Increasing availability of GPS-enabled devices technically enables a broad variety of people to participate in the volunteered geographic information (VGI) movement and to collect and share information about places and spatial entities. But in order to be useful, geo-data has to be correctly classi ed, and inexperienced users need assistance to be able to provide correctly classi ed information, because the classi cation system is complex and not always intuitive. In this paper, we propose a natural classi cation approach for spatial entities based on speech recognition and ontological reasoning to allow users to contribute geo-data with as little barriers as possible.
2
To enable systems to make correct use of the collected data, it has to be classi ed correctly. For example, cartographic renderers can only draw and label objects with correct style if the entities follow a certain speci cation. The classi cation of geo-data is complex and often ambiguous. For example, the type of a street or the function of some grass covered ground may remain unclear to the contributor. Trained contributors know how to apply a classi cation system correctly; for non-experts or casual contributors, the lack of this knowledge marks a barrier: most of the tools to collect, contribute, and classify geo-spatial data are complex systems requiring high technical a nity and skills. Moreover, even for experts, repeated classi cation of objects can become tiresome, leading to the danger of incompletely speci ed data.
Places have di erent facets for di erent people. Namely, the same place can
have very di erent functional roles depending on who is looking at it [
Research on VGI and Human-Computer Interaction (HCI) is increasingly
addressing the technological gap between potential contributors and the existing
data collection applications (e.g., [
Ontological reasoning for spatial classi cation
When we allow users to annotate spatial entities by means of natural language
rather than by using a prede ned catalogue, we have to expect a signi cant
mismatch between what users think the entity is and what the classi cation system
allows to describe. In [
Consider the following situation: some member of a development project wants to contribute data about the distribution of small backyard sh ponds which have been installed to minimize the lack of protein supply in poor areas of developing countries. This user is not educated to use a geographic classi cation system and is not aware of the proper term within a system like CityGML2, OSM, ATKIS3, or the OS MasterMap4.
If the user now labels the backyard sh ponds with the term \ sh pond", none of the above mentioned systems will recognize it as a valid entity. Without a proper classi cation, however, the data remains useless as it cannot be rendered or addressed by other algorithms.
To be able to match natural concepts of spatial entities with spatial classication systems, we propose a reasoning system as illustrated in Figure 2. The goal of the proposed reasoner is to identify the closest conceptual match in the classi cation system with the naturally spoken term. The term should not just be replaced, but the link between the spoken term and the linked term in the classi cation system is kept for further re nement of both the classi cation system
and the reasoner's capabilities. A main ingredient to make this re-classi cation possible is an abstraction layer on top of existing GIS classi cations, namely the meta-ontology GeoMO sketched in the next section. 3.2
The meta-ontology GeoMO and the OntoHub repository
OntoHub. Existing ontology repositories such as BioPortal5 lack the ability to
host heterogeneous ontologies in the sense of being formulated in ontology
languages other than OWL. As not all relevant ontologies will be OWL ontologies
(Dolce, e.g., is formulated in rst-order logic) we host our ontologies at
OntoHub6. Users of OntoHub can upload, browse, search and annotate basic
ontologies written in various standard ontology languages via a web frontend (see [
„Fish pond“
Search Space Reducer OSMonto Dolce OpenCYC YAGO
OntoHub WordNet GeoMO
GeoMO. The role of the meta-ontology GeoMO is twofold: rst, the mediation between human everyday concepts of space and spatial entities that should be matched against existing geo-spatial classi cations, and secondly, to translate between di erent classi cation systems such as OSM, ATKIS, OS MasterMap, CityGML, etc. For OSM, we have already designed OSMOnto, an automatically
generated ontology of OSM tags [
The GeoMO ontology, on a technical level, results from a colimit operation on the ontologies re ecting the classi cation systems of the participating GISs (we mentioned the OSMOnto ontology above, being one component), together with knowledge (i.e. term mapping, subsumptions between terms, etc.) about their relationship. Such mappings are part of the OntoHub infrastructure.
Here is a simple example illustrating the functionality of GeoMO. The OS MasterMap might contain the category s (i.e. `water structure | manmade'), whilst OSM might use the term t (i.e. `water body'). GeoMO establishes the subsumption s v t, i.e. the term t is more general than s. If the user now expresses the term `Fish pond' with spoken, natural language, the term is translated by available speech recognition into a proccessable term. The Concept Store uses this term for a lookup in WordNet and identi es the synonym w. Moreover, OpenCYC will tell us that this synonym w is in fact a special case of s, an o cial category in the OS MasterMap classi cation scheme. Finally, GeoMO can infer that t can be used as a more general category for labeling `Fish pond', without any user interaction. 3.3
A sketch of the Architecture of MapIt The reasoner depicted in Figure 2 will work as follows: the smartphone translates the spoken term \ sh pond" via a standard speech recognition module into parsable text. The detected term \ sh pond" is then send to the Search Space Reducer (SSR). The function of the SSR is to cut down the search space in a context-sensitive way: as we are in a geographic domain, we only want to query ontologies or parts of ontologies dealing with spatial objects and activities related to them. This situation allows the SSR to ignore a signi cant amount of entries, like facts about artists, movies, books, vehicles, etc.
After checking for the existence of the term in the target classi cation (in
this case OSM) and GeoMO. If both do not contain a direct correspondence,
the reasoner looks up the Concept Store. A core component of the Concept Store
is illustrated in Fig. 3. It illustrates the implementation of a work ow, previously
developed in [
check no
USER FALCON
PELLET
HETS
Merged Ontology select matching configuration extract modules
Modules Ontology Matching Graph
Matching Configuration compute colimit Alignment Specification produce formal specification their combination. This work ow is in particular essential for the construction of GeoMO, as the compatibility of mappings between the terms used in the various GIS ontologies has to be veri ed. We here brie y introduce these tools.
The ontologies to be matched and aligned are taken from OntoHub. As
matching system we use Falcon [
The MapIt architecture carefully integrates existing ontologies and reasoning systems and aims at an enhanced classi cation technology for geo-data. We expect that MapIt, once realized as a system, has the potential to lower the barrier of contribution of VGI tag data to OpenStreetMap or any other geographic classi cation catalogues. Currently, tagging in OpenStreetMap mostly happens at geographical level, and much less at a higher ontological level, e.g., concerning activities or individual perception, or place usage of users. This situation could greatly improve using MapIt.
Support by the German Research Foundation (DFG) is gratefully acknowledged. 11 See http://owlapi.sourceforge.net 12 See www.informatik.uni-bremen.de/co /hets