=Paper=
{{Paper
|id=Vol-210/paper-4
|storemode=property
|title=Classification-based Situational Reasoning for Task-oriented Mobile Service Recommendation
|pdfUrl=https://ceur-ws.org/Vol-210/paper4.pdf
|volume=Vol-210
|dblpUrl=https://dblp.org/rec/conf/ecai/LutherFSFN0K06
}}
==Classification-based Situational Reasoning for Task-oriented Mobile Service Recommendation==
https://ceur-ws.org/Vol-210/paper4.pdf
Classification-based Situational Reasoning
for Task-oriented Mobile Service Recommendation
Marko Luther1, Yusuke Fukazawa2, Bertrand Souville1, Kunihiro Fujii2
Takefumi Naganuma2, Matthias Wagner1, Shoji Kurakake2
Abstract. We study the case of integrating situational reasoning system features a task knowledge base, which contains semantic de-
into a mobile service recommendation system. Since mobile Inter- scriptions of potential activities and links to corresponding services
net services are rapidly proliferating, finding and using appropriate that may be helpful. Although this system enables effective service
services requires profound service descriptions. As a consequence, retrieval, it behaves passive in requiring a users initial input to trigger
for average mobile users it is nowadays virtually impossible to find the problem solving process.
the most appropriate service among the many offered. To overcome In this paper we propose a proactive extension of our basic system
these difficulties, task navigation systems have been proposed to that suggests tasks and services actively, without the need for initial
guide users towards best-fitting services. Our goal is to improve the user input. This is achieved by the integration of a situation engine
user experience of such task navigation systems by adding context- and a situation-based task filter, meant to expose only those tasks that
awareness (i.e., to optimize service navigation by taking the user’s are relevant for a user in a given situation. Taking the user’s situation
situation into account). In this paper we propose the integration of into account avoids the necessity of an initial task query. This leads
a situational reasoning engine that applies classification-based infer- to a considerable improvement of the navigator’s usability, especially
ence to context elements, gathered from multiple sources and rep- for non-expert users who are often not willing to input queries.
resented using ontologies. The extended task navigator enables the The abstract characterization of a user’s situation is computed by
delivery of situation-aware recommendations in a proactive way. Ini- inference mechanisms on several pieces of context information gath-
tial experiments with the extended system indicate a considerable ered from multiple context sources [20]. We formulate high-level
improvement of the navigator’s usability. qualitative context elements in the Web Ontology Language (OWL)
[22] and concrete situations as instances within the assertional com-
ponent (Abox) of a situation ontology. To profit from sound, com-
1 Introduction plete and high-performance classifiers such as FaCT++ [31], Pel-
let [30] and Racer [12], we restrict ourselves to the OWL DL frag-
Within the growing market for mobile Internet, NTT DoCoMo is to- ment of OWL. To separate concerns we assume that probabilistic as-
day providing services to over 50 million mobile phone subscribers pects of context representation and reasoning are dealt with at lower
in Japan. The majority of these users enjoy widely diverse contents representation levels applying bayesian networks or fuzzy logics.
such as entertainment services (ring-tone downloads, games, etc.), The rest of this paper is organized as follows. After discussing re-
transaction services (money transfer, airline reservation, etc.) and in- lated work in the field of ontology-based context reasoning in the
formation services (weather forecast, maps and local information, next section, we introduce our task-based service navigator appli-
etc.) through DoCoMo’s high-speed 3G mobile network. Already to- cation together with some usage scenarios in Section 3. The overall
day, the number of commercial i-mode sites – DoCoMo’s brand of system architecture that underlies the application is presented in Sec-
mobile Internet services – ranges in the region of many tenth of thou- tion 4 and the details on our approach to context representation and
sand. With 4G networks at the horizon that promise still substantially classification-based reasoning are given in Section 5. In the closing
higher bandwidth for data transmissions, the market for services with section we report on our experiences gained from this development.
rich content is expected to expand further.
Key to support such growth is the availability of intelligent service
platforms that mediate between services and users by observing the 2 Related Work
users’ activity. These platforms have to assist the user in selecting Several projects consider the use of ontologies as a key requirement
the most appropriate service from the fast growing service pool to for building context-aware applications. Closely related to our ap-
support their real world activities, anytime and anywhere. proach is the work done in the CALI project [16] as it explores the
Our previously developed task-based service retrieval system for use of Description Logics (DL) [1] and the associated inferencing. To
the non-expert mobile user makes it easy to retrieve appropriate ser- overcome the limitations of pure DL-based reasoning, a hybrid ap-
vices for tackling the users challenges in managing his or her every- proach is proposed. However, our earlier experiments [24] indicate
day life [25]. The term task refers here to “what the user wants to that the suggested loose coupling of a DL reasoner with an external
do” as an expression of the users current activity. Furthermore, the generic rule engine leads to serve performance problems. To achieve
completeness both reasoners have to be applied successively until no
1 DoCoMo Euro-Labs, Landsbergerstr. 312, 80687 Munich, Germany
{luther,souville,wagner}@docomolab-euro.com
new facts have been derived. Furthermore, it remains unclear how
2 NTT DoCoMo Inc., 3-5 Hikari-no-oka, Yokusuka, Kanagawa, 239-8536 Japan consistency can be guaranteed taking both the knowledge base and
{y-fukazawa,naganuma,kurakake}@netlab.nttdocomo.co.jp the rule base into account.
Felica Reader-Writer is installed near the gate at Tokyo [2] M. Luther et al.: Situational reasoning – a practical OWL use
station (like the mobile Suica system that is currently case. In Proc. of the 7th Int. Symposium on Autonomous
deployed by Sony and NTT DoCoMo for JR East[4]) and Decentralized Systems (ISADS'05), 2005.
that it delivers location information to the mobile phone [3] http://www.nttdocomo.co.jp/english/p_s/i/felica/index.html
via its Felica tag whenever the user puts it close to the [4] http://www.jreast.co.jp/suica/
Reader-Writer device as shown in Fig. 2(b).
The scenario of our demonstration is as follows.
Dawson Campbell, the main character, and his colleague
Fiona Davidson are at Tokyo station one afternoon taking
the train to another facility of their company located
outside the city. At first, Dawson Campbell passes the
gate at Tokyo station as shown in Fig. 2(b). The task-list (a) Gat at Tokyo station (b) Device
Figure 2. Felica Dawson passing the gate
associated with the location concept “Station” appears on
Dawson's cellFigure phone 1. and includesService
Situation-aware the entries "Prepare to
Recommender
ride a train", "Buy souvenirs", "Meet someone at the To detect the user’s location we further assume that the cell phones
station" etc. as shown in Fig. 2(c). While displaying the of Dawson, Fiona and Gordon are equipped with Felica3 contact-less
task-list, Dawson's phone connects to the situational RFID tags, enabling a two-way communication with Sonys Felica
reasoningOtherengine
approaches and suchupdates
as CONON Dawson's location to[4]
[32] and SOUPA/CoBra Reader-Writer devices. Whenever a user puts his phone close to a
solely rely on rule-based reasoning
“Tokyo station“. No task-list is shown on Fiona's which cannot be completecellfor Felica Reader-Writer device (e.g., to make a mobile payment at a
OWL (not
phone at this moment.even for OWL Lite [5]) and easily leads to undecidability, train gate) the recommender application retrieves the corresponding
Fewasseconds
generic rules can be used to simulate role value maps [11].
later, Fiona Davidson passes the same gate location information as a semantic description of this place (cf. Fig-
CONON is an OWL DL encoded upper-context ontology for per- ure 2). Since Sony and NTT DoCoMo just started to deploy their
at Tokyo station (Fig. 2(d)). Fiona's phone connects to the
vasive computing applications defining almost 200 concepts. Rule-
situation mobile Suica4 system for JR East at all stations in the Tokyo region,
reasoning is used toserver
reasoning and uploads
derive high-level Fiona's new
context information and to this assumption is not a fiction but reality.
location
check("Tokyo station").
its consistency. In turn,
To cope with the theobserved
situation reasoner
delay of several
infersseconds
that Dawson Campbell (c) Dawson's
After phone the
having passed displaying the task-list
gate at Tokyo station,suited for station
Dawson’s phone dis-
caused by the reasoning and process,Fiona
complexDavidson
reasoningare tasks plays a basic list of tasks, associated with the concept Station. This
both are
located
computed offline. However, this approach is not feasibleThe
at Tokyo station, traveling together. in our list may include entries such as “Prepare to ride a train”, “Buy sou-
situation reasoning
dynamic setup. engine refers to the situation ontology, venirs”, “Meet someone” etc. While displaying this task-list, Daw-
and then SOUPA, finds
another thatOWL the relation designed
DL ontology between Dawsonap-
for ubiquitous son’s phone connects to the situational reasoning engine and updates
Campbell andis Fiona
plications, about theDavidson
same size as is thecolleague. Dawson's
CONON ontology. Its ex- his location to Tokyo station.
situation
tensionis CoBra-Ont
reasonedisbased used byon time ("afternoon"),
a context broker architectureplaceto real- Before having passed the gate, no tasks are shown on Fiona’s
("station") and relation
ize a scenario where people ("colleague").
on a universityIncampus this come
case,together
the (d) Fiona
phone. Oncepassing the gate
her location has been detected, a connection to the rea-
reasoned situation
for a meeting. To limitbecomes
the reasoning BUSINESS
overhead causedand this
by importing soning engine is established and her current location is updated.
judgment
standardis ontologies,
then passed to concepts
single both Dawson's
are mapped and to Fiona's cell
foreign ontology As a result, the situation reasoner infers that Dawson Campbell
phone and service navigation server. Both Dawson andcor-
terms. Still, the SOUPA ontology is of a rather high-complexity and Fiona Davidson are travelling together, based on their proximity
responding
Fiona's cell phone SHOIF(D),shows the because it contains
reasoned nominals.
results as shown in at the station. In addition, a lookup in the knowledge base reveals
An interesting approach to speed up the rule-based inferencing
Fig.2(e). Service Navigation server acquires the task-list that Dawson and Fiona are colleagues and that the scene takes place
that oniscomplex ontologies is to determine relevant contexts required to
determined from both reasoned situation at a weekdays afternoon.
answer queries using the query-tree method [17]. It remains to be Because Dawson is located at a public place during office hours
("Business") and the place ("station"), and then sends the
seen how this method extends to our classification-based approach.
acquired task-list to both Dawson's and Fiona's cell phone together with colleagues, his situation is classified as a business sit-
uation. His phone shows the inferred situation together with a cor-
(Fig.2(f)).
responding list of filtered tasks (shown on the left part of Figure 1).
The3 second demo scenario
Situation-aware ServiceisRecommendation
as follows. Dawson
To further specify his needs, Dawson may select one of the recom-
Campbell and his father in law Mark Buchanan are at mended tasks (“go to destination” in this case) and finally invoke an
Tokyo Westation
build on during an afternoon
a task-oriented to go system
service navigation somewhere [25] thatbysup- associated service (as shown on the right part of Figure 1).
train.ports
In this case,inthe
the user inferred
finding situation
appropriate is "Private",
services by querying and a rich (e) Both
Let phones
us assume displaying
another the inferred
situation situation
taking place at theBUSINESS
same location.
corresponding
task ontologytask-lists
that represents appearcommon on sense
both knowledge
Dawson's andtyp-
about
Mark'sicalcell phone.
complex tasks. Situation 2: Private Meeting at Tokyo Station
The key pointofofthisthese
The usage basic scenarios
task navigator isisthat the delivered
as follows. After having Dawson Campbell arrives on a Saturday around noon at the Tokyo
task-lists
specifiedarea task-oriented
tailored toquery the such
different
as “go touserthemesituations,
park” a list of main station where Mark Buchanan, his father in law, is awaiting
"Business"
tasks thatormatch
"Private", even
this query if both
is sent to theplace
mobileand timeNow
device. arethe him. They plan to shop for a birthday present for Dawson’s wife.
the same,
user canstation
select in thethis
mostcase.
appropriate task and a corresponding de-
Demo requirements:
tailed task-model is displayed LAN access point,
accordingly. In a either wireless
final step, associated
This situation is classified as private family meeting, because it takes
or wired LAN
services canis beOK.invoked by establishing an Internet connection to the
place during leisure hours and only relatives are in the proximity.
actual i-mode services.
In this case, the situation-aware recommender application suggests
Figure 1 shows the user interface of the situation-aware variant
References
of the basic service recommender. To explain its functionality, let us
tasks that are related to private activities such as “go to movie the-
ater”, “go shopping”, etc.
[1] T.assume
Naganuma and S.
the following Kurakake: Task Knowledge Based
situation.
The key statement of these scenarios is that task-lists are actually
Retrieval for Service Relevant to Mobile User's Activity, In (f) Both phones displaying the task-list
- associated with the
tailored to different situations of the user, even if some context con-
Proc. Situation
of the 4th Int. Semantic
1: Important Web
Business Conference
Meeting (ISWC’05),
at Tokyo Station situation BUSINESS
ditions are the same (location in this case). In this respect, our system
Y.Gil etTwo
al. (Eds.), LNCS
travellers, 3729,
Dawson pp.959-973,
Campbell and his2005.
boss Fiona Davidson,
Fig.2
facilitates users to access Demoservices
the mobile Sequence
that fit best to their cur-
arrive on a Friday morning at the Tokyo main station. Gordon rent situation, purely based on qualitative context information.
Green, a project partner, is already waiting for them at the plat-
form. The group is looking for a quick transfer to the airport. 3
4
2
Situation Engine Task Navigator We refer to an ontology as a logical theory accounting for the in-
classify
Context Sensor Data
tended meaning of a formal vocabulary, i.e. its ontological commit-
- location (address, place)
- attendee
ment to a particular conceptualization. Therefore, the decidability of
the selected ontology language is crucial. The OWL DL fragment of
Situation Context Situation Task the OWL fulfills this requirement, is highly expressive and has the
Ontology Reasoner classification Ontology
result potential to become the standard ontology language for the Seman-
tic Web. Its selection as the ontology language of choice resulted in
Context Situation-based
Enrichment Task Filter
the construction of high-quality ontologies (i.e., ontologies that are
proven consistent by fully automatic inference engines that are avail-
"raining"
"afternoon"
able for OWL DL). It is important to note that we do not propose
"night" "colleague" Task List
the ontologies described hereafter as the main representation format
Environmental Social Qualitative for all aspects of context modeling, as ontologies are limited to the
Data Relationships Time
formulation of qualitative aspects and the available inference engines
Context Management
are generally weak in handing large amounts of data efficiently.
The context ontologies are composed of eight interrelated compo-
Figure 3. Architecture
nents defining more than 300 concepts, 200 properties and 300 indi-
viduals. They provide a general vocabulary for temporal and spatial
concepts, agents as well as devices. Being informed by the vCard
4 Architecture standard, the iCalendar representation and the FOAF (Friend-of-a-
friend) format, an extension for the precise modeling of complex so-
Figure 3 depicts the overall system architecture. The implementation cial relations has been developed. All component ontologies are inte-
contains two main parts, the situation engine and the task navigator. grated by a situation ontology that defines a top-level concept named
The situation engine receives context information that has been Situation (cf. Figure 4). This concept is refined by concepts such as
collected by the task navigator on the mobile device. Furthermore, Private and Business by referring to concepts and relations defined
this information is enriched by context artifacts, such as environ- in the component ontologies.
mental data, social relations between companions and a qualitative We exemplarily sketch the OWL definitions of two typical situa-
representation of time, all gathered form a distributed network of tions using standard DL syntax [1]. A person’s situation is classified
context providers. Thereupon, an axiomatized situation instance is as Business, if he is either located at a business place (such as an
constructed and sent to the inference engine. According to the world office) or at a public place (e.g., a train station) during office hours.
knowledge encoded in the situation ontology, this instance is clas-
sified and the inferred situation is propagated back to the task nav- Business := Situation u (∃ location . Business place t
igator. A subcomponent of the task navigator, the task filter, detects (∃ location . Public place u ∃ time . Office hour))
the most appropriate task nodes within the task ontology by match-
ing the derived situation with the task-specific categories. Finally,
a representation of the resulting task list is constructed by the task A person is participating a family meeting if he or she is in a private
navigator and presented to the user on his mobile device for further meeting situation where all participants are relatives.
navigation and service selections.
The task ontology stores descriptions for abstract as well as con- Family meeting := Situation u (∀ company . Relative)
crete tasks and their interrelations as semantic descriptions. Large
and abstract tasks are thereby described by sequences of smaller sub- Situational reasoning is realized using a DL reasoning engine that
tasks. In addition, abstract tasks are annotated with enabling context classifies concrete individual situations w.r.t. the ontology. Let us
conditions and concrete tasks are linked to appropriate information consider the Situation 1 introduced in Section 3. First, each piece
services via Uniform Resource Identifiers. The task structures are of context information such as the location (Tokyo station), the time
defined in terms of the process model of the OWL-S ontology [21]. (Sunday morning), and all companions (Dawson’s boss Fiona and
Each task node is represented as a service class and categorized ac- his project partner Gordon) are represented in terms of vocabulary
cording to the high-level context concepts such as Business meeting, formalized by the context ontologies. This requires the mapping of
defined within the situation ontology. The context conditions describ- sensed quantitative data to qualitative representations (e.g. a time-
ing the applicability of a task node are thereby encoded as corre- stamp is mapped to an individual in the Abox representing a Fri-
sponding OWL-S service profiles. More details about our task ontol- day morning). The qualitative representations are enriched by the
ogy can be found elsewhere [26]. world-knowledge formalized in the component ontologies and are
combined to an Abox individual in the situation ontology.
Computed by the reasoning engine, the direct concept type for the
5 Context Representation and Classification situation instance according to Scenario 1 is Important meeting. In
We adopted the IST MobiLife5 Context Management Framework [7] this case, the location of the scene is a public place (as tokyo station
to achieve interoperability between context sources from diverse do- is an instance of the concept Station, which in turn is a subconcept of
mains by defining an XML-based context meta model. The elements Public place) during office hours (as the individual friday morning
of this meta model are linked to ontologies that define the basic con- is classified as Office hours) and the main actor Dawson is accompa-
textual categories, used to represent qualitative aspects of context in- nied by his supervisor and a business partner. Similarly, the situation
formation. instance constructed for Scenario 2 is classified as Family meeting
as it takes place at a public location during leisure time and only
5 http:\\www.ist-mobilife.org relatives are detected in the proximity of Dawson.
3
Private_place ⊔
(Public_place ⊓ Leisure_time)
Private ⊓ Meeting ⊓ Private_meeting ⊓
Private company (Relative ⊔ Friend) company Relative
Private_meeting Family_meeting
Situation Meeting
company ≥ 1
Business_meeting Important_meeting
Business Business ⊓ Meeting ⊓ Business_meeting ⊓
Business_place ⊔
company (Colleague ⊔ Business_partner) company Supervisor
(Public_place ⊓ Office_hour)
Figure 4. Situation Ontology Fragment
The situational reasoning process described above is supported by syntactic as well as the semantic level is necessary for referencing
deductions in all component ontologies. For example, the agent on- entities in another ontology without inheriting all of its complexity.
tology specifies in detail the semantics of social relations between Furthermore, our modeling of context ontologies would benefit from
people. Based on the knowledge encoded within the ontology, it can additional constructs such as qualified cardinality restrictions and a
be inferred that two persons (like Dawson and Fiona) are colleagues, richer object property structure that would allow the specification
taking into account the transitivity of this relationship in case they of reflexive, irreflexive, symmetric and anti-symmetric properties as
have a common colleague. Similarly, even if no direct relation be- well as property chains and disjoint property axioms. Reasoning sup-
tween Dawson and Mark is specified it can be inferred that Mark is port for the DL-safe fragment [23] of SWRL [14] and for concrete
Dawson’s father in law (defined to be the father of the spouse of a domains on user defined datatypes would allow us to further enhance
person), because Dawson’s wife Madeleine is known to be the child the quality of our situation engine. While concrete domain reasoning
of Mark. In this case, the subproperty and inverse property specifica- and support for SWRL is already available in some inference en-
tions within the agent ontology enable this logical inference: wife is gines, and most of the requested additional language constructs are
defined as a subproperty of spouse and father is the inverse of child. part of the OWL 1.1 draft6 created by the ad-hoc OWL community,
an improved import mechanisms as given by the E-connection mech-
anism [10] and implemented in Pellet is not included.
6 Discussion At first, we experimented with the DIG interface to realize the
We integrated a situational reasoning engine into a real-world mo- communication between our application and the inference engine.
bile service application. Our classification-based approach relies on However, DIG 1.1 does not support the removal of specific axioms
ontology technology for the representation and reasoning on context making it necessary to re-submit the complete ontology for each re-
information. As the scalable management of data is not a core fea- quest to our situation engine. This is especially awkward for our ap-
ture of pure ontology-based context management and typical context plication where only a very small part of the assertional knowledge
models are usually rather large, we restricted its scope to high-level changes between two requests. As active members of the informal
qualitative context elements. Lower-level context information is rep- DIG 2.0 working group7 we therefore propose a modular extension
resented according to an XML-based meta model and managed sepa- to the interface that supports incremental reasoning and retraction.
rately. The arising reasoning problems are answered by a Description Unfortunately, current reasoner typically only provide some kind of
Logic (DL) [1] inference engine that provides complete reasoning batch-oriented reasoning procedure. A notable exception is Racer
support for the decidable fragment of OWL. which offers low-level retraction support for most of its statements.
The use of the standard representation language OWL and the Still, because of the lack of algorithms for appropriately handling in-
standardized reasoner interface DIG [2] (a stateless HTTP-based pro- cremental additions as well as retractions, Racer initiates a complete
tocol with XML syntax) enabled us to directly compare the influence reclassification after each change in the ontology. Initial empirical re-
of different context ontologies and reasoners on the overall system sults, performed with an experimental version of Pellet, indicate that
performance. We observed that the inference technology as imple- incremental classification algorithms for SHOIN (D) can be quite
mented in modern DL reasoners made significant progress during effective [28].
the last years. Novel optimization techniques enabled a tremendous The ability to handle simultaneous requests is one of the key re-
increase in performance, and also the coverage was greatly extended. quirements in our dynamic mobile setting. However, current infer-
By now most systems can be accessed via DIG, and support nomi- ence engines do not implement any transaction management. Only
nals as well as Abox reasoning directly. FaCT++ and Pellet support for Racer, support for dispatching, load balancing and caching of
SHOIQ(D) (OWL DL extended by qualified cardinality restric- OWL-QL [6] queries is available via the RacerManager [8]. As
tions) and RacerPro supports SHIQ including approximated nomi- OWL-QL does not support modifications of an ontology, we had to
nals and reasoning with concrete domains. implement our own transaction management system that enables the
Nevertheless we observed several limitations in the available tech- sharing of reasoning resources between requests, but avoids the ne-
nology (see [18] for details). The import mechanism of OWL, which cessity to maintain a separate knowledge base for each user.
brings all triples into the importing ontology, has a limited use for the 6 hhttp://www-db.research.bell-labs.com/user/pfps/owli
sharing and reuse of ontologies. An appropriate mechanism on the 7 hhttp://homepages.cs.manchester.ac.uk/∼seanb/digi
4
It has been observed before [17][32] that the delay caused by REFERENCES
ontology-based inferencing easily becomes a major obstacle for re-
[1] F. Baader et al., The Description Logic Handbook, Cambridge Univer-
alistic applications. This is especially problematic for ontologies that sity Press, Cambridge, January 2003.
constantly change, because well-established optimization techniques [2] S. Bechhofer, R. Möller, and P. Crowther, ‘The DIG Description Logic
such as tabling (used in various rule-based inference engine) cannot Interface’, in Proc. of the Int. Workshop on Description Logics, (2003).
be applied. As a consequence of the high worst-case complexity of [3] J. Carrol et al., ‘Jena: Implementing the Semantic Web recommenda-
tions’, Technical Report HPL-2004-146, HP Labs, (2004).
expressive DLs, such as SHOIN (D) underlying OWL DL, mod- [4] H. Chen, T. Finin, and A. Joshi, ‘The SOUPA ontology for pervasive
ern DL reasoners implement a suite of optimization techniques to computing’, in Ontologies for Agents, eds., V. Tamma, S. Cranefield,
achieve acceptable performance. The efficiency of implementations and T. Finin, Springer, (July 2005).
on concrete cases depends therefore on the applicability of optimiza- [5] K. de Bruin and D. Fensel, ‘Owl-’, WSML Deliverable D20.1, (2005).
[6] R. Fikes et al., ‘OWL-QL’, Journal of Web Semantics, (2005).
tions, which varies with the language features in use. For example,
[7] P. Floréen, M. Przybilski, P. Nurmi, J. Koolwaaij, A. Tarlano, M. Wag-
the use of domain and range restrictions can lead to cycles in a Tbox ner, and M. Luther et al., ‘Towards a context management framework
for which termination of the tableaux algorithm can only be ensured for MobiLife’, in Proc. of the IST Summit, (June 2005).
by blocking. However, known blocking strategies for SHOIN are [8] J. Galinski et al., ‘Development of a server to support the formal Se-
less effective if inverse roles are involved. On the other hand, if nom- mantic Web query language OWL-QL’, In Horrocks et al. [15].
[9] Y. Gil, E. Motta, R. Benjamins, and M. Musen, eds. Proc. of the 4th
inals do not occur in an ontology blocking can be realized more effi- Int. Semantic Web Conference, volume 3729 of LNCS. Springer, 2005.
ciently [13]. Therefore we avoid the use of standard ontologies, such [10] B. Grau, B. Parsia, and E. Sirin, ‘Combining OWL ontologies using
as the SHOIF(D) entry sub-ontology of time [27]. It has to be seen E-connections’, Journal of Web Semantics, 4(1), (2005).
how the recently suggested techniques for optimizing DL reasoning [11] B. Grosof, I. Horrocks, R. Volz, and S. Decker, ‘Combining logic pro-
grams with Description Logic’, in Proc. of the Int. WWW Conf., (2003).
in the presence of nominals [29] perform in practice.
[12] V. Haarslev and R. Möller, ‘Racer: A core inference engine for the Se-
We optimized our initial ontologies by removing nominals and mantic Web Ontology Language (OWL)’, in Proc. of the 2nd Int. Work-
most of the domain and range restrictions. Furthermore, we reduced shop on Evaluation of Ontology-based Tools, pp. 27–36, (2003).
the number of loaded axioms and objects (especially Abox individ- [13] V. Haarslev, R. Möller, and M. Wessel, ‘Description Logic inference
uals) and axioms by splitting the ontology in small components and technology: Lessions learned in the trenches’, In Horrocks et al. [15].
[14] I. Horrocks and P Patel-Schneider, ‘A proposal for an OWL rules lan-
by separating ontologies in A- and Tboxes to cope with the limits of guage’, in Proc. of the Int. WWW Conf., pp. 723–731. ACM, (2004).
the OWL import statement. This step resulted in a performance gain [15] I. Horrocks et al., ed. Int. Workshop on Description Logics, July 2005.
of up to 1,5 seconds per request. Furthermore, we compared differ- [16] D. Khushraj and O. Lassila, ‘CALI: context awareness via logical in-
ent retraction strategies using Racer. The simplest form of retraction ference’, in Proc. of the Workshop on Semantic Web Technology for
Mobile and Ubiquitous Applications, (November 2004).
is reloading of ontologies and can be accelerated by either loading
[17] J. Lee, I. Park, and S. Hyun, ‘Application-oriented context pre-fetch
from a pre-classified image or by cloning an ontology in memory. method for improving inference performance in ontology-based context
For small Aboxes cloning outperformed true retraction realized with management’, in Workshop on Contexts and Ontologies, (2005).
forgot statements. However, the strategy performed best was to keep [18] Th. Liebig, M. Luther, O. Noppens, M. Paolucci, M. Wagner, and F..
situation individuals up to a certain number (about 20 in our case) in von Henke, ‘Building applications and tools for OWL’, in Proc. of the
OWL Experiences and Directions WS, (November 2005).
the Abox before cloning a fresh pre-loaded Abox. Of course, keep- [19] M. Luther, S. Böhm, M. Wagner, and J. Koolwaaij, ‘Enhanced presence
ing individuals and axioms in the Abox is only possible if they do tracking for mobile applications’, In Gil et al. [9].
not influence later classifications. [20] M. Luther, B. Mrohs, M. Wagner, S. Steglich, and W. Kellerer, ‘Situ-
The time to compute our comparable simple reasoning problems ational reasoning – a practical OWL use case’, in Proc. of the 7th Int.
Symp. on Autonomous Decentralized Systems, (April 2005).
is dominated by the communication overhead caused by the reasoner
[21] D. Martin et al., ‘OWL-S: Semantic Markup for Web Services’, W3C
interface. Accessing Racer via its native API using TCP is about 1,5 Member Submission, The OWL Serivces Coalition, (November 2004).
times faster then the access via HTTP/DIG and even 2 times faster [22] D. McGuinness and F. van Harmelen, ‘OWL Web Ontology Language
then the access realized with the triple-oriented framework Jena2 [3]. overview’, W3C Recommendation, (February 2004).
Naturally, we achieved the best performance by using the Pellet rea- [23] B. Motik, U. Sattler, and R. Struder, ‘Query answering for OWL-DL
with rules’, in Proc. of the Int. Semantic Web Conf., (2004).
soner running in the same Java virtual machine and this way com- [24] B. Mrohs, M. Luther, R. Vaidya, and M. Wagner et al., ‘OWL-SF – a
pletely avoiding any external communication. distributed semantic service framework’, in Proc. of the Workshop on
Because existing performance results of DL reasoners are often Context Awareness for Proactive Systems, pp. 67–77, (June 2005).
limited to static Tbox classification, we plan to perform a detailed [25] T. Naganuma and S. Kurakake, ‘Task knowledge based retrieval for
service relevant to mobile user activity’, In Gil et al. [9], pp. 959–973.
analyze of the influence of different retraction strategies for dynamic
[26] T. Naganuma and M. Luther et al., ‘Task-oriented mobile service rec-
assertional reasoning, to compare the performance of interfaces and ommendation enhanced by a situational reasoning engine’, in Proc. of
to test the effect of the ontology size and complexity on realistic rea- the 1st Asian Semantic Web Conference (ASWC’06), (2006). To Appear.
soning tasks. By that we hope to gain inside on how to further opti- [27] F. Pan and J. Hobbs, ‘Time in OWL-S’, in Proceedings of the AAAI
mize our situation engine. Spring Symposium on Semantic Web Services, (2004).
[28] B. Parsia et al., ‘Towards incremental reasoning through updates in
Our current prototype has only a limited support for automatic OWL-DL’, in Reasoning on the Web WS, (2006). To Appear.
context acquisition. We plan to advance the prototype towards the [29] E. Sirin, B. C. Grau, and B. Parsia, ‘From wine to water: Optimizing
use of more actual context information from the real world. Planed Description Logic reasoning for nominals’, in Int. Conf. on the Princi-
extensions will combine GPS-based location information with the ples of Knowledge Representation and Reasoning, (2006). To Appear.
[30] E. Sirin and B. Parsia, ‘Pellet: An OWL DL reasoner’, in Proc. of the
RFID-based context tags we use currently for location tracking, as
Int. Workshop on Description Logics, pp. 212–213, (2004).
well as or short distance wireless communication technologies such [31] D. Tsarkov and I. Horrocks, ‘Ordering heuristics for Description Logic
as Bluetooth to detect people in proximity [19]. reasoning’, in Proc. of the 19th Int. Conf. on AI (IJCAI’05), (2005).
[32] X. Wang, Q. Zhang, T. Gu, and H. K. Pung, ‘Ontology-based context
modeling and reasoning using OWL’, in Proc. of the Workshop on Con-
text Modeling and Reasoning (PerCom’04), pp. 18–22, (March 2004).
5