This paper describes the modelling and publication of part of the etymological information in the Lexicon der indogermanischen Verben, an etymological dictionary of verbs attested in ancient Indo-European languages, as Linguistic Linked Open Data. The lexicon has been made interoperable with a set of lexical and textual linguistic resources for Latin in the Lila Knowledge Base.
can really reach its potential only when it is made interoperable with other (types of) resources. Today, interOver the past decades, several linguistic resources for operability between linguistic resources can be obtained historical languages have become available in digital for- by describing and publishing their data according to the mat. This has given scholars the chance to access and principles of the Linked Open Data paradigm [3]. As a exploit them in a quicker and deeper way. consequence, in recent years, the amount of linguistic
In particular, several diferent linguistic resources are resources published as Linked Open Data has been raised
available for Latin today. They consist of textual corpora, substantially, as witnessed by the growing size of the
such as the LASLA corpus1 for Classical Latin and the Linguistic Linked Open Data Cloud (LLOD Cloud)3. In
Index Thomisticus Treebank [
Due to the centuries-long lexicographic tradition for Although LiLa currently makes quite a number of linthe Latin language, its lexical resources comprise a num- guistic resources for Latin interoperable, there is still a ber of etymological dictionaries. A dictionary is defined large set of digitized materials to interlink in the Knowlas etymological when it contains information about the edge Base. Among them is an important etymological etymology of its entries, that is about their origin and his- dictionary, the Lexicon der indogermanischen Verben (LIV) torical development: for Indo-European (IE) languages, [4]. This paper describes the process of transforming etymological dictionaries often put their entries in rela- information contained in this dictionary into a LLOD tion with reconstructed Proto-Indo-European (PIE) roots, resource, linked to LiLa. Section 2 describes the LiLa minimal lexical units to which the dictionary’s entry and architecture and the LIV structure. Section 3 details the further related formations may be traced back. It is also modelling of the resource, and the linking process. Secoften explained by which morphological processes the tion 4 describes some possible examples of exploitation attested word has been formed from the root. and interaction. Finally, Section 5 discusses conclusions
Even if the available resources (for Latin and beyond) and sketches the future work. provide a huge amount of linguistic information, at the present day, their full exploitation is still hindered by their isolation. In fact, most resources can be accessed only individually, and cannot interact.
Isolation of resources is an issue because each resource
the Lexicon der indogermanischen Verben 2.1. The LiLa Knowledge Base
Latin, containing FAIR [6] linguistic resources, published as LOD. As usual in LLOD, structural interoperability
4https://lila-erc.eu. between resources is based on the Resource Description dictionary of Latin and other Italic Languages [16, 17] and Framework (RDF) [7], which is the data model used for a resource of principal parts of Latin words, PrinParLat the Semantic Web [8]. Conceptual interoperability [9] is [18]. achieved by using common ontologies built and adopted by the LLOD community, such as the Ontolex Lemon 2.2. The Lexicon der indogermanischen model5 and the OLiA ontology6 [10].
LiLa is built around the so-called Lemma Bank, which Verben contains a set of more than 200k Latin lemmas, taken The Lexicon der indogermanischen Verben, also known from the database of the morphological analyser LEM- as LIV, is an etymological dictionary of verbs attested in LAT [11] and constantly extended. Each lemma of the ancient Indo-European languages. After the first edition, Lemma Bank is a gateway between the diferent linguis- curated by Helmut Rix [19] and published by Reichert tic resources linked to the Knowledge Base, starting from Verlag in 1998, a second edition was published in 2001 the assumption that words (indexed by their lemmas) with additions and corrections by Martin Kümmel and can be used as the point of contact between textual re- Helmut Rix [4]. sources (which are made of occurrences of words), lexical The LIV is the main reference work for Proto-Indoresources (which describe words) and NLP tools (which European verbal roots and contains three types of inforprocess words). mation:
Each entry in the LiLa Lemma Bank is an instance of ontolex:Form7. In particular, the lila:Lemma8 is a • Reconstructed Proto-Indo-European verbal form that can be linked to an ontolex:LexicalEntry9 roots, which coincide with the entries of the dicvia the property ontolex:canonicalForm10, which tionary and are provided with their presumed identifies the canonical form used to represent a lexical lexical meaning and their phonological structure. entry. Every other realization of a word is linked to the For each root, the corresponding index in the Inlexical entry via the property ontolex:lexicalForm11. dogermanisches etymologisches Wörterbuch [20] Each lemma and each form may also be described with is specified as well. other properties, which give information, for example, • Reconstructed Proto-Indo-European priabout phonetic representation, Part of Speech (POS) tag- mary verbal stems, which are either root forging and other grammatical features. mations or are formed by adding to the roots
Lexical resources are connected to LiLa by primary afixes that mainly express categories of linking the ontolex:LexicalEntry of each aspect and actionality. The meaning of the stem resource to the lila:Lemma via the property is usually not specified. ontolex:canonicalForm. Once a linguistic re- • Word forms that are historically attested in source is linked to the KB via the Lemma Bank, all ancient IE languages, which show how the the interoperable resources can be queried together, Proto-Indo-European stems evolved in the varusing a SPARQL endpoint12 also through a user-friendly ious daughter languages. Each attested form is interface13. provided with its lexical meaning in the respec
The textual resources connected to LiLa so far include tive language. At the end of certain entries are more than 3,5M words from Latin texts of diferent eras, sometimes listed innovative verbal stems that such as the LASLA corpus14, the Index Thomisticus Tree- may ultimately be traced back to Proto-Indobank [12], containing works of Thomas Aquinas, and European roots, but are unlikely to directly reflect UDante, a Universal Dependencies15 treebank for Dante Proto-Indo-European verbal stems, having been Alighieri’s Latin works [13]. The lexical resources of created according to language-specific productive LiLa include a derivational lexicon, Word Formation Latin patterns. [14], a manually checked subset of the Latin WordNet connected to a valency lexicon [15], the Etymological 5https://www.w3.org/2016/05/ontolex/. 6https://acoli-repo.github.io/olia/. 7http://www.w3.org/ns/lemon/ontolex#Form. 8https://lila-erc.eu/lodview/ontologies/lila/Lemma. 9http://www.w3.org/ns/lemon/ontolex#LexicalEntry. 10http://www.w3.org/ns/lemon/ontolex#canonicalForm. 11http://www.w3.org/ns/lemon/ontolex#lexicalForm. 12https://lila-erc.eu/sparql/. 13https://lila-erc.eu/query/. 14https://www.lasla.uliege.be/cms/c_8508894/fr/lasla. 15https://universaldependencies.org/.
sists of a spreadsheet containing information from the LIV extracted and structured by Thomas Olander, with the collaboration of Simon Poulsen and Anders Richardt Jørgensen, and shared with us by the authors. For each of the 7,888 LIV entries the spreadsheet records its root, stem and attested forms.
The LIV is copyrighted by the Reichert Verlag, and LiLa is not authorized to reproduce the full content of the dictionary. The publisher has however agreed to allow us to model the basic etymological relations between the PIE roots, the stems and the Latin words and stems, provided that explicit bibliographical attribution is given to the linguistic reconstruction.
This is the information that we modelled to be linked to the LiLa Knowledge Base, as described in the following section.
Making linguistic resources interoperable means using a shared set of vocabularies for knowledge description, as defined in specialized ontologies, to represent the information contained in them. The process of linking, on the other hand, aims to connect this information to a wider network of data, so that a meaningful context is provided16. Within the network of LiLa, this step means that all entries of a lexical resource must make reference to the canonical forms of the Lemma Bank, as described above.
This section details how we modelled our target information from the LIV, how we applied such modelling to the publication of these data as LOD and how we linked them to the LiLa collection.
In applying this model to the LIV data, it is crucial to define what the “linguistic elements” connected via etymological links are. The previously mentioned lexical resources rested on a simple model where the etymological links involved only Latin lexical entries and etymons from a source language, so that e.g. the Latin word abacus19 was the target of a link that had its source in the Greek etymon ábax ‘reckoning board’.
In the LIV, on the other hand, relations are established between: • Inflected forms of a historical language (e.g.
Latin). In the case of Latin, those forms are used in the LIV as placeholders for all forms derived from the same stem; so, for instance, the Latin 1st-person perfect fidi stands for all forms from the perfect stem of the verb findo ‘to cleave, split’; • the PIE stems, to which the inflected forms and stems of Latin (and other languages) must be traced back; • the PIE root that underlies the PIE stems.
ical relations between a PIE and a Latin stem (the latter represented by a Latin inflected form). While the PIE 3.1. Modelling root (e.g. *bheid-) and the Latin target lexical item (e.g. In the lexical resources linked to LiLa, etymological in- ifndo , inclusiv“e of all its stems) can be conceptualised as formation has been expressed using the lemonEty ex- lexical entries, the stems and the word forms must be tension of the Ontolex-Lemon model [21]. This ontology described using concepts from other vocabularies. was used in LiLa to represent loanwords from Greek [22] For the Latin forms and stems we reused the individuals of the class Stem20 provided by PrinParLat, a lexical IatnadlicfoLratnhgeuEatgyems o[l1o6g,i1ca7l].Dictionary of Latin and the other resource listing all Latin “principal parts”. Principal parts suiTtahbelesetot oexfpcrleassssetsheanedtymproolpoegritciaelsinoffolrmemaotinoEntoyf tahree aorfealsletthseofotihneflerctpeadrawdoigrdmfocremllss fcraonmbewihnifcehrrtehde21c.ontent LIV too, but, compared to the aforementioned dictionar- For the perfect stem, PrinParLat already includes all ies in LiLa, a more complex modelling and a series of forms linked to their Stem therein, which could thus be extensions are also required. immediately reused. As for the present stems, however,
The lemonEty ontology establishes etymological the related forms were not available, and had therefore relations between instances of the Ontolex’s class to be generated and linked to their Stem via the property LexicalEntry. In particular, a special subclass called ontolex:lexicalForm22.
Etymon is reserved for lexical items of the source lan- Some specific information provided by the dictionary guage that are introduced in order to explain the history that could not be represented with any of the available of the entries in the target language.
Two core classes of lemonEty that are particularly important are Etymology17, and EtyLink18. The former “reifies the whole process of etymological reconstruction as scientific hypothesis” [ 17, p. 22]. Etymological links, on the other hand, connect “linguistic elements” from the source language to the corresponding elements of the target.
19http://lila-erc.eu/data/lexicalResources/IGVLL/id/ LexicalEntry/abacus.
20https://lila-erc.eu/lodview/ontologies/prinparlat/Stem. This class will be defined as a sub-class of morph:Morph, when the module will be released.
21https://github.com/CIRCSE/PrinParLat.
22Note that, following the conventional definition of a form in PrinParLat, which is more granular than the one used by LiLa, we chose to recreate the form for the same paradigmatic slot that is traditionally chosen as lemma. PrinParLat creates a unique form for each graphical or spelling variant of a form; lemmas in the Lemma Bank, on the other hand, unify them under a single form. In many cases, this choice results in a duplicated form for the 1st-person singular present indicative verb.
16This is the fifth and final step in Berners-Lee’s five-star rating system: see https://www.w3.org/DesignIssues/LinkedData.html.
17http://lari-datasets.ilc.cnr.it/lemonEty#Etymology.
18http://lari-datasets.ilc.cnr.it/lemonEty#EtyLink. modules required the creation of ad hoc classes. In par- word form, but rather to a whole lexical root that is analticular, some Latin stems that trace back to PIE roots, but ogous to LiLa’s lexical base [5, p. 191]. This morpheme are unlikely to directly reflect a PIE stem, are classified by represents a lexical element that is neither a prefix nor the LIV as Neubildungen, that is ‘innovations’, since they a sufix and is shared by all members of a derivational have been created according to language-specific produc- family. Comparably, for instance, the Latin hyphenated tive patterns. These innovations cannot be traced back form -pleo in the entry *ℎ1- is used in the LIV as a to a PIE stem, so that no etymological link can be created. placeholder for all the possible Latin verbs that can be We therefore created a specific class Innovation, which formed adding diferent preverbs to the same base (e.g. contains all those innovative stems. compleo ‘to fill up’, depleo ‘to empty’, expleo ‘to fill up’...).
Moreover, some Latin entries are defined by the dictio- In those cases where the LIV uses this notation, we nary as ‘remodelings’ (Umbildungen): their stems may be chose to create one lexical entry for each verb connected traced back to PIE stems, but have been reshaped follow- to the corresponding lexical base in LiLa (e.g. the ‘base ing language-specific productive patterns (e.g. Latin fo- of pleo’25). dio has lost the first syllable of the PIE reduplicated stem Once the lexical entries had been created, the correct *bhé-bhodhh2/bhdhh2-). The remodeled Latin stems are stems in the PrinParLat resource were easily identified now defined as instances of the new class Remodeling. by leveraging the advantages of the LOD model. In fact, each LiLa’s lemma is linked to the appropriate stems via 3.2. Linking an instance of the PrinParLat class of Flexeme26; the stems for a lexical entry are therefore easily recoverable LIV provides etymological information for other IE lan- once the LiLa lemma is known. guages in addition to Latin. Since, however, LiLa is limited to Latin resources, we restricted our attention only 3.2.1. A LIV lexical entry linked to LiLa on entries where a connection to Latin forms was explicitly mentioned. The Figure 1 is taken from the LodLive27 visualization
In total, we identified 550 Latin forms linked to PIE in LiLa. It shows an example of how a LIV lexical entry roots, 354 of which corresponded to the main lemma (glubo ‘to peel’) was modelled and linked to the Lemma of a verb; the remaining 196 were instead analysed as Bank. inflected forms. On the left side of the figure is the LIV LexicalEntry
The forms were analysed with the UDPipe pipeline23, glubo, which is linked to LiLa’s lemma glubo via the in order to perform the POS tagging of all forms and property canonicalForm: this simple but crucial link lemmatization of 196 inflected forms. The results were allows us to connect the LIV etymological relations with manually checked, which confirmed a good accuracy the other resources of the Knowledge Base. of 97% for POS-tagging (only 11 cases were incorrectly Then, the LIV LexicalEntry is connected via the tagged), but much lower performances for lemmatization property lexicalRel28 to two PrinParLat Latin stems, (87 out of 196, i.e. 44%). the present stem glub- and the perfect stem glups-. Each
For each of the remaining lemmas in the manually stem is connected with a Latin form: the present form corrected set, we created a lexical entry in our new et- glubo is part of the LIV resource, and is thus linked to ymological resource. The canonical forms of these en- the present stem via the property lexicalForm; on the tries were identified by matching the lemma strings with other hand, the perfect form glupsi is part of the PrinParthe written representations24 of the lemmas in the LiLa Lat resource, and is thus linked to the perfect stem via Lemma Bank. In 143 cases, manual disambiguation was the property consistsOf. needed, as the query returned more than one possible We then link the Latin stems to their Proto-Indomatch. In one case, it was not possible to link the form European ancestors. Each of the two Latin stems is in (ta¯to¯d) to any lemma in the Lemma Bank: we decided fact the etyTarget of an EtyLink (Etymology link: pres against adding the invariable form to the Lemma Bank glubo and Etymology link: perf glupsi), which connects and instead created a LIV LexicalEntry ta¯to¯d, without them to their etySource, that is the corresponding PIE connecting it to any lemma. stem (for the present *g/“gléubh-/g/“glubh-, and for the
Moreover, a set of 11 entries required a special treat- perfect ?*g/“gl¯´eubh/g/“gléubh-s-“). These etymological links ment. For a series of entries, like for instance *ℎ1- [4, “ “ p. 482], in fact, the LIV does not point to a single Latin verb2s5hthttapt:/a/rlielao-nerlyc.eduo/cduamtae/nidt/ebdasine/l1e0x7i.coNnosteotfhMatewdieevexacllLuadteind,tshuocshe as the Du Cange dictionary [23].
26http://lila-erc.eu/ontologies/prinparlat/Flexeme. 23https://lindat.mf.cuni.cz/services/udpipe/. 27“LodLive project provides a demonstration of the use of Linked 24In Ontolex Lemon, written representations are the diferent Data standards (RDF, SPARQL) to browse RDF resources”. http: graphical variants of a form. See http://www.w3.org/ns/lemon/ //en.lodlive.it/. ontolex#writtenRep. 28http://www.w3.org/ns/lemon/vartrans#lexicalRel.
reify the etymological relations that the LIV postulates between the stems, and constitute the bridge between Latin and PIE. The modelling and linking work has, as shown above,
On the right side of the figure is the PIE symmetri- benefited greatly from the advantages provided by the cal counterpart of the model. The PIE root *g/“gléubh- LOD paradigm. The re-use of the lemmas from the LiLa (which is an individual of the class Etymon, subclas“s of Lemma Bank as canonical forms for the LIV entries has the class LexicalEntry) is linked to the two PIE stems allowed us to retrieve the stems from PrinParLat, as well via the property lexicalRel, in the same way as the as all words derived from a handful of selected lexical LexicalEntry is linked to the Latin stems. bases.
Finally, the generic etymological relation between the The exploration of the full set of words derived via PIE root and the Latin lexical entry is reified by the the regular Latin word-formation rules from verbs of IE Etymology class: this class establishes a link between origin can be extended to all the entries in the LIV, beyond them via the properties etymon and etymology, respec- the 11 entries explicitly marked as bases in the dictionary. tively. The Etymology is also connected with the two In fact, 342 regular entries of the LIV currently linked EtyLink, thanks to the property hasEtyLink, and thus to a LiLa lemma are connected to a lexical base. Via this constitutes a central crossroad between the LIV lexical relation, we can access a set of 4,019 other verbs. Also, items. by leveraging the links in the LiLa network to textual resources, we can easily access the earliest occurrences in the corpora.
The full network of the resources linked to LiLa allows for even more advanced inquiries in historical linguistics liv:5 liv:147 liv:157 liv:252 liv:167 liv:260 Base of facio Base of ago Base of fio Base of fero Base of capio Base of eo and in the study of Latin lexicon. Table 129 reports the most productive bases connected to entries in the LIV, with the number of other verbs linked to each lexical base (note that the lemma of the LIV entry was excluded from the calculation).
As can be seen, the most productive words are some of the most common verbs belonging to the oldest IE substratum of Latin, like facio ‘to do, make’, fero ‘to bring’ or capio ‘to seize, take’. Indeed, a joint query between the two dictionaries with IE etymologies, viz. LIV and [16], and the lexical bases in LiLa confirm that the bases that have at least one lemma that is traced back to PIE are considerably richer and more productive than those lexical families without any inherited lexemes. While the former have on average 23.70 members, the latter display only an average of 4.86 members. This fact can be easily explained considering that the latter group is mostly made up by loanwords, which are generally technical terms (especially from the Greek scientific or technical lexicon), and tend to be more specialised and less productive in terms of word formation.
The two dictionaries combined provide now information on PIE etymologies for 1,473 lemmas: 1,393 are connected to entries in the Etymological Dictionary of Latin and the Other Italic Languages [16, 17], 355 in the LIV. In particular, 275 lemmas are shared by the two resources: for these entries, it is therefore now possible to use LiLa to compare the approach to etymological reconstruction by the LIV and the dictionary by de Vaan.
Knowledge Base with etymological information about the processes that, starting from PIE roots, have led to the formation of the Latin word forms. Given the highly lexically-based nature of the architecture of the Knowledge Base, this makes the linking of the LIV an important achievement of the LiLa project.
29The namespace liv in Tab. 1 refers to the URL http://lila-erc. eu/data/lexicalResources/LIV/id/LexicalEntry/.
erable with that of the several other lexical resources currently interlinked through LiLa and can be queried together with the textual data provided by the Latin corpora published in the Knowledge Base.
In collecting and publishing as LOD the wealth of different digital resources for Latin built so far, an important challenge is to impact the scholarly community that has long been using the data provided today by these resources. The web-based interoperability among resources permitted by the LiLa Knowledge Base makes it possible to exploit such wealth of (meta)data like never before, in terms both of the quantity of the (meta)data under analysis and of the quality of the process leading to their retrieval. Interlinking through the Knowledge Base a set of data from the LIV, a reference lexical resource for the communities of Classicists and Historical Linguists, is expected to help overcome the challenge of making the use of LiLa a daily presence in the life of scholars who work in the fields of Classics and Historical Linguistics.
Finally, it is worth considering that the LIV provides etymological information not only about the Latin word forms, but, for each PIE root, it also reports a set of word forms which reflect the same root in several other IE daughter languages. The availability of this information allows for substantial research work to be performed in the near future. Indeed, by applying the principles of the Linked Data paradigm and reusing the same vocabularies adopted in LiLa to interlink the distributed linguistic resources for Latin, it is now possible to move one step further from the Latin language and aim to make interoperable word forms from several IE languages, by using the collection of PIE roots provided by the LIV as a kind of pivot resource to interlink them all.
from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme – Grant Agreement No. 769994. We thank the Reichert Verlag publisher for granting us permission to use some data from the LIV. We thank Matteo Pellegrini for his valuable insights about PrinParLat.