This paper describes the steps taken to include data from the Lewis & Short bilingual Latin-English dictionary into the Knowledge Base of linguistic resources for Latin LiLa. First, data were extracted from the original XML and matched with entries in LiLa, overcoming ambiguities and structural inconsistencies in the source. Subsequently, senses were modelled using the Ontolex Lemon Lexicographic module (lexicog), so that they could be included in the LiLa Knowledge Base and thus made interoperable with the (meta)data of the linguistic resources for Latin therein interlinked.
Since 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
Particularly 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
Copyright © 2021 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
1See, for example, Musisque deoque for Classical Latin
poetry
However, 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
meta2Index Thomisticus Treebank
3Such as, for instance, valency and subcategorisation
lexica
4https://www.clarin.eu. 5https://logeion.uchicago.edu/lexidium. dictionary that allows to query together the lexical entries of several dictionaries for Ancient Greek and Latin, while Corpus Corporum6 is a metacollection 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.
Syntactic 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
Attaining 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
Based 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.
The lexical resources for Latin linked so far
to LiLa include a word formation lexicon
Net and a valency lexicon
This 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 2.1
The 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
In 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
worklfow used for the Greek-English Lexicon
Perseus’ 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.
In 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).
Entries 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.
8One example is the app Diogenes for querying corpora of Greek and Latin texts: https://d.iogen.es/.
9The digital edition is available from the repository of the Perseus DL and is distributed under a CC BY SA 4.0 license: https://github.com/PerseusDL/lexica.
10See https://tei-c.org/release/doc/teip5-doc/en/html/ref-entryFree.html. 2.2
The 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.
In 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
partof-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
As 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
Ontolex 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.
In order to overcome these issues, the Ontolex
community has developed a specific extension of
the ontology called the “OntoLex lexicography
module” or lexicog
As 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
11http://www.w3.org/ns/lemon/ontolex#c anonicalForm.
12https://www.w3.org/ns/lemon/lexicog#.
Lexicographic 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.
It 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
The 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
Figure 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.
The 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).
The 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 13http://lila-erc.eu/data/lexicalResour ces/LewisShort/Lexicon.
14https://lila-erc.eu/query/, and https: //lila-erc.eu/sparql/. 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
Perhaps 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.
One 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
The 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.
This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme – Grant Agreement No. 769994.