In this paper we present a methodology to provide visitors, in smart regions, additional cultural heritage attractions based on prior museum visits using user models and Linked Open Data. Visitor preferences and behavior are tracked via a museum mobile guide and used to create a visitor model. Semantic models and Linked Open Data support the representation of regional assets as Cultural Objects. The visitor model preferences are exploited using a graph similarity approach in order to identify personalized opportunities for visitors by filtering relevant Cultural Objects.
In this short paper we show a blueprint how semantic models and Linked Open Data
(LOD) support the representation of regional assets in order to identify categories of
opportunities for visitors based on different personal characteristics determined by
previous visits. Having a broad infobase from which to cull possibilities is an arduous
task that can benefit from automation. Due to the overwhelming number of
possibilities, it is important to personalize the Cultural Heritage (CH) experience. When
considering what is requires from a smart, personalized system, it becomes clear that the
reasoning process of the system has to focus on identifying opportunities for
intervention. When and how to intervene and what information to deliver/service to offer.
Having a user model, a context model, and a model of the cultural objects are
essential for successful support. These can lead to the interaction of museums and places of
cultural heritage to create mega-tourist experience (similar to Verbke and Rekom [
We describe our methodology: First we use exhibits in a museum (we use Castle Buonconsiglio in the Trentino Region as examples throughout this paper) and tag them using semantic concepts. Then a mobile museum guide is used to track visitors. Based on this data a user model is developed consisting of characteristics and preferences. We then use a dataset of Cultural Objects using an ontological representation of the domain to cull opportunities. Visitor Preferences are used to filter which Cultural Objects are relevant, and Characteristics are used to determine whether an event or cultural heritage place is desired. Context is used to filter for proximate locations weather conditions, opening times, etc. Again characteristics are used to determine how best to present this information to the visitor. 2
In this section we review two technology areas: User Modeling and personalization in CH, and Linked Open Data and Semantic relatedness.
According to Ardisonno et al[
Public agencies collect organize and manage a vast amount of data. Local and
European projects aims to deliver data as freely available, reusable and distributed
without any restriction, the so call Open Data. As part of these initiatives, tourism and
cultural heritage datasets have been published as Open Data. Semantic Web
technologies and in particular the Linked (Open) Data paradigm, introduced by Sir Tim
Berners-Lee in 2006 [
2 http://www.w3.org/2001/sw/wiki/OWL triples. On top of RDF and OWL, SPARQL Query Language3 is used to query and retrieve information stored as triples thus allowing and facilitating access to the so called Web of Data. DBpedia4, can be seen as the ontological version of Wikipedia, its the core of the Linked Open Data cloud.
In the Natural Language Processing area, semantic relatedness between terms or
concepts can be computed using two main approaches: (1) defining a topological
graph similarity using ontologies and computing the minimal graph distances between
terms, (2) using statistical methods and word co-occurrence in a corpus and
calculating the correlation between words. “WikiRelate!" [
The mobile guide, at each position of interest (POI), presents a list of relevant media assets. The mobile guide system logs: the POI, which assets are chosen how long they viewed the asset, and in general how long did they stay at the point of interest. We collect two types of information, the first in order to determine general personal characteristics and the second in order to determine specific topic interests. In general we use movement styles, to predict user characteristics (such as personality). We use time viewing presentations in order to determine user topic preferences.
In order to characterize the user we make use of his general movement activities. We use the following statistics: 1) NumberOfPOIsVisted (NPV) – number of positions where a person stayed more than 9 seconds as detected and logged by the mobile guide's positioning system. Nine seconds is a number we have used for previous analysis and has provided good results. 2) POIsWherePresentationsSeen (PPS) – the number of positions where the visitor viewed at least one media asset connected to that position as computed from the logs of the mobile guide. 3)NumberOfPresentationSeen (NPS) – the total number of media assets the visitor viewed as computed from the logs of the mobile guide.
The system uses annotated internal and external information about cultural places and events. Internal information is taken from catalogues or websites and is used by the mobile guide app to describe user preferences by storing the relevant topics related to exhibits the user has visited and liked. External information is imported from available Open Data about museums and cultural events and enriched in the domain ontology, using knowledge from the Linked Open Data cloud (DBpedia dataset). Data is stored using a domain ontology for tourism called eTourism5. The ontology covers methodological and practical aspect of services (hotels, B&B, etc.), cultural objects (museum, cultural places, etc.) and events. It is used as a vocabulary model to map external Open Data into RDF triples validated by the ontology concepts. For the present work we have developed a specific module of the eTourism ontology named Cultural Objects Ontology (coo) that covers (1) properties (such as topic, keywords, geographical information) of museums or events, exploits the semantic identity with LOD/DBpedia concepts (using owl:sameAs predicates) and implements (2) user profile types and topics of interests selections.
For each museum source, we extract - as a first step, keywords from exhibits of the Castle Buonconsiglio museum. We exploit the semantic relatedness implementing the graph similarity approach. We annotate keywords - for each description, and we disambiguate them to DBpedia concepts using DBpedia Spotlight APIs6. We filter out all the not relevant concepts and we then obtain a bag of concepts (related to cultural heritage) like the following: {dbpedia7:Trentino, dbpedia:Prehistory, dbpedia:Ancient_Rome, dbpedia:Middle_Ages, dbpedia:Hunter-gatherer, dbpedia:Upper_Paleolithic, dbpedia:Bronze_Age} In DBpedia, each concept is related to a category using the property dcterms:subject, then each category is part of a hierarchy structure with nodes connected via skos:broader properties. For example the below two DBpedia concepts have as dcterms:subject the DBpedia topic categories: 1) Last_glacial_period (dcterms:subject) ->{Climate_history, Glaciology, Holocene, Ice_ages} 2) Ancient_Rome (dcterms:subject) ->{Ancient_history, Ancient_Rome, Civilizations} 5 Currently under development at ICAR-CNR within the framework of the national project Dicet-InMoto-Orchestra, (http://www.progettoinmoto.it). 6 http://spotlight.dbpedia.org 7 Prefix for http://dbpedia.org/resource/
For the second step, we extract from the DBPedia SPARQL endpoint, for each concept, the topic categories of the DBpedia taxonomy. As result we obtain a wider bag of DBpedia topic categories describing each museum exhibit. Using the hierarchical structure of categories is thus possible to discover similarities among concepts that have ancestor categories in common.
As external sources, we take the Open Data set delivered by the Italian Cultural Heritage Minister8 (MIBAC) and we map these objects using the coo ontology; then, for each object, we exploit the same process applied for the internal resources, in order to annotate and extract the corresponding bag of topics. As a result, we obtain a list of information for each MIBAC Cultural Object (cultural place or event), as in the following example: foaf:name = “Memorie della Grande Guerra”, coo:mainCategory = http://dbpedia.org/resource/Category:History Bag of Concepts (dcterms:description) -> {1918_disestablishments, Aftermath_of_World_War_I, Austria-Hungary, Austria_articles needing attention, States_and_territories_established_in_1867, Anoxic_waters, Backarc_basins, Contemporary Italian_history, History_of_Austria-Hungary, History_of_modern_Serbia, Wars_involving_Italy, World_War_I } In order to select suitable Cultural Objects candidates for the user, we define a metric to measure the semantic distance between the user profile tags and the available cultural objects tags. As a first step, we measure the shortest path distance between each of the m topic categories in the bag of topics of the user profile and the coo:mainCategory topic of the suitable candidates (see table 2), and we reduce candidates cardinality by applying an upper threshold to the distance. After this step, we refine the result by calculating (via SPARQL queries on the DBpedia endpoint) the shortest path between the user bag of topics (m) and the suitable candidates bag of topics (n) on the remaining subset of cultural objects. Its important to underline that when computing the distance measure between topic categories we also take into account, for each hop of the shortest path, the number of outgoing links of the node: the more outgoing links a node has (to other DBpedia taxonomy nodes) the less it is specific. Broad connected nodes receive low weights while nodes with less outgoing connection will get higher values. We use each pairwise distance as a component of a normalized vector of distances, we evaluate, for each museum or http://dbunico20.beniculturali.it/DBUnicoManagerWeb/#home
Prefix for http://dbpedia.org/resource/Category: event an average normalized distance for each m user category and we sum all these distances to define the relatedness of each cultural object. Again an empirical threshold on distance is applied to retain a limited number of candidates. 3.2
Use of characteristics
Using behavior types we can tailor the amount and presentation of information. For example for ants and butterflies we can give ten items. For grasshoppers and fish we may only give two items. Ants and grasshoppers may be given places while butterflies and fish may be given events. Additional personalization may be possible. 4
Our current metric of semantic relatedness doesn't take into account whether the user profile bag of topics is representative of a sufficiently broad range of museums categories to cover their cultural preferences. To balance this, when all/most of the user preferences are of the same topic area (e.g. Prehistory), one or more among suggested items could be chosen from a minor topic category, to elicit variation in user interests. Our current research involves, the implementation of the methodology to the Old City and the Tower of David Museum in Jerusalem, and the evaluation of the user model and the semantic suggestions results.