This project represents a computational framework for historical linguistics. The general purpose of the CoToHiLi project is to integrate expert knowledge and computational power to address cognate identification, cognate-borrowing discrimination, Latin protoword reconstruction and semantic divergence. The goal of the project is twofold: 1) to automate certain parts of the traditional work-flow of the comparative method (such as the collection of data or the automatic alignment based on predefined or inferred rules), and 2) to bring new insights or avenues of investigation, which might not be easily accessible otherwise (e.g., the automatic identification of patterns and regularities in large amounts of data). The project will provide tools for the main Romance kernel group (French, Italian, Portuguese, Romanian, Spanish), as well as Latin. The methodologies and computational tools proposed could also serve as a basis for further development for other comparable language families, including less studied languages, with scarce resources available.
2018) explored the potential correlation of genetic and linguistic distances, starting from what he called The general purpose of the CoToHiLi1 project is Darwin’s last challenge: “If we possessed a perfect to integrate expert knowledge and computational pedigree of mankind, a genealogical arrangement of power to address the following topics: cognate iden- the races of man would aford the best classification tification, cognate-borrowing discrimination, Latin of the various languages now spoken throughout proto-word reconstruction and semantic divergence. the world; and if all extinct languages, and all inOur project is focused on the Romance languages termediate and slowly changing dialects, were to (French, Italian, Portuguese, Romanian, Spanish), be included, such an arrangement would be the and will provide tools for the main Romance kernel only possible one” (see also [2]). Given that the group and for Latin. The duration of the project is socio-economical and cultural factors are some of 3 years, starting from January 2021. the motivations for borrowing from one language
The research problems that we address are sig- to another [1, 3], the topic of this research project nificant on multiple levels. From a scientific point facilitates reconstructing certain aspects related to of view, any advance in historical linguistics is of society and culture for groups of people speaking paramount cultural importance, being inherently a given proto-language, and gaining insights into connected with human history (“each word a his- their past social interactions and into their social tory”, cf. [1]). Longobardi (LanGeLin project, 2012- and cultural practices [3]. Moreover, establishing the direction and source of borrowing is important to our understanding of the social relations between the groups involved. From a technological perspective, as linguistic change is the most visible at the lexical and semantic level, computational tools can be designed to serve both aspects, for instance to automatically identify related words and to assess the semantic change. Even though historical lexicology has leveraged technological advances, and some pioneering work was initiated on various steps of the work-flow (cognate identification, proto-word where reconstruction), historical semantics has not sufiupdates: ciently benefited from the advances in computer science. Yet, by drawing special attention to the English. Our aim is to track the semantic change of semantic divergence occurring in pairs of cognates, words in Latin and across multi-languages, in the we could both take a few steps forward towards Romance language family, for the first time, with a unitary theory of semantic change, and improve the substantial purpose of looking for common patpractical applications such as automatic translation terns characterizing the overall semantic divergence systems or language e-learning systems, aware of cases. Additionally, we intend to explore the stafalse friends and related phenomena. tistical properties of the word embedding vectorial spaces [9, 10].
The innovation of the project consists in integrating linguists’ knowledge with new computational meth- The methodologies and computational tools we proods in a unified framework, to address important pose could extend their applicability not only to problems from historical linguistics, enabling ex- various linguistic branches in the Indo-European perts to provide input and feedback throughout the family, but also to less studied languages or linwhole development process, in the pre-processing, guistic families. Such advances could provide new annotation, feature engineering, training and evalu- answers in historical and social sciences, given that ation phases: lexical and semantic change is a key source of clues regarding both the dynamics of cultural interactions 2.1. Identification of related words between groups in the past, and the technological innovations and exchanges that have taken place We aim at going one step further than the cur- across space and time [3]. Moreover, the semantic rent state-of-the-art methods by: a) proposing a data provided by the CoToHiLi project could be more in-depth analysis, by identifying the direction of great help for the cognitive sciences and neuroof the borrowings and b) automatizing the whole sciences [11], to the extent that they can ofer a new process as a pipeline that, given a pair of input perspective on our brain mechanisms. words, provides an automatic analysis regarding the As for the socio-economic impact of the CoTorelationship between them [4, 5, 6]. HiLi project, in the context of the increasing number of attempts to create automatic tools designed 2.2. Latin proto-word reconstruction for linguistic comprehension, our computational devices could support Romance intercomprehension To improve previous results, we intend to use more by bringing into light the common linguistic fearecent techniques [7], such as conditional random tures, as well as the semantic relations between the ifelds (CRF) for sequence labelling and deep learn- Romance cognates or borrowings. Such an advance ing, in particular character-level neural networks. can prove its usefulness in the constant eforts to The alignment technique, which stands at the foun- improve the automatic translation systems. dation of our approach, will be improved by an heuristic for choosing the best alignment. We also address the challenging problem of multiple 4. Methodology alignment (finding an alignment for more than two words), in order to be able to extract knowledge from cognate sets in multiple languages. Another promising line of research is to make use of the more recent Latin resources, such as The Latin Diachronic Database [8].
For the first two objectives, our methodology is
focused on two main aspects: creating clean datasets
and developing computational methods for
achieving the proposed research tasks. For the Romance
languages there are already some existing resources
(for cognates, for borrowings and for proto-word
reconstruction), but they are scattered, incomplete,
2.3. Diachronic semantic divergence or with uncertain availability (cf. [12, 13]). Thus,
Semantic change is a continuous and complex pro- datasets do not have to be built from scratch, but
cess ([1] presents no less than 11 types of semantic the data need to be harmonized, verified and
enchange), which has been recently studied in the hanced where necessary, in order to become a
benchcontext of distributional semantics theory. Vec- mark in the domain. By using computational tools,
torial representations of word meaning (word em- corroborated by the direct intervention of classical
beddings) have been used for tracking semantic linguists, we have already built a significant part
shifts across diferent time periods, especially for of the database, representing the starting point for
3. Impact
the computational methods that are being devel- automatically discriminating between inherited and
oped. We have continued with the alignment of borrowed Latin words. We have introduced a new
word pairs. Given the lack of an unanimously ac- dataset and investigated the case of Romance
lancepted alignment method [
For the third objective, in order to identify seman- scribed the types of errors that occurred during tic shifts across time periods as well as languages, the automated cognate identification process and we leverage vector space representations of meaning, manually corrected them. Additionally, we labeled or word embeddings, relying on traditional models the English cognates according to their etymology, such as word2vec and FastText [15, 16], as well as separating them into two groups: old borrowings experimenting with state-of-the-art language mod- and recent borrowings. On this curated dataset, els such as BERT [17]. The method consists of we analysed word properties such as frequency and building vectorial semantic representations for the polysemy, and the distribution of similarity scores words in each of the target languages, based on the between cognate sets in diferent languages. We multilingual corpora, and then obtaining a shared automatically identified diferent clusters of English multilingual semantic space. This will allow us to cognates, setting a new direction of research in cogcompute semantic distances between cognates as nates, borrowings and possibly false friends analysis well as analyze the statistical and the linguistic in related languages [10, 19]. properties of words whose meanings have diverged.
The available corpora are unequal from one language to another; for instance, the Royal Spanish 6. Conclusions Academy provides an exhaustive diachronic corpus of its language, whereas for Romanian we only have Drawn within a computational framework, the Coaccess to a scarce data-base, composed of a fairly ToHiLi project addresses key concerns of historilimited number of old texts. In order to ensure the cal linguistics centered on the Romance languages, accuracy of our analysis, in this stage of the project, such as cognate identification, cognate-borrowing we use mainly lexicographic resources, as well as discrimination, Latin protoword reconstruction and data-bases built for the contemporary stage of each semantic divergence, towards which we have taken language (such as multilingual Wikipedia2). a few steps forward by performing various experiments. At this stage of the project, we analyze only the main five Romance languages (French, Ital5. Current Results ian, Portuguese, Romanian, Spanish), but as we advance we intend to include other Romance idioms For the first two objectives, we have started building as well. We predict that the methodologies and datasets of cognates and borrowed words for the computational tools proposed will also serve as a Romance languages [18]. This first step relies on basis for further development for other comparable dictionaries that contain etymological information language families, including less studied languages, (e.g., for Romanian we use 13 dictionaries available with scarce resources available. in digital format). We have proposed a new method