Intelligent Systems for Tourism - Geographisches Institut › md › chemgeo › geog › gis › ... · 2017-05-23 · learning new things, ... importance of an interdisciplinary

T r e n d s & C o n t r o v e r s i e s

Intelligent Systems for Tourism

“Is E-Commerce Dead, Past Its Prime,or Just Resting?”

This is the title of a recent call forpapers for the Journal of IT Theory andApplication. The special issue aims todiscuss the problems of unfulfilled busi-ness forecasts and related stock marketplunges. However, despite these prob-lems, online transactions are increasing—at least in some sectors, such as the traveland tourism industry. In fact, tourism isthe leading application in the B2C (busi-ness-to-consumer) arena. Although theslow economy and current political devel-opments have negatively influenced e-commerce, it is still flourishing in thetourism sector.

In the first quarter of 2002, travel andtourism accounted for a total turnover ofapproximately US$7 billion—an increaseof 87 percent from the same quarter in2001 (see www.comscore.com/news/online_travel_q1_041602.htm). Consumersmade 46.7 million hotel reservationsworldwide in 2001, netting US$12.9 bil-lion in revenue, according to the HotelElectronic Distribution Network Associa-tion (see www.hedna.org). In addition,32 percent of US travelers this year haveused the Internet to book travel arrange-ments (see www.nua.com/surveys/index.cgi?f=VS&art_id=905357908&rel=true),and further growth is expected.

Tourism industry featuresThe tourism industry has specific fea-

tures that explain its importance for eco-nomic (regional) development and itsinclination toward IT systems.1

Tourism is a leading industry world-wide, representing approximately 11percent of the worldwide GDP (accord-ing to the World Travel & Tourism Coun-cil’s tourism satellite account method1).There will be approximately one billioninternational arrivals in the year 2010(according to the World Tourism Orga-nization, www.world-tourism.org). Fur-thermore, tourism represents a cross-sector (umbrella) industry, includingmany related economic sectors such asculture, sports, and agriculture, whereover 30 different industrial componentshave been identified that serve travelers.In addition, tourism greatly influencesregional development, owing to its SME(small- and medium-sized enterprises)structure and relatively small entrancebarriers. For example, in the EuropeanUnion, the hotel and restaurant sectoraccounts for more than 1.3 million enter-prises. This is approximately 8.5 percentof the total number of enterprises, and95.5 percent of these enterprises aresmall (with one to nine employees).2

Also, because tourism is based onmobility, the supply and demand sideforms a worldwide network, where pro-duction and distribution are based oncooperation. In addition, it is an informa-tion-based industry, so the tourism prod-uct is a confidence good, where at themoment of decision-making, only infor-mation about the product—not theproduct itself—is available.

The problem with these statistics is thatthey refer to different meanings andvarying definitions of e-business and e-

commerce. Some definitions distinguishbetween the two, while others viewthem as the same, and all have their ownvariables and measurement methods.3

Even more problematic is that the defini-tions are all transaction- and business-oriented. They ignore that the Web isalso a medium for creating communities,learning new things, and having fun—things that don’t always result in busi-ness. The Web also encourages user inter-action; users can build their own sites toshare their travel experiences. Thus,another traveler—rather than hotel man-agement or a travel agency—might pro-vide the most valuable information abouta vacation resort.

So, where does AI come into play?Whereas other industries have a

stronger hold on doing things tradition-ally, the travel and tourism industry hasalways been open to new technologies.For example, back in the 1960s, airlinecentralized reservation systems wereamong the first worldwide computernetworks.

In addition, companies traditionallyoutside the tourism field are enteringthe sector, mainly from IT and media sec-tors. Industry features (mainly that ITand media are information-based busi-nesses and are umbrella industries)might explain this trend, or even thechange in consumer behavior. For exam-ple, consumers use IT not only for infor-mation gathering but also to order ser-vices over the Internet. A new type ofuser is emerging who doesn’t just tryone or two services but all kinds of traveland leisure services. Such users don’tmind becoming their own travel agents,but given the extensive use of distri-buted systems on the Internet, therecomes the urgent need to find, combine,

Steffen Staab, University of Karlsruhe,[email protected] Werthner, ITC-irst Research Cen-ter, [email protected].

NOVEMBER/DECEMBER 2002 1094-7167/02/$17.00 © 2002 IEEE 53

54 computer.org/intelligent IEEE INTELLIGENT SYSTEMS

and sift through the right pieces ofinformation intelligently.

Today, AI-based developments in the field are at the forefront, such asindividualized pricing (priceline.com),reversed multiattribute auctioning(mytraveldream.com), recommenda-tions in bundling products (as describedlater), Semantic Web applications (Har-monise.org), and mobile applications(described later). In fact, IT develop-ments and research have induced muchchange in this industry. We can expectthis innovation to continue—at boththe business level (such as dynamic marketstructures and prices) and the technol-ogy level. In addition, the IT and tourismfield represents the nucleus of a newindustry that will produce new products,skills, and jobs.

Challenges for intelligentsystems in tourism

The industry’s dominant features areits heterogeneous and worldwidedistributed nature and its strong SMEbase (especially SMEs in tourist destina-tions). Another inherent characteristic ismobility, where the entire tourist lifecycle is integrated with the respectivesupplier processes (see Figure A).

Obviously, suppliers’ processes crosscompany borders, leading to enhancedB2BC (business-to-business and business-to-consumer) applications, enforcingcooperation between companies, andsupporting mobile communication withthe consumer. Given such a framework,future systems should

• Be heterogeneous, distributed, andcooperative

• Enable full autonomy of the respectiveparticipants

• Support the entire consumer life cycleand all business phases

• Allow dynamic network configurations• Provide intelligence for customers and

suppliers (interfaces and tools) as wellas in the network (which would leadto a set of different services)

• Be scalable and open (with respect togeographical and functional extension)

• Focus on mobile communication (andthe notion of ambient intelligence),enabling multichannel distribution

Ambient intelligence, in which the sur-roundings become the interface, is at the

focus of European research (see www.cordis.lu). We can define it as the conver-gence of ubiquitous computing and com-munication and intelligent user-friendlyinterfaces. Such systems should beembedded, personalized, adaptive, andanticipatory, and they should provideaccess for everybody, anywhere, at anytime. Whereas today the dominant modeof interaction is lean-forward (that is,tense and concentrated interaction withthe computer), it will become laid-back(relaxed and enjoyable). People shouldenjoy computer interaction for travelplanning, and technology should moveto the background.

Opportunities for researchThe e-tourism market’s dynamics and

the requirements of future systemsemphasize e-tourism’s importance andraise several technical research issues:

• (Semantic) interoperability and medi-ated architectures (where we can dis-tinguish between system integrationand the semantics issue, related to theproblem of too much information)

• E-business frameworks supportingprocesses across organizations—virtual organizations

• Mobility and embedded intelligence• Natural multilingual interfaces (also

novel interface technologies)• Personalization and context-based

services• Information-to-knowledge transfor-

mations—data mining and knowl-edge management

The tourism domain is also an excellentexample of the trend toward person-alized services and a complex marketmechanism. It reflects users becoming apart of product creation. So, researchersmust also study nontechnical issuesrelated to markets and users, such as

• Dynamic market and network structures • Pricing and market design • Design and experimenting business

models • User decision modeling and usage

analysis

These research issues underline theimportance of an interdisciplinaryapproach. Many different disciplinesshould contribute, including computerscience, management science, economics,law, statistics, sociology, and psychology.

In this issueThe following essays tackle some of

these thorny issues and stress the need toprovide a holistic approach. For instance,recommender systems that simply pro-pose itineraries without consideringthe trust a user must invest will fail.Francesco Ricci elaborates, arguing thatrecommender systems should not only fil-ter information but also offer completelynew suggestions. Analogously, AlexanderZipf discusses how location-based servicesfor tourists must go beyond adding ageographical parameter to some data-base query. They must consider the per-sonal context to provide an adequatelocation-based adaptation to maps, itin-eraries, or products. Furthermore, sup-porting the user requires providing aglobally coherent view of tourism services(Ulrike Gretzel and Daniel R. Fesenmaierprovide a general analysis of this topic,and Cécile Paris offers a concrete pro-posal). Finally, Craig Knoblock points outthat users want not only advanced plan-ning but also a system that will take careof their itineraries before, during, andafter their travels.

—Steffen Staab and Hannes Werthner

References

1. H. Werthner and S. Klein, InformationTechnology and Tourism—A ChallengingRelationship, Springer-Verlag, New York,1999, p. 323.

Beforetrip

Onsite

Service delivery

Aftertrip

Planning

Marketing

Sales

Monitoring

Relationshipcommunity

Consumerlife cycle

Supplierprocesses

Figure A. The tourist life cycle and suppliers’ processes.

Travel Recommender Systems Francesco Ricci, eCommerce and TourismResearch Laboratory

Recommender systems are commonlydefined as applications that e-commercesites exploit to suggest products and provideconsumers with information to facilitatetheir decision-making processes.1 Theyimplicitly assume that we can map userneeds and constraints, through appropriaterecommendation algorithms, and convertthem into product selections using knowl-edge compiled into the intelligent recom-mender. Knowledge is extracted fromeither domain experts (content- or knowl-edge-based approaches) or extensive logsof previous purchases (collaborative-basedapproaches). Furthermore, the interactionprocess, which turns needs into products, ispresented to the user with a rationale that

depends on the underlying recommenda-tion technology and algorithms. For exam-ple, if the system funnels the behavior ofother users in the recommendation, itexplicitly shows reviews of the selectedproducts or quotes from a similar user.

Recommender systems are now a popularresearch area2 and are increasingly used bye-commerce sites.1 For travel and tourism,3

the two most successful recommender system technologies (see Figure 1) areTriplehop’s TripMatcher (used by www.ski-europe.com, among others) and Vaca-tionCoach’s expert advice platform, Me-Print (used by travelocity.com).

Both of these recommender systems try tomimic the interactivity observed in traditionalcounselling sessions with travel agents whenusers search for advice on a possible holidaydestination. From a technical viewpoint, theyprimarily use a content-based approach, in

which the user expresses needs, benefits,and constraints using the offered language(attributes). The system then matches theuser preferences with items in a catalog ofdestinations (described with the same lang-uage). VacationCoach exploits user profilingby explicitly asking the user to classify him-self or herself in one profile (for example, asa “culture creature,” “beach bum,” or “trailtrekker”), which induces implicit needs thatthe user doesn’t provide. The user can eveninput precise profile information by complet-ing the appropriate form.

TripleHop’s matching engine uses amore sophisticated approach to reduceuser input. It guesses importance of attri-butes that the user does not explicitly men-tion. It then combines statistics on past userqueries with a prediction computed as aweighted average of importance assignedby similar users.4

NOVEMBER/DECEMBER 2002 computer.org/intelligent 55

2. H. Werthner et al., “Information SocietyTechnologies for Tourism,” Report of the Strategic Advisory Group on the 5thFramework Program on Information Soci-ety Applications for Transport and Associ-ated Services, 1997.

3. H. Werthner, “Just Business—Shouldn’tWe Have Some Fun?” Proc. 3rd Int’l Conf.Electronic Commerce & Web Technologies

(Proc. ECWEB-DEXA), Springer-Verlag,New York, 2001.

Hannes Werthner is head of the eCommerceand Tourism Research Lab at the IRST ResearchCenter, a professor at the University of Trento,and founder of the eCommerce CompetenceCenter in Vienna. His research activities coverdecision support systems, simulation, artificialintelligence, and Internet-based information

systems, especially in the field of tourism. Heearned an MS and PhD in computer sciencefrom the Technical University Vienna. He is amember of the strategic advisory board forthe European research program IST, acts as theeditor in chief of the journal InformationTechnology and Tourism, and is HonoraryPresident of the International Federation forIT and Travel/Tourism (IFITT). Contact him [email protected].

Figure 1. (a) Ski-Europe and (b) Travelocity destination recommendation tools.

(a) (b)

Caveats and limitationsNeither system supports the user in build-

ing a “user defined” trip, consisting of oneor more locations to visit, accommodations,and plans to visit additional attractions (amuseum, the theater, and so forth). Althoughtravel planning is a complex decision process,these systems support only the first stage—deciding the destination.

Researchers have proposed severalchoice models,5 which identify two groupsof factors that influence destination choice:personal features and travel features. Thefirst group contains both socioeconomicfactors (such as age, education, and income)and psychological and cognitive ones(experience, personality, involvement, andso forth). The second group might listtravel purpose, travel-party size, length oftravel, distance, and transportation mode.These various factors affect all stages ofthe traveller’s decision-making process,which is a complex constructive activity.

Another reason why these systems focuson destination selection relates to the filter-ing (content-based) approach. Even if wecould apply the same filtering technology toother tourism objects, such as cruises, thesystem would have to describe a catalog ofcruises—that is, build a catalog using aselected set of features (decision variables).The approach does not scale unless we pur-sue a costly knowledge-engineering activityfor each product type. So, these systemsmust have a particular catalog—in this case,a catalog of destinations—which requiresextensive domain knowledge and must bebuilt for the particular application. Currently,the focus is on destinations because they arerather stable, reusable concepts (many rec-ommender systems can exploit the same des-tinations knowledge base).

Pure collaborative filtering approachesdo not suffer from this problem, but, unfor-tunately, we cannot readily implement themin the travel domain. The major issue is thecomplexity of travel objects; we can’t sim-plify a trip to the point where two travellers’trips are the same. Surely two people havebought the same book, but it is less likelythat two people have experienced the sametrip. This points to a basic requirement ofCF approaches: one user’s purchase historymust be comparable to that of another. Thus,one user’s travel list must somehow overlapthat of another user. One approach could beto simplify the travel description to a certainpoint—for instance, representing just the

destination—but then we will discover thatthe already visited destinations are insuffi-cient to predict the next one. Additional con-text information must be included, so wemust query the user about the content of hisor her trip. Hybrid approaches that combinecontent- and collaborative-based approacheswill more likely succeed.6

Broadening the scopeGoing back to the basic recommendation

process (moving from needs to productswith explanations), this apparently linearprocess is far from being straightforward inthe real world.

Catching user needs and decision stylesRecommender systems struggle to catch

user needs, and companies have imple-mented different approaches to tackle thisissue. Amazon.com, for instance, immedi-ately recognizes the user’s identity and rec-ommends a book, without asking for anyuser input. In contrast (similar to the twotravel recommender systems mentionedearlier), www.activebuyersguide.com in-volves a user searching for a vacation in amultistage interaction. First, the site asksabout the vacation’s general characteristics(type of vacation, activities, accommoda-tion, and so forth). Second, it asks for detailsrelated to these characteristics, then fortradeoffs between characteristics. Finally,it recommends destinations. Both approacheshave drawbacks, but an adaptive approach,where questions are fine-tuned as thehuman–machine interaction unfolds, hasmore potential.

Researchers have recently argued thatrecommender systems should support mul-tiple decision styles.5 The DieToRecs rec-

ommender (a case-based travel planningsystem) supports these decision styles byletting the user enter the system throughthree main doors: iterative single-itemselection, complete travel selection, andinspiration-driven selection.

Iterative single-item selection lets themost experienced user efficiently navigate inthe potentially overwhelming informationspace. The user can select whatever productshe or she likes and in the preferred order,using the selections done up to a certainpoint (and in the past) to personalize the nextstage. For example, if the user selects a par-ticular destination, that destination is used torecommend a particular accommodation.

Complete travel selection lets the userselect a personalized travel plan that bun-dles items available in the catalog. The per-sonalized plan is constructed “reusing” thestructure of travels built by other users insimilar sessions.

Inspiration-driven selection lets the userchoose a complete trip by means of a sim-pler user interface (icon based) and an inter-action that is as short as possible. The tech-nology behind this approach is provided byintegrating case-based reasoning with inter-active query refinement. Interactive queryrefinement allows a more flexible dialoguemanagement—the system tackles failuresdue to over- or underspecified user needs,suggesting precise repair actions (constraintrelaxation or tightening, respectively). Case-based reasoning provides the framework tocast a recommendation session into a case-and similarity-based ordering of both com-plete trips and single products.6,7

Generating recommendationsThe mechanistic idea that from needs

(problems), the recommender’s intelligentalgorithm can deduce the right products(solution) is far too simple. Marketers statethat needs can be created such that productscan be sold. This motivates the suggestionpath in Figure 2. Products shown on a Website can help create needs by offering exam-ples to users who might not have enoughexperience to formulate the query as the rec-ommender system might require (see, forexample, www.activebuyersguide.com). Inother words, an effective travel recommendersystem should not only notice the user’s mainneeds or constraints in a top-down way butalso allow the exploration of the option spaceand support the active construction of userpreferences (in a bottom-up way).


Recommendation path

Suggestion path

Product

Needs Rationale

Figure 2. Recommendation and suggestion paths.

Recent research has emphasized thischange of perspective, defining it as naviga-tion by proposing.8 In this approach, the sys-tem shows the user examples of products,selected from those that the initial queryretrieved. The user can choose a product asthe current best choice, which updates theinitial query and lets the recommender iden-tify a new set of suggestions. The relevancefeedback technique used in informationretrieval (for example, Rocchio’s method)has influenced this approach, which basi-cally injects new constraints or terms—extracted from the selected item or a corre-sponding cluster—into the original query. Inaddition, the approach is conversational inthat it supports either a multistage interactionor a dialog that interleaves needs elicitationwith products.9 In multistage interaction,example recommendations elicit user needsby exploiting a dialog control component,which poses only focused questions, deter-mined by the previous interaction steps.

Speaking the right languageAs I mentioned earlier, recommender sys-

tems must carefully manage the human–machine dialogue such that even a naive usercan effectively use the system. Rephrasing auser-centered design slogan: “Recommendersystems are about people, not machines.”Thus, usability issues, such as choosing theproduct description language, come to thefore. For instance, asking if the user needs a“hot shoe” or a “manual white balance” in adigital camera could be a “hard to say” ques-tion for a naive photographer.

A recommender system’s ultimate effective-ness relies on its algorithms and their ability toextract useful and novel products from the cata-log.10 However, even if the recommendationsare useful, users will struggle if the help systemis poor, the item descriptions are too terse, or thesite navigation support is confusing. Systemusability is such an important issue that even arecommendation that is not useful but correct(for example, a place already visited) canincrease a user’s trust in the system—a neces-sary condition for recommendation acceptance.

Recommender systems could becomelearning environments or simpler informa-tion presentation tools, but we must designthem to support surplus learning and userbehavioral changes; again, usability comesfirst. Furthermore, the interaction and inter-face design can deeply affect the user’sdecision-making process. Different designchoices can induce distinct decision strate-

gies and influence the user’s affective state(emotions, level of involvement, quality ofthe flow experience) in peculiar ways.

Recommender systems emerged initiallyas filtering tools, where the primary concernwas to discard, in a large database of prod-ucts, items inappropriate to user needs.Now, experiences with real recommendersystems and research prototypes show thatthe user tasks and functions supported bysuch systems are much more varied. Wethus should focus on new support functionsfor expanding the user’s horizon.

AcknowledgmentsI thank Fabio Del Misser, Elena Not, and

Hannes Werthner for comments on an earlierversion of this essay. Thanks to Daniel R. Fesen-maier and Josef Mazanec for their inspiring cri-tiques and encouragement. This work is partiallyfunded by the CARITRO foundation (underContract “eCommerce and Tourism”) and theEuropean Union’s Fifth Research and Technol-ogy Development Framework Programme(under Contract DIETORECS IST-2000-29474).

References

1. J.B. Schafer, J.A. Konstan, and J. Riedl, “E-Commerce Recommendation Applications,”Data Mining and Knowledge Discovery, vol.5, nos. 1–2, Jan.–Apr. 2001, pp. 115–153.

2. F. Ricci and B. Smyth, “Recommendation andPersonalization in eCommerce,” Proc. Adap-tive Hypermedia 2002 Workshop (AH 2002),Univ. of Malaga, Malaga, 2002.

3. H. Werthner and S. Klein, Information Tech-nology and Tourism—A Challenging Rela-tionship, Springer-Verlag, New York, 1999.

4. J. Delgado and R. Davidson, “KnowledgeBases and User Profiling in Travel and Hos-pitality Recommender Systems,” Proc. 9thInt’l Conf. Information and Comm. Tech-nologies in Tourism (ENTER 2002), K. Woe-ber, A. Frew, and M. Hitz, eds., Springer-Ver-lag, Heidelberg, Germany, 2002, pp. 1–16.

5. D. Fesenmaier et al., Tourist Decision Model,tech. report D2.2 DieToRecs IST-2000-29474, EU IST project, 2002; http://dietorecs.itc.it/PubDeliverables/D2.2-V1.0.pdf.

6. F. Ricci and H. Werthner, “Case-BasedQuerying for Travel Planning Recommenda-tion,” Information Technology and Tourism,vol. 4, nos. 3–4, 2002, pp. 215–226.

7. D.R. Fesenmaier et al., “DieToRecs: TravelAdvisory for Multiple Decision Styles,” to bepublished in Information and Communica-tion Technologies in Tourism 2003, A. Frew,ed., Springer-Verlag, New York, 2003.

8. H. Shimazu, “ExpertClerk: Navigating Shop-pers Buying Process with the Combination ofAsking and Proposing,” Proc. 17th Int’l JointConf. Artificial Intelligence (IJCAI 2001),Morgan Kaufmann, San Francisco, 2001, pp.1443–1448.

9. M. H. Göker and C. A. Thomson, “Personal-ized Conversational Case-Based Recommen-dation,” Proc. Advances in Case-Based Rea-soning: 5th European Workshop (EWCBR2000), Springer-Verlag, New York, 2000, pp.99–111.

10. K. Swearingen and R. Sinha, “Beyond Algo-rithms: An HCI Perspective on RecommenderSystems,” Proc. Recommender Systems: Papersfrom the 2001 ACM SIGIR Workshop, 2001,http://cs.oregonstate.edu/~herlock/rsw2001.

Adaptive context-awaremobility support for touristsAlexander Zipf, European Media Laboratory

As mobile devices decrease in size,weight, and price and increase in power,storage, connectivity, and positioning capa-bilities, tourists will increasingly use themas electronic personal tour guides. However,to make such mobile tourist services a suc-cess, a range of factors must work together,from technical issues (such as bandwidth,positioning availability, and supportedinteraction paradigms) to user interface andsecurity issues. We must also considerissues such as the availability of accurate,timely, and localized data, end-user costs(business models), and trust.

Location awareness for mobileusers

Resolving these issues becomes moreurgent as time-to-market gains importance.However, the danger exists of investing a lotof money into solutions that tourists willnot accept. For example, many companieshave already started developing mobile cityinformation and navigation systems tar-geted at tourists (in particular, during theUniversal Mobile Telecommunications Sys-tem (UMTS) hype in Europe). These com-panies often claim to provide personalized,location-aware solutions, but using buzz


words does not assure that the lessons aboutpersonalization and context-awareness havebeen thoroughly learned. Personalizationand localization are important prerequisitesfor successful tourist applications, but wemust further combine them to better inte-grate contextual information.

A common assumption is that a systemoffering only the most relevant information ina given situation (determined by place, time,task, interests, and so forth) will be more suc-cessful than a system offering only defaultinformation. I argue not only that context-aware information includes standard touristinformation (for example, on sights, hotels,and restaurants) but also that supporting ser-vices such as tour planning and dynamicmaps must profit from personalization andcontext-awareness. So, we must developlocation-aware proactive tips or aggregatedservices, such as personalized tour proposals,that geographic information systems (GIS)offer. This, in turn, will lead to new (user- andcontext-) adaptive mobile GIS services fortourists.

Mobile systems for tourists can stronglybenefit from the power of GIS. Informationinteresting to tourists is location-dependentby nature, and GIS can offer this data in alocation-aware way. The tourist’s positionthen acts as a filter and parameter for sys-tem queries.

In addition, location awareness is a keyfactor for mobile commerce’s success,because it can contribute to a system’s easeof use in many ways. GIS can handle spatialand topological queries, allowing navigationand route finding. More advanced GIS datamodels also let us store and retrieve histori-cal information, which gives much morepower to possible queries regarding a regionor site’s development and history—knowl-edge that an electronic tour guide should beable to provide. Also, in many queries, thesystem must handle fuzzy and user-specificmeasures such as “interesting,” “ugly,” or“within reach.”

Standard GIS functionality can imple-ment personalization and context awarenessthrough adaptive map generation, personal-ized tour proposals, or context-aware proac-tive tips, as has been realized in severalresearch projects (such as Crumpet, DeepMap, and others).

Adaptive map generationWhen orienting yourself in a foreign area

or searching for some kind of business or

sight, maps are of great value. They can rep-resent large amounts of information about the area of interest in a single picture—in a(potentially) easily comprehensible form.However, to facilitate the correct reading andunderstanding of a (sometimes confusing)map, we must design it properly. This is still achallenge for AI and smart systems, becausemap design is a complex task involving cog-nitive and psychological aspects.

We must properly design a map to manyfactors, from technical conditions (screensize, network bandwidth, and so forth), inter-ests, socioeconomic parameters, and therecipient’s cognitive abilities, to task and usepurpose (for example, if the user wants toreceive information on topographic, histori-cal, or other thematic aspects or if the map issolely used for navigation purposes). Obvi-

ous examples include map style, color, anduse of symbols. Furthermore, we should beable to adapt the map’s scale, alignment, andsize according to the user’s current location,travel method and velocity, or interests. Firstattempts to model the requirements anddevelop a framework for adaptive map pro-duction are under way, but we need improve-ments.1 We have yet to formalize all parame-ters or discover all psychological or cognitiverelationships, and known rules often contra-dict each other. This includes contradictionsbetween maps needing to show the correctgeometry of objects and needing to general-ize objects in specific scales and work withlimited displays. Oftentimes maps mustchange the object’s position, change theboundary’s appearance, and omit details.

Personalized tour proposalsA conventional tour planner computes a

path through a street network, given two or

more locations on the network. A tour pro-posal component’s task is more complex. Ittries to suggest individual sight-seeing toursaccording to the tourist. Recent develop-ments find user-optimal tours that mustrespect hard time constraints via differentheuristic approaches related to the prize-collecting traveling salesman problem.2

Several possibilities exist for includinguser interests in a tour-proposing algorithm.First, we must identify the set of possibleattributes that could influence the choice fora particular section of a route. Such attributesinclude hard restrictions (or physically given attributes) and dynamic and soft user-specific parameters. Their weighting canvary strongly from one person to another orthrough time. Such parameters could includeaesthetic aspects, the area’s social milieu,dislike of traffic, or a preference for niceviewpoints. Such attributes are rarely avail-able in street databases or GIS, nor are theseparameters’dependencies modeled formally.

Context-aware proactive tipsA proactive spatial-context module gives

tips based on the user’s location and interestsregarding nearby objects of interest. Market-ing quickly adopted this idea of pushinglocation-dependent information (advertise-ments) to users, who were not always happyto receive this information. So, it is crucialto integrate fine-grained user and contextmodels into such a service to raise accep-tance and keep tourists from turning off such a feature. The system must use morethan the user’s position and the location ofobjects to deliver suggestions. Even resolv-ing what “nearby” means to the user in thecurrent situation involves a wide range of personal parameters and contextual information. Parameters that might influ-ence the definition of “nearby” include

• The user’s physical condition• Available transportation• The weather• The task• Travel speed • Familiarity with the region• The region’s structure• Terrain

In addition, we must draw the user’s atten-tion to these nearby objects of interest non-intrusively to help ensure acceptance. Thisis a usability challenge, and we must testand evaluate different possibilities.


Personalization and localization

are important prerequisites for

successful tourist applications,

but we must further combine these

to better integrate contextual

information.

Geodata handling on mobiledevices

How do we get these services to run onmobile devices? To move the described GIScomponents from heavy servers to thinclients, we must develop lightweight versionsof these components. The European MediaLaboratory has implemented a first prototypeproviding geodata-handling-and-processingcapabilities such as topological queries (forexample, “Is it within …?” or “Does it inter-sect …?”) on a PDA using a spatial accessmethod using an R-Tree.3 Interoperability isalso important, because clients might needto work together with different back-endsystems. So, the data model is based on anopen standard—the OpenGIS Consortium’sGeography Markup Language. Through suchdevelopments, we can realize future applica-tions that might use client-side resources toperform GIS services. This would also reducedependency on network availability. Next wemust consider intelligent prefetching andcaching strategies in combination with loca-tion awareness and resource (network, CPU)adaptivity. Further research on update andsynchronization strategies or optimized pro-tocols for adaptive geodata transmission overwireless networks would also be helpful.

To choose relevant information accord-ing to interests and context, we need fine-grained user and context models and inter-action histories or the ability to analyze auser’s movements. Only a few systems cur-rently try to offer all this. We still need toprove that the effort to collect and exploitthis information will pay off economicallyand in terms of improved usability andacceptance.

This leads to a remaining issue for adap-tive mobile services: the availability (or lackthereof) of appropriate (timely, correct) data(also geodata) with appropriate semanticmetadata. Usable content is generally impor-tant—the best adaptation techniques won’thelp much if the content is insufficient. Wecan view the lack of semantic understandingas a possible threat to the success of adap-tive location-aware systems. Although theSemantic Web initiative is developing thenecessary technologies,4 it might still be awhile before semantically enriched data iswidespread. Incorporating context is impor-tant, but so is further researching sensors andcontext data.

Finally, we can’t neglect security and pri-vacy. Spatial privacy is a major issue forspatially enabled tourism services, becauseusers usually only provide their position inexchange for some benefit. However, local-ization techniques now let system providersobtain information without much userintervention, and some party in the valuechain will always be able to track the user’slocation. This has raised the awareness ofpossible fraud and created the need forlegal and technical solutions.

AcknowledgmentsThe Klaus-Tschira Foundation and the Crum-

pet project (EU-IST-1999-20147) supported thiswork. I thank all my colleagues at the EuropeanMedia Laboratory and partner institutions fortheir valuable input.

References

1. A. Zipf, “User-Adaptive Maps for Location-Based Services (LBS) for Tourism,” Proc. 9thInt’l Conf. Information and Comm. Technolo-gies in Tourism (ENTER 2002), Springer-Ver-lag, Heidelberg, Germany, 2002.

2. M. Jöst and W. Stille, “A User-Aware TourProposal Framework Using a Hybrid Opti-mization Approach,” to be published in Proc.10th ACM Int’l Symp. Advances in Geo-graphic Information Systems, ACM Press,New York, 2002, pp. 329–338.

3. S. Schmitz, A. Zipf, and H. Aras, “OpenGML-Based Mobile Geodata-Handling forPDAs,” to be published in Proc. Annual Conf.Int’l Assoc. Mathematical Geology (IAMG2002), Terra Nostra—Schriften der AlfredWegner Stiftung, Berlin, 2002.

4. S. Staab, “Ontologies’ KISSES in Standard-ization,” IEEE Intelligent Systems, vol. 17, no.2, Mar./Apr. 2002, pp. 70–71.

Building Narrative Logic intoTourism Information SystemsUlrike Gretzel and Daniel R. Fesenmaier,University of Illinois at Urbana-Champaign

Tourism is one of the top e-commercecategories and one of the most experientialand complex products sold online. Neitherholistic sensory experience nor complexitylends itself to prevailing Web site designand its underlying computing structures.Consequently, tourism experiences arealmost exclusively captured as pieces ofinformation that can be described in func-

tional terms and thus easily translated intodatabase structures.

You would assume that the interface’s rolewould be to reintegrate these informationfragments into consistent wholes; however,online encounters of tourism information arecurrently restricted to interactions with inter-faces that more or less directly mirror theontology of the database systems to whichthey are connected. The database principle’sdominance in tourism information systemdesign becomes apparent when you look atthe search options and result displays thatthese systems offer. Users typically mustexpress their information needs about traveldestinations (accommodation, transportation,attractions, activities, and so forth) as highlystructured queries or choices among searchoptions that more or less reflect the rows andcolumns in which the data is stored. Evenwhen systems support natural languagequery, database logic still largely drives theoutput’s structure. The resulting display ofbits and pieces of data in the form of itemlists or collections of hyperlinks can onlymeet specific, functional information needs.It thus fails to reflect the complexity oftourism information’s role.

We intend to challenge the databaseapproach’s dominance in tourism interfacedesign by reflecting on its limitations in termsof effectively conveying relevant informationabout holistic vacation experiences. We sug-gest exploring narratives as a means to orga-nize and display tourist information in a waythat can communicate the different aspectsof tourism experiences, including sensoryand emotional components. Integratingnarrative principles should lead to tourisminformation systems that can better meet amultitude of informational needs, provideinformation that more closely matcheshuman memory, and support tourism deci-sion-making throughout its various stages.

Searching for tourismexperiences

It has long been recognized that experi-ences form the tourism industry’s founda-tion. However, experience’s role in con-sumption (including before, during, andafter a purchase) is only now being consid-ered for building effective marketing strate-gies. Travel activities such as skiing, hiking,shopping, and so forth provide the founda-tion for travel experiences, while the tourismindustry acts as an experience facilitator. Inaddition, the setting in which activities occur


contributes substantially to the nature ofthe experience (see Figure 3). Some havesuggested that the memories stored as aresult of travel experiences are key to attract-ing and retaining visitors. Furthermore,researchers have argued that stories—theprominent mechanism for communicatingexperiences—provide the path throughwhich the tourism industry can build andextend markets.

Marketing and tourism scholars haveargued that consumers often evaluate prod-ucts more on experiential aspects than onobjective features such as price and avail-ability.1 So, experiential information is notonly entertaining and stimulating but alsoessential to the travel decision-makingprocess, because it lets consumers under-stand and evaluate aspects of the travelproduct that cannot be easily described infunctional terms or expressed as monetaryvalues. Consequently, experiential infor-mation responds to the need for a holisticunderstanding of the specific travel experi-ence to be evaluated.

Despite the travel decision-makingprocess’s sequential nature, whereby trav-elers move step by step through a series ofhierarchically organized decision compo-nents, the information assimilated to serveas the basis for the various subdecisionsmust eventually make sense as a whole.2

Information presented as unrelated items ina list or under separate categories makes itdifficult for consumers to construct thiscohesive picture of a travel experience.Furthermore, narrative situations—such asquerying family and friends or consulting atravel agent—dominate traditional travel-information search. These human travelinformation providers typically supply con-textual information and emphasize particu-lar experiential aspects in a way that lets theinformation seeker establish mental con-nections among the various trip elements.

Whereas existing tourism information systemsappear to successfully provide functionalinformation for specific components of traveldecisions, they fail to address the need forholistic, experiential, and conversationalways of communicating travel information.

Database thinking in tourisminformation systems

The database logic’s appeal lies in itsclarity and suitability for computationalpurposes. Lev Manovich defines the data-base as a conceptual way to represent theworld as a list of items.3 Interacting with adatabase is a linear experience that differsconsiderably from viewing films or playingcomputer games. Database records are oftendisplayed in arbitrary order or according totheir relevance with regard to a certain searchtopic. Furthermore, interfaces followingdatabase logics essentially communicateinformation in fragmentary pieces. Althoughusers can make mental connections betweenitems that are displayed in a list, it requiresadditional cognitive effort. And if the userfails to establish connections, the numberof items that he or she can successfullyremember is rather limited. Whether theseconnections are made and how they are inter-preted remain outside the tourism informa-tion system’s control.

A lack of interpretation of the relation-ship between items is less problematic forunidimensional search concepts—for exam-ple, a search for room rates in a specifichotel. However, tourism experiences aretypically multifaceted, and cognitively sep-arating vacation aspects without losingcoherence and meaning is often impossible.Searches for information on entire vacationtrips are problematic from a database per-spective because they are open-ended,vague, ill-defined, multidimensional, andunconstrained. Interfaces that simply provideaccess to databases, and that feature queries

and information displays modeled afterdatabase structures, fail to acknowledge thecomplexity of tourism experiences.

The issue is to create an interface that canadd relevance to the information it displaysby supporting users in their efforts to imag-ine coherent experiences. A growing streamof research in psychology and AI suggeststhat narratives might connect seeminglyunrelated items.4 Narrative interfaces cantranslate the underlying database into a dif-ferent kind of user experience that is notonly more entertaining but also more infor-mative, because it helps the user derive con-textual information necessary to interpretcoherent experiences.

Narrative concepts asorganizing principles

We can describe narratives as eventsequences that create a cause-and-effecttrajectory of seemingly unordered items.3,4

They not only allow for meaningful con-nections between pieces of information butalso simultaneously afford the addition ofemotional content and sensory details.Thus, they can convey great quantities ofinformation, especially of experientialnature, in a format that users can quicklyand easily assimilate.5

Recognizing the importance of narrativesfor communicating tourism information isnot an issue of believing that narrative isthe only organizing principle in humanmemory. Rather, it acknowledges the expe-riential nature and complexity of vacationsand the importance of narratives as a meansof communicating tourism experiences.Human beings can organize experienceinto narratives that help them make senseof the world.6 The focus of narratives is onmeaning and relevance, not on precision.Furthermore, narratives support mentalimagery more successfully than other textgenres.7 Thus, narratives provide guidancein terms of interpreting search results butleave room for imagination.

Given the narrative framework’s inclu-sive nature, the significance of restructuringWeb site visits into narrative experiencesbecomes apparent from both a human–computer interaction and a marketing per-spective. If narratives are a closer match tohuman knowledge and communicationstructures in the travel domain, then theyshould more effectively educate peopleabout fuzzy or complex travel-related situ-ations. Also, because narratives can link


Activities

Socialsetting

Attractions

Actualexperience Memories Fantasy

storiesSense of

attachment

Figure 3. The sequence of travel experience.

items into logical and consistent wholes,they can better represent bundles of infor-mation in contrast to single-item concepts.Furthermore, the inherent entertainmentvalue of narratives promises to engage Website visitors at a much higher level thaninterfaces that are direct representations ofdatabase structures. Finally, by providing asequential path, plot, or storyline, narrativeapproaches can potentially facilitate naviga-tion through unknown knowledge territory.

We should view the narrative conceptas an integral part of the tourism experience,which begins with the information searchprocess. Thus it should be understood as anunderlying process for travel informationsearch rather than an imposed design princi-ple. Its integration into Web sites in general,and tourism information systems in particu-lar, goes beyond adding yet another story.The importance of narratives lies in applyingthe narrative concept to both information dis-plays and navigational space to provideusers with destination information thataffords immediate comprehension, sense-making, and ultimately high personal rele-vance.

Researchers have proposed alternativestrategies for integrating narrative designconcepts into tourism information systemsthrough the means of story matching or nar-rative cues provided through an interactivedisplay.8 Other domains have already usednarrative design; online gaming and socialagent development, for instance, have suc-cessfully integrated narrative principles intotheir designs. Tourism information systemscould greatly benefit from adopting the nar-rative approaches developed in these areas.The ultimate goal of tourism informationsystem design is to provide an electronicenvironment that can meet all travel-relatedinformation needs, including searches forthe purpose of satisfying hedonic needs andthe need for coherence.

References

1. C.A. Vogt and D.R. Fesenmaier, “Expandingthe Functional Information Search Model,”Annals of Tourism Research, vol. 25, no. 3,July 1998, pp. 551–578.

2. D.R. Fesenmaier and J. Jeng, “Assessing Struc-ture in the Pleasure Trip Planning Process,”Tourism Analysis, vol. 5, no. 1, 2000, pp. 13–27.

3. L. Manovich, The Language of New Media,MIT Press, Cambridge, Mass., 2001.

4. R. Schank and R. Abelson, “Knowledge andMemory: The Real Story,” Knowledge andMemory: The Real Story, R.S. Wyer, ed.,Lawrence Erlbaum Associates, Hillsdale,N.J., 1995, pp. 1–86.

5. N. Gershon and W. Page, “What StorytellingCan Do for Information Visualization,”Comm. ACM, vol. 44, no. 3, Mar. 2001, pp.31–37.

6. M. Mateas and P. Sengers, “Narrative Intelli-gence,” Proc. AAAI Fall Symp. NarrativeIntelligence, AAAI Press, Menlo Park, Calif.,1999, pp. 1–10.

7. W.F. Brewer, “Imagery and Text Genre,” Text,vol. 8, no. 4, 1988, pp. 431–438.

8. U. Gretzel and D.R. Fesenmaier, “Stor-e-telling in Destination Recommendation Sys-tems: Concepts and Implications of NarrativeDesign,” User Modeling and Decision Mak-ing in Travel and Tourism Emergent Systems,Electronic Commerce and Tourism ResearchLaboratory, Trento, Italy, 2002, pp. 5–7.

Information Delivery forTourismCécile Paris, Commonweath Scientific andIndustrial Research Organization,Mathematical and Information Sciences,Australia

Tourism has not escaped the information-overload trend, and we see an increasingnumber of tourism portals providing infor-mation filtered by users’ specific requestsand preferences. The information providedis diverse—including travel planning, routedescriptions, and advice on sites to visit—and the filtering mechanisms often requiresophisticated strategies.

Unfortunately, this is only half of theproblem. Once collected, to facilitate under-standing and use, the information must bepresented in a manner that is appropriate andnatural, tailored to the unique context. Thiscontext might include the fact that users arerestricted either in their real estate (for exam-ple, the size of their handheld devices) ormedium (for example, a phone only allowsspeech). In addition, to gather information ona topic, users typically must issue numerousqueries (about hotels, the weather, possibleactivities, history, and so forth), with the bestresults achieved when they know about rele-vant Web sites and their structure.

Here, I argue that filtering is not enoughand that tailored delivery includes deliver-ing appropriate and relevant information ina coherent and meaningful way. Deliveringa coherent document is an effective way togive users information they can reasonabout and act on, thus potentially turninginformation retrieval systems into morepowerful communication tools.

I begin by examining two approaches toinformation delivery: information filteringand natural language generation. Mostsystems in the tourism domain use IF, but Iargue that coupling it with NLG will pro-duce more coherent and relevant output.

Filtering vs. languagegeneration

The most common way to deliver informa-tion responsive to users’needs is based oninformation retrieval (IR) techniques thathave an added filter. On the Web, we typi-cally obtain a user’s information by askinghim or her to complete a form. A system thenadapts the content of the pages presented tothe user. (Similarly, when delivering infor-mation to a user on the go, the user profile isstored, and filtering happens given the addi-tional location information that a GPS pro-vides.) This type of system employs informa-tion-filtering techniques and appends theretrieved answers into a fixed documentstructure. This approach is fairly rigid, oftendoes not produce a document that is coherentas a whole, and requires the user to issuenumerous queries to satisfy his or her infor-mation needs.

The other approach uses NLG and user-modeling techniques to form a text by follow-ing typical rules of discourse answering ageneral-information need. These rules ensurethat the system includes and coherently orga-nizes all relevant information. This approachis flexible and ensures that the resulting docu-ment is coherent and akin to a naturally occur-ring document. Furthermore, this approachavoids having the user issue numerous spe-cific individual queries. However, the datasources are often restricted to those con-structed for the system.

In the tourism domain, where large data-bases and documents already exist, the sec-ond approach is unrealistic, if it requiresmanually reengineering the data sources.Thus, the IR approach is typically the oneadopted. However, delivery would be moreuseful if the resulting output were organizedcoherently. In the tourism domain, such a


document could be a travel guide. Indeed,simply providing a list of items (for exam-ple, as the output of a retrieval engine), inno natural order—or at least, not in an orderthat makes sense to a user—and requiringthe user to issue a number of queries, is notenough. This is true even if items have beenfiltered to consider a user’s profile.

Achieving coherence In our work at CSIRO, we have followed

this approach, coupling NGL techniqueswith IR. Given a discourse goal that repre-sents a desired effect on the user’s mentalstate and knowledge (such as to familiarizethe user with his or her travel location), oursystem—the Virtual Document Planner(VDP)1 and Tiddler, its application to thetourism domain2—explicitly plans a text.Tiddler obtains relevant information fromheterogeneous data sources, on the basis ofits discourse rules and constrained by theuser model. Tiddler also organizes theretrieved information as a coherent whole,potentially synthesizing it when required.

The system achieves coherence by exploit-ing coherence relations representing how sec-tions of text relate to each other. We useRhetorical Structure Theory, which lendsitself to computational planning and has beenused in many generation and multimedia sys-tems.3–7 Our text planner uses a library of dis-course plans, indicating how to achieve a dis-course (or communicative) goal.3 Such aplanner selects, synthesizes, and assemblesonly relevant content for the user and coher-ently presents the information. Typical ofthis approach, 3, 5–6 the discourse plans aredesigned on the basis of a corpus analysisand, in this domain, represent a travel guide’sprototypical structure and contain the infor-mation typically included in a travel guide.Documents produced by following theseplans thus resemble a travel guide, with tai-lored information based on the user model.

Complementary to discourse planning isthe problem of pulling out bits of informa-tion from various sources and combiningthem to form a valid document. Instead ofmanually building the knowledge base, asgeneration-based approaches often do,3,5 orautomatically building the knowledgebases from other sources,8 we use IR tech-niques to retrieve information from hetero-geneous databases and Web pages. In par-ticular, we use Norfolk, a scripting languagedeveloped for synthesizing virtual docu-ments from databases and existing Web

pages.9,10 Norfolk constitutes the interfacebetween the discourse planner and the datasources, retrieving the information the dis-course component requests and pulling ittogether, forming XML fragments.

Coupling an NLG approach to an IR sys-tem has several advantages. First, the out-put’s overall organization and content iscoherent, reflecting naturally occurringdocuments and discourse. This output isthus more natural to the reader. Indeed,Tiddler produces a (tailored) travel guidethat looks like traditional guides and thatusers can take along or use as a startingpoint to get additional information. Thisprovides more effective output than merelyfiltering information.

Second, our architecture is such that the

tailored guide’s overall organization andcontent is constant regardless of the deliv-ery medium. We achieve this by decou-pling the content and organization planningfrom the presentation planning, the stage atwhich the medium is considered. It ensuresthat, should users choose a different mediumas they move from one setting to another(such as from a desktop to a handhelddevice), the information is still accessibleand easy to browse.11

Third, the system can participate in ameaningful and natural interaction (or dia-log) with the user, because it understandswhat it produces and can reason about theuser’s further requests, in the context of thecurrent discourse. It achieves this by keeping adiscourse tree at the end of the planningprocess. This intermediate structure repre-sents the document content and contains theintermediate discourse goals and the coher-ence relations that hold among them.3

Fourth, the system pulls together infor-

mation addressing the user’s informationneed, as opposed to answering specificqueries. This alleviates the need for theuser to issue numerous queries.

Although the results of our initial evalua-tions were not statistically significant owingto a lack of participants, they indicated that,as a whole, users prefer tailored documents.2

We have had difficulty, however, proving thattailoring is a more effective communicationtool using quantitative measures reflectingactions (for example, did the users use theinformation?) as opposed to qualitative mea-sures reflecting user judgment (did the usersfind the output relevant?). Effectiveness ismuch harder to test than, for example, preci-sion, recall, or readability, because it requiresa task’s context. In turn, the tourism domain(and many others) requires that we tracknumerous participants over an extended timeperiod. These conditions are difficult (andcostly) to satisfy.

We also must acknowledge concernsabout systems that tailor (or filter) informa-tion. First, can someone hijack the technol-ogy and turn it into a marketing trick? Thereis a delicate balance between providing rel-evant information and pushing a company’sproduct. Similarly, the goals of providingthe user with useful information and appro-priately highlighting a product or serviceprovide conflicting constraints for systemimplementation. Having the latter overridethe former will likely annoy users and pushthem to stop using the system.

A similar concern can be raised regard-ing coherence and the discourse strategies Idiscussed: a discourse goal might be toconvince the user of something, thus influ-encing the user’s opinion. Furthermore, tai-loring information to a user model mightbe too prescriptive and restrictive, prevent-ing serendipity. It is still possible, however,to develop mechanisms that allow for anelement of surprise or that let users bypassthe tailored information. Finally, user pri-vacy is an important issue, with systemsstoring user information.

AcknowledgmentsThese colleagues participated in the design

and development of the Virtual Document Plan-ner and Tiddler: Nathalie Colineau, Francois


Coupling an NLG approach to an IR

system has several advantages.

First, the output’s overall

organization and content is

coherent, reflecting naturally

occurring documents and discourse.

Paradis, Anne-Marie Vercoustre, Stephen Wan,Ross Wilkinson, and Ming Fang Wu. I thankRoss Wilkinson, Ming Fang Wu, and Keith Van-der Linden for their feedback on this essay.

References

1. N. Colineau and S. Wan, “Mobile Delivery ofCustomized Information Using Natural Lan-guage Generation,” Monitor, vol. 26, no. 3,Sept.–Nov. 2001, pp. 27–31.

2. C. Paris et al., “Generating Personal TravelGuides—And Who Wants Them?” Proc. Int’lConf. User Modeling (UM 2001), LectureNotes in Computer Science 2109, Springer-Verlag, New York, 2001, pp. 251–253.

3. J.D. Moore and C.L. Paris, “Planning Text forAdvisory Dialogues: Capturing Intentionaland Rhetorical Information,” ComputationalLinguistics, vol. 19, no. 4, 1993, pp. 651–694.

4. W.C. Mann and S.A Thompson, “RhetoricalStructure Theory: Towards a Functional The-ory of Text Organization,” Text, vol. 8, no. 3,1988, pp. 243–281.

5. E. André and T. Rist, “Designing CoherentMultimedia Presentations,” Proc. Human-Computer Interaction: Software and Hard-ware Interfaces (HCI 93), Elsevier, Amster-dam, 1993, pp. 434–439.

6. B. De Carolis et al., “The Dynamic Genera-tion of Hypertext Presentations of MedicalGuidelines,” The New Rev. Hypermedia andMultimedia, vol. 4, 1998, pp. 67–88.

7. N. Green, G. Carenini, and J. Moore, “A Prin-cipled Representation of Attributive Descrip-tions for Generating Integrated Text and Infor-mation Graphics Presentations,” Proc. 9thInt’l Workshop Natural Language Genera-tion, Assoc. for Computational Linguistics,Ontario, Canada, 1998, pp. 18–27.

8. R. Dale et al., “Dynamic Document Delivery:Generating Natural Language Texts onDemand,” Proc. 9th Int’l Conf. and WorkshopDatabase and Expert Systems Applications(DEXA 98), IEEE Press, Piscataway, N.J.,1998, pp. 131–136.

9. F. Paradis, “Information Extraction and Gath-ering for Search Engines: The TaylorApproach,” RIAO (Recherche d’InformationsAssistée par Ordinateur), 2000.

10. F. Paradis, A.M. Vercoustre, and B. Hills, “AVirtual Document Interpreter for Reuse ofInformation,” Proc. Electronic Publishing, Lec-ture Notes in Computer Science 1375,Springer-Verlag, New York, 1998, pp. 487–498.

11. D. Chincholle, “Designing Effective MobileServices on Small Communication Devices,”Proc. OzCHI 2000: Interfacing Reality in theNew Millennium (OzCHI 2000), CHISIG,Sydney, 2000.

Agents for Gathering,Integrating, and MonitoringInformation for TravelPlanning Craig Knoblock, University of SouthernCalifornia

The standard approach to planning busi-ness trips is to select the flights, reserve ahotel, and possibly reserve a car. Choosingthe airports, deciding whether to park at theairport or take a taxi, and deciding whetherto rent a car are often ad hoc decisions basedon past experience. The time and effort re-quired to make more informed decisionsusually outweighs the cost. Similarly, oncewe’ve planned a trip, many of us forgetabout it until a few hours before the flight.We might check the status of our flights oruse a service that automatically sends up-dated flight information, but other thanthat, most of us cope with problems as theyarise. Beyond flight delays and cancel-lations, there are a variety of possible eventsthat travelers would like to anticipate, butagain, the cost and effort required to moni-tor these events are not usually deemedworth the trouble. Schedules can change,prices can decrease after purchasing aticket (if a ticket price goes down, manytickets can be returned and repurchased fora small fee), flight delays can result inmissed connections, and hotel rooms andrental cars can be given away because atraveler arrives late.

To address these issues, at USC wedeveloped the Travel Assistant,1 an inte-grated travel planning and monitoring sys-tem. This system provides an interactiveapproach to making travel plans where allthe information required to make informedchoices is available to the user. For exam-ple, if the user is deciding whether to parkat the airport or take a taxi, the system com-pares the cost of parking and the cost of ataxi given the choice of airport, the selectedparking lot, and the traveler’s starting loca-tion. Likewise, when the user is decidingwhich airport to fly into, the system not only provides the cost of the flights but alsodetermines the cost of ground transporta-tion at the destination. Once a traveler hasplanned a trip, the system monitors variousaspects of the trip using a set of infor-mation agents that can attend to details forwhich it would be impractical for a humanassistant to monitor. For example, beyondsimply notifying a traveler of flight delays,

an agent also sends faxes to the hotel andcar rental agencies to notify them of adelay and ensure that the room and car willbe available. Likewise, when a travelerarrives in a city for a connecting flight, anagent notifies the traveler of any earlierconnecting flights and provides the flights’arrival and departure gates.

These innovations in travel planning and monitoring are made possible by twounderlying AI technologies. The first is theHeracles interactive constraint-based plan-ner,2 which captures the interrelationshipsof the data and user choices using a set ofconstraint rules. Using Heracles, we caneasily define a system for interactivelyplanning a trip. The second is the Theseusinformation-agent execution system,3,4

which facilitates the rapid creation of effi-cient information gathering and monitoringagents. These agents provide data to Hera-cles and keep track of information changesrelevant to the travel plans. On the basis ofthese technologies, we have developed acomplete end-to-end travel planning andmonitoring system that is in use today.

Should I drive or take a taxi? Heracles integrates a wide variety of

travel-related data from Web sources toprovide the data that users need to plantheir trips. It uses information agents(described later) to provide real-timeaccess to any of the online sources relatedto travel. The system starts by checking atraveler’s calendar to identify upcomingtrips. The user then selects a trip, and thesystem interactively helps the user plan thetrip, providing the choices of flights, hotels,ground transportation and so forth. Foreach decision, the system makes a recom-mendation that the user then selects oroverrides. Relationships about the differentchoices are made explicit using constraintssuch that choices in one part of a plan areimmediately reflected in other parts. Forexample, if the traveler decides to departfrom Long Beach Airport instead of LosAngeles International Airport, then the sys-tem would immediately retrieve new direc-tions to the airport and update when thetraveler should leave his or her house onthe basis of both the distance and updatedflight time. Figure 4 shows the initial infor-mation about driving to the airport and theresult of changing to a different airport.The propagation of updates occurs automa-tically as part of the constraint network.


Consider the choice of parking your carat the airport or taking a taxi. The appropri-ate choice depends on a variety of factors,and most people make suboptimal decisions

based on simple heuristics. Figure 5 showsan example constraint network for makingthese types of decisions, where the systemcalculates the total cost of parking your car,

based on the selected parking lot and thenumber of days you will be gone, and com-pares this to the cost of taking a taxi, whichdepends on the distance and taxi rate.Instead of simply guessing about thesetypes of choices, the system can carefullyevaluate them and make a recommendation,which the user can still override.

Besides integrating the data to help usersevaluate tradeoffs, the system also makesmore information available directly to theuser. For example, instead of simply choos-ing arbitrarily between two flights, the sys-tem can tell the user the types of plane, howthe seats are configured, and the flight’s on-time performance. If the system has access tothe user’s frequent flier accounts, it can eventell the user which airline to fly to accumu-late enough miles to qualify for a free ticket.All of this information is organized into theconstraint network to give users the informa-tion they need to make informed decisions toplan their trips.

You gave away my room butcharged my card?

The Travel Assistant uses information


Figure 4. Choices are propagated during the planning process.

computeDuration

getParkingRate

multiply

selectModeToAirport

DepartmentDateDepartureDateMar 15,2001

ReturnDate

OriginAddress

DestinationAddress

FindClosestAirport

DepartureAirport

Mar 18,2001

Duration

ParkingTotal

4 days

ParkingRate

ModeToAirport

$16.00/day

Taxi

LAXgetTaxiFare

getDistance

Distance

TaxiFare$23.00

15.1 miles

$64.00

Figure 5. The constraint network for comparing taking a taxi versus driving.

agents to support both planning and moni-toring. Although information agents are sim-ilar to other types of software agents, theirplans are distinguished by a focus on gather-ing, integrating, and monitoring data fromdistributed and remote sources. To efficientlyperform these tasks, we use Theseus.

Theseus provides a framework for theefficiently executing these agents. The sys-tem is based on a streaming dataflowmodel, which means it executes the agent’soperations as information streams into thesystem. This is critical in a networked envi-ronment, where the main delay in execut-ing an agent is waiting for informationfrom remote sources. A dataflow systemmaximizes the operations’ parallelism byexecuting as many of the actions as possi-ble in parallel. A streaming dataflowsystem sends results from one action (forthe next action to process) even before thefirst action has completed. By exploitingthese two types of parallelism, the systemissues information requests as early as pos-sible to minimize the execution time.

There are two types of agents written inTheseus: information agents and monitor-ing agents. Information agents provide allthe data needed in Heracles. Each informa-tion agent extracts information from a spe-cific Web site. These agents take a particu-lar information request and navigate to theappropriate page on a Web site, locate therequired information, and return it as anXML document for processing by anotheragent or application. We build agents forextracting data from Web sites usingmachine-learning techniques to train anagent by example on which information toextract from a given Web site.5,6

The monitoring agents are built on top of

the information agents and perform theirtasks at regular intervals. Table 1 shows a setof example messages from the monitoringagents in the Travel Assistant. These agentsoften must maintain state to keep track ofpreviously returned results. For example, totrack prices or schedule changes, the agentsmust know about previous prices and sched-ules. The monitoring agents also need tocommunicate with people, so they supportthe ability to send email, text messages, orfaxes. For example, as Table 1 shows, whena flight is delayed or will arrive after 5 pm,an agent sends a fax to a hotel to ensure itdoesn’t give away the hotel room.

Isn’t this being donecommercially?

Most commercial systems for travel plan-ning take the traditional approach of provid-ing tools for selecting flights, hotels, and carrentals in separate steps. There are two inte-grated approaches to this problem. The firstis a system called MyTrip from XTRA On-line. On the basis of personal calendar infor-mation, the system automatically produces acomplete plan that includes the flights, hotel,and car rental. Once it has produced a plan,the user can then edit the system’s individualselections. Unlike with the Travel Assistant,the user cannot interactively modify the plan,such as constraining the airlines or departureairports. Also, MyTrip is limited to only theselection of flights, hotels, and car rentals.

The second approach, which i:FAOSwitzerland is commercially developing,uses constraint satisfaction to find a completeitinerary.7 However, this system assumesthat all the relevant data has already beenretrieved before the constraint satisfactionprocess and does not address how to inter-

leave information gathering with constraintsatisfaction to handle the enormous amountof potentially relevant information.

For monitoring a trip, some commercialsystems (such as the one run by United Air-lines) provide basic flight status and notifica-tion. However, these systems do not actuallytrack changes in the flight status (they merelynotify passengers a fixed number of hoursbefore flights), and they do not notify hotelsabout flight delays or suggest earlier flights orbetter connections when unexpected eventsoccur (such as bad weather).

While the Travel Assistant provides a useful set of functionalities, we couldimprove it in many ways. First, a naturalextension would be to more tightly integratethe monitoring and planning capabilities, sothat when a flight is cancelled or delayedthe system would automatically provide analternate travel plan. Another limitation isthat the set of monitoring tasks are fixed bythe system builders. We are working on theAgent Wizard, which is a user interface thatwould let an end user specify his or her ownagents. For example, a user might want tospecify an agent that monitors the FAA siteand sends notification whenever there ismore than a 30-minute flight delay in theuser’s connecting city so that he or shecould consider alternate routes. Or anotheruser might want to define an agent thatrecords real-time flight data every five min-utes for a user’s flight to build a detailedmap of the actual flight path. With the hugeamount of information available on theWeb, there is no end to the set of monitor-ing agents that users could specify.


Table 1. Actual messages monitoring agents send.

Agent Message type Message text

Flight status Flight delayed Your United Airlines flight 190 has been delayed. It was originally scheduled todepart at 11:45 AM and is now scheduled to depart at 12:30 PM. The new arrival time is 7:59 PM.

Flight status Flight cancelled Your Delta Air Lines flight 200 has been cancelled.

Flight status Fax to a hotel Attention: Registration Desk. I am sending this message on behalf of David Pynadath, who has a reservation at your hotel. David Pynadath is on United Airlines 190, which is now scheduled to arrive at IAD at 7:59 PM. Since the flightwill be arriving late, I would like to request that you indicate this in the reservationso that the room is not given away.

Airfare Airfare dropped The airfare for your American Airlines itinerary (IAD - LAX) dropped to $281.

Earlier flight Earlier flights The status of your currently scheduled flight is: # 190 LAX (11:45 AM)–IAD(7:29 PM) 45 minutes late. The following United Airlines flight arrives earlier than your flight: # 946 LAX (8:31 AM)–IAD (3:35 PM) 11 minutes late.

AcknowledgmentsThanks to my colleagues who participated in

the development of the Travel Assistant: JoseLuis Ambite, Greg Barish, Steve Minton, MariaMuslea, and Jean Oh. This material is based onwork supported in part by DARPA and the AirForce Research Laboratory under agreementF30602-00-1-0504, and in part by the Air ForceOffice of Scientific Research under grant num-ber F49620-01-1-0053.

References

1. J.L. Ambite et al., “Getting from Here toThere: Interactive Planning and Agent Exe-cution for Optimizing Travel,” Proc. 14thConf. Innovative Applications of ArtificialIntelligence (IAAI 2002),AAAI Press, MenloPark, Calif., 2002, pp. 862–869.

2. C.A. Knoblock et al., “Mixed-Initiative,

Multi-Source Information Assistants,” Proc.World Wide Web Conf., ACM Press, NewYork, 2001, pp. 697–707.

3. G. Barish et al., “A Dataflow Approach toAgent-Based Information Management,” Proc.2000 Int’l Conf. Artificial Intelligence (IC-AI2000), CSREA Press, Athens, Ga., 2000, pp.657–664.

4. G. Barish and C.A. Knoblock, “SpeculativeExecution for Information Gathering Plans,”Proc. 6th Int’l Conf. Artificial IntelligencePlanning and Scheduling (AIPS 2002), AAAIPress, Menlo Park, Calif., 2002, pp. 184–193.

5. N. Kushmerick, Wrapper Induction for Infor-mation Extraction, tech report UW-CSE-97-11-04, Dept. of Computer Science and Eng.,Univ. of Washington, Seattle, 1997.

6. C.A. Knoblock et al., “Accurately and ReliablyExtracting Data from the Web: A MachineLearning Approach,” IEEE Data Eng. Bulletin,vol. 23, no. 4, Dec. 2000, pp. 33–41.

7. M. Torrens, B. Faltings, and P. Pu, “Smart-Clients: Constraint Satisfaction as a Paradigmfor Scaleable Intelligent Information Sys-tems,” Constraints, vol. 7, no. 1, Jan. 2002,pp. 49–69.


Francesco Ricci is a senior researcher and thetechnical director of the eCommerce andTourism Research Lab at ITC-irst. His researchinterests include recommender systems,constraint satisfaction problems, machine learn-ing, case-based reasoning, and software archi-tectures. He received his PhD in mathematicsfrom the University of Padova. Contact him [email protected].

Alexander Zipf is a research associate at theEuropean Media Laboratory in Heidelberg, Ger-many. Recent projects he’s worked on includeSmartKom, on multimodal and mobile human-computer interaction, and the IST EU projectCrumpet (creation of user-friendly mobile ser-vices personalized for tourism), which is onuser- and context-adaptive agent-based location-based services. He studied mathematics and

geography at the Universities of Heidelberg and of Manchester, and hereceived his PhD from the University of Heidelberg. He is member ofthe International Federation for IT and Travel & Tourism, and of severalworking groups on mobile and geographical information systems inGermany. Contact him at the European Media Laboratory–EML, Schloss-Wolfsbrunnenweg 33a, 69118 Heidelberg, Germany; [email protected];www.eml.org/english/homes/zipf/zipf.html.

Daniel R. Fesenmaier is a professor and thedirector of the National Laboratory for Tourismand eCommerce, University of Illinois atUrbana-Champaign. In addition, he directsresearch in the areas of internet marketing andthe development of knowledge-based decisionsupport systems for tourism organizations. Hisprimary research interests are in tourism mar-keting, information search, and consumer deci-

sion making. He received his PhD in geography from the University ofWestern Ontario. Contact him at the Nat’l Laboratory for Tourism andeCommerce, Univ. of Illinois at Urbana-Champaign, 104 Huff Hall,1206 South Fourth St., Champaign, IL 61820; [email protected].

Ulrike Gretzel is a PhD student at the Institute ofCommunications Research and a research assis-tant at the National Laboratory for Tourism andeCommerce, University of Illinois at Urbana-Champaign. Her research efforts focus on theintegration of information technologies in tourismorganizations and the communication of tourismexperiences through online tourism advertising,with a focus on embodied cognition and narrative

design. She received her master’s with emphases in tourism marketingand international business from the Vienna University of Economics andBusiness Administration. Contact her at the Nat’l Laboratory for Tourismand eCommerce, Univ. of Illinois at Urbana-Champaign, 104 Huff Hall,1206 South Fourth St., Champaign, IL 61820; [email protected].

Cécile Paris is a principal senior research scientistand R&D group leader at CSIRO/Mathematicaland Information Sciences. Her research interestsinclude language technology and multimedia pre-sentation, user modeling, human–computer interac-tion, authoring tools, and intelligent tutoringsystems. She received her BA in computer sciencefrom the University of California, Berkeley, and anMS and a PhD in computer science from Columbia

University. She is the chair of CHISIG, the Australian Computer HumanInteraction Special Interest Group of the Ergonomics Society of Australia.Contact her at [email protected]; www.cmis.csiro.au/Cecile.Paris.

Craig Knoblock is a research associate profes-sor at the University of Southern California and asenior project leader at the Information SciencesInstitute. He is also the chief scientist for FetchTechnologies. His research interests include infor-mation agents, information integration, automatedplanning, machine learning, and constraint rea-soning. He received his BS in computer sciencefrom Syracuse University and his MS and PhD in

computer science from Carnegie Mellon. He leads the InformationAgents Research Group. Contact him at Univ. of Southern California,Information Sciences Inst., 4676 Admiralty Way, Marina del Rey, CA90292; [email protected]; www.isi.edu/~knoblock.

C o m i n g N e x t“Web Services: Been There, Done That?”