Comparing categories among geographic ontologies

Transcription

1 Computers & Geosciences 31 (2005) Comparing categories among geographic ontologies Marinos Kavouras, Margarita Kokla, Eleni Tomai Cartography Laboratory, School of Rural and Surveying Engineering, National Technical University of Athens, Zografos Campus, Athens 15780, Greece Received 16 June 2004; received in revised form 27 July 2004; accepted 27 July 2004 Abstract Numerous attempts have been made to generate semantic mappings between different ontologies, or create aligned/integrated ones. An essential step towards their success is the ability to compare the categories involved. This paper introduces a systematic methodology for comparing categories met in geographic ontologies. The methodology explores/extracts semantic information provided by categories definitions. The first step towards this goal is the recognition of syntactic and lexical patterns in definitions, which help to identify (a) semantic properties such as purpose, location, cover, and (b) semantic relations such as hypernym, part of, has-parts, etc. At the second step, a similarity measure among categories is applied, in order to explore how (the) extracted properties and relations interrelate. This frameworkenables us to (a) better understand the impact of context in cross-ontology mappings, (b) evaluate the quality of definitions as to whether they respect mere ontological aspects (such as unambiguous taxonomies), and (c) deal more effectively with the problem of semantic translation among geographic ontologies. r 2004 Elsevier Ltd. All rights reserved. Keywords: Geographic ontologies; Semantic properties; Semantic relations; Similarity 1. Introduction A close inspection of existent geographic categorizations or geographic data exchange standards shows that although they often refer to apparently similar categories, they use different semantics due to different contexts. This Babel Tower makes the association process and the establishment of an aligned or integrated ontology (Sowa, 2000) very problematic. There have been numerous attempts to deal with the problem of ontology integration and semantic interoperability (Wache et al., 2001; Vckovski et al., 1999; Uitermark, 2001; Kokla and Kavouras, 2001). In this endeavor, it is essential to understand, reveal and resolve existing heterogeneities. In earlier attempts, similarities Corresponding author. Fax: address: mkav@survey.ntua.gr (M. Kavouras). and heterogeneities between geographic categories were identified on the basis of their common attributes (Kavouras and Kokla, 2002). Such attributes were defined based on experts knowledge on the features involved. This approach however involved a great degree of subjectivity, which may contradict the intentions of the original designer. Since definitions have been recognized as an important source of semantics, it was decided to exploit their potential. Besides their semantic value (Jensen and Binot, 1987; Klavans et al., 1993; Swartz, 1997), definitions are rich in different kinds of knowledge such as lexical, world, encyclopedic and semantic (see SIGLEX workshop, in Barriere, 1997). In addition, they are often the only objective available source we can rely on, especially in existing geographic data collections. Furthermore, any findings concerning the semantic completeness of definitions in describing categories would greatly help to form better definitions /$ - see front matter r 2004 Elsevier Ltd. All rights reserved. doi: /j.cageo

2 146 M. Kavouras et al. / Computers & Geosciences 31 (2005) in new classifications as described by Tomai and Kavouras (2004). The purpose of the present research is to identify semantic information from definitions and to enrich the representation of categories with semantic properties and relations, such as those reported by Kokla and Kavouras (2002), in order to disambiguate geographic categories. The ability to represent and visualize the degree of semantic similarity with concept mapping tools (Skupin, 2002; Tomai and Kavouras, 2002) can greatly facilitate the entire process. For tackling these semantic heterogeneities, we explore similarities/dissimilarities of two well-known geographic ontologies, i.e., CORINE LC, MEGRIN, and one lexical ontology WordNet, which includes geographic categories. Another aspect of the research is the representation of semantic similarity, in order to identify semantic heterogeneities, and therefore facilitate interoperability. In the domain of geographic information science, few approaches have attempted to model similarity. As far as similarity of geographic entity classes is concerned, Rodrı guez et al. (1999) and Rodrı guez and Egenhofer (2003) have proposed a computational method for assessing similarity between two ontologies by a similarity function that compares distinguishing features of the entities involved, such as parts, functions and attributes. Our approach to establish semantic similarity among categories from different geographic ontologies exclusively uses semantic information, which can be derived from categories definitions. In summary, this paper compares category definitions, determines heterogeneities, portrays semantic similarity, and overall prepares the ground for the integration process. In the rest of the paper, Section 2 presents the characteristics of the ontologies employed. Semantic relations and properties are described in Section 3. Section 4 is concerned with the computation and visualization of semantic similarity. The results of this workare analyzed in Section 5, and finally some conclusions are drawn in Section Case-study ontologies 1 European Environmental Agency: CORINE Land Cover Methodology and Nomenclature. COR0-part1/en, As mentioned above, several problems of association and integration are encountered when trying to compare categories from distinct repositories of geographic information. In this research, we identified semantic relations properties from the following categorizations: CORINE LC 1 is a categorization intended to provide consistent localized geographical information on land cover of the member states of the European Community, by using satellite data. CORINE Land Cover has a three-level hierarchy of categories. The upper level consists of 5 categories, the middle level of 15, and the lowest one of 44 categories. GDDD-Geographical Data Description Directory (MEGRIN s GDDD) 2 contains information on available digital geographic information from Europe s National Mapping Agencies (NMAs). Layer names, feature type names, and feature attribute types names correspond to the nomenclature used in the DIGEST Feature and Attribute Coding Catalogue (FACC). WordNet 3 is a lexical database for the English language, whose design was based on current psycholinguistic theories. At the context level, the first two categorizations are considered as domain ontologies, while WordNet is a general lexical ontology. At the level of formality, all three categorizations can be considered as terminological ontologies (Sowa, 2000), since they contain categories specified by definitions expressed in natural language, associated by subtype/supertype relations. Furthermore, the first two can be characterized as light ontologies or taxonomies since they establish classifications. Categories at the lowest (and more detailed) hierarchical level were examined for CORINE LC. In addition, for the sake of simplicity and clarity, the study was restricted only to a small, yet representative set of categories from the three ontologies, properly selected to account for a range of heterogeneities encountered between geographic categorizations. Therefore, the categories selected for the experiment were: CORINE LC s categories 4 (wetlands) and 5 (water bodies), MEGRIN s category hydrography, WordNet s definitions for the related category terms. Therefore, we ended up with definitions of 29 category types (Table 1). The term category type refers to categories that are found in different ontologies under the same term (name of the category) but exhibit differences in their definitions or the contexts under which they are used. 3. Determination of semantic relations and properties The rationale behind this research is to determine semantic information from definitions and to enrich the 2 MEGRIN s PETIT project. 3 WORDNET a Lexical Database for the English Language, Cognitive Science Laboratory, Princeton University.

3 M. Kavouras et al. / Computers & Geosciences 31 (2005) Table 1 Category types used in our approach Ontology CORINE land cover MEGRIN WordNet Category type Peat bog Water course Water body Salt marsh Saline Intertidal flat Coastal lagoon Estuary Sea and ocean Inland marsh Bog Canal Lake/pond Salt marsh Salt pan Watercourse Body of water Bog Canal Lake Pond Salt pan Watercourse Watercourse Marsh Estuary Sea Ocean Lagoon representation of categories with semantic properties and relations in order to reveal similarities and heterogeneities. The field of natural language processing develops methodologies for automatic extraction of semantic information from definitions. According to Jensen and Binot (1987), definitions include a wealth of knowledge expressed in natural language, which can be analyzed by natural language processing systems. Definitions are a kind of text with special structure and content. They are rich sources of scientific knowledge of a domain. In geographic ontologies, definitions are the primary and usually the only descriptions of category terms, since other elements that could contribute to the semantic definition of geographic categories (e.g., properties, functions, axioms) are either missing or superficially described. Research on definitions is seeking ways to exploit the wealth of information latent in this special kind of text. Definitions of geographic categorizations are usually comprised of two parts: the genus and the differentiae. The genus or hypernym is the superordinate term of the defined category term. For example, in the definition: hotel: a building where travelers can pay for lodging and meals and other services, building is the genus of category hotel. The differentiae are other elements of the definition apart from the genus, which differentiate words with the same genus. Thus, in the definition: skyscraper: a very tall building with many storeys, skyscraper has the same genus (i.e., building ) with hotel, but they are distinguished by the differentiae (e.g., where travelers can pay for lodging and meals and other services and tall, with many storeys ). The methodology adopted here, for analyzing definitions and extracting semantic information, was introduced by Jensen and Binot (1987) and further pursued by Ravin (1993) and Vanderwende (1995). This approach consists in the syntactic analysis of definitions and the application of rules, which examine the existence of certain syntactic and lexical patterns. Patterns take advantage of specific elements of definitions, in order to identify a set of semantic properties relations and their values based on syntactic analysis. Patterns applied in the genus part of the definition extract the hypernym or is-a relation. Patterns applied in the differentiae part extract other semantic relations such as is-part-of, has-parts, adjacent-to, etc., as well as semantic properties such as purpose, location, time, size, etc. Therefore, it was necessary to specify semantic relations and properties used in geographic definitions. For that reason, different geographic ontologies, standards and categorizations (e.g., CYC Upper Level Ontology, WordNet, CORINE Land Cover, DIGEST, SDTS, etc.) were analyzed in order to identify patterns, which are systematically used to express specific semantic relations and properties. The most commonly used are shown in Table 2. Besides general semantic elements (e.g., PURPOSE, CAUSE, TIME, etc.), other context-specific semantic elements were also identified. For example, categories relative to hydrography are described by semantic elements, such as nature (natural or artificial) and flow (flowing or stagnant). The pattern for the extraction of the semantic relation PURPOSE (Vanderwende, 1995) is: If the verb used (created, intended, prepared, provided, etc.) is post modified by a prepositional phrase with the preposition for, then create a PURPOSE relation with the head(s) of that prepositional phrase as the value. For example, a PURPOSE property is extracted from the definition: canal: a manmade or improved natural waterway used for transportation (MEGRIN), with value transportation. The methodology for extracting semantic information is used to decompose definitions of geographic categories into a set of semantic properties relations and

4 148 M. Kavouras et al. / Computers & Geosciences 31 (2005) Table 2 Examples of semantic properties and relations Semantic properties Purpose Cause Location Time Material-cover Size Semantic relations Is-a Is-part-of Has-part Adjacent-to Surrounded-by Associated-with their corresponding values. This formalized semantic information is further used to disambiguate similar categories by explicitly and objectively identifying similarities and heterogeneities between them. More specifically, if the methodology for extracting semantic information is used for analyzing category lake as defined by MEGRIN: lake/pond: a body of water surrounded by land, the following semantic properties and relations are determined: HYPERNYM with value body, MATERIAL with value water and SURROUNDED-BY with value land. Respectively, from the analysis of same category type as defined by WordNet: lake: a body of (usually fresh) water surrounded by land, the same semantic properties relations and values are determined. Therefore, it is evident that the two ontologies equivalently define the category lake (Table 3). If, however, the above methodology is used for the analysis of category ditch as defined by the same ontologies (MEGRIN and WordNet), the resulting semantic properties relations and values reveal heterogeneities between the definitions of the homonymous categories (Table 4). Table 5 shows the complete set of semantic information (properties and values) that can be identified in the definitions of the 29 categories from the three different ontologies Findings The presence of hypernyms in definitions may express the is-a relation, but the values of hypernyms in definitions differ significantly. A sample of 29 categories from two geospatial ontologies and one lexical database, which all refer to waterbodies and watercourses and coincide in naming (11 naming terms for the 29 categories), present 19 distinct hypernyms. Furthermore, as far as the hypernymic relation is concerned, we can state the following: o CORINE s 10 categories that belong to 4 categories of the intermediate level, which further up belong to 2 categories of the superordinate level, have 9 distinct hypernyms in their definitions. Definitions do not properly address the taxonomic structure of the hierarchy, i.e., genera of category terms do not necessarily coincide with their superordinate category terms. Suggestively, we pinpoint two cases of inconsistency. (a) Definitions are circular (water courses are water coursesy). (b) The use of distinct terms, which could refer to the same hypernyms, for instance, the terms area stretch, zone, expanse, etc. o MEGRIN s 6 category definitions also have 5 distinct hypernyms; one definition is circular (water course is a coursey). o In WordNet this kind of inconsistency is absent (the hypernymic relation is correctly addressed in the definitions). o CORINE s hypernyms do not match those of WordNet at all. All water bodies (such as lagoon, Table 3 Determination of semantic information for category lake HYPERNYM MATERIAL SURROUNDED BY Lake (MEGRIN) Body Water Land Lake (WordNet) Body Water (usually fresh) Land Table 4 Determination of semantic information for category ditch HYPERNYM PURPOSE SIZE NATURE Ditch (MEGRIN) Channel Irrigation or drainage Ditch (WordNet) Waterway Small Natural

5 Table 5 Properties and values of categories as identified in their definitions in three ontologies Categories Semantic information (properties and VALUES) Hypernym Nature Use/purpose Material-cover Is part of Form Size Location Surrounded Condition-state morphology by (attribution) CORINE LC Inland marsh Land Low-lying Flooded (TIME: in winter) Saturated (MATERIAL-CAUSE: water TIME: all year round) Peat bog (Peat) land Decomposed moss and vegetable matter Salt marsh Area Vegetation Low-lying, above the high-tide line Susceptible to flooding (MATERIAL-CAUSE: sea water) Saline Salt-pan (Salt) Active or in process of abandonment Intertidal flat Expanse Water course Water course Natural or artificial Water body Stretch Natural or artificial Coastal Stretch lagoon Estuary Sea Mouth and ocean Zones Water drainage channel Mud, sand or rock (Water) Water Salt or brackish water River Between high and low water marks Coastal areas Seaward of the lowest tide limit MEGRIN Bog Area Soil rich in plant residue Canal Waterway Manmade Transportation (Water) or improved natural Lake/pond Body Water Land Salt marsh Depression Natural Salt encrusted In arid/semi-arid clayey soil regions Salt pan Area Natural surface Flat salt deposits Generally unvegetated Watercourse Course Natural (Water) Flowing Poorly drained periodically flooded M. Kavouras et al. / Computers & Geosciences 31 (2005) ARTICLE IN PRESS

6 Table 5. (continued ) 150 Categories Semantic information (properties and VALUES) Hypernym Nature Use/purpose Material-cover Is part of Form Size Location Surrounded Condition-state morphology by (attribution) WordNet Body of water Part Water Earth s surface Bog Ground Decomposing vegetation Wet spongy Canal Strip Boats or irrigation Water Long and narrow Lake Body (Usually fresh) water Land Pond Lake Small Salt pan Basin Salt and gypsum Shallow In a desert region Watercourse Channel Natural or artificial Watercourse Body Natural Running water On or under the earth Marsh Land Grassy vegetation Low-lying Lagoon Body Water Cut-off from land (ACTOR) a reef of sand or coral Estuary Part Fresh or salt water River Wide Near the sea Sea Division Salt water Ocean Large (Partially) enclosed ACTOR: land Ocean Body Water Hydrosphere Large M. Kavouras et al. / Computers & Geosciences 31 (2005) ARTICLE IN PRESS

7 M. Kavouras et al. / Computers & Geosciences 31 (2005) estuary, sea and ocean, water body, watercourse of category 5) in CORINE LC are defined using terms that refer to two-dimensional hypernyms while in WordNet they are defined using as hypernym the term body that refers to three-dimensional physical objects. This distinction indicates that CORINE LC is taking a map view because it classifies land cover, not geographic entities. (a) CORINE LC is a land cover ontology, subsequently semantic property material-cover is present in most definitions of its categories; therefore definitions in existent geospatial ontologies (esp. taskor domain ontologies) are context driven. (b) The same semantic property, however, is also present in the remaining ontologies (only two of the category definitions do not contain lexical information for that semantic relation). Semantic property nature (artificial/manmade) is addressed in only 7 definitions out of the total 29. Semantic property purpose is present only in 3 out of 29 definitions. This is because natural entities do not have purposes in contrast to artificial ones. Semantic properties such as size and form/ morphology are not adequately included in definitions either: 3/29 and 4/29, respectively. This is very low, considering that geospatial categories are expected to significantly possess properties about size and morphology. The importance of the meronymic semantic relation has-part, or is part-of is not widely addressed in definitions; only 5 of the total 29 definitions present such information, 4 of which belong to WordNet (out of a total of 13). Semantic properties location and surrounded by (both of them denoting topology) are met in 12 definitions only, which again seems low. Both realizations are contrary to what is generally expected about the presence of mereotopologic relations in geographic ontologies (Casati et al., 1998). Semantic property time is also absent to a wide extent from definitions. 4. Determination and visualization of semantic similarity In order to determine the similarity between two categories, we take into account the values of the properties/relations they possess. If the values of a given semantic property or relation coincide, then the two category types are similar in terms of that property/ relation. If the values of a property/relation are distinct, then similarity between the two categories is equal to zero. The similarity measure S between two categories a, b is set by the ratio model (based on Tversky s similarity measure): C Sða; bþ ¼ A þ B þ C ; where C is the number of properties/relations which categories a and b share, but also exhibit common values for, A is the number of properties/relations of category a but not of b, and B is the number of properties of category b but not of a (examples can be found in Table 6). As it can be understood, the ratio is bounded between 0 and 1, the former denoting complete dissimilarity, and the latter, coincidence of entities. In special cases, to assess the similarity appropriately; Compound nouns, such as peatland (does not exist in WordNet) the hypernym of peatbog (Table 6), were identified and used as adjective+noun (peat land) and not as a compound word, so the hypernym was taken to be land instead. Similar terms were grouped to diminish the range of values of certain properties. Consider, for instance, the values water, fresh water, brackish water, salt water of property material-cover for canal, lake, and coastal lagoon, respectively. When establishing similarity, the value for these categories was taken as water. In order to visualize the different ontologies, we use multi-dimensional scaling (MDS) (Kruskal and Wish, 1978). The method uses a similarity/dissimilarity matrix to project the data into the projection space, which in our case is a two-dimensional space. MDS is a dimensionality reduction method that represents multidimensional data sets by using a stress function; therefore, distances among data reflect the corresponding (dis) similarities. The value of the stress function is an indicator of the goodness-of-fit of the result. The higher its value, the more the distortion imposed on the visualization of the entities; therefore, distances are Table 6 Similarity for categories: lake and (peat) bog based on Table 5 Categories Similarity S Categories Similarity S 1 Lake (MEGRIN) S 1;2 ¼ 1:000 Peatbog (CORINE LC) S 1;2 ¼ 0:333 2 Lake (WordNet) Bog (WordNet)

8 152 M. Kavouras et al. / Computers & Geosciences 31 (2005) Bog Bog Salt pan Salt pan Marsh Salt marsh Watercourse Canal Watercourse Estuary Watercourse Body of water Lake/Pond Sea Canal Lake Ocean Pond Lagoon MEGRIN WordNet Fig. 1. Visualization output for three ontologies. greater than the corresponding dissimilarities. The output is a scatter plot of the data where similar entities are close in the representation space while dissimilar ones are far away. The visualization result is shown in Fig Interpretation of results As mentioned before, the output of the MDS is the set of coordinates for the examined caterory types. Subsequently, a clustering method is used to form groups of categories that enjoy common properties relations and values. We are then able to explore whether differentiations in naming denote the same category, while sameness in naming but differentiation of the categories definitions denote distinct ones. In the current approach, we used a hierarchical clustering method to examine which way the three distinct ontologies contribute in the formation of common upper-categories in a unified schema (Fig. 2). The analysis of properties/relations and their values (in the findings of Section 3) indicated whether and which ontological assertions (such as unambiguous taxonomic structure) can be derived from definitions. As a result, the following guiding principles for the definitions of categories in geospatial ontologies could be useful: Basic ontological semantic relations (meronymy, hypernymy, hyponymy) should be present in definitions due to their expressiveness and rich semantics. Category definitions in ontologies should address the taxonomic structure of the categorization correctly. Any inconsistency between the definition s hypernym and the superordinate category term itself (when the categorization is hierarchical) presents a misconstruction in representing the hierarchy. Definitions should account for the so-called special features of geospatial categories such as morphology, location/topology, which, according to the previous analysis, does not seem to be the case. Domain and taskontologies have context-driven definitions, which is not a drawback. These definitions however, should not contradict general knowledge of the given categories and should reflect to some extent the way they are construed by humans, otherwise they are superficial and not widely accepted. 6. Conclusions and further work The research presented focuses on the determination of semantic information from definitions of geographic categories in order to identify and formalize similarities and heterogeneities. Visualization of semantic similarity proves to be a very useful tool for the association of similar categories. Portraying similarities/dissimilarities in a projection space gives us a concrete measure of the heterogeneity of distinct ontologies. We can then draw inferences

9 M. Kavouras et al. / Computers & Geosciences 31 (2005) Fig. 2. Resulting clusters showing heterogeneities among same category types (terms) in different ontologies. as to what extent different ontologies can be integrated, about the associations between category types, concerning the comparison of categories definitions in a cross-ontological examination. The purpose of the present workwas to demonstrate the difficulty in dealing with category semantics. It also presents an alternative to customary approaches, which manually determine similarities and heterogeneities between category types, mainly based on similarity between category terms. However, similarity in terms does not necessarily imply equivalent category types. Besides superficially dealing with categories, such approaches usually result in misapprehending the intentions of the original designers. On the contrary, the present workformalizes semantic information immanent in definitions of category types. Definitions are usually the basic available and semantically rich feature of geographic data collections and they reflect the intentions of original designers. The result of semantic similarity determination and visualization can be used as a pre-processing step to semantic integration of geographic categorizations. Finally, it should be realized that the approach was not intended to produce perfect results. Emphasis was put on objectivity and automation (avoiding ad hoc manual procedures and subjective experts knowledge). Furthermore, any consequential imperfect results have a value of their own, for they reveal (and provide an opportunity to fix) imperfections of the original taxonomy definitions, as well as help engineer better ontologies in the future. Acknowledgments This workhas been partially supported by the Heraclitus Research Programme b of the Hellenic Ministry of National Education. The authors are also indebted to the anonymous reviewers for their very constructive comments. References Barriere, C., From a children s first dictionary to a lexical knowledge base of conceptual graphs. Ph.D. Dissertation. Simon Fraser University, Vancouver, BC, Canada, 339 pp. Casati, R., Smith, B., Varzi, A., Ontological tools for geographic representation. In: Guarino, N. (Ed.), Formal Ontology in Information Systems. IOS Press, Amsterdam, pp Jensen, K., Binot, J.L., Disambiguating prepositional phrase attachments by using on-line dictionary definitions. Computational Linguistics 13 (3/4), Kavouras, M., Kokla, M., A method for the formalization and integration of geographical categorizations. International Journal of Geographical Information Science 16 (5), Klavans, J., Chodorow, M., Wacholder, N., Building a knowledge base from parsed definitions. In: Jensen, K.,

10 154 M. Kavouras et al. / Computers & Geosciences 31 (2005) Heidorn, G., Richardson, S. (Eds.), Natural Language Processing: The PLNLP Approach. Kluwer Academic Publishers, Dordrecht, The Netherlands. Kokla, M., Kavouras, M., Fusion of top-level and geographical domain ontologies based on context formation and complementarity. International Journal of Geographical Information Science 15 (7), Kokla, M., Kavouras, M., Extracting latent semantic relations from definitions to disambiguate geographic ontologies. In: GIScience 2002 Abstracts, Second International Conference on Geographic Information Science. Boulder, CO, pp Kruskal, J.B., Wish, M., Multidimensional scaling. Sage University Paper Series on Quantitative Applications in the Social Sciences, Number Sage Publications, Newbury Park, CA, 96 pp. Ravin, Y., Disambiguating and interpreting verb definitions. In: Jensen, K., Heidorn, G.E., Richardson, S.D. (Eds.), Natural Language Processing: The PLNLP Approach. Kluwer Academic Publishers, Dordrecht, The Netherlands, pp Rodríguez, A., Egenhofer, M., Determining semantic similarity among entity classes from different ontologies. IEEE Transactions on Knowledge and Data Engineering 12 (2), Rodríguez, A., Egenhofer, M., Rugg, R., Assessing semantic similarities among geospatial feature class definitions. In: Vckovski, A., Brassel, K., Schek, H.-J. (Eds.), Interoperating Geographic Information Systems, Second International Conference, INTEROP 99, Zurich, Switzerland. Lecture Notes in Computer Science, vol Springer, Berlin, pp Skupin, A., A cartographic approach to visualizing conference abstracts. IEEE Computer Graphics and Applications 22 (1), Sowa, J.F., Knowledge Representation: Logical, Philosophical, and Computational Foundations. Brooks Cole Publishing Co., Pacific Grove, CA 594 pp. Swartz, N., Definitions, dictionaries, and meanings, Tomai, E., Kavouras, M., Sharpening vagueness: identifying, measuring, and portraying its impact on geographic categories. In: GIScience 2002 Abstracts, Second International Conference on Geographic Information Science. Boulder, CO, pp Tomai, E., Kavouras, M., From onto-geonoesis to onto-genesis : the design of geographic ontologies. GeoInformatica 8 (3), Uitermark, H.T., Ontology-based geographic data set integration. Ph.D. Dissertation. Deventer, The Netherlands, 139 pp. Vanderwende, L., The analysis of noun sequences using semantic information extracted from on-line dictionaries. Ph.D. Dissertation. Faculty of the Graduate School of Arts and Sciences, Georgetown University, Washington, DC, 312 pp. Vckovski, A., Brassel, K., Schek, H.-J. (Eds.), Interoperating geographic information systems. Second International Conference, INTEROP 99, Zurich, Switzerland. Lecture Notes in Computer Science, vol Springer, Berlin, 327 pp. Wache, H., Vögele, T., Visser, U., Stuckenschmidt, H., Schuster, G., Neumann, H., Hübner, S., Ontology-based integration of information a survey of existing approaches. In: Proceedings of IJCAI-01 Workshop: Ontologies and Information Sharing, Seattle, WA, pp