We describe experiments done in using the lemon model for encoding lexicographic data we got from different sources with distinct coverages. Our focus is on delivering statements on lessons learned and on questions that still should be discussed in the lemon community, as a possible input for forthcoming versions of the model.
In recent years, we have been experimenting with the use of the Lexicon Model for Ontologies (lemon)1 for representing lexicographic data. lemon has been developed within the W3C Ontology-Lexica community group2, building on the version that was first proposed in the context of the European R&D project “Monnet”3. Members of the W3C community group started to investigate, in close cooperation with members of the ENeL Cost Action4, on how this model could be used and possibly extended for the purpose of the encoding of rich lexicographic data in the context of the Linguistic Linked Open Data (LLOD) framework5. 1 See https://www.w3.org/2016/05/ontolex/ [accessed 13.07.2017, like all other URLs mentioned in this paper] 2 https://www.w3.org/community/ontolex/ 3 http://cordis.europa.eu/project/rcn/93713_en.html. See for the first version of lemon: http://lemon-model.net/ 4 ENeL stands for “European Network of e-Lexicography”. See http://www.elexicography.eu/ for more details on this COST Action. 5 See http://linguistic-lod.org/ for more details.
The studies we present in this paper are dealing with the “Wörterbuch der
bairischen Mundarten in Österreich” (WBÖ)6 and with the “Algemeen Nederlands
Woordenboek” (ANW)7. Details on the first study are described in [
The original aim of lemon was to provide for a rich linguistic description for natural language expressions used in knowledge resources, like taxonomies or ontologies. This linguistic grounding includes the formal representation of morphological and syntactic properties of lexical entries as well as the specification of the meaning of these lexical entries with respect to available knowledge resources, more specifically formal ontologies. This relation between lexical entries and their meaning to be found in external ontologies is specified in a core module, which is realising the so-called ontology-lexicon interface (ontolex). Figure 1 below is giving a graphical overview of the ontolex module.
Dictionary of the Bavarian Dialects in Austria: https://www.oeaw.ac.at/acdh/projects/wboe/. General Dutch Dictionary: http://anw.inl.nl/ The full list of lemon modules comprises the following items8:
Ontology-lexicon interface (ontolex), which is the core module of lemon Syntax and Semantics (synsem) Decomposition (decomp) Variation and Translation (vartrans)
Linguistic Metadata (lime) 3
Our first experiment dealing with the encoding of lexicographic data in lemon involved the “Wörterbuch der bairischen Mundarten in Österreich“ (WBÖ)9, which is a large-scale dictionary documenting (spoken) languages used in Austria and neighboring areas such as South Tyrol. WBÖ describes a very large varietal spectrum of the language spoken in the regions it covers. In addition to “Austriacisms”, the publication also contains standard German keywords, which have developed a different range of meaning in the dialects. 3.1
The basis we are working on is an XML representation of the electronic version of the WBÖ dictionary. As the WBÖ is primarily aimed at scientists it uses for its descriptors complex and rich representation forms. So for example, the strings for representing headwords in WBÖ can include information about pronunciation or word formation properties10, etc. A first issue for the encoding of the WBÖ headwords in lemon-ontolex concerns thus the status of this highly specialized and compact representation form. As the lemon model considers usually the “lemma form” of a lexical entry as its canonical form11, we would tend to give to the corresponding instance of the ontolex:Form class, as the range of the associated object property ontolex:canonicalForm, the lemma form of the original headword as the value of datatype property ontolex:writtenRep, without the special characters used with the original headword. And we would encode for example the pronunciation properties as a value of the related datatype property “ontolex:phoneticRep”12. Thus marking explicitly all the information related to a headword by corresponding properties.
The issue we have is how to encode the original headword. We do not consider it as being potentially an “other form”, in the sense this feature has been introduced in lemon-ontolex (and similar also in other vocabularies, like SKOS-XL with the “alter8 Taken from https://www.w3.org/2016/05/ontolex/ 9 Dictionary of the Bavarian Dialects in Austria: https://www.oeaw.ac.at/acdh/projects/wboe/ 10 An example of such a headword is “(Ge)pâcht”, where the parentheses mark a prefix and the circumflex refers to a pronunciation property. 11 See also http://www.w3.org/ns/lemon/ontolex#canonicalForm 12 See Figure 1 for the roles played by those properties. native label”). We can encode the original headword as a “hidden form”, but there would be a need to associate a comment to this representation, stating that this was the original form used in the dictionary. In general, the issue we are dealing with concerns the wish to keep information about the way the data was represented in the original source. In case we can decompose all the lexical information included in the rich and compact original representation of the headword and port it to different elements of lemon, we might then just add a property referring to the original headword string.
We used only the ontolex module for the lemon encoding of WBÖ, and as
described in [
We see in the current specifications of the “vartrans” module13 that it covers among others both diatopic and diachronic lexical variations, but the specifications also encourage the use of external lexical vocabularies for marking temporal information on the usage of a word, this being also relevant for synchronic lexicography. It will be important to reach a consensus on which vocabulary to use for indicating temporal and local information, similar to the use of the LexInfo vocabulary14 for marking many morphological and syntactic properties of lexical entries within ontolex. 3.2
A very interesting resource associated with WBÖ is provided by the (“Fragebögen”) (questionnaires) that were used for interviewing people in different regions of Austria on which words they use for expressing a specific concept. The questionnaires included concepts and related definitions that could also represent various senses associated with one term. Those questionnaires form an important part of the very rich documentary material that was used for the creation of the WBÖ, and which contains an estimated 4 million individual sheets15. This documentary material was entered and digitized between 1993 and 2011 in the so-called database of the Bavarian dialects in Austria (DBÖ)16.
As those questionnaires represent a very rich combination of conceptual and related lexical information, we also investigated the possibility of encoding this information in lemon-ontolex. It turned out that the possibility to link lexical entries (and also lexical senses) to a lexical concept, which is itself a member of a SKOS 13 See for more details on this module: https://www.w3.org/2016/05/ontolex/#variationtranslation-vartrans 14 See http://www.lexinfo.net/ for more details. 15 This information is taken from the German Wikipedia page: https://de.wikipedia.org/wiki/W%C3%B6rterbuch_der_bairischen_Mundarten_in_%C3%96 sterreich 16 See https://wboe.oeaw.ac.at/projekt/beschreibung/ scheme17, is very suitable to represent the type of information that is encoded in both the questionnaires and the WBÖ dictionary. The questionnaire has been encoded in the SKOS vocabulary and the ontolex property “isEvokedBy” is linking the concepts of the questionnaires to the lemon-ontolex lexical entries, while the ontolex property “lexicalizedSense” is linking the concepts to corresponding WBÖ lexical senses. It is unclear, if we can still speak of a mental abstraction or unit of thought of lexical entries (as the introduction of the “LexicalConcept” class in ontolex was aiming at), but the fact is that we can efficiently relate the conceptual background that was developed as a basis for the creation of a dialectal dictionary to the Lexical Entries in the lemon representation of this dictionary.
Based on our piloting work, the questionnaires are now completely conceptually interlinked and all sources made available for conceptual based discovery within the project exploreAT! 18 3.3
An additional artefact that was used for crafting the dictionary is a (huge) set of paper slips, on which the field lexicographer was indicating the answer of the interviewed persons, with some metadata (location, time, and any other comments). Those paper slips are for sure an important element of cultural heritage, but besides this they also offer – together with the questionnaires – a view on the methodology, the “workflow” and the material used. We do not foresee to encode this data in lemon-ontolex. We rather propose to build for this combination of the two artefacts, questionnaires and paper slips, a model using SKOS, and linking the concepts of those SKOS schemes to the lexical entries we have in lemon-ontolex. All the dialectal variants that were encoded in the paper slips (and those already present in WBÖ) will then be encoded in the lemon module “vartrans”. 3.4
Our aim in the WBÖ case was to develop a series of methodological prototypes of a machine-readable and modular version of the lexicographical work, aiming at making it available in the Linguistic Linked Open Data framework. We also suggest ways for encoding in SKOS some information on data material that was used for the creation of the original dictionary. But for this, the open question remains if we should aim at creating a new module in lemon, which is collecting all those aspects of a lexicographic work, and not only the lexical knowledge expressed in the dictionary.
Our learnings and prototypes are further explored within the project exploreAT! to open up the resources for multicultural, multilingual knowledge discovery. 17 See again for details Figure 1. 18 exploreAT! exploring austria´s culture with the https://exploreat.grial.eu/dashboard (last accessed: June, 12th 2017). language glass.
A second source of data we have been working with is the XML representation of the
online version of the Algemeen Nederlands Woordenboek (ANW). ANW is an
online, corpus-based, scholarly dictionary of contemporary standard Dutch in the
Netherlands and in Flanders, describing the Dutch vocabulary from 1970 onwards
(see [
The ANW is a digitally-born dictionary. It is very rich data encoded in a format that is already very abstract. With the addition of the semagram framework, the ANW includes also some accompanying data structure reflecting the conceptual world to which lexical entries are related. A nice aspect of the ANW is that it also refers to an external resource for the information on etymology. Thus a high level of modularity and connectivity is already realized.
Like a number of scholarly dictionaries, ANW has a large number of senses associated to the entries (at least for certain categories of entries – nouns, verbs, etc.). As this repository of senses is large and complex, a discussion arose if we should not introduce in lemon a sub-sense hierarchy, which is present in ANW, where a numbering strategy is used for representing the hierarchy of senses. We note that a former version of the lemon model19 included the notion of “subsense”. But the property defined in this version was meant to describe the composition of senses needed for describing the composition of senses resulting from the argument structure of a verb (or another category introducing syntactic arguments). In the new version of lemon, resulting from the W3C community group discussions, this aspect is dealt with by the synsem module and the notion of “subsense” has disappeared.
For the time-being we suggested to have in lemon-ontolex a flat list of senses and to interrelate those by the use of corresponding properties, such as lexinfo:hypernym, lexinfo:partHolonym, lexinfo:substanceMeronym, lexinfo:pertainsTo, etc. As a matter of fact, it turned out that the LexInfo vocabulary is here (and in general) very helpful for describing relational properties of lexical senses. But we kept the original ANW numbering for naming the senses object in lemon-ontolex. Extending the study we describe in this paper to other scholarly dictionaries will certainly help in getting final decisions on this issue.
For the porting of the verbs included in the ANW we also started to investigate the use of the synsem module. A first comment would be that we are not sure about the necessity to use the ontology mappings, as described in the specification document20. It seems to introduce a higher level of complexity. And the need to introduce a frame element for each verb seems also to introduce a lot of redundancies. But we have no alternative solutions for the time being. Our tentative modelling of the verb “eten” (to eat) looks like in the following: 19 http://lemon-model.net/lemon-cookbook.pdf 20 See again https://www.w3.org/community/ontolex/wiki/Final_Model_Specification :lex_eten_47968 rdf:type ontolex:Word ; lexinfo:partOfSpeech lexinfo:verb ; ontolex:canonicalForm :form_eten_infinitive ; ontolex:sense <http://tutorial-topbraid.com/anw#sense_eten0.1> synsem:synBehavior :eten_frame_1 ; <http://tutorial-topbraid.com/anw#sense_eten0.1> rdf:type ontolex:LexicalSense ; skos:definition "iets als voedsel tot zich nemen; iets opeten; iets nuttigen"@nl ; #(take something as food, eat something, …) ontolex:isLexicalizedSenseOf :Semagram_activiteit ; ontolex:isLexicalizedSenseOf :Semagram_handeling ; ontolex:isSenseOf :lex_eten_47968 ; synsem:objOfProp :eten_frame_obj_1 ; synsem:subjOfProp :eten_frame_subj_1 ; ; :eten_frame_1 rdf:type lexinfo:TransitiveFrame ; rdf:type synsem:SyntacticFrame ; lexinfo:directObject :eten_frame_obj_1 ; lexinfo:subject :eten_frame_subj_1 ; rdfs:comment "one syntactic frame for the Dutch verb \"eten\""@en ;
rdfs:label "transitief eten"@nl ; :eten_frame_subj_1 rdf:type lexinfo:Subject ; rdf:type synsem:SyntacticArgument ; rdfs:comment "A subject of the eten_frame" ; rdfs:label "subject 1 for eten_frame"@en ; ontolex:concept :SemaGram_dier ; ontolex:concept :SemaGram_mens ; :eten_frame_obj_1 rdf:type lexinfo:DirectObject ; rdf:type synsem:SyntacticArgument ; rdfs:comment "An object of the eten_frame"@en ; rdfs:label "object 1 for eten_frame"@en ; ontolex:concept :SemaGram_voedsel ;. ; . :OntoMap_eten_1 rdf:type synsem:OntoMap ; rdfs:comment "Mapping the syntacic frame eten_1 with semantics" ; rdfs:label "OntoMap_eten_1@en}" ; synsem:objOfProp :eten_frame_obj_1 ; synsem:ontoMapping <http://tutorialtopbraid.com/anw#sense_eten0.1> ;
synsem:subjOfProp :eten_frame_subj_1 ;
But as mentioned above, we have the feeling that this representation is getting too complex.
We also tested the decomposition module for encoding Dutch compounds with the example word being “wijnfles” (bottle of wine). As above for the “eten” verb, we just display the current code, so that the reader can get a concrete idea of the possibility offered by this “decomposition” module of lemon: :wijnfles rdf:type ontolex:MultiWordExpression ; <http://lemon-model.net/lexinfo_anw#articleType> "\"de\"" ; lexinfo:gender lexinfo: commonGender ; lexinfo:partOfSpeech lexinfo:commonNoun ; lexinfo:partOfSpeech lexinfo:noun ; rdf:_1 :comp_wijn_1 ; rdf:_2 :comp_fles_1 ; <http://www.w3.org/ns/lemon/decomp#constituent> :comp_fles_1 ; <http://www.w3.org/ns/lemon/decomp#constituent> :comp_wijn_1 ; <http://www.w3.org/ns/lemon/decomp#subterm> <http://dictionary_lemon/anw#lex_wijn_182155> ; <http://www.w3.org/ns/lemon/decomp#subterm> :lex_fles_18089 ; ontolex:sense <http://tutorial-topbraid.com/anw#sense_wijn1.3> ontolex:sense <http://tutorial-topbraid.com/anwentry#sense_wijn1.0> ; . :lex_fles_18089 rdf:type ontolex:Word ; <http://lemon-model.net/lexinfo_anw#articleType> "\"de\"" ; lexinfo:gender lexinfo:feminine ; lexinfo:gender lexinfo:masculine ; lexinfo:partOfSpeech lexinfo:commonNoun ; lexinfo:partOfSpeech lexinfo:noun ; ontolex:canonicalForm :form_fles_singular ; ontolex:otherForm :form_flessen_plural ; :comp_fles_1 rdf:type <http://www.w3.org/ns/lemon/decomp#Component> ; <http://www.w3.org/ns/lemon/decomp#correspondsTo> <http://dictionary_lemon/anw#lex_wijn_182155> ;
<http://www.w3.org/ns/lemon/decomp#correspondsTo> :lex_fles_18089 ; . :comp_wijn_1 rdf:type <http://www.w3.org/ns/lemon/decomp#Component> ; <http://www.w3.org/ns/lemon/decomp#correspondsTo> <http://dictionary_lemon/anw#lex_wijn_182155> ;
<http://www.w3.org/ns/lemon/decomp#correspondsTo> <http://tutorial-topbraid.com/anw#sense_wijn1.0> ; .
The only addition we suggest here, is to add the possibility to have a sense as the value of the “correspondsTo” property, as this can be seen for the element “comp_wijn_1” above. The point being that in this case the component “wijn” can only refer to the generic use of the word, which is the one covered by the Lexical Sense “sense_wijn1.0”, as can be seen below: <http://tutorial-topbraid.com/anw#sense_wijn1.0> rdf:type ontolex:LexicalSense ; skos:definition "alcoholhoudende drank, verkregen door gisting van het sap van druiven of van andere vruchten, met een middelmatig alcoholgehalte van doorgaans ongeveer 12 procent; alcoholhoudende drank van gegist druivensap" ; ontolex:isLexicalizedSenseOf :Concept_325624 ; ontolex:isLexicalizedSenseOf :Concept_Stofnaam ; ontolex:isSenseOf :lex_wijn_182155 ;
ontolex:reference <https://www.wikidata.org/wiki/Q282> ; ontolex:usage lexinfo:massNoun ; ontolex:usage lexinfo:singular ;
This way, it seems that we can cover most (if not all) of the relevant conceptual and lexical elements included in the ANW offer. A remaining question being if one needs to introduce in lemon a hierarchy of senses, or rather, like we opted for now, if one can adopt a single listing of Lexical Senses and to explicitly mark the relation among them by the use of a possibly extended LexInfo vocabulary. 5
It is our conviction that a lot of the original dictionary data, in different formats and with different coverages, can be accurately modelled with the lemon modules. Relevant lexicographic information that is not directly related to the description of the entries (in the sense of providing knowledge about the words) can be designed in or re-used from models external to lemon, but a consensus building on the best vocabularies to be used will be needed in this case.
And our current intuition is that elements in lemon should not include (deeper) hierarchical structures but represent the relation between elements of the lexicon by the use of specialized properties. We think this is an aspect that should be discussed and possibly fixed within the W3C Ontology-Lexica community.