In this paper we report our ongoing work on the application of semantic technologies and machine learning to Integrated Tourism in the Apulia Region, Italy, within the Puglia@Service project. Integrated Tourism can be de ned as the kind of tourism which is explicitly linked to the localities in which it takes place and, in practical terms, has clear connections with local resources, activities, products, production and service industries, and a participatory local community. Integrated Tourism thus needs ICTs that should go beyond the mere technological support for tourism marketing, di erently from most approaches in eTourism research (see [1] for a comprehensive yet not very recent review). In this paper, we report our experience in supporting Integrated Tourism services with Semantic Technologies (STs) and Machine Learning (ML). The work has been conducted within Puglia@Service,1 an Italian PON Research & Competitivity project aimed at creating an innovative service infrastructure for the Apulia Region, Italy. The paper is structured as follows. Section 2.1 shortly describes a domain ontology for Integrated Tourism, named OnTourism, which has been modeled for being used in Puglia@Service. Section 2.2 brie y presents a Web Information Extraction (WIE) tool, named WIE-OnTour, which has been developed for populating OnTourism with data automatically retrieved from the Web. Section 2.3 illustrates some of the Semantic Web Services (SWSes) which have been de ned on top of OnTourism for supporting Integrated Tourism in Apulia. Section 3 outlines an application scenario for a ML tool, named Foil-DL, to better adapt the automated composition of these services to user demands. Section 4 concludes the paper with nal remarks and directions of future work.
2.1 from the viewpoint of Integrated Tourism because, e.g., it lacks concepts modeling the reachability of places. In Puglia@Service, we have decided to build a domain ontology, named OnTourism, 3 more suitable for the project objectives and compliant with the OWL 2 standard. It consists of 379 axioms, 205 logical axioms, 117 classes, 9 object properties, and 14 data properties, and has the expressivity of the DL ALCOF (D).
The main classes of the terminology are Site, Place and Distance. The rst is the root of a taxonomy which covers several types of sites of interest (e.g., Hotel and Church). The second models the places where sites are located at. The third, together with the object properties hasDistance and isDistanceFor and the data properties hasLengthValue/hasTimeValue, allows to represent the distance relation between sites with values in either length or time units. Distances are further speci ed according to the transportation means used (see, e.g., the class Distance on Foot ). Other relevant classes in the terminology are Amenity (with subclasses such as Wheelchair Access) and Service (with subclasses such as Bike Rental ) that model, respectively, amenities and services available at the accommodations. Finally, the terminology includes the o cial 5-star classi cation system for hotel ranking. 2.2
WIE-OnTour is a wrapper-based WIE tool implemented in Java and conceived for the population of OnTourism with data concerning hotels and B&Bs available in the web site of TripAdvisor4. The tool is also able to compute distances of the extracted accommodations from sites of interest (e.g., touristic attractions) by means of the Google Maps5 API. Finally, the tool supports the user in the speci cation of sites of interest.
Instantiations of OnTourism for the main destinations of urban tourism in Apulia have been obtained with WIE-OnTour. Here, we consider an instantiation for the city of Bari (the capital town of Apulia). It contains 34 hotels, 70 B&Bs, 106 places, and 208 foot distances for a total of 440 individuals. The distances are provided in time and length on foot and have been computed with respect to Basilica di San Nicola and Cattedrale di San Sabino (both instances of Church and located in Bari). The restriction to foot distances is due to the aforementioned preference of Integrated Tourism for eco-mobility. 2.3
In Puglia@Service, we have de ned several atomic services in OWL-S on top of the aforementioned domain ontologies, travel and OnTourism. For example, city churches service returns the churches (o.p. of type Church) located in a given city (i.p. of type City ) whereas near attraction accomodations service returns all the accommodations (o.p. of type Accommodation) near a given attraction (i.p. 3 http://www.di.uniba.it/~lisi/ontologies/OnTourism.owl 4 http://www.tripadvisor.com/ 5 http://maps.google.com/ of type Attraction). Note that closeness can be de ned on the basis of distance either in a crisp way (i.e., when the distance value is under a xed threshold) or in a fuzzy way (i.e., through grades of closeness). In both ways, however, the computation should consider the transportation means used as well as the measure units adopted according to the OnTourism ontology.
In Puglia@Service, we intend to obtain composite services by applying
methods such as [
In Puglia@Service, automated service composition will be enhanced by exploiting
users' feedback. The idea is to apply ML tools in order to induce ontology
axioms which can be used for discarding those compositions that do not re ect the
preferences/expectations/needs of a certain user pro le. Here, we illustrate this
idea with an application scenario which builds upon the accommodation rating
provided by TripAdvisor's users. More precisely, we consider the task of
accommodation nding. This task strongly relies on a classi cation problem aimed at
distinguishing good accommodations from bad ones according to the amenities
available, the services o ered, the location and the distance from sites of interest,
etc. In order to address this classi cation problem, we need ML tools able to
deal with the inherent incompleteness of Web data and the inherent vagueness of
concepts such as the closeness. One such tool is Foil-DL [
As an illustration of the potential usefulness of Foil-DL in the Puglia@Service context, we report here a couple of experiments concerning the ltering of results returned by the SWSes reported in the previous section for the case of Bari. We set up a learning problem with the class Bad Accommodation as target of the learning process. Ratings from TripAdvisor users have been exploited for providing Foil-DL with positive and negative examples. Out of the 104 accommodations, 57 with a higher percentage (say, over 0:7) of positive users' feedback are asserted as instances of Good Accommodation, whereas 15 with a lower percentage (say, under 0:5) are asserted as instances of Bad Accommodation. The latter, of course, play the role of positive examples in our learning problem. Syntactic restrictions are imposed on the form of the learnable GCI axioms.
In the rst experiment, we have not considered the distances of the accommodations from the sites of interest. With this con guration, Foil-DL returns just the following GCI with con dence 0:5: Bed_and_Breakfast and hasAmenity some (Pets_Allowed) and hasAmenity some (Wheelchair_Access) subclass of Bad_Accommodation The GCI suggests that B&Bs are not recommended even though they provide disabled facilities. It can be used to lter out from the result set of wheelchairaccess accommodations service those accommodations which are classi ed as bad.
In the second experiment, conversely, we have considered the distances of the accommodations from the sites of interest. With this con guration, Foil-DL returns the following GCI with con dence 1:0 hasAmenity some (Bar) and hasAmenity some (Wheelchair_Access) and hasDistance some (isDistanceFor some (Bed_and_Breakfast) and isDistanceFor some (Church)) subclass of Bad_Accommodation The GCI strenghtens the opinion that B&Bs are not recommendable accommodations for disabled people whatever their distance from the churches is.
As a further experiment, we have restricted our analysis of accommodations in Bari to only B&Bs. Starting from 12 positive examples and 39 negative examples for Bad Accommodation, Foil-DL returns the following two GCIs with con dence 0.154 and 0.067 respectively: hasAmenity some (Pets_Allowed) and hasAmenity some (Wheelchair_Access) subclass of Bad_Accommodation hasAmenity some (Bar) and hasAmenity some (Wheelchair_Access) subclass of Bad_Accommodation which con rm that B&Bs should not be recommended to disabled tourists. 4
In this paper we have reported our ongoing work on the use of STs and ML for Integrated Tourism in Apulia within the Puglia@Service project. Though developed for the purposes of the project, the technical solutions here described are nevertheless general enough to be reusable for similar applications in other geographical contexts. Notably, they show the added value of having ontologies and ontology reasoning (including also non-standard inferences like induction as exempli ed by Foil-DL) behind a Web Service infrastructure.
For the future we intend to carry on the work on the application of Foil-DL to the automated service composition. Notably, we shall consider the problem of learning from the feedback provided by speci c user pro les.