Abstract
We introduce lemonUby, a new lexical resource integrated in the Semantic Web which is the result of converting data extracted from the existing large-scale linked lexical resource UBY to the lemon lexicon model. The following data from UBY were converted: WordNet, FrameNet, VerbNet, English and German Wiktionary, the English and German entries of OmegaWiki, as well as links between pairs of these lexicons at the word sense level (links between VerbNet and FrameNet, VerbNet and WordNet, WordNet and FrameNet, WordNet and Wiktionary, WordNet and German OmegaWiki). We linked lemonUby to other lexical resources and linguistic terminology repositories in the Linguistic Linked Open Data cloud and outline possible applications of this new dataset.
Introduction
Recently, the language resource community has begun to explore the opportunities offered by the Semantic Web, lead by the formation of the Linguistic Linked Open Data (LLOD) cloud and an increasing interest in making use of Linked Open Data principles in the context of Natural Language Processing (NLP) and Linguistics [7]. The use of RDF supports data integration and offers a large body of tools for accessing this data. Furthermore, the linked data approach gives rise to novel research questions in the context of language resources and their application.
For lexical resources, data integration has been in the focus of interest for many years, resulting in numerous mappings and linkings of lexica, as well as standards for representing lexical resources, such as the ISO 24613:2008 Lexical Markup Framework (LMF) [13]. In this context, the LLOD cloud can be considered as a new data integration platform, enabling linkings not only between lexical resources, but also between lexical resources and other language resources.
We extend the LLOD cloud by a new lexical resource called lemonUby1
While many lexical resources have already been included in the LLOD cloud, e.g., [3,10,18,21,22], the LLOD cloud is still missing a large-scale lexical resource rich in lexical information on verbs, including aspects such as syntactic behaviour and semantic roles of a verb’s arguments. Such information is crucial for lexicalizing relational knowledge, e.g., the relation
The new resource lemonUby addresses this gap: Along with resources for word-level semantics (WordNet [12], English and German Wiktionary,3
The core of the lemon model.
There has been significant work towards integrating lexical resources using RDF and Semantic Web principles [6], and many resources are already available as Linked Data. Yet, representing lexical resources in RDF does not per se make them semantically interoperable. Consider, for instance, existing conversions of WordNet and FrameNet [24,27], where a simple mapping to RDF is provided, and augmented with OWL semantics so that reasoning could be applied to the structure of the resource. However, the formats chosen for the RDF versions of WordNet and FrameNet are specific to the underlying data models of WordNet and FrameNet. Although these lexicons are complementary resources [1], it is difficult (i) to link them in this form on the Semantic Web, and (ii) to use them as interchangeable modules in NLP applications.
In order to overcome this difficulty, the lemon model [19] was proposed as a common interchange format for lexical resources on the Semantic Web. lemon has its historical roots in LMF and thus allows easy conversion from LMF-like, non-linked data resources. It links to data categories in annotation terminology repositories, and most of all, it realises a separation of lexicon and ontology layers, so that lemon lexica can be linked to existing ontologies in the linked data cloud.
This core model is illustrated in Fig. 1, which defines the basic elements used by all lexica published as linked data. In addition to this there are a number of modules used to model linguistic description, syntax, morphology and relationships between lexica.5
More detail of the model and descriptions of the modules can be found at
lemon has been used as a basis for integrating the data of the English Wiktionary with the RDF version of WordNet [18]. lemon’s similarity to the WordNet model made this conversion straight-forward, with only the need for a slight change in modelling to accommodate inflectional variants of lexical entries.
UBY is both a network of interlinked lexical-semantic resources and a project on continuous integration and linking of lexical resources for NLP applications. It is motivated by the observation that an essential requirement in NLP is the availability of a wide range of lexical resources that can be used for many different NLP tasks. In a continuous process, such resources are integrated into UBY by means of (i) making them interoperable and (ii) linking them to other resources in UBY at the sense level.
In UBY, interoperability is achieved by standardizing lexical resources according to UBY-LMF [8,9], a lexicon model which is an instantiation of LMF, specifically designed for NLP. The lexicon model UBY-LMF has been developed to fully cover a wide range of heterogeneous lexical resources without information loss, which resulted in a fine-grained model of lexical information types (documented by data categories from ISOcat,6
The mapping from UBY-LMF to lemon is motivated by an increase in interoperability with the Semantic Web and its resources, thereby making it available to a new group of potential users and novel applications. Beyond this, mapping UBY-LMF to lemon is an interesting task per se, because lemon links lexical resources and ontologies, whereas UBY-LMF is not related to any ontology. Another benefit is that LMF is not an open standard (in the sense that its specification is not freely available), while lemon fully complies with open data and open access principles.
To automatically convert data from UBY-LMF (extracted from UBY) to lemon, we performed a mapping of UBY-LMF elements to lemon concepts and properties. This involves two aspects: Formalizing and mapping UBY data structures in a Semantic-Web compliant way, and the actual conversion of UBY data.
Modeling data categories: ubyCat
In UBY-LMF, most of the terms used and many features (e.g., the feature “partOfSpeech” and the corresponding attribute values) are linked to community-maintained data categories in ISOcat.7
However, ISOcat is not a formal ontology, but only a semistructured collection of terms, and while it serves as a repository of definitions, it does not provide a formal data model that can be applied to a resource: ISOcat contains doublets created by different data providers, and such superficially similar categories may actually have incompatible definitions, e.g., gerundive [DC-1294] is an “adjective formed from a verb” (excluding verbal nouns), whereas gerundive [DC-2243] is a “non-finite form (…) other than the infinitive” (including verbal nouns). Hierarchical relations between ISOcat terms are possible, but not obligatory, and when compared with a full-fledged ontology, ISOcat terms that represent superconcepts for a bundle of features do not distinguish relational and categorial aspects.
To render UBY/ISOcat data categories in a formal data model, we created the OWL/DL ontology ubyCat8
With respect to data categories, ubyCat is linked to the Ontologies of Linguistic Annotations (OLiA) [4,5],10
The actual conversion of UBY data was achieved by means of an XML style sheet transform11
The XSL file can be downloaded at https://raw.github.com/jmccrae/lemon.api/master/src/main/resources/xslt/ubylmf2lemon.xsl.
The following data from UBY were converted: WordNet, FrameNet, VerbNet, English and German Wiktionary, the English and German entries of OmegaWiki, as well as links between pairs of these lexicons at the word sense level:
cross-lingual automatic sense links between WordNet and German OmegaWiki [14], and manual inter-language links already given in OmegaWiki.
Number of triples and links for each resource
The XML style sheet mapping between the two lexicon models UBY-LMF and lemon revealed a number of differences between them. Most differences are due to the fact that lemon is a model for ontology lexica where the lexicon and ontology layers are kept separate. Thus, sense representations in lemon primarily consist of references to the associated ontology where a rich and domain-specific sense definition is provided. UBY-LMF, on the other hand, represents fine-grained sense information in the lexicon itself.
In line with previous authors [18], synsets in UBY-LMF (provided in WordNet and OmegaWiki) are considered to be SKOS concepts and referenced like ontology classes in lemon. Hypernym relationships stated in the lexicon are mapped directly to broader/narrower relationships in the ontology.
The “SemanticPredicate” class in UBY-LMF is used to represent semantic frames from FrameNet [9]; frames group senses which evoke the same kind of situation with participants taking over particular roles. Thus, senses in a FrameNet frame are semantically related, but not synonymous; e.g., the verbs love and hate are both in the same FrameNet frame. In lemon this relation is represented as subclass of the general lexical sense and linked to the rest of the lexicon by a broader (sense) relationship.
VerbNet verb classes and their hierarchical organization correspond to “Subcat[egorization]FrameSet” in UBY-LMF. VerbNet classes group verbs that share the same syntactic subcategorization frame, semantic roles, selectional restrictions, and semantic predicate. Although the resulting verb classes are semantically coherent, the semantic relatedness of verb senses in a VerbNet class is much more distant than in FrameNet, e.g., the verbs believe, swear and doubt are in the same class. Verb classes in lemon are represented by a broader sense as well, so lemon collapses the distinction between FrameNet-style and VerbNet-style sense groupings.
Translations to other human languages which are encoded as string-valued lexical items in the “Equivalent” class are mapped as a datatype property on the sense, in line with the UBY representation [17] rather than attempting to construct a cross-lingual linking for these terms as in the original resources (OmegaWiki and Wiktionary). Finally, provenance information is added to each resource at the root URL, including the source resource and version, time of creation and a link to the FOAF profile of the executor. Further examples of usage and modelling are available from the website.14
See, e.g.,
We linked lemonUby to other related resources in the LLOD cloud in order to make it accessible from existing datasets on the Semantic Web.
Linking ubyCat with repositories of linguistic terminology
We linked the data categories in ubyCat with the OLiA Reference Model, which provides formal definitions of annotation schemes for various linguistic phenomena in about 70 languages as OWL/DL ontologies, and which is further linked with other terminology repositories such as ISOcat and GOLD [11].
The ubyCat ontology plays an important role for the linking of UBY data categories to OLiA, because it defines grammatical concepts used in UBY-LMF in a formal data model. The linking between ubyCat and the OLiA Reference Model is implemented by subClassOf relationships between its concepts and the Reference Model. In OLiA, this mechanism is generally applied to physically separate resource-specific, interpretation-independent information (e.g., from an annotation scheme, hence the term ‘OLiA Annotation Model’) from resource-independent terminology (provided by the OLiA Reference Model) and its interpretation in terms of the latter (‘Linking Model’).
The separation of interpretation and interpretation-independent information by means of a declarative Linking Model15
Another advantage of an ontological formalization of linguistic terminology is that it is more expressive than a plain hierarchy or list of terms. An example for a complex linking is the verb form annotation ‘ingForm’, used for verb forms like talking in the resource. This category represents a language-specific merger of present participles (he is
It should be noted that such a complex linking requires the use of operators like and and or in the linking, to capture this information, OLiA ontologies employ OWL/DL. The direct mapping between annotations and reference concepts originally advocated for ISOcat and GOLD cannot represent this information.
Through OLiA, grammatical information from lemonUby is interoperable with other LLOD resources linked to either OLiA or any of the terminology repositories it is linked with, including GOLD and ISOcat.
Number of external links created between lemonUby and other resources in the LLOD cloud
As UBY is derived from existing lexical resources, the simplest links to create are those to other RDF versions of the resources that compose UBY. For WordNet, these links are simply created by mapping the data of UBY, which uses WordNet 3.0, to the linked data version of WordNet 3.0.17
In addition, we provided links at the lexical entry level to two existing resources that are also widely used, i.e., RDF WordNet 2.0 [27] and an RDF instantiation of Wiktionary.18
For the RDF WordNet 2.0, we created such a simple linking at the lexical entry level, because a linking at the sense level based on the sense identifiers is not possible due to different WordNet versions using different sense identifiers.
The linking of the Uby version of WordNet and the RDF export of Wiktionary at the lexical entry level is based on the assumption that corresponding word classes share the same part-of-speech category in both resources. While this is mostly true, we found a few word classes where we had to manually unify diverging part-of-speech categories; these word classes were initially not covered by the linking. For instance, for acronyms (e.g. IBM), Wiktionary assigns “Initialism” as a part-of-speech, whereas WordNet counts these as nouns. Statistics for all mappings including the coverage of the linkings are given in Table 2.
As a web resource, lemonUby provides two main possible applications: firstly as a resource to allow better lexical description of ontology entities and secondly as a resource for NLP applications.
As a wide coverage resource containing lexical information, it would be possible to link existing ontologies and terminology vocabularies to lemonUby and then obtain rich information about the linguistic usage of a given term. As an example, consider the cross-lingual linking of verb senses across English and German available in lemonUby. Verb senses in the German OmegaWiki can be enriched by semantic role and selectional preference information from VerbNet and FrameNet via the cross-lingual linking between German OmegaWiki and WordNet, and via the WordNet–VerbNet and WordNet–FrameNet linking. Currently, no German lexica with these information types are freely available for research purposes.
As a resource for NLP applications, lemonUby can be easily deployed into existing applications [26]. The wide variety of lexical information types it offers, ranging from taxonomic relationships (e.g., hyponomy) and translations to fine-grained lexical-syntactic information, makes it an attractive resource for many different NLP tasks, such as Entity or Predicate Disambiguation, to name only two.
Recently, it has been shown that lemonUby is a valuable resource for machine translation [20]. By mining translations from lemonUby and combining them with the Moses statistical machine translation, the performance of the resulting machine translation system improved.
Furthermore, all the data from lemonUby is available via a SPARQL endpoint.19
We presented lemonUby, a new linked data resource which combines data from the standardized lexical resource UBY with the principled model, lemon, for representing lexical data on the web. This resource provides not only rich information about many lexical entries in two languages, it is also linked at the word sense level both within its component resources and to other language resources on the web.
We see two main directions for future research on lemonUby: first, the further linking of this dataset both within lemonUby and to other resources on the web, and second, the use of lemonUby in RDF-based NLP applications in order to gain a deeper understanding of the added value of linked lexical resources for NLP.
Acknowledgments
The work on lemon was developed in the context of the Monnet project, which is funded by the European Union FP7 program under grant number 248458, the CITEC excellence initiative funded by the DFG (Deutsche Forschungsgemeinschaft). The work on UBY has been supported by the Volkswagen Foundation as part of the Lichtenberg-Professorship Program under grant No. I/82806. The research of Christian Chiarcos has been partially supported by a German Academic Exchange Service (DAAD) funded PostDoc fellowship at the University of Southern California. We thank Silvana Hartmann, Sebastian Hellmann, Michael Matuschek, Christian M. Meyer, and Zijad Maksuti for their contributions to this work. Our special thanks go to the three reviewers Tadej S̆tajner, Vojtĕch Svátek and Fabian M. Suchanek for their detailed and valuable feedback.