A huge amount of datasets on the Semantic Web are linked to a few datahubs, the most prominent of which is DBpedia. What makes the exploitation of DBpedia challenging for natural language-based applications, however, is that such NLP applications require knowledge about how the ontology elements are verbalized in natural language. In order to provide such knowledge at the required scale and thereby leverage the use of DBpedia in di erent applications, we construct a lexicon for the DBpedia 2014 ontology by means of existing automatic methods for lexicon induction. It contains 11,998 lexical entries for 574 di erent properties in three languages: English, German, and Spanish. Just like DBpedia provides a hub for Semantic Web datasets, this lexicon can provide a hub for the lexical Semantic Web, an ecosystem in which ontology lexica are published, linked, and re-used across applications.
The amount of datasets on the Semantic Web is evermore increasing, and large
part of these datasets are linked to central hubs. The biggest and arguably most
important of those hubs is DBpedia [
For an example of a lexical entry consider the property http://dbpedia. org/ontology/spouse from the DBpedia 2014 ontology. This property expresses
that two persons are married to each other. The property can be expressed in natural language by the following expressions (lexical entries).
{ X is the husband of Y e.g. Barack Obama is the husband of Michelle Obama. { X is the wife of Y e.g. Michelle Obama is the wife of Barack Obama. { X is married to Y e.g. Barack Obama is married to Michelle Obama.
All these lexical entries can be interpreted as expressing the spouse property. We say that the property spouse is the reference of the lexical entry.
To our knowledge, DBlexipedia is the rst wide-coverage lexicon of DBpedia. It contains 11,998 lexical entries for 574 di erent RDF-properties from the DBpedia 2014 ontology in three languages: English, German, and Spanish.
In contrast to an earlier manually crafted DBpedia lexicon [
The remainder of the paper is structured as follows. In the next section, we present our approach to generating ontology lexica, followed by a description of the resulting lexicon in Section 3. We conclude with perspectives for future work. 2
The lexicon published on http://dblexipedia.org is the result of applying
M-ATOLL2 [
In the following we brie y explain the main corpus-based approach as well as an extension of it that deals with a special case of adjective entries. We call this second approach label-based, as we use the label of a property in combination with machine learning techniques to generate the entries. 2.1
M-ATOLL takes as input an ontology and a dependency parsed text corpus in
the target language. For DBlexipedia, the input was the DBpedia 2014 ontology
and a Wikipedia text corpus parsed for three target languages: English, German,
and Spanish. As dependency parser we use the MaltParser [
M-ATOLL performs three steps in order to nd lexicalizations of ontology properties in the accompanying text corpus: 1. Retrieving all triples for a given property from the ontology. For example, the results for the property spouse include the triple <Barack Obama,spouse, Michelle Obama>.
2. Retrieving all sentences from the parsed text corpus which contain mentions of the subject and object of the triples discovered in Step 1. 3. Searching for prede ned patterns in those sentences, in order to extract candidate lexicalizations of the property.
So far, M-ATOLL covers entries that describe transitive verbs (e.g. to cross), intransitive verbs with a prepositional object (e.g. to live in), relational nouns with prepositional object (e.g. capital of ), and relational adjectives (e.g. similar to) in all three languages: English, German, and Spanish. Important to note is that one entry can have multiple references, e.g. the relational noun entry village in3 can refer to hometown, birthplace, and location, among others. 2.2
Examining the lexical entries created by the above approach and the kind of lexicalizations needed for question answering, for example, it is clear that MATOLL does not yet cover all relevant patterns. Consider the request Give me all female Danish politicians. In this case a lexical entry for female is needed, which w.r.t. DBpedia refers to all individuals that are related to the resource Female by means of the property gender. Similarly, Danish refers to all individuals that are related to the resource Denmark by means of the property country or birthPlace.
As these adjective lexicalizations are not handled by the standard
corpusbased approach, we extended M-ATOLL with a dedicated module, extending
[
Resulting entries are, for example, the above mentioned female4, and blue5.
For more information about this approach and the implemented features, see
[
This approach currently works only for English, but will be soon adapted to German and Spanish. 3
The dataset we present in this paper is the result of the above approaches and contains entries in three languages: English, German, and Spanish. Table 1 shows how many properties were lexicalized for each language. Overall, 574 di erent properties were lexicalized, but note that not every property is covered for each language. This is due to the fact that for some properties no sentences are found 3 http://dblexipedia.org/LexicalEntry_village_as_Noun_withPrep_in 4 http://dblexipedia.org/LexicalEntry_female_as_AdjectiveRestriction 5 http://dblexipedia.org/LexicalEntry_blue_as_AdjectiveRestriction in the text corpus that match the prede ned lexicalization patterns. For English, 567 properties were lexicalized, 224 by the corpus-based approach and 445 by the label-based approach. For German and Spanish, 145 and 50 properties, respectively, were lexicalized using the corpus-based approach (the label-based approach is currently not implemented for these languages).
Table 2 shows the number of entries generated by the di erent approaches for the di erent languages. Overall our dataset contains 11,998 entries. It is important to mention that one lexical entry can lexicalize multiple properties.
English German Spanish
Corpus-based Label-based Total 224 445 567 145 n.a. 145 50 n.a. 50
Attached to the entries we also publish meta-data, in particular about provenance, specifying from which pattern an entry was created (for the corpus-based approach), with which frequency, and with which con dence (for the label-based approach). In the future we also intend to include example sentences for each entry (for the corpus-based approach) and the set of corresponding features which led to the entry (for the label-based approach).
Moreover, the lexical entries are linked to dbnary6 and lemon UBY 7, considering the canonical form and the part of speech as relevant information for comparison.
The dataset is published at http://dblexipedia.org, which was generated by a modi ed version of YUZU 8. The website enables a user to browse through the lexical entries and supports search over them. The whole dataset can also be downloaded at http://dblexipedia.org/public/all.nt.gz (as N-Triples). In
addition to the dataset, the most current version of M-ATOLL is available at http://dblexipedia.org/public/MATOLL.jar.
We compared the published entries of the English corpus-based approach
with those of the manually created DBpedia lexicon [
The main limitation of the presented dataset is the small coverage of the DBpedia 2014 ontology, containing around 2796 properties. However, for 1424 properties of this ontology, no data is available at the o cial DBpedia SPARQL endpoint. Considering only properties with at least one received data item, our very rst release of the lexicon covers already 42% of the properties. 4
In this paper we presented the rst multilingual, automatically generated lexicon
for DBpedia, covering 574 properties from the DBpedia 2014 ontology. The
evaluation showed that on average half of the generated entries are correct. In order
to further improve the quality of the resulting lexica across languages, we will
work on the adaptation of the label-based approach to German and Spanish, and
in the future also Japanese. Moreover, we plan to evaluate to which extent an
increase in coverage can improve tasks like question answering (see, e.g. [
This work was supported by the Cluster of Excellence Cognitive Interaction Technology CITEC (EXC 277) at Bielefeld University, which is funded by the German Research Foundation (DFG).