As the body of knowledge available as linked data grows, so does the need to provide methods that make this knowledge accessible for humans. Such methods usually require knowledge about how the vocabulary elements used in the available ontologies and datasets are verbalized in natural language. This has lead to much interest in the development of models and frameworks for publishing ontology lexica as linked data. In this paper we describe a process for the manual development of such lexica in lemon format and illustrate some of the key challenges involved. As a proof of concept, we provide a manually created English lexicon for the DBpedia ontology and describe its rst release.
As the body of knowledge available as linked data grows, so does the need to
provide methods that make this knowledge accessible for humans, for example
by systems that transform natural language questions into SPARQL queries,
systems that generate natural language paraphrases of SPARQL queries, or systems
that generate verbalizations of a given ontology or natural language summaries
from RDF datasets. All such systems require knowledge about how the
vocabulary elements used in the available ontologies and datasets are verbalized in
natural language, in particular covering di erent verbalization variants, possibly
in multiple languages. Although standards such as RDFS and SKOS, or models
such as OTR [
In order to capture linguistically rich information about verbalizations of
simple and complex elements of an ontology or dataset, for example specifying
arguments as optional or capturing restrictions on the usage of verbalizations,
lexical knowledge is needed. Moreover, this lexical knowledge should become part
of the linked data cloud itself, in order to avoid that it has to be recreated by
every application that wants to use it. The lemon model1 [
As proof of concept for a lexical layer enriching the linked data cloud, we manually developed an English lexicon for the DBpedia ontology in lemon format. In this paper we describe the release of the rst version of this lexicon, illustrate the methodology that has lead to its creation and summarize the challenges in creating such a resource. Finally, we give an outlook on future plans and invite NLP and Semantic Web researchers to use the lexicon, improve and extend it, and to create similar lexical resources for other domains. 2
The DBpedia ontology2 currently comprises 359 classes and 1,775 properties. Since the manual creation of lexical entries is an e ort-intensive process, our approach to creating a lexicon for the DBpedia ontology is an iterative one. The rst step consists in covering all classes and those properties that are most frequent with respect to the number of occurrences in triples in the DBpedia dataset. Later, we will successively extend the lexicon to also cover the tail of less frequent properties, ideally with support from the community.
The rst release of the English DBpedia 3.8 lexicon comprises lexicalizations of 354 classes and 300 properties. The covered classes are complete except for one abstract class (PersonFunction) and a few classes without any instances (e.g. NoteworthyPartOfBuilding). The covered properties comprise all those with more than 10,000 occurrences in the DBpedia dataset. Excluded were the Wikipedia-speci c ones (e.g. wikiPageRevisionID and thumbnail), abstract properties (e.g. leaderFunction), and properties for which no straightforward lexicalization was found (e.g. the datatype property strength, relating a battle to the numerical sizes of the involved military units).
The rst release of the lexicon thus covers 98 % of the classes and approximately 20 % of the properties. We plan to successively extend the current lexicon,
so that a second release will cover all properties with at least 1,000 occurrences (752 properties, i.e. 42 %), and a third release will cover all properties with at least 100 occurrences (1,112 properties, i.e. 63 %).
The lexicon currently contains 1,217 entries (443 class lexicalizations and 774 property lexicalizations), which amounts to approximately 1.8 entries per ontology concept (1.3 per class and 2.4 per property). The distribution of the number of lexicalizations per entry is depicted in Figure 1, where the x-axis speci es the number of entries, up to 5, and the bars specify how many concepts have this number of lexicalizations.
All lexical entries were created
manually by two of the authors, fa- 277
miliar with both DBpedia and lemon,
partly by manually selecting the most
frequent patterns from the
WikiFramework [
Most of the entries were constructed using a domain-speci c language3 for
common lemon design patterns [
Of all lexical entries, 54 could not be captured by the lemon design patterns but only by writing lemon RDF triples. This is mostly the case for constructions, such as X has Y inhabitants as verbalization of population, or X consists to Y percent of water as verbalization of percentageOfAreaWater, and for entries with a compound meaning, i.e. a sense that consists of several subsenses. An example is the verb link to as in The Autostrada A19 links Palermo to Catania, which refers to both routeStart and routeEnd. Both constructions and compound meanings in patterns will be subject of future developments. 3 https://github.com/jmccrae/lemon.patterns
In case a lexical entry verbalizes a concept that is not named in the ontology, it is de ned in the lexicon. In total, the lexicon speci es 89 classes and 2 properties in addition to the DBpedia concepts. An example are the classes of all male or female persons, which are not part of the DBpedia ontology but can straightforwardly be de ned as the restriction classes of all things with gender either Male or Female, e.g.: lex:Female rdf:type owl:Restriction ; owl:onProperty dbpedia:gender ; owl:hasValue resource:Female .
These classes can be verbalized as man and woman. Additionally they can serve as domain or range restrictions, as e.g. in the verbalization daughter of, referring to the property child with its object restricted to the class Female, as shown in Figure 3.
Other common cases of compound meanings are properties for which not the verbalization of the property itself but rather the verbalization of the property together with a particular object is relevant. Examples are nationalities (e.g. Dutch being the class of all persons related to the resource Netherlands through the property nationality), occupations (e.g. Surfer being the class of 1 2 3 4
dbpedia : conservationStatus ," EX ")
all persons related to the resource Surfing through the property profession), and religions (e.g. Buddhist being the class of all persons related to the resource Buddhism through the property religion). Another example is the datatype property conservationStatus, which characterizes a species as endangered, extinct, or the like. Since the conservation status code is not meaningful for non-experts, verbalizations such as endangered and extinct, as given in Figure 4, should be included in the lexicon.
Although our DBpedia lexicon currently covers only the top 30 % of all URIs in the DBpedia ontology, it can already prove useful for NLP applications. As a rough estimation, our lexicon provides verbalizations for 91 of the 104 di erent URIs used in the QALD-34 query set for question answering over linked data (i.e. for 95 %).
Note that the provided lexicon includes only schema but no instance data.
For lexicalizations of the nearly 3 million individuals, for example the DBpedia
lexicalization dataset [
The rst version of the DBpedia lexicon is released at http://lemon-model. net/lexica/dbpedia_en/, under the Creative Commons license CC BY 3.05. It is accessible as open source on GitHub, at https://github.com/cunger/ lemon.dbpedia, allowing others to improve and extend the lexicon as well as to port it to other languages. In the future, we plan to release the lexicon also under lemon source6, a collaborative web interface for creating and editing lexica. Possibly even the DBpedia community could be involved in lexically enriching the ontology in multiple languages when editing the ontology wiki.
In order to get an idea of the coverage of the lexicon and its usefulness for NLP applications like question answering and natural language generation, we provide a demo at purl.org/3dlt/demo.
In future releases we will focus much more on a semi-automatic approach to
creating lexical entries, as described in [
Furthermore, we are creating a rst version of a German and Spanish DBpedia
lexicon, based on an automatic translation of the English lexical entries (cf. [
We described the rst release of a manually constructed English lexicon for the DBpedia ontology in lemon format. The focus is on high quality entries, covering especially those lexicalizations that di er from the provided RDFS label, and those that verbalize complex constructs and thus cannot yet be handled by automatic lexicalization methods.
We hope that the rst release of the DBpedia lexicon serves to prove the usefulness of rich lexical knowledge for NLP applications, and inspires NLP and Semantic Web researchers to improve and extend the lexicon, port it to other languages, as well as build and share lexica for other domains, thereby enriching the linked data cloud with a lexical layer.
Acknowledgment This work was partially funded within the EU projects PortDial (FP7-296170) and Monnet (FP7-248458).