Linked data related to places o ers a superior collection of information to base search and recommendation functionality on in eTourism visit planning as well as location-aware mobile applications. Besides places interesting in themselves, through linked data it is possible to discover places interesting only through association, such as being the venue for a concert by an artist with an interesting genre. However, in order to harness this collective data source, challenges relating to data heterogeneity, quality, scale, and indexing and querying complexity must be resolved. In this paper, the TravelSampo visit planning and mobile application is presented, which tackles these issues. Using the system, queries describing both simple and complex interests can be run over some 17 million places of interest from over 20 vastly heterogeneous sources.
Location-aware mobile devices are becoming increasingly commonplace. This has lead to a multitude of mobile applications to search for e.g. events, places of interest or services near the user's physical location. On the other hand, many eTourism web applications also now allow people to design travel plans online, picking sites to visit and exporting visit lists to their phone's navigator software.
The TravelSampo project is an attempt to harness linked data as a source
of material for an application to help travellers nd content relevant to them,
both in planning as well as during a trip. As compared to existing non-linked
data solutions as well as similar linked data systems such as DPBedia Mobile [
Particularly, the TravelSampo system takes into account that there are multiple ways in which a location may be of interest to a user. First, the place itself may have some quality of interest, such as being a church, or being a church in the gothic style. On the other hand, the place may also be of interest only through a more or less direct association, such as being the venue for an interesting event or having been the birthplace of a painter with a style of interest. In addition, a place may be of interest by virtue of the services o ered there, such as Internet access.
This variety of ways in which data can be both interesting as well as encoded necessitates a exible architecture for querying locations of interest. The strength of this is that the application should ultimately be able to cater to a wide variety of interests, from people looking for nearby museums through music fans interested in concerts by Norwegian heavy metal bands to freegans searching for dumpsters near big supermarkets without nearby surveillance cameras. At the same time, this scale of heterogenuity causes severe problems in both integrating the content as well as providing e cient and intelligent search and recommendation services and user interfaces on top of it.
In the current demonstration system of TravelSampo, some 17 million
locations have been loaded into the system, integrating information from over
20 vastly di erent datasets of places, places of interest, and content making
places interesting through association, such as ction taking place in real-world
locations, or the birthplaces of famous artists. Included are for example the
huge datasets of DBPedia [
In this paper, the TravelSampo application is presented rst through its user interface. After that, the challenges faced and solutions developed in integrating, mapping and making usable the disparate heterogeneous data sources are described. Finally described are the indexing and querying interfaces created that make possible the complex queries required to provide the advanced functionality of the TravelSampo application. 2
The TravelSampo application has two distinct interfaces. The web interface is used to plan the trip beforehand and to examine and share the trip afterwards. The mobile interface is used during the trip to nd the destinations and to get more information about them. 2.1
A typical user would be someone who is going on a trip to a new city. Before the trip he can use the planning interface to nd out what kind of cultural destinations and events the city o ers during his trip. The destinations can be searched with di erent levels of complexity. In the simplest case our user is interested in churches, which are places themselves. Our user is also interested in wall climbing, which is a service located in a place. And nally he's interested in modern art, which is a topic of an exhibition held in a place. The application can handle all these searches.
The user has found a couple of churches, a sport center with wall climbing and an exhibition of modern art and now he can save them to his destination shelf which can be accessed during the trip in the mobile interface. The lters used to nd these destinations can also be saved to be used on other trips either in the planning interface or the mobile interface. The visit planning interface is not yet implemented but the shelves and lters can be produced and used in the mobile interface. 2.2
During his trip the user can use the mobile application to nd the previously saved destinations. The starting page of the mobile interface, depicted in gure 1 has a list of nearby destinations (1.A) with their types, distance to them and interesting instances associated to them which can be ltered using the lter menu (1.B). This menu contains the users destination shelves that are relevant in his current location as well as his personal lters and some prede ned ones.
The user can now use the shelf containing his destinations to access their pages and get a route map to the destination as well as information about it, links to associated instances and a button to mark the destination for future reference. As the user is in the destination page in the context of the destination shelf there are also buttons to browse through other destinations of the shelf.
If the user nds he has more time than he expected he can go back to the main page and use his saved lters to nd for example more churches or instances related to modern art. This can also be used to quickly nd interesting destinations or useful services on the vicinity without prior planning.
There will also be a possibility to use free text search to nd destinations (1.C). The location area (1.D) shows the user's current address and allows him to reload it using the geolocation capabilities of the mobile device or set it manually. On the top right corner there is a link to the user's pro le page or a login page if the user hasn't already logged in (1.E). 3
As already stated, the TravelSampo data repository contains some 17 million locations sourced from over 20 vastly di erent datasets. These sources are described in table 1, stating the general type of data sourced from each provider, the number of location and reference items in the data sources, as well as some example content types in the data thus gained. As can be seen from the table, the TravelSampo system contains a truly heterogeneous mix of data of di erent types, sources and schemas.
Particularly interesting in analyzing the data sources integrated is the fact that the boundaries between geographic place names, places of interest and services are not crisp. For example, the general place name registries contain not only hills and swamps, but also areas of sporting services, churches and abandoned police posts, while on the other hand the locations in the Espoo nature site point of interest database are precisely hills and swamps of special interest. Further, the Helsinki City service database for example contains both o ce entities such as childcare services as well as located services such as swimming pools, which also feature in point of interest databases.
Based on this observation, the TravelSampo system was designed not to discern between these sources for locations at all, but to treat all location information as equal. This puts the burden of discovering whether a location is a point of interest to someone on the information gathered for that location. Among direct indicators of interest, primary among them is the type of the location. To be able to use this across all the place data in the TravelSampo system, it was decided to attempt to build a single uni ed place and place of interest type ontology POIO from all the type ontologies used in the various data sources.
This was done semiautomatically so that rst all place type labels were compared automatically, already yielding several hundred equivalency mappings. Then, these mappings were examined by hand, and a large number of spurious mappings rejected, while an equally large amount of new mappings and subclass relations were curated, until all place types could be found under a single root. Table 2 relates the numbers of distinct place types in the constituents of the POIO ontology, as well as the size of the nal ontology. In total, of the 2499 concepts in the nal ontology, 62% (1539) were found to be shared between at least two source ontologies.
Besides di ering in place types, the datasets also di ered vastly in terms of modeling and level of content description. For example, the RKY database of culturally signi cant milieu contains areas of cultural interest de ned as polygons. However, most of these areas are actually collections of multiple points of interest, which are not modeled separately at all. On the other hand, most of the other databases listed do not model areas at all, but only provide centerpoints for even large features. Even worse, it is often di cult to automatically deduce when a location actually refers to a notable mass of land, such as an amusement park, instead of a small point, such as a statue.
As regards services, in the vast majority of data and data sources used in the TravelSampo project, the services described are those that can be described indirectly through place type, such as being a restaurant or a pharmacy. However, in the Helsinki City data source the services o ered at a particular location are described separately, for example noting if a particular library o ers Internet access, has a scanner or loans AV equipment.
In the case of the TravelSampo system, particularly as it was making use of many automatically converted and dynamically updating data sources, it was decided that these heterogeneities in content modeling could not be uni ed, at least without losing information or the expressivity of the original data, but would have to be resolved at the query construction level. Fortunately, it seems that some quite general mapping rules could be made to facilitate this, for example linking services and events described as separate resources to the places they are provided or help in, or linking a culture site with no direct description to the compound description of the larger area it was found in.
As can be gleamed from the table listing the TravelSampo data sources, events such as concerts and exhibitions were identi ed as a particularly interesting non-direct element signaling a place of interest. That is, particularly for cultural applications, often one is for example not interested in a museum per se, but in the exhibitions that are on display in that museum at present.
Now, events are particularly dynamic sort of data. At the time of creating
the TravelSampo system, there were no sources for current and coming event
information in RDF. However, there were multiple sources from which such could
be gleamed in other formats, such as comma separated values, JSON or RSS.
The event content for TravelSampo thus comes from a converter pipeline that is
capable of being run at regular intervals, or by request. This pipeline is actually
a more general one, called Harava [12], created in the FinnONTO project1 [
A major problem in the event data sources to be used in the project however was that none of them contained any machine-processable descriptions of the topics related to the event, such as the style of an exhibition or the artist. This problem was also evident in some of the point of interest data sources. For example, in the OpenStreetMap data on Helsinki, there is an object of type \memorial", which only in its textual description says that 1) it is actually a statue and 2) it depicts the runner Paavo Nurmi.
To overcome these limitations, automated information extraction services were integrated into the TravelSampo architecture and the Harava pipeline, which could then extract relevant entities such as people, organizations and places as well as general content keywords from the textual descriptions of the events and other data items.
Because the information extraction tools were con gured to use the whole vast TravelSampo database as a source for keywords, they are usually able to pick up a huge number of potentially related instances. The problem then became more of ltering these potential instances to the most important and factual ones. Fortunately, here the project could make use of the open source Maui information extraction tool [10], which has been previously shown to be humancompetitive in selecting primary topic keywords from text. 4
After getting all the di erent data sources together, one nds a large amount of overlap between them. Besides the same places occurring in many of the place databases, also events typically are entered in more than one of our dynamically updated event sources.
The indexing system used in the TravelSampo project is capable of inferring
and resolving ontology language equivalency statements transparently. Thus,
mapping between the di erent datasources does not need to happen at indexing
time, but can be done centrally and iteratively in the global TravelSampo data
space through generating RDFS, OWL and SKOS equivalencies. Actually, this
task becomes one with the general task of mapping di erent RDF materials to
each other, and could use any of the readily available ontology and instance
mapping tools [
Due to this, all but certain mappings are also relegated into this late stage of processing, so that for example all keywords found in the data sources during pipeline processing are created as resources in the data source's own namespace, instead of being equated with ontology concepts directly. This also makes sure
that no information is lost and no errors introduced in indexing, due to e.g. the keywords used not being found in the reference vocabularies, or being translated to a wrong concept based on improper fuzzy reasoning.
As already said, the semantic enrichment done to the materials through information extraction tools is also done in this global data space. This ensures that, for example, when searching for concepts from textual descriptions, the algorithms have a maximal amount of content available from which to draw matches.
The transparent resolution of equivalency statements also means that any erroneous mappings can be undone easily after the fact by just removing the RDF triple specifying the bad mapping. The tasks of verifying and improving the resource mappings generated, as well as verifying automatic enrichments, can be done in the TravelSampo ecosystem through the SAHA metadata editor [9] created in the FinnONTO project, which has special support for going through annotations marked as suspect. The marking of such annotations can either be done originally in the enrichment process, or at a later date by utilizing heuristic or schema-based quality assessment rules.
For this latter task, the FinnONTO architecture contains the semantic content validator service VERA4. The output that VERA produces is not a list of errors per se, but rather a list of possible problems that an expert user can assess, and modify the schema or data as needed. The report also contains general statistics about the data, such as language de nition usage, so it can also be used for a general analysis instead of validation.
In this way, the dynamically updated content of the TravelSampo portal can be iteratively improved and corrected as the system is running, right when problems are discovered, allowing focusing on the areas most critical to e cient use. 5
Our stated choice of semantic integration by mapping properties and resources in RDF required that the triple-store used had to support easy and e cient resolution of both equivalency as well as subsumption relations, as those were the primary means used to map content.
In fact, in the custom triple-store implemented for TravelSampo, both of these are done transparently. As an example, a query for \?s rdfs:label ?o" would return also all skos:prefLabel and skos:altLabel triples, as well as any custom schema properties marked as equivalent to any of these. A query for \?s rdf:type foaf:Agent" on the other hand returns also instances of all the subclasses of foaf:Agent. For ease in additional processing, a uni ed view to the data is also provided, where all URIs in an equivalency set in the source are replaced with a single canonical version. This way, anyone processing the results of such inferred queries need not themselves repeat the equivalency calculation.
In the material used for TravelSampo, a total of 11 million equivalency sets were discovered, touching 25 million resources out of a total 350 million.
In addition to subsumption and equivalency inference, the triple-store of TravelSampo also includes support for quickly discovering all location resources annotated with geo-coordinates inside a speci ed bounding-box, as well as all other resources related to those locations. The same is done for any temporal entity resources such as event times and further resources related to them. These are all functionalities that were needed in the various user interfaces of the TravelSampo system. Similarly, e cient text search is provided for searching 1) objects by their labels, 2) objects by their literal attributes and 3) objects by the labels associated with their object attributes. The last index is used in the general text search interface of TravelSampo, so that one can for instance query by the string \Pyhajarvi" and be quickly returned all objects that relate to any of the 50 or so lake Pyhajarvis of Finland.
In order to cater to the complexity and heterogeneity of the data sources used in TravelSampo, the indexing and querying system also has to be able to e ciently query quite complex patterns. For example, the intent is that for example the query \Finnish electronica near Helsinki in the following week" would match a concert by Jimi Tenor at the Helsinki Ice Hall, because Jimi Tenor is a Finnish electronica artist playing there in the speci ed timeframe. However, inside the data model, this is quite a complex pattern, as visualized in gure 2.
To answer the query, rst, each resource with a label matching any of the keywords must be found, as well as those matching the temporal and spatial constraints. This results in a result set with (among others) the nationality Finnish, the genre Electronica, the concert event and the location of Helsinki ice hall. Then, all resources relating to these or their subconcepts must be added to the result set. This results in (among others) the artist Jimi Tenor (who is Finnish) and the album Intervision (which has a genre of downtempo, which is a subgenre of electronica).
Finally, all resources that are not already locations must be mapped to any that they refer to, and nally an intersection taken between all locations found to reveal the nal result. In this case, such mappings must also be done iteratively. While the concert event relates directly to the location, but the artist and the album are still two and three steps away, respectively. To obtain the nal result, one must follow rst the link from album to the artist, and then from the artist to the event, which then nally leads to the location.
To resolve this, the search functionality in the TravelSampo backend was split into multiple stages, each taking in SPARQL queries. First, multiple \select" queries are run, one for each incoming keyword, temporal and spatial constraint, acting on a dedicated index. Using the index, it is easy to e ciently return not only resources matching the spatial and temporal constraints, but also any resource that is related in any way to them, or a literal or another resource with a label matching a particular text query. In addition, this index also performs subclass inference. Thus, from this stage, in the case of the example queries one coordinates: near Helsinki
Location: Helsinki ice hall
Nationality:
Finnish temporal: next week
Event: Jimi Tenor concert
Artist: Jimi Tenor
Album: Intervision
Genre: Electronica
...
Genre: Downtempo would already get the artist Jimi Tenor and the album Intervision, in addition to the more direct resource hits.
Then, \mapping" queries are run separately and iteratively for each select query result set. In the example, these would map for example any albums to artists, artists to events and events to locations. After this, the system automatically takes an intersection of the mapped results returned from each select query. A further \ lter" query is also run. In TravelSampo, this makes sure that only locations ever make it to the nal result set returned.
After the result set is nally obtained, it is paged and returned. This can still be manipulated by a \grouping" query. This can be used to ensure that for example a set amount of both event locations and culturally signi cant locations matching a particular query are returned. To make sure all information to be shown in the search listing for each matched resource is included (such as images, event details, etc.), the system still runs any given \describe" queries for each returned resource, before nally returning answers.
Because of the e cient indexes of TravelSampo as well as caching of e.g. the mapping query results, the average processing time for even these complex queries is still 100-400 milliseconds on a modern desktop server. 6
While still a work in progress, the TravelSampo system already demonstrates the potential for a much richer way of searching for points of interest. In developing the system, multiple issues were identi ed.
Firstly, locations may be of interest not only through their immediate properties, but through quite long chains of associations. Secondly, it is hard to isolate points of interest from other general locations.
In processing actual databases for use in the TravelSampo system, the lack of machine-processable content keywords in most currently available datasets was identi ed to be a major problem. In the TravelSampo system, this was addressed by integrating state of the art information extraction tools into the system.
In order to enhance precision and recall in searching the heterogeneous datasets, key class ontology level reference resources in the TravelSampo system such as point of interest types were mapped to each other by hand. However, another requisite part of an integration architecture such as TravelSampo is still the support for iterative, automatic mapping of the instances and keyword concepts in the di erent datasets pouring in, sometimes dynamically each day. An equally important feature is the ability of human editors to correct these mappings.
Finally, the TravelSampo system and the datasets loaded into it highlight the complexity of queries needed to cater to complex needs, while demonstrating that answering such queries e ciently even on massive data sources is still quite possible.
Acknowledgements This research is part of the Semantic Ubiquitous Services Project (SUBI) 2009{2011, funded by the Finnish Funding Agency for Technology and Innovation (Tekes) and a consortium of 18 companies and public organizations, pre-eminently Turku { European Capital of Culture 2011. Some work has been funded also by the Finnish Cultural Foundation. 8. Kalfoglou, Y., Schorlemmer, M.: Ontology mapping: the state of the art. Knowl. Eng. Rev. 18, 1{31 (January 2003), http://portal.acm.org/citation.cfm?id=975027.975028 9. Kurki, J., Hyvonen, E.: Collaborative metadata editor integrated with ontology services and faceted portals. In: Workshop on Ontology Repositories and Editors for the Semantic Web (ORES 2010), the Extended Semantic Web Conference ESWC 2010, Heraklion, Greece. CEUR Workshop Proceedings (June 2010) 10. Medelyan, O.: Human-competitive automatic topic indexing. Ph.D. thesis, University of Waikato, Department of Computer Science (2009) 11. Ruotsalo, T., Makela, E., Kauppinen, T., Hyvonen, E., Haav, K., Rantala, V., Frosterus, M., Dokoohaki, N., Matskin, M.: Smartmuseum: Personalized Contextaware Access to Digital Cultural Heritage. In: Proceedings of the International Conferences on Digital Libraries and the Semantic Web 2009 (ICSD2009). Trento, Italy (2009) 12. Suominen, O., Hyvonen, E.: Expressing and Aggregating Rich Event Descriptions.
In: Proceedings of the 6th Workshop on Scripting and Development on the Semantic Web (2010) 13. Wilson, M., Russell, A., Smith, D.A., Owens, A., Schraefel, M.: mSpace mobile: A mobile application for the semantic web. In: Proceedings of the ISWC 2005 End User Semantic Web Interaction Workshop (2005), http://eprints.ecs.soton.ac.uk/11101