A CONTEXT-BASED MODEL FOR WEB QUERY
REFORMULATION
Ounas Asfari, Bich-Liên Doan, Yolaine Bourda
Department of Computer Science, SUPELEC, 3 rue Joliot-Curie, Gif-sur-yvette, France
Jean-Paul Sansonnet
LIMSI-CNRS, University of Paris 11, Orsay, France
Keywords: Query Reformulation, Context, Task modeling, Information Retrieval, Personalization.
Abstract: Access to relevant information adapted to the needs and the context of the user is a real challenge in the
Web Search, owing to the increase of heterogeneous resources on the web. In most of cases, user queries are
shortened and ambiguous, thus we need to handle implicit needs or intentions that are behind these queries.
For improving user query processing, we present a context-based method for query expansion that
automatically generates context-related terms. Here, we consider the user context as the current state of the
task that the user is undertaking when the information retrieval process takes place, thus State Reformulated
Queries (SRQ) are generated according to the user task state and the ontological user profile to provide
personalized results in a particular context.
1 INTRODUCTION
In the present, it has become common to seek daily
information on the web, including such tasks as
using information retrieval system for shopping,
travel booking, academic research, and so on.
Understanding the user task is critical to improve the
processing of user needs. The increase of mobile
devices (such as PDA, cellular phone, laptop…)
including diverse platforms, various work
environments, have created new considerations and
stakes to be satisfied. Thus, it is now expected to use
the mobile device anywhere to seek information
needed to perform a current task, but the problem is
that the classic information retrieval systems provide
same results for different needs, contexts, intentions
and personalities, so too many irrelevant results are
provided, it is often difficult to distinguish context-
relevant information from secondary information or
even noise. Thus the results provided for mobile
users to perform tasks must be related to the context.
The user context can be assimilated to all factors
that can describe his intentions and perceptions of
his surroundings (Mylonas et al., 2008). These
factors may cover various aspects: environment
(light, services, people…), spatial-temporal
(location, time, direction…), personal
(physiological, mental, professional …), social
(friends, colleagues…), task (goals, information
task), technical etc. Many studies (Mylonas et al.,
2008), (Sieg et al., 2007) try to take into account the
user context but the problems to be addressed here
include how to represent the context, how to
determine it at runtime, and how to use it to
influence the activation of user preferences. It is
very difficult to take into consideration all the
contextual factors in one information retrieval
system, so the researchers often define the context as
certain factors (location for example). In this paper
our definition of the context is that the context
describes the user current task, its changes over time
and its states, i.e. we take into account the task
which the user is undertaking when the information
retrieval process occurs.
Recent studies have tried to dynamically enhance
the user query with the user’s preferences by
creating a user profile for providing personalized
results (Micarelli et al., 2007). However, a user
profile may not be sufficient for a variety of queries
of the user. One disadvantage of automatic
personalization techniques is that they are generally
applied out of context. Thus, not all of the user
interests are relevant all of the time, usually only a
288
Asfari O., Doan B., Bourda Y. and Sansonnet J..
A CONTEXT-BASED MODEL FOR WEB QUERY REFORMULATION.
DOI: 10.5220/0003100802880293
In Proceedings of the International Conference on Knowledge Discovery and Information Retrieval (KDIR-2010), pages 288-293
ISBN: 978-989-8425-28-7
Copyright
c
 2010 SCITEPRESS (Science and Technology Publications, Lda.)

subset is active for a given situation, and the rest
cannot be considered as relevant preferences.
User query is an element that specifies an
information need, but queries, especially short one
as mobile user queries, do not provide a complete
specification of the information need. Many relevant
terms can be absent from queries and terms included
may be ambiguous, thus queries must be processed
intelligently to address more of the user’s intended
requirements. Typical solution includes expanding
query representation that refers to methods of query
reformulation, i.e., any kind of transformation
applied to a query to facilitate a more effective
retrieval. This paper presents a method to
reformulate user queries depending on the user
profile, containing his interests, together with the
user context which is considered as the actual state
of the user current task in order to provide
personalized results in context. We combine
linguistic knowledge about query using WordNet
and semantic knowledge using ODP ontology (Open
directory Project www.dmoz.org) and knowledge
about user (user profile and user task context) into a
single framework in order to provide the most
appropriate answer for a user’s information needs in
the search time and task state.
The rest of the paper is organized as follows:
Section 2 shows the related work; Section 3 presents
our models to reformulate user’s queries; section 4
presents the system architecture; Section 5 shows the
experimental study; Finally, Section 6 gives the
conclusion.
2 RELATED WORK
Many studies have been employed to expand the
user query in information retrieval, as far as we
know these studies don’t depend on the user task, in
this paper we depend on a task model for expansion
the user query, thus in section 2.1, we describe
related work where the query expansion had been
investigated. In section 2.2, we review studies where
task model had been used.
2.1 Query Expansion
Query expansion is the process of augmenting the
user’s query with additional terms in order to
improve results by including terms that would lead
to retrieving more relevant documents. Many works
have been done for providing personalized results by
query reformulation. Approaches based on the user
profile for query enrichment have been proposed,
this process consists in integrating elements of the
user profile into the user’s query (Koutrika et al.,
2004). The limitation of these approaches is that
they do not take into consideration the user context
to activate elements from the user profile.
Studies on query reformulation by relevance
feedback are proposed, the aim is to use the initial
query in order to begin the search and then use
information about whether or not the initial results
are relevant to perform a new query (Lv and Zhai,
2009). Because relevance feedback requires the user
to select which documents are relevant, it is quite
common to use negative feedback. Furthermore the
techniques of disambiguation aim to identify
precisely the meaning referred by the terms of the
query and focus on the documents containing the
words quoted in the context defined by the
corresponding meaning. But this disambiguation
may cause the query to move in a direction away
from the user’s intention and augment the query
with terms related to the wrong interpretation.
Many approaches, like (Bhogal et al., 2007), try
to reformulate the web queries based on semantic
knowledge by using ontology in order to extract the
semantic domain of a word and add the related terms
to the initial query, but sometimes these terms are
related to the query only under a particular context.
Others use sense information (WordNet) to expand
the query (Navigli and Velardi, 2003). In this paper,
we propose a hybrid query expansion method that
automatically generates query expansion terms from
the user profile and the user task. In our approach we
exploit both a semantic knowledge (ODP Ontology)
and a linguistic knowledge (WordNet) to learn the
user’s task, and we exploit UML states diagram to
represent the user current task.
2.2 Task Model
One aspect of characterizing user’s contexts is to
consider the tasks which have led them to engage in
information retrieval behavior. Users use documents
to understand a task and solve a specific problem.
Thus, when a user begins a task, he searches the
information that will help solve the problem at hand.
Various researchers have demonstrated that the
desired search results differ according to types of
tasks. According to (Terai et al., 2008) two types of
tasks: Informational task which involves the intent to
acquire some information assumed to be present on
one or more web pages; transactional task which is
based on the intent to perform some web-mediated
activity. The approach (Freund et al., 2005) proves
that the nature of the task has an impact on decisions
of relevance and usefulness. In the approach
A CONTEXT-BASED MODEL FOR WEB QUERY REFORMULATION
289

(Luxenburger et al., 2008) a language model of a
user task is defined as a weighted mixture of task
components: queries, result sets, click stream
documents, and browsed documents. Most existing
systems do not integrate user needs with the
characteristics of the relevant task state as the
execution of the task progresses.
3 MODELS AND ALGORITHMS
Our aim is to provide context-based personalized
results by improving the user web-queries
intelligently. We consider the user current task as a
contextual factor. Here we will describe our models
for detecting the user current task, constructing an
ontological user profile and generating the
reformulated queries.
3.1 General Language Model
We construct here a new general language model for
query expansion including the contextual factors and
user profile in order to estimates the parameters in
the model that is relevant to information retrieval
systems. In the language modeling framework, a
typical score function is defined in KL-divergence as
follows (Bouchard and Nie, 2006):
Score ( Q , D ) =
∑
∈Vt
P ( t | θ
Q
) log P ( t | θ
D
) ∝ − KL ( θ
Q
||
θ
D
)
(1)
Where: θ
D
is a language model created for a
document
D, θ
Q
a language model for the query Q,
generally estimated by relative frequency of
keywords in the query, and V the vocabulary.
P (t|θ
D
): The probability of term t in the document
model, P (t|θ
Q
): The probability of term t in the
query model.
The basic retrieval operation is still limited to
keyword matching, according to a few words in the
query. To improve retrieval effectiveness, it is
important to create a more complete query model
that represents better the information need. In
particular, all the related and presumed words should
be included in the query model. In these cases, we
construct the initial query model containing only the
original terms, and a new model SRQ (state
reformulated queries) containing the added terms.
We generalize this approach and integrate more
models for the query. Let us use θ
Q
0
to denote the
original query model, θ
Q
T
for the task model, θ
Q
S
for
the contextual state model, and θ
Q
U
for a user profile
model. θ
Q
0
can be created by MLE (Maximum
Likelihood Estimation). Given these models, we
create the following final query model by
interpolation:
(2)
Where: X= {
0, T, S, U} is the set of all component
models and
i
a
(with
1=
∑
∈ Xi
i
a
) are their mixture
weights, Thus the (1) becomes:
(3)
Where:
(4)
is the score according to each component model.
3.2 Constructing Task Model
The task model is used to detect and describe the
task performed by the user, when he submits his
query to the information retrieval system. In this
paper we depend on study questionnaires (W. White
and Kelly, 2006) which were used to elicit tasks that
were expected to be of interest to subjects during the
study. A generic classification was devised for all
tasks identified by all subjects, producing the
following nine task groupings:
Academic Research; News and Weather;
Shopping and Selling; Hobbies and Personal
Interests; Jobs/Career/Funding; Entertainment;
Personal Communication; Teaching; Travel.
We generate a UML states diagram for each task
in order to detect the changes in the task-needs over
time and for describing all the sequences of the
performed task. This generated diagram contains the
task states and at least one attribute for each one.
Accordingly, an index is built for: the terms of the
tasks, the terms of its states including the state
attributes, and the related task concepts from ODP.
Thus this index consists of r terms. We will use this
index when using the term vector model.
The user task can be identified automatically by
taking advantages of existing linguistic resources
(WordNet) and semantic resources (ODP Ontology)
for assigning a task to user query. For that, we apply
the following algorithm:
Let q a query submitted by a specific user at the
current task denoted A
*
. This query is composed of
n terms; it can be represented as a term vector:
q = 〈 t
1
, t
2
, ….,t
n
〉
For this query q a current task A
*
is built by a single
term vector:
A
*
= 〈 a
s1
, a
s2
, ….,a
si
〉
Where: a
S1
, a
S2
, …a
Si
the terms that represent the
state attributes of the task states s
1
, s
2
, …s
i
for the
KDIR 2010 - International Conference on Knowledge Discovery and Information Retrieval
290

current task A
*
. For example, if the actual state is
“Find a Restaurant”, then the state attribute will be
“Restaurant” and a value from the user profile (such
as vegetarian) will be assigned to this state attribute
in order to personalize the query.
The initial query q is parsed using WordNet in
order to identify the synonymous terms:
q
w
= 〈 t
w1
, t
w2
, ….,t
wn
〉
The query q
w
is queried against the ODP
ontology in order to extract a set of concepts
(c
1
,c
2
…,c
m
with m≥n) that reflect the semantic
knowledge of the user query. These concepts of the
user query and its sub-concepts are represented as a
single term vector:
C
q
= 〈 c
1
, c
2
, ….,c
m
〉
Then the concepts are compared with the
previous nine tasks, to do this, we compute the
similarity weight between C
q
and the proposed nine
tasks, depending on the task index which is
previously explained:
SW (A
1
) = Cos ( C
q
, A
1
)
SW (A
2
) = Cos (C
q
, A
2
)
SW (A
9
) = Cos ( C
q
, A
9
)
Finally, the task A
*
corresponding with the
maximum similarity weight (Max (SW (A
*
))) is
automatically selected as the current task. Figure 1
illustrates the various vectors.
Figure 1: Representation of the tasks and the query as term
vectors.
Where: query terms: t
1
, t
2
, ….,t
n
. Terms of task
index: t
1
, t
2
, ….,t
r
. Terms of task state attributes: a
S1
,
a
S2
, …,a
Si
. Each term's weight is computed using tf
* idf weighting scheme.
3.3 Contextual State Model
The contextual state model is responsible for
determining and analyzing the actual state of the
current task. We suppose that the different states of
the current task are modeled using an UML state
diagram. There is at least one relevant attribute a
si
for each detected state S
i
. Because mobile device
moves with the user, it is possible to take into
account the actual task state in which the user is in
when submitting certain queries to the information
retrieval system IRS. Such contextual information
may come automatically from various sources such
as the user’s schedule, sensors, entities that interact
with the user; it may also be created by the user.
According to our assumption, we have defined 9
UML state diagrams for the main pre-defined tasks.
After the user's query is submitted to our platform,
the related task is assigned automatically to the user
query and a set of SRQ (State Reformulated
Queries) related to each state is presented to the
user. The user is then asked to choose the
appropriate SRQ according to his state. Finally, the
contextual model will follow the UML state diagram
to present the next SRQ.
3.4 Ontological user Profile Model
Ontology is a formal representation of a set of
concepts within a domain and the relationships
between those concepts. Concepts (or classes or
categories) appear as nodes in the ontology graph. A
user profile is a collection of personal data
associated to a specific user. The Ontological user
profile is constructed by the representation of the
user profile as a graph of related concepts of the
ODP ontology, inferred using an index of user
documents. Here, a dynamic ontological user profile
is considered as semi-structured data in the form of
attribute-value pairs where each pair represents a
profile’s property. The properties are grouped in
categories or concepts using ODP Ontology. In the
proposed ontological user profile the annotating of
each concept is done by giving value for each
attribute in the ontology concept based on an
accumulated similarity with the index of user
documents, a user profile is created consisting of all
concepts with non null value. Using ontology as the
basis of the profile allows the initial user behavior to
be matched with existing concepts in the domain
ontology and relationship between these concepts.
When the ontological user profile is created, its
query-related concepts must be activated. This is
done by mapping the query context C
q
=〈c
1
,c
2
,…,c
m
〉
on this ontological user profile (note that, the query
context is calculated during the construction of the
task model). This allows activating for each query
context concept its semantically related concepts
from the ontological user profile, following our
contextual approach depending on the relevant
propagation (Asfari, 2008). Hence, the relevant user
profile attributes that are determined by the previous
activated concepts are found. This attributes with its
values are used to reformulate the user query.
q
t
1
A
2
t
2
t
r
θ
A
1
A CONTEXT-BASED MODEL FOR WEB QUERY REFORMULATION
291

3.5 SRQ Model (State Reformulated
Queries)
Query expansion is the process of adding relevant
terms to the original query (Asfari et al., 2009).
However, in a more general sense, it also refers to
methods of query reformulation, Thus we look for a
relevant terms to use it in query expansion, that
means we look for terms that are related to the
query, the user, and the task state in the same time
and don’t contain unrelated terms. The initial user
query is reformulated depending on these relevant
terms in order to produce SRQ (State Reformulated
Query) to improve the retrieval performance. The
two aspects for producing SRQ are: query expansion
and query refinement.
Query Expansion: the initial query is expanded with
tow type of generated terms:
- The terms that represent the state attributes,
from UML state diagram, for the current task A
*
(denoted a
S1
, a
S2
, …,a
Si
) One state attribute for
each task state.
- The query-relevant attributes from the
ontological user profile with its values. (<a
u1
,
va
u1
>, < a
u2
, v a
u2
>, …,<a
uj
, v a
u3
>)
Query Refinement: Sometimes irrelevant attributes
may be present in the selected user profile concepts.
In order to keep only the relevant user profile
attributes for the current task state S
i
, we compare
between these generated attributes and the current
state attributes, next we exclude from the generated
user profile attributes these non similar with the state
attributes. We must also exclude the duplicated
terms if they exist in the resulting SRQ. Finally SRQ
is built according to the syntax required by the used
search engine in order to submit the SRQ and to
provide back results to the user.
Let q= {t
1
, t
2
..., t
n
} initial query which is related
to task at hand. The state reformulated query in the
task state S
i
and for a specific user profile P
j
is:
S
i
RQ<Q,P
j
,S
i
> , The relevant results D
i
in the states
S
i
are produced by applying S
i
RQ<Q,P
j
,S
i
> on an
information retrieval system. We expect that the
results D
i
in the task state S
i
are more relevant than
the normal results produced by using the initial
query q in S
i
.
4 SYSTEM ARCHITECTURE
Figure 2 illustrates the system architecture. It
combines the several models described in the
previous section: the task model, the contextual state
model, the ontological user profile model and the
SRQ model.
Query Q:= {t
1
, t
2
, …, t
n
}
WordNet
ODP Ontology
Domain Knowledge
User Profile
{c
1
, c
2
, …, c
m
}
{au
1
, au
2
, ..au
j
} {as
1
, as
2
, ..as
i
}
Sim
SRQ
Results
State1
State2
State I
UML State diagram
Tasks:={A
1
, A
2
, .., A
I
}
Contextual
Application
Sensor
User Task State
Contextual model
for XML retrieval
Figure 2: System architecture.
5 EXPERIMENTAL STUDY
Here we first suppose that the queries we are
considering are related to current task at hand and
secondly, the tasks are modeled by UML state
diagrams. Our system works depending on the
following practical consequent steps:
When the user submits his query in our platform,
the task model will assign a task for this query as the
first step. Next, the UML state diagram for this task
is retrieved. The system then uses the attributes
associated with each state (in UML) and the user
profile attributes for producing the relevant terms,
next the irrelevant terms are excluded, finally, the
reformulated query denoted SRQ is submitted to
Google to retrieve the relevant results. For example,
let q: “Trip to Paris”, the task model assigns the task
“Travel” to this query, and then the contextual state
model allows the proposition of several task states
that are represented in UML state diagram as shown
in the figure 3. Next the system can present the
following SRQ:
- S
1
RQ :{Paris + Airline OR Book ticket OR
Inexpensive}
- S
2
RQ :{Paris + hotel+2 star OR single}
- S
3
RQ :{Paris+ Monuments OR Weather OR
plan OR Metro}
- S
4
RQ :{Paris+ restaurant OR Italian Cuisine OR
Vegetarian Food}
- S
5
RQ:{Paris + Photos}
- S
6
RQ: {Paris+ News + Weather}
The evaluation of such systems is complicated
due to the dynamic aspect of the system
environment. So, we performed two manual
evaluations, one to evaluate the detected task and
KDIR 2010 - International Conference on Knowledge Discovery and Information Retrieval
292

another to evaluate the SRQ (State Reformulated
Queries):
We asked 10 different users to submit 3 queries
(for doing different tasks), the system then detects
the task for each query. Next the users are asked if
the tasks were similar to their tasks or not. We then
got nearly 21 out of 30 positive responses (70%). To
evaluate the SRQ queries we asked the 10 users to
submit different queries and we applied each one to
the Google search engine at the different states of
the task which was proposed by our task model. We
reformulated these queries by adding the relevant
terms and then we reapplied them at the states using
the same search engine. We compared the first 20
retrieval results produced in the two cases (by
queries q and queries SRQ).
Results: we calculated the average number of
relevant pages by queries q and SRQ on the first 20
results (P@20). We noticed that the precision of the
relevant results using the initial query q is 0.17 and
0.59, respectively, by using SRQ queries which were
reformulated depending on the current task state and
user profile.
Conc urrent Substate
Before the Trip
Book the ticket
Hotel Reservation
preparation of the program
Concurrent Substate
After the Trip
news
pictures on the web
Concurrent Substate
During the Trip
find restaurant
Figure 3: Shows an example of a “travel task” that is
modeled by UML state diagram.
6 CONCLUSIONS
In this paper, we have proposed a hybrid method to
reformulate user queries depending on a dynamic
ontological user profile and user context for
producing State Reformulated Queries (SRQ). The
user context is considered as the actual state of the
task that he is undertaking when the information
retrieval process is performed. We have constructed
a general architecture that combines several models
for query expansion: the task model, the contextual
model, the user profile retrieval model and SRQ
model. We exploit both a semantic knowledge (ODP
Ontology) and a linguistic knowledge (WordNet) to
learn user’s task, and we exploit a UML states
diagram for this task to learn user current state. We
have also constructed a new general language model
for query expansion including the contextual factors
and user profile. We have illustrated on an
experimental study that the results obtained by SRQ
queries are more relevant than those obtained with
the initial user queries in the same task state. As a
future work, we plan to evaluate this method by
creating a test collection.
REFERENCES
Asfari, O., Doan, B. L., Bourda, Y. Sansonnet, J. P.:
Personalized access to information by query
reformulation based on the state of the current task and
user profile, In: The Third International Conference
on Advances in Semantic Processing, Malta, 2009.
Asfari, O.: Modèle de recherche contextuelle orientée
contenu pour un corpus de documents XML. In:
CORIA 2008: 377-384, France.
Bouchard, H., Nie, J. Y.: Modèles de langue appliqués à la
recherche d’information contextuelle, In: Conf. en
Recherche d’Information et Applications (CORIA),
Lyon, 2006.
Bhogal, J., Macfarlane, A., Smith, P.: A review of
ontology based query expansion, Information
Processing and Management. In: an International
Journal, v.43 n.4, p.866-886, July, 2007.
Freund, L., E. G. Toms, and Charles, L. A. Clarke:
Modeling task-genre relationships for IR in the
workplace, In: SIGIR 2005, at Salvador, Brazil.
Koutrika, G., Ioannidis, Y. E., Personalization of Queries
in Database Systems, In: Proceedings of 20th
International Conference on Data Engineering, 2004.
Lv, Y., Zhai C.: Adaptive Relevance Feedback in
Information Retrieval, IN: CIKM, 2009, Hong Kong.
Luxenburger, J., Elbassuon, S., Weikum, G.: Task-aware
search personalization, In: Proceedings of the 31st
annual international ACM SIGIR, Singapore, 2008.
Micarelli, A., Gasparetti, F., Sciarrone, F., and Gauch, S.:
Personalized Search on the World Wide Web. P.
Brusilovsky, A. Kobsa, and W. Nejdl (Eds.), In: The
Adaptive Web, LNCS 4321, Berlin, 2007.
Mylonas, Ph., Vallet, D., Castells, P., Fernandez, M., and
Avrithis, Y., Personalized information retrieval based
on context and ontological knowledge. In: Knowledge
Engineering Review, Cambridge University, 2008.
Navigli, R., Velardi, P.: An Analysis of Ontology-based
Query Expansion Strategies, In: Workshop on
Adaptive Text Extraction and Mining at the 14th
Conference on Machine Learning, Croatia 2003.
Sieg, A., Mobasher, B., and Burke, R.: Representing
context in web search with ontological user profiles.
In: Proceedings of the 6th International Conference on
Modeling and Using Context, Denmark, August 2007.
Terai, H., Saito, H., Takaku, M., Egusa, Y., Miwa, M., and
Kando, N.: Differences between informational and
transactional tasks in information seeking on the web.
In: Proceedings of the Second Symposium IIiX, 2008.
W. White, R., Kelly, D.: A Study on the Effects of
Personalization and Task Information on Implicit
Feedback Performance, In: CIKM’06, 2006, USA.
A CONTEXT-BASED MODEL FOR WEB QUERY REFORMULATION
293