1. Introduction

1Since the pioneering times of 1949, when the Jesuit Roberto Busa persuaded Thomas Watson Sr., CEO of IBM, to fund his project aimed at processing the Latin texts of Thomas Aquinas with computers (Jones 2016), scholars in the areas of Computational Linguistics, Literary Computing and Digital Humanities have built a plethora of linguistic resources for both modern and historical languages.

2Particularly over the last two decades, many and diverse linguistic resources have been made available for Latin. These consist in corpora of texts spanning different eras and genres1, dependency treebanks2 and lexica3. These digital resources join the large set of textual and lexical resources that were created over the centuries for Latin: textual collections, thesauri, lexica, glossaries and mono/bilingual dictionaries. Among the latter, we could mention, for instance, the Oxford Latin Dictionary (Glare 1968), the Dictionary of medieval Latin from British sources (Ashdowne, Howlett, and Latham 1975), the Forcellini lexicon (Forcellini and Facciolati 1871) and the still under construction Thesaurus Linguae Latinae (Ehlers 1968), many of which are today accessible also in digital format.

3However, the impact of these digital resources on the everyday work of classicists is still limited. On the one side, this is due to the still existing divisive dichotomy between “traditional” Humanities and computational approaches. On the other, it is a matter of fact that classicists are not yet put in the best condition to fully exploit all available resources for ancient languages, as these are currently scattered across the web in uncommunicative blocks, using different query languages, data formats, annotation criteria and tagsets. The last decade has seen a number of exploratory solutions to tackle the sparseness of linguistic resources. Among them, the European infrastructure CLARIN4 represents a common hub where data and metadata of resources collected in single repositories (at national level) can be searched (through the so-called Virtual Language Observatory) and processed with different tools (through the CLARIN Language Resource Switchboard). As for Classical languages, Logeion5 is a meta-dictionary that allows to query together the lexical entries of several dictionaries for Ancient Greek and Latin, while Corpus Corporum6 is a meta-collection that allows searches across more than twenty different corpora for Latin. However, what such initiatives still lack is to provide a real interoperability between distributed resources, which would result in interaction at both syntactic (structural) and semantic (conceptual) level.

4Syntactic interoperability is defined as ‘the ability of different systems to process (read) exchanged data either directly or via trivial conversion’, using a common data model consisting of shared protocols and data formats. Semantic interoperability, on the other hand, is ‘the ability to automatically interpret exchanged information meaningfully and accurately in order to produce useful results’, by using a set of common linguistic data categories defined in ad-hoc ontologies (Ide and Pustejovsky 2010).

5Attaining syntactic and semantic interoperability between distributed linguistic resources is the objective of the Linguistic Linked Open Data (LLOD) community, which applies the principles of the Linked Data paradigm (Bizer et al. 2008) to the (meta)data contained in linguistic resources. As for Classical languages, the LiLa Knowledge Base (KB)7 (Passarotti et al. 2020) makes textual and lexical resources for Latin interact through a commonly used data model, called the Resource Description Framework (RDF) (Lassila et al. 1998), and ontologies developed and shared by the LLOD community. In this way, the linked resources become interoperable with each other as well as with those for other languages described following the same structural and conceptual principles.

6Based on a large collection of “canonical forms” (lemmas) - the so-called “Lemma Bank”, LiLa achieves interoperability between resources by linking all those entries in lexical resources and tokens in corpora that point to the same lemma in the LiLa collection.

7The lexical resources for Latin linked so far to LiLa include a word formation lexicon (Pellegrini et al. 2021), a polarity lexicon (Sprugnoli et al. 2020), an etymological dictionary (Mambrini and Passarotti 2020) and a joint resource providing a manually checked subset of the Latin WordNet and a valency lexicon (Mambrini et al. 2021). The most recent among the LiLa connections is the bilingual Latin-English dictionary by Charlton Lewis and Charles Short . The inclusion of this type of lexicon in LiLa was much needed, as no resource providing semantic information consisting of translations and definitions was available in the network of connected resources before. Since Lewis & Short is the first lexical resource of its kind included in LiLa, the process of its linking to the KB opened a number of LLOD-related challenges.

8This paper describes how such challenges have been tackled and is organised as follows: Section 2 describes the Lewis & Short dictionary in its main characteristics. Section 3 discusses the ontologies involved in the modelling phase, the challenges that need to be overcome in the representation of the linguistic data as LLOD (3.1), and the strategies adopted to represent the dictionary entries using the chosen vocabularies (3.2). Finally, Section 4 discusses conclusions and highlights directions for future work.

2. The “Lewis & Short” Dictionary

2.1 The Printed and Digital Dictionary

9The Latin Dictionary, curated by Ch. T. Lewis and Ch. Short and commonly referred to as the “Lewis & Short” (L&S), was published by Harper and Oxford University Press in 1879 (Lewis and Short 1879). Though based on previous work by German scholars, it remained a standard in Latin lexicography in the English-speaking world until it was superseded by the Oxford Latin Dictionary (Glare 1968).

10In the digital age, its importance rests on two grounds. On the one hand, its relevance for the history of Classical Scholarship is undeniable. On the other hand, also on account of its copyright status, as the dictionary belongs now to the public domain, the L&S has quickly become one of the most used and best curated digital Latin dictionaries on the web. Following the same workflow used for the Greek-English Lexicon (Liddell, Scott, and Jones 1940), the Perseus Project has developed a widely used digital edition of the dictionary based on the standards of the Text Encoding Initiative (TEI) (Rydberg-Cox 2002). The digital L&S has been incorporated in the word-search tools available on the Perseus website and in a series of other desktop and web applications.8

11Perseus’ TEI edition is the point of departure of our work.9 Though its publication was a remarkable achievement, this electronic text is not exempt from occasional flaws and inconsistencies, which had to be taken into account.

12In the digital edition, entries from the L&S are based on an XML encoding of the whole dictionary. The XML structure, albeit not always consistent, offers the following information about each word:

1. Entry: the headword. Entries are encoded within the TEI element <entryFree> and are 51,596 in total.10

2. Information about inflection, encoded as attributes in the XML and visualised in the output reproducing the customary descriptions for Latin dictionaries, e.g. a masculine noun of the second declension (e.g. gallus ‘cock’) is followed by the genitive singular ending of the word (‘i’), and the abbreviation for gender ‘m.’ (e.g. gallus, i, m.).

3. Etymological or derivational information, encoded within the same element <etym>.

4. Sense(s): these act as containers where the meaning of the word is matched with a number of representative citations from Classical Latin sources. Each citation is accompanied by its canonical reference (e.g. “Cic. Sen. 8, 26” for a reference to Cicero, De Senectute, chapter 8, paragraph 26).

13Entries can contain what we call “sub-entries”, words that are not given a record of their own, but are discussed within another entry. Usually, these sub-entries consist of lexicalised present and past participles like, for example, adolescens ‘young man’ – sub-entry of adolesco ‘to grow up’; another instance is the substantivised forms of adjectives, such as verum ‘the truth’ – sub-entry of verus ‘true’. Sub-entries are encoded within the <sense> element and followed by the same type of inflectional information structured as the main entries.

2.2 Linking the L&S to LiLa

14The LiLa KB includes about 200,000 canonical forms, each of which is described by a series of properties that record the part of speech (PoS), the full morphological description and the inflectional category. Also, the data property “written representation”, defined in the ontology Ontolex (see Section 3.1), registers all the attested spellings of any lemma. Publishing a lexical resource as LLOD within LiLa means to both represent its information using the appropriate standards and vocabularies (Section 3.1) and to link the dictionary entries to the right form in LiLa by matching the lemmas used to index the records to the appropriate form in the KB.

15In order to achieve the latter goal, firstly we had to normalise the spelling of the L&S dictionary lemmas by removing upper case initials and substituting j with i and v with u in order to mirror LiLa’s conventions. Then, after mapping part-of-speech and inflectional information between resources, we extracted 31,142 1:1 matches, 2,998 1:N matches and 4,553 1:0 matches, on the basis of the tuple written representation - PoS. The latter group was subsequently matched only on the basis of graphical representation, at which point we obtained 946 1:1 matches and 50 1:N matches. Of the remaining 3,557 unmatched entries, 1,289 were successfully analysed by the morphological analyser Lemlat (Passarotti et al. 2017), leaving 2,239 definitely unmatched entries. After resolving multi-word spellings and graphical variants, the unmatched entries were all added to the LiLa Lemma Bank, while 1:N matches were manually disambiguated and matched to the relevant lemmas.

3. Modelling Lexical Entries

3.1 LiLa, Ontolex and lexicog

16As said, the LiLa KB for Latin resources is built around a collection of canonical forms that can be used both as head words of dictionaries or as “targets” for the lemmatisation of corpora (Passarotti et al. 2020). These lemmas are modelled using the Ontolex ontology, a now de facto standard of the LLOD community (Cimiano et al. 2020; McCrae et al. 2017). In particular, lemmas in the LiLa KB are defined as forms of words that are linked (or are ready to be linked) to lexical entries via the property “canonical form” of the Ontolex ontology.11

17Ontolex provides several classes and properties to describe the relationships that lexical entries have with, on the one hand, the grammatical forms attested in language and, on the other, the senses and the meanings of words. The core Ontolex module, however, imposes a series of restrictions that make its classes and properties ill-suited to represent the information in most standard dictionaries. The class Lexical Entry from the core Ontolex module, for instance, is inadequate to represent entries that license multiple syntactic interpretations, such as words that are registered in a dictionary as both adverb and conjunction. Subentries like the noun verum from the adjective verus, formed by a process of substantivisation from the word in the main entry, would also produce a mismatch between the dictionary and the lexical entry. Finally, the L&S, as most dictionaries, defines the senses of all but the most simple words by grouping them in sense clusters; those clusters are generally organized into hierarchies with multiple levels of nesting, from the most general to the most specific sense, a structure for which Ontolex has no suitable representation.

18In order to overcome these issues, the Ontolex community has developed a specific extension of the ontology called the “OntoLex lexicography module” or lexicog (Bosque-Gil and Gracia 2019).12 The module is explicitly designed to capture the structural information expressed in a lexicographic resource and is primarily intended to support the conversion of lexicographic data that are not native to Ontolex. Retro-digitised dictionaries like the L&S are thus a perfect use case.

19As said, lexicog focuses on the structural properties of dictionaries and does not attempt to convey any lexical, or indeed linguistic information, which are left to the classes and properties of Ontolex. The most important of these structural elements introduced in the vocabulary is that of the Lexicographic Entry. In lexicog, an entry is a container that represents a lexicographic article or record as it is arranged in the source (Bosque-Gil and Gracia 2019). Thus, while a lexical entry (as defined in Ontolex) is an item in the lexicon of a given language, a lexicographic entry is a record in a linguistic resource that documents or discusses some properties of a given lexical item.

20Lexicographic entries are a special subset of a larger class called Lexicographic Component. Apart from whole dictionary articles (the entries), components can be used to represent senses, sense groups or subentries (like the substantivised verum) within lexicographic entries.

21It is important to stress once again that components represent only structural units; all linguistic information that is conveyed within these units must be expressed using Ontolex. The property lexicog:describes provides a link between the two dimensions, so that a lexicographic entry can be said to describe a lexical entry (as defined in Ontolex). In the same way, the lexicographic components that discuss a sense of a word or introduce a subentry, describe that specific lexical sense (as defined in Ontolex) or another lexical entry.

3.2 Lexicographic and Lexical Entries in the L&S

22The LLOD version of the L&S linked to LiLa is now available online in the LiLa KB.13 The entries can also be searched using LiLa’s query interface and SPARQL endpoint.14

23Figure 1 shows a visualisation of how the information from a sample entry, the adjective hosticus in the L&S dictionary, is represented in LiLa. In particular, the interplay between the linguistic and structural information is reflected in the complex relation between the lexical and lexicographic entries.

Figure 1: An entry in the LiLa’s representation of the L&S

24The L&S distinguishes two senses for the word: “belonging to an enemy, hostile” and “belonging to a stranger, foreign”. Following the Ontolex approach, these meanings are represented by the two ‘triangles’ between the lexical entry (the light green node on the left), the concepts evoked by the word (gray-blue nodes), and the senses, labeled 0 ad 1, that mediate between them (greenish-yellow nodes).

25The lexical entry is described by a lexicographic entry, identified by the id n21014 (inherited from the TEI XML file of the Perseus DL), while a specific lexicographic component describes each of the two senses (n21014_0 and n21014_1, respectively). What is particularly relevant is that the component n21014_0, which corresponds to the sense “hostile”, is linked to a sub-component that describes the lexical entry of the noun hosticum, a substantivised usage of the neuter adjective that means “the enemy’s territory”. That section of the entry that discusses the subentry “hosticum”, which is itself a section of the paragraph dedicated to the first sense, is thus linked (via the “describes” property) to a different lexical entry.

4. Conclusions and Future Work

26Perhaps even more than for any other modern language, a great number of lexical resources, either bi- or monolingual, is available for Latin, many of which have already been digitised and disseminated on the web. In this paper, we described a model of how this huge wealth of information can be published using the modern standards of the Semantic Web. The greatest advantage of this approach is that all the lexical resources published according to the same data model can be integrated in a wider network of linguistic information, along with the other digital resources that are connected to it. In the case of the L&S in LiLa, the Latin lexical entries of the bilingual dictionary can be queried together with the information about the same words provided by the other linguistic resources linked to the lemmas in the KB.

27One example of the fruitful interactions between resources is the possibility to investigate the polysemy of words in relation to their derivation, as recorded in the Word Formation Latin resource, which is also linked to LiLa (Litta, Passarotti, and Mambrini 2020). The adjective hosticus of Figure1, for instance, clearly inherits its two main senses (‘hostile’ and ‘foreign’) from the same polysemy of the noun hostis ’stranger’ or ’enemy’, from which it is derived. At the same time, while other resources in LiLa describe the senses of words, such as the Latin WordNet (Franzini et al. 2019; Mambrini et al. 2021), the complex relations between those senses (whether, for instance, one sense is interpreted as a specialised derivation from another) is generally available only in traditional lexical resources like the L&S.

28The solutions we found to address the challenges raised by the representation of the L&S in LLOD will be reused when we will link further bilingual, as well as monolingual, dictionaries of Latin to the KB. Including such lexical resources in LiLa is an important achievement, as it makes it possible for the KB to interact with linguistic (meta)data for languages other than Latin. Undoubtedly, such an inter-linguistic (re)use of distributed resources is one of the objectives of the LLOD community, to which LiLa contributes by steadily providing it also with new (kinds of) linguistic resources represented in LLOD.