This document is aimed to guide in the process of creating a Linked Data (LD) version of a wordnet, and may be of interest for other kinds of resources. These guidelines contain advice on the vocabularies selection, RDF generation process, and publication of the results. As result, the converted language resource is more interoperable and easily accessible on the Web of Data by means of standard Semantic Web technologies. This document describes the models used and the recommended format (WN-JSON-LD) for representing wordnets on the Web that is based on lemon. We then describe the InterLingual Index, a common resource for providing interlingual links between wordnets, which is adminstered by the Global WordNet Association.

This document was published by the Best Practices for Multilingual Linked Open Data community group. It is not a W3C Standard nor is it on the W3C Standards Track.

There are a number of ways that one may participate in the development of this report:

• Mailing list: public-bpmlod@w3.org
• Wiki: Main page
• More information about meetings of the BPMLOD group can be obtained here
• Source code for this document can be found on Github.

What is WordNet?

WordNet is one of the most widely used lexical resources within natural language processing. From the time since the first version of WordNet was released, many resources have been produced that represent complementary information to WordNet or extend it to other languages .

WordNet is a large lexical database of English nouns, verbs, adjectives and adverbs. Word forms are grouped into more than 117,000 sets of (roughly) synonymous word forms, so called synsets. These are interconnected by bidirectional arcs that stand for lexical (word-word) and semantic (synset-synset) relations, including hyper/hyponymy (tree-oak), meronymy (tree-branch), antonymy (long-short) and various entailment relations (buy-pay, show-see, untie- tie).

WordNet’s synsets and its network structure yield a rough measure of semantic similarity among words and concepts in terms of synset membership as well as the number of arcs separating synsets. Due to its availability under open licenses, WordNet has become a popular tool for Word Sense Disambiguation (WSD) and Natural Language Processing in general. WordNets have been built for around 100 different languages. Most are mapped onto the Princeton WordNet, enabling translation on the lexical level as well as cross-lingual WSD and applications. WordNet continues to evolve both in terms of coverage and representation of meaning. Recent enhancements include the addition of internet language and partially compositional multi word units. Finally, WordNet has been mapped to formal ontologies, including SUMO and KYOTO.

Selection of vocabularies

lemon

lemon is a model that has been proposed for the representation of lexicons relative to ontologies. As such, this model is well suited to the representation of semantic networks such as WordNet and defines many useful features for linking a WordNet to wider objects in the Semantic Web/Linked Open Data Cloud. lemon models lexicons by means of a core consisting of the following elements:

• A lexical entry which represents a single word or multi-word unit.
• A lexical sense, representing a meaning of that word, which contains a reference to a concept in the ontology.
• Forms, which are inflected versions of the entry, and associated with a string representation.

SKOS and Custom Vocabulary

In addition, to using the lemon vocabulary to model the semantics of WordNet, we use the SKOS vocabulary as this is better suited to WordNet’s structural model than a formal ontology language such as OWL. Furthermore, we introduce a new vocabulary at http://wordnet-rdf.princeton.edu/ontology# to include properties found only in WordNet.

RDF Generation

Data modelling

An example of the modelling a single word and synset and links to other resources

It is not trivial to apply lemon to the case of a WordNet as there is no clear ontology in WordNet. Clearly, WordNet’s words can be regarded as lemon lexical entries and the word senses correspond well to lemon’s lexical senses. WordNet has lemmas and a separate list of variants of these, and as such we recommend creating a canonical form for each lemma and a form object for each of these variants. Since there is currently no indication in WordNet of what grammatical properties these variants have, we do not recommend attaching additional properties to these variants/forms. As lemon is a model for ontology-lexica, the main question is what the reference of the lexical senses should be. We recommend regarding WordNet’s synsets as ontological references, but instead of assigning them a formal ontological type (e.g., class, property or individual), we introduce a new type Synset as a subclass of Concept in SKOS.

This allows the nature of synsets to be captured without ontologizing the semantic network. Similarly, we introduce relations such as hypernymy, meronymy etc. as new properties rather than attempt to relate them to existing ontological properties such as OWL’s subClassOf. In order to capture the new properties, an ontology has been created at http://wordnet-rdf.princeton.edu/ontology

URI design

Another key question concerns the identifiers to use for each element in the data. In order to enable browsability, we recommend that each entry and each synset has its own page and thus it own URI. We do not recommend distributing WordNets as a single flat file as they are large and this impedes their access on the web. You should assign new identifiers using the existing identifiers in your wordnet. Furthermore, as wordnets have released several versions and are still under development, we consider it important to include the version number in the URI. As such, we recommend the following scheme for URIs, as exemplified below:

• Each lexical entry is represented by means of the URL-encoded lemma and then a dash followed by the part-of-speech as a single letter (i.e., ‘n(oun)’, ‘v(erb)’, ‘a(djective)’, ‘r (adverb)’, ‘adjective s(atellite)’ or ‘p(article)’).
• Senses and forms in the model use the entry URI and add a fragment identifier. For forms for which there is no previous identifier in WordNet, we use CanonicalForm and Form-n where n is a number. For senses, the fragment is the index of the senses and the part of speech.
• Synsets are similarly are identified by a number consisting of 8 or 9 digits corresponding to offset codes in the WordNet database The 9 figure codes include an extra initial digit for part-of-speech, followed by a dash and the part of speech as a single letter.

WordNet JSON

In combination with the Global WordNet Association, we have developed a set of guidelines for creating wordnets as linked data. This can be done in either the LMF-compliant WN-LMF format or using the JSON-LD enabled schema below. The JSON-LD schema consists of the following elements

• Lexicon the root of a document should be a lexicon or a lexical entry that consists of a number of Lexical Entries
• A Lexical Entry is a single word or multiword expression as defined by lemon
• Each lexical entry must have exactly one Lemma it may have any number of other forms. Each of these must have a single Written Representation
• It may have other properties such as part-of-speech, see documentation at http://globalword.github.io/schemas/JSON-LD.md
• Any number of senses each of which should link to synsets and give a gloss plus any lexical relationships such as hypernym

Linking to the Interlingual Index

1. Your resource has an license that allows redistribution with only attribution. Non-commercial and Share-A-Like license are not accepted.
2. The resource is in WordNet LMF or WordNet JSON
3. Every synset in your wordnet has a definition in English
4. Every synset is either linked to the ILI as either a synonym, hypernym, antonym or meronym
5. You do not create any duplicates of concepts already in the ILI (this is checked by automatic methods)
6. Definitions are written according to the quality guidelines on the Global WordNet Assocation webpage