Geographic names have always been important identifiers. People typically use names and not coordinates to identify geographic features. Therefore to establish identity beyond coordinates, name disambiguation is required to identify the exact geographic feature that is denoted by a name. This paper introduces an ontology-based approach to disambiguate geographical names in texts. The ontology defines the central conceptual basis of our approach and is used to rank geographic features based on disambiguation rules that take into account structural information contained in the ontology (e.g. population of a town), as well as textual indicators contained in the text at hand. Our first evaluation on a subset of the well-known Reuters 21578 corpus indicates promising results both in terms of precision and recall.
H.1.m [Miscellaneous]: ontology; H.3.3 [INFORMATION STORAGE AND RETRIEVAL]: Information Search and Retrieval: selection process; I.7.m [DOCUMENT AND TEXT PROCESSING]: Miscellaneous: Document Assignment
Geography has become a popular theme on the Web. Applications like Google Maps are easy to use and incentivise people to publish data tagged with geographical identifiers, i.e. coordinates in form of longitude and latitude, to visualize their data in maps and other representations of geography.
When talking about geographic identifiers, however, people use geographic names to denote a certain coordinate and look up information pertaining to this coordinate using names. Geographic names, like all names, are often highly ambiguous. For example, the name San Jose refers to 1724 different coordinates in the two largest publicly available database of geographic names GeoNet and GNIS (cf. Figure 1).
Figure 1 shows the disambiguation of geographic names is a concrete and relevant task that is needed (I) to attach the appropriate identifier to data containing a geographic name (II) to resolve the appropriate identifier from a geographic name when users retrieve information.
Our paper presents a novel method to disambiguate geographical names based on an ontology. The ontology incorporates data from publicly available geographic gazetteers - the Geonames databases GNS and GNIS (cf. section 2.1) - as well as common world / linguistic knowledge that has been obtained from WordNet. The ontology is used as a gazetteer to map references to (multiple ambiguous) geographic feature candidates from text.
Our disambiguation approach chooses among those feature candidates by dealing with three types of ambiguity: (I) multi-referent ambiguity, when two different geographic locations share the same name; (II) name variant ambiguity, when the same location has different names; and (III) geoname-non geoname ambiguity, where a location name could also stand for some other word such as a person name or nouns, e.g. Metro as the city in Indonesia vs. Metro as the subway system.
Our approach establishes a ranking among feature candidates recognized in a text. To this extent, we are leveraging further information about the geographical name in the given document as well as scoring rules that prefer certain concepts in the ontology over others, e.g. geographic names for cities dominate geographic names for forests or lakes. Thirdly, data available from the ontology, particularly the population of cities is used to give preference to certain candidates to prefer large towns over small towns. These rules reflect natural heuristics as employed by humans. For example, ”Paris” - without further information - more likely refers to the capital of France rather than the small town Paris in Texas.
The paper is organized as follows: Section 2 describes the ontology as well as the data sets used to build the ontology. Section 3 describes our approach of disambiguating geographic names using the ontology. Section 4 describes our evaluation corpus and discusses the results of our evaluation. We conclude with a discussion of related work in section 5 followed by a summary of our contribution and an illustration of next steps in Section 6.
Our ontology is derived from two publicly available data
sources for geographic names:
• The GEOnet Names Server (GNS) [
These data sets have been frequently used as gazetteers by the NLP community. Besides providing a list of names, the data contains several additional information items (that can be used for disambiguation). All names are classified in feature types (e.g. city, park, forest, ...) and information about the containing (political) administrative regions (e.g. county, state and country ) is provided. For unique identification, coordinate information stored with the name can be used.
Figure 1 visualizes the place name disambiguation distribution characteristic for the GeoNet- and GNIS databases. While a geographic name has in average 4.4 different meanings, some names derived from Spanish saints refer to more than 1000 different locations. Table 1 summarizes the size of both data sets in terms of populated places (without location name variants) and the full names of locations in both data sets.
WordNet nouns ambiguous with geo names not ambiguous with geo names absolute relative 114.648 100% 13.169 11,5% 101.479 88,5%
WordNet is a well known lexical resource about English words, which is often used in natural-language processing and information retrieval applications. The core concept in WordNet is the synset. A synset groups words with synonymous meaning. Ambiguity of words is represented by the mapping a Word to multiple Synsets. We are using WordNet 2.0 which contains 114.648 nouns1. 11,5% of those nouns intersect with the geographic names in GNIS and GeoNet (cf. Table 2) and constitute Geo-Non Geo ambiguities. 2.3
Our ontology is created from the above data sources and is represented in the well-known OWL format. We currently only use a subset of the axioms and class descriptions possible in OWL. For our purpose the features already present in RDF Schema suffice. Our ontology uses the following structures:
O := (C, ≤C , R, σ) • two disjoint sets C and R whose elements are called classes and properties, resp., • a partial order ≤C on C, called class hierarchy or taxonomy, • a function σ : R → C ×C called signature of a property All classes are serialized as owl:Class, while all properties are serialized as owl:ObjectProperty or owl:DatatypeProperty.
main and its range by dom(r) := π1(σ(r)) and range(r) := π2(σ(r)).
c2 is a superclass of c1.
then c1 is a direct subclass of c2 and c2 is a direct superclass of c1. We note this by c1 ≺ c2.
Definition 3. A knowledge base is a structure
KB := (CKB , RKB , I, ιC , ιR)
• a set I whose elements are called instances (or
identifiers typically denoted by URI’s),
1[
Ref C (s) := {c ∈ C | (s, c) ∈ Ref C } Ref C−1(c) := {s ∈ S | (s, c) ∈ Ref C } .
An ontology with lexicon is a pair
(O, Lex )
IL := (SI , RI ) where O is an ontology and Lex is a lexicon for O.
KB := (CKB , RKB , I, ιC , ιR) is a pair • a set SI whose elements are called names for instances, • a property RI ⊆ SI × I called lexical reference for instances.
A knowledge base with lexicon is a pair
(KB , IL) where KB is a knowledge base and IL is an instance lexicon for KB .
The lexicon is serialized as RDF triples using rdf:label as the predicate in the triple. 2Which itself has further subclasses such as Sea, Stream or River (not visible in Figure 2) 3http://www.w3.org/TR/2006/ WD-WordNet-rdf-20060619/
The knowledge base of the ontology is created by iterating over all WordNet noun word senses and populating the NounWordSense and NounSynset classes with the exception of word senses of hyponyms of the synset “city” and “country”4. Additionally we iterate over all GeoNet and GNIS entries, while instating the appropriate subclass of GeographicFeature that is indicated by the feature classification of the entry at hand. At the same time properties are established to the related classes such as Country, Coordinate etc.
During the iterations to establish the knowledge base, the (instance) lexicon of the ontology IL and Lex is created from all names (and name variants) in the geographic databases as well as all nouns in WordNet. Additionally, we have populated the lexicon with names that reference classes, e.g. the word ”Sea” establishes a reference to the class ”Sea”, as well as names that reference instances, e.g. the ISO codes for countries reference the respective instance that represents the country.
Our approach for disambiguation establishes a ranking among feature candidates recognized in text. We are leveraging information about the geographical name in the given text and use the ontology as a gazetteer for instance identification. Additionally we utilize scoring rules that prefer certain concepts in the ontology over others, e.g. geographic names for cities dominate geographic names for forests or lakes. The scoring rules reflect natural heuristics as employed by humans and are based on definitions available in the ontology.
The disambiguation process involves several steps: 1. Spotting candidates for geographical identifiers in text utilizing the ontology as a gazetteer 2. Narrowing candidates using surrounding textual information (textual disambiguation) 3. Ranking candidates using ontological information (ontology-based disambiguation) 3.1
To spot candidates, we first transform a given text into a bag of word representation. Therefor a document is represented as a vector D consisting of the terms (t1, . . . , tn) of the document. The original ordering of the terms within the document is preserved. In the formation of D we rely on the built-in gazetteers of the NLP tool, which offer means to spot person names, organization names and stop words5. Such elements spotted by the NLP tool are not part of D. Additionally, we have defined several grammars that suppress the recognition of consecutive terms. For example, the recognition of a person’s first name through the NLP tool will automatically consider the following term as a last name and not add both terms to D.
The ontology is used as a gazetteer utilizing the instance lexicon IL and obtaining references to candidate instances 4Otherwise we would duplicate this information in the ontology 5We utilize the well-known stop word list of the SMART project ftp://ftp.cs.cornell.edu/pub/smart/english. stop i ∈ I using the relation RI . We obtain the set of candidate geographic identifiers in the document D using a function cand(ti) := {i ∈ I|(ti, i) ∈ RI }6. In a similar fashion we identify a set of concepts by utilizing the signs for concepts SC in the lexicon Lex : con(ti) := {c ∈ C|(ti, c) ∈ Ref C }. Note, that cand(ti) will also include references to the nongeographic names which have been incorporated from WordNet.
The second step is textual disambiguation where textual patterns allow to narrow candidates of ambiguous geographic names. We first traverse the term vector D with a window of two consecutive terms (ti,ti + 1) (cf. Figure 3). If candidates for ti refer to instances of geographic feature and candidates for ti + 1 refer to instances of administrative regions or vice versa, the combination of ti and ti+1 is used to narrow the set of candidates for ti to those geographic features that are located in the given administrative region ti+1. For example, the text “Paris, France” will thereby lead to the selection of the French capital.
In a next step we traverse the term vector D for a second time with a larger window of 11 consecutive terms (ti−5, . . . , ti, . . . , ti+5) (cf. Figure 4) and narrow candidates based on geographic feature classes. If candidates for a given term ti are instances of any class referenced by con(tj), where i − 5 ≤ j 6= i ≤ i + 5, the set of candidates is narrowed to the instances of these classes. For example, the text ‘Notthingham forest’ will lead to the selection of the forest in the UK over the city of Notthingham. 3.3
The remaining candidate sets are then ranked utilizing weights that are attached to the classes of the ontology. These weights are propagated to the instances in the candidate set and used to rank those instances. The instance with the highest rank is chosen. If the weight of an instance is negative, we disregard the term as geographic name.
Table 3 shows the weights attached to concepts as used in our evaluation: 6This is a simplified definition. References to instances constituted by multiple consecutive nouns in the text are supported by combining those nouns.
Class weights from table 3 are transitively propagated to their subclasses via ≤C . Class weights are then propagated to class instances via the instantiation function ιc. For all instances of the class PopulatedPlace the number of inhabitants (divided by thousand) is added to the weight of the instance.
Note, that we use negative weights for the WordNet classes. This ranks names from WordNet down and at the same time disregards all terms that have only a WordNet meaning. Finally, we select the instance i ∈ cand(ti) with the maximum score (or randomly choose an instance if there are several instances with the equal maximum score).
For example, the term Lancaster (without any further information such as county or country) would yield a score of 3129 for Lancaster (CA) (a populated place with 129 tsd. inhabitants in California) and a score of 3047 for Lancaster (UK) (the well-known city in the UK with 47 tsd. inhabitants), while all Lancaster counties in the US would only get a score of 1000, and the Lancaster trough in the Lancaster sound in the north of Canada receives -10 points being a subclass of undersea. Hence, the algorithm chooses Lancaster, CA as the instance that provides the most likely geo reference.
To additionally allow to determine a page focus, i.e. choose one geographic reference among all geographic references in the text to describe the main geographic focus of the text, we simply count term occurrences and multiply the weight of the instances with the number of occurrences in the text. We then choose the instance with the maximum weight among all candidate sets.
While Leidner [
We evaluate our approach on both corpora and consider a result as correct, if it corresponds exactly to the user’s input.
Test Run Precision Recall Weights WordNet senses Administrative Region Country Hypsographic Locality PopulatedPlace Road Spotplace Hydrographic Undersea Vegetation 0 -20000 -10000 1000 1000 1000 3000 3000 3000
10 1 10 1000 500 1000 3000 10000 3000
5 10 5 1000 2500 1000 10 10 10 -10 1 -10 10 20 10
We evaluate the success of our approach using the classical precision and recall measures.
• precision = |Gres∩Grel|
|Gres| • recall = |Gres∩Grel|
|Grel| Here, Gres is the set of geographic references identified by the algorithm and Grel is the set of relevant geographic references of the corpus as identified by the annotator. We have carried out three test runs with different weights.
The first run did not weight WordNet terms negatively and therefore leads to a low precision of 40%. This is due to the fact that many terms are incorrectly considered as geographic names. These false positives are constituted by common words, such as the aforementioned example ‘Metro’. The recall is at an acceptable level of 86,7%. Despite the use of such a large gazetteer as the one constituted by our ontology, annotators where (semantically) correctly tagging several organizations such as the ‘Bank of England’ as well as tagging adjectives such as ‘French’ with a geo-reference to the country, while this information is not utilized in our approach.
In the second run, we weight all WordNet terms with a high negative score. Similarly, we choose to increase the weight of populated places, as cities are very common in the corpus. This improves the precision tremendously to almost 70%.
The third run tests the sensitivity of precision to the weight of populated places and WordNet WordSenses and indeed the precision goes back slightly when decreasing the weights of those two classes. Unfortunately, we were not able to improve the precision above the level of test run II in several iterations with other weights.
Looking at the results, we could observe that the identification of country names has much greater precision due to absence of ambiguities. Here, only misspellings and other noise in the data prevent 100% precision. 4.2
Inspired by the good results for country name assignment, we were curious about the results for page focus assignment limited to country names. While we can achieve 100% precision our recall is limited to 90,9%. Hence, our simple approach to count the occurances of country names already leads a very high recall, only in seldom cases the human annotators chose to identify countries with lower frequency as the page focus of the annotated texts.
Ontologies have been used in geographic information
system to merge data from different sources [
The Perseus project [
[
The disambiguation method described in [
MetaCarta [
The disambiguation approach of [
We have captured many of the ideas for disambiguation heuristics presented in the above-mentioned papers in our disambiguation rules. Our use of ontologies generally increases the flexibility and allows to easily incorporate further knowledge into the disambiguation process. Similarly, it allows to flexibly change disambiguation rules that are defined on top of the ontology, e.g. to provide the recognition of person names. 6. 6.1
We have presented a novel approach for disambiguation of geographical names based on ontologies, which allow us to formulate ranking rules based on concepts and utilize the lexicalization of the ontology as a gazetteer. The geonames ontology constructed for our purpose of disambiguation is used as a gazetteer to map textual references to instances and classes in the ontology. Disambiguation rules in form of rankings are based on attaching weights to concepts and propagating weights to the instances of concepts. Data from the ontology can be utilized for the ranking as well (such as the population of towns). The ontology also finally establishes unique references for geographic identifiers in order to be reused in other applications in form of URI’s. This provides a basis for both data integration as well as natural extension points of the ontology. 6.2
Going forward we see several next steps: First, we want to evaluate our approach large scale by disambiguating geographic references in arbitrary RSS feeds and utilizing this information to filter news articles by regions referenced. Technically, RSS feeds will be tagged by geographic identifiers.
Second, as part of this application, we will incorporate means for users to provide feedback on the correctness of the disambiguation. This information will be used to improve our algorithms using statistical machine learning techniques. We will also use the ontology beyond its current gazetteer function and display information items related to ontology instances (such as country and state located in, neighboring geographic features, facts such as population, altitude, alternate spellings in different languages, etc.). We also plan to utilize the extensibility of the ontology by incorporating further information sources such as the dbpedia7 project, which has created an ontology for facts from Wikipedia. The 7http://dbpedia.org/ currently used expressiveness power of the ontology will be extended by considering more than the currently used object property relations between concept nodes for determination of the correct meaning of the geographic identifiers.
Thirdly, in the process of spotting the meaningful candidate terms for geographic references, the sliding window algorithm will be further improved. Additional to the estimation of an accurate window size, we also consider the use of further text patterns to improve the introduced algorithm.
Concerning the suggested ranking algorithm, we plan a dynamically adjustment of the ontology weight values. In a next step we are going to an iteratively refinement of the initial measure through consideration of already estimated documents over the corpus. We also plan to improve our page focus algorithm by consideration of spatial distances between the identified geographic entities.
Beyond this application, we will look for further disambiguation objectives. Many cases of incorrect assignment of geographic names are due to organization names such as ”Texas Instruments”. We will again try to tackle this issue by extending the ontology, using it as a gazetteer and defining appropriate heuristics. We assume this will also allow us to improve the page focus task in many cases, as our annotators have assigned page focus based on organization information, e.g. ”Bank of England” → ”UK” or ”OECD” → ”Europe”.
Acknowledgments. We thank our colleague Frank Kleiner for his technical support and our students for the annotation support.