steps such as

identifying if a search expression is either a compos-

ite word or two independent words have already been

performed.

Algorithm 2 (Search Algorithm).

1. The user inserts a search keyword;

2. For each keywork{

3. Look in the topic map for matching the keyword;

4. If system finds topic(s){

5. Disambiguate the topics using their weights;

6. Expand the search using the disambiguated

topic’s associations;

7. Identify the terms in the ontology corresponding

to the topics;

8. }else{

9. Look in the ontology for terms marching the

keyword;

10. If system does not find term(s){

11. Ends search;

12. }else{

13. Identify those terms;} } }

14. Recover and list to the user the resources annotated

by the terms identified terms;

For the searches driven by the context informa-

tion, the ﬁrst step is, if necessary, to disambiguate the

keyword. The disambiguation is done by selecting

the topics with the higher weight in the topic map,

since they are likely the ones the user is interested

at. After selecting a topic, it is necessary to verify

how the user correlates that topic with the others. For

this veriﬁcation, the associations departing from the

selected topic are ordered according to their weights,

from higher to lower weights. The search is then ex-

panded, considering also the terms in the ontology re-

lated to the topics on the other end of those associa-

tions. This process is exempliﬁed in Example 2 and

illustrated in Figure 3.

Example 2. Take the results from Example 1. Con-

sider that the user performed a new search for

São Paulo. Through the weight of the topics

named São Paulo the system can infer that the

user means city(São Paulo). By also consid-

ICEIS 2008 - International Conference on Enterprise Information Systems

158

ering the weights from the associations originating

at the topic city(São Paulo), it can also be

inferred that he is most likely trying to ﬁnd infor-

mation about Guarulhos and airport and not

about the city of São Paulo. This last inference is

only possible because of the association established

between city(São Paulo) and Guarulhos or

airport, which is not represented in the ontology,

only in the user’s topic map.

city(São Paulo), airport, Guarulhos

Other results for the keyword São Paulo

city(São Paulo), Guarulhos

city(São Paulo), city

...

Resources related to city(São Paulo)

Doc

Annotated to:

Doc

Annotated to:

Figure 4: Results from the search from Example 2.

6 CONCLUSIONS

The use of personalized knowledge views mapped

to an ontology reduce the tension between usability

and consensual knowledge and reach the

Semantic

Collaboration

stage of the Web. This paper

presents a proposal to (i) gradually build the user’s

view and keep it updated, by gathering contextual

information from his interactions with a semantic

search system; (ii) represent the user’s context as a

weighted topic map, whose topics are connected to

speciﬁc connotations described in an ontology; (ii)

use these weighted topic maps to infer probable user

intentions, based on his previous choices, in order to

make a ranked list of search results tailored for the

speciﬁc user.

At the current stage, experiments are being

planned and the data necessary for them are being

collected. Thus, there are no experimental results

available yet. Nevertheless, many of the assumptions

made so far can be veriﬁed by experiments realized

in related works, such as (Challam et al., 2007), (Sieg

et al., 2007) and (Michlmayr et al., 2007).

Currently, the topic map representing a particular

user’s view only allows concepts and instances de-

scribed in the underlying ontology as topics. One

challenge is to ﬁnd a trade-off between the level of in-

dependence given to the users to express their knowl-

edge views and the limitations and cost of the fa-

cilities to establish semantic bridges between these

views. Thus, allowing topics with no corresponding

term in the ontology and using these topics to evolve

the ontology is a theme for future work. The visual

presentation, browsing and edition of the user’s topic

map are also considered for future developments.

ACKNOWLEDGEMENTS

This work is supported by CNPq (grant

48139212007-6), CAPES and FEESC.

REFERENCES

Challam, V., Gauch, S., and Chandramouli, A. (2007).

Contextual search using ontology-based user proﬁles.

Conference RIAO2007.

D’Agostini, C., Fileto, R., Dantas, M. A. R., and Gauthier,

F. O. (2007). Inferring user’s intentions through con-

text. Sessão de Posters do Simpósio Brasileiro de Ban-

cos de Dados (SBBD), pages 7–10.

Garshol, L. (2004). Metadata? Thesauri? Taxonomies?

Topic Maps! Making Sense of it all. Journal of Infor-

mation Science, 30(4):378.

Mangold, C. (2007). A survey and classiﬁcation of seman-

tic search approaches. International Journal of Meta-

data, Semantics and Ontologies, 2(1):23–34.

Mani, A. and Sundaram, H. (2007). Modeling user con-

text with applications to media retrieval. Multimedia

Systems, 12(4):339–353.

Michlmayr, E., Cayzer, S., and Shabajee, P. (2007). Tech

report: Hpl-2007-72: Adaptive user proﬁles for enter-

prise information access. Technical report.

Naeve, A. (2005). The Human Semantic Web-Shifting from

Knowledge Push to Knowledge Pull. International

Journal of Semantic Web and Information Systems,

1(3):1–30.

Novak, J. and Gowin, D. (1984). Learning How to Learn.

Cambridge University Press.

Park, J. and Cheyer, A. (2006). Just For Me: Topic Maps

and Ontologies. Lecture Notes in Computer Science,

3873:145–159.

Sieg, A., Mobasher, B., and Burke, R. (2007). Ontological

user proﬁles for personalized web search. Proceedings

of the 5th Workshop on Intelligent Techniques for Web

Personalization, Vancouver, Canada, July.

CONTEXTUAL SEMANTIC SEARCH - Capturing and using the User’s Context to Direct Semantic Search

159