3.1 Computational Approaches to Historical Linguistics

Computational approaches to historical linguistics can be roughly divided into those pertaining to language classification (e.g. Reference Ringe, Warnow and TaylorRinge et al. 2002), chronology (e.g. Reference Gray and AtkinsonGray & Atkinson 2003), cognate detection (e.g. Reference KondrakKondrak 2002), comparative reconstruction (e.g. Reference Hewson, Anderson and JonesHewson 1974) and the simulation of phonological (Reference BakerBaker 2008) and analogical change (Reference SkousenSkousen 1989).Footnote ¹ They usually involve quantitative methods to calculate the relationship between the languages or to date the chronology of their split from the proto-language, as well as algorithms to align cognates, reconstruct proto-forms or simulate sound changes (including artificial neural networks for analogical changes).

Although the use of quantitative methods is not new in contemporary linguistics – they were already being used in the 1930s – the application of computational methods in historical linguistics, like those used in evolutionary biology, represents a novel approach which is gaining adherents but is mostly regarded as problematic by traditional historical linguists. This is partly because of the history of quantitative approaches, which included methods such as glottochronology and lexicostatistics that are now found to be largely unreliable (cf. Reference Bergsland and VogtBergsland & Vogt 1962) or controversial at the very least (cf. Reference HoijerHoijer 1954). Additionally, many linguists approached such methods with extreme caution because they involved handling and converting the data to a machine-readable format, using statistical algorithms which seemed to generate a “black-box” effect rather than explain the results, and comparing language development to the replacement of genetic material in evolutionary biology. Compounding the problem was the fact that many of these early computational approaches were actually implemented by computational biologists rather than linguists.

Even though the methods and the data used have often not been applied very methodically or carefully (especially in the early works dealing with linguistic classification and chronology), there is no doubt that computational methods can be useful in linguistics, simply because computers can analyse masses of data in a short space of time and without errors. This might not be helpful in all aspects of historical linguistics, but it will certainly make it easier to check, for example, the coherence of a hypothesis put forward, to test competing hypotheses or the results of the application of a sound change etc. However, it has to be borne in mind that:

1. 1. The computer is only a tool and the results yielded will always ultimately depend on the quality of the algorithm, the input data and how they are converted into a machine-readable format (together with all the judgements made by the researchers at this point).

2. 2. Research results from using computational methods have to be interpreted, and the method itself is usually meant as a supplement to traditional methods, not a replacement.

In this chapter, I will outline the history of computational approaches to historical linguistics, concentrating on those concerning language classification and chronology, and then describe the three major contrasting approaches to the classification and chronology of the Indo-European languages. Following this, I will present the method used for the computational replication of sound change, which is a useful tool for testing the relative chronology of sound changes and may have some bearing on the grouping of the Indo-European languages. I conclude with some perspectives for future work.

3.2 Sources and Origins

Although most of the contemporary computational approaches to language classification and chronology stem directly from the methods used in evolutionary biology (cf. Reference Dunn, Bowern and EvansDunn 2015), they seem to have their indirect roots in the earlier quantitative approaches which used statistical and mathematical methods to calculate different aspects of language change and comparison.Footnote ² This clearly stems from the ongoing search for more objective bases to support the traditional arguments from historical linguistics, which were often seen as subjective as they depended largely on the assumptions of the scholars who performed the research.

One of the earliest quantitative approaches was a method devised by the Polish anthropologist Jan Reference CzekanowskiCzekanowski (1927), who tried to calculate the similarities between the Indo-European languages and present the results in a numerical way. Czekanowski approached the Polish linguist Jerzy Kuryłowicz for a list of characteristic binary features which could be used in distinguishing the different subgroups of Indo-European (at first twenty, then twenty-two) and proceeded to create the two-feature contingency tables in which those features were counted for every language that was being compared. Then he used Pearson’s tetrachoric correlation formula known from statistics. This allowed him to present a distance matrix of the Indo-European languages. His approach made its way to the United States through his student in anthropology, Stanisław Klimek, who came to study with Alfred Kroeber in the 1930s. Czekanowski’s method was adopted by both Reference Chrétien and KroeberAlfred Kroeber and Charles Douglas Chrétien (1937), who tried to count the similarities between the Indo-European languages using a broader range of features (seventy-four in total) taken from Meillet’s monograph on the Indo-European dialects (Reference Meillet1922). However, their results were criticised (cf. Reference SafarewiczSafarewicz 1948) for being biased from the very start because of the use of data based on a work which itself intended to prove the groupings of, for example, Italo-Celtic (Reference MeilletMeillet 1922). Probably the most famous approach was championed in the 1950s by Morris Reference SwadeshSwadesh (1952) in the form of glottochronology. Although promising at first, it was harshly criticised for working on the initial assumption that there is a constant rate of change in languages. Reference EmbletonEmbleton (1986) tried to further enhance the computerised methods of lexicostatistics and glottochronology and concluded that

3.3 Computational Approaches to Language Classification and Chronology

As mentioned above, most of the contemporary computational approaches to language classification and chronology stem from the methods used in biological sciences.Footnote ³ The computational approaches used in evolutionary biology were applied to linguistics in the late 1990s and began to be used in the early 2000s. They usually use statistical Bayesian inference to infer phylogenies. This kind of work has become very popular and has already been applied to different families of languages.

An interesting distance-based approach was pioneered by Søren Wichmann and his team (Reference Wichmann, Holman and Brown2018) in the Automated Similarity Judgement Program and the corresponding ASJP database (Reference Wichmann, Holman and BrownWichmann et al. 2018). The database includes a list of forty basic words for more than 5,000 languages and can be used, for example, to date when the languages in one family split away from each other. Because it uses the Levenshtein distance and lexical data, it is often regarded sceptically by linguists (cf. Reference GreenhillGreenhill 2011).Footnote ⁴ Additionally, there are some errors in the database itself. In the word list for Latin, for example, there is no vowel length present, and the words are transcribed inconsistently: wenire ‘to go’ is transcribed with /w/ whereas viya ‘road’ is transcribed with /v/.

One of the most controversial aspects of computational (or more accurately statistical) approaches to language classification and chronology is the fact that they are heavily based (often even exclusively based) on lexical data. In contrast, the standard procedure in traditional historical linguistics is the analysis of phonological and morphological features, and this is probably the main reason that many traditional historical linguists are generally very sceptical about using computational approaches for language classification and chronology. As pointed out by Gerhard Jäger and Johann-Mattis List in their recent comparison of traditional and computational methods, the crucial difference between the classical comparative method and the approaches adopted by computational historical linguistics is that “the comparative method strives to reconstruct the true history of languages in their entirety while statistical approaches search for probable or at least useful models of the observed patterns in some well-defined partial range of data”.Footnote ⁵

3.4 Application to Indo-European Studies

Apart from the lexicostatistical approach employed by Reference Dyen, Kruskal and BlackDyen et al. (1992) using the 200-word Swadesh lists for ninety-five languages and assuming a similar rate of change in all languages, probably the most famous computational classification of Indo-European languages was developed in 2002 by a team of experts combining linguistics (Don Ringe, Ann Taylor) and computer science (Tandy Warnow) in the project on Computational phylogenetics in historical linguistics (with contributions from statistician Steven Evans).Footnote ⁶ Using 22 phonological, 13 morphological and 259 lexical features as coded characters, they were able to produce a tree with a “perfect phylogeny” algorithm that tracked the branching of twenty-four ancient and medieval Indo-European languages. However, the phylogeny was not quite perfect since the position of Germanic could not be determined. As it turned out from subsequent work, which included language contact (Reference Nakhleh, Ringe and WarnowNakleh, Ringe & Warnow 2005), this was due to the fact that Germanic was apparently in contact with the other branches and therefore did not fit the “perfect phylogeny”.

Probably the most controversial computational approach to the subgrouping and chronology of the Indo-European languages was that adopted by Reference Gray and AtkinsonRussell Gray and Quentin Atkinson (2003). In their research, Gray and Atkinson used the word lists of basic vocabulary for eighty-seven Indo-European languages compiled by Reference Dyen, Kruskal and BlackDyen et al. (1992) along with Dyen et al.’s cognancy judgements and applied Bayesian inference to establish the dates for linguistic divergence of the languages analysed. They employed the algorithms for estimating the divergence time of DNA from evolutionary biology calibrated to the dates of the languages’ known split times. Using this technique, they were able to generate a tree in which the estimated dates of divergence of the particular groups of Indo-European languages were essentially in line with Colin Renfrew’s theory on the Anatolian origin of Indo-European languages (Reference RenfrewRenfrew 1987). The method was further expanded using phylogeographic approaches by Reference BouckaertBouckaert et al. (2012), with the results also pointing to an Anatolian origin.

The work of Reference Gray and AtkinsonGray & Atkinson (2003) and Reference BouckaertBouckaert et al. (2012) was challenged by a team from the University of California, Berkeley (Reference Chang, Cathcart, Hall and GarrettChang et al. 2015). They tried to use the same method but with the addition of ancestry constraints, i.e. information relating to the fact that Latin is the parent language of Romance etc. Their research indicated that the chronology of the Indo-European splits was significantly shorter than previously thought and roughly in accordance with the dating of the so-called steppe hypothesis in archaeology as regards the homeland of the speakers of Proto-Indo-European. Although the authors claim that “the agreement between our findings and the independent results of other lines of research confirms the reliability of statistical inference of reconstructed chronologies” (Reference Chang, Cathcart, Hall and GarrettChang et al. 2015: 194), there are problems inherent in the Bayesian inference itself (cf. Ringe in Chapter 4 and the comments by Reference Nichols and WarnowNichols & Warnow 2008: 785 on Bayesian methods).

At present, it is difficult to say whether computational methods can give us a reliable chronology of the splits of the individual branches.Footnote ⁷ It is clear, though, that they can provide us with a reliable indication of how the trees (topology) branched, as the work by Reference Ringe, Warnow and TaylorRinge et al. (2002) has shown. It is also striking that most of the approaches identify the traditionally assumed subgroups (Indo-Iranian, Balto-Slavic, Italo-Celtic, Graeco-Armenian). Even with careful calibration, perhaps the best we can hope for is a very rough estimate. However, even this should be corroborated by linguistic, archaeological, genetic and historical data.

3.5 Computational Replication of Sound Change (Computerised Forwards Reconstruction)

Another approach which could yield promising results in the reconstruction of the Indo-European family tree, especially with regard to the relative chronology of sound changes, is the replication of sound change or computerised forwards reconstruction (cf. Reference Sims-WilliamsSims-Williams 2018).Footnote ⁸ The procedure of reconstructing forwards is not unknown in traditional historical linguistics. For example, Calvert Reference WatkinsWatkins (1962: 5) mentions it in his monograph on the Indo-European origins of the Celtic verb, where he employs the method to see how a Proto-Indo-European form would be regularly continued in Celtic. The method was also adopted by Ives Reference Goddard, Melchert and H. JasanoffGoddard (1998: 183) in his analysis of the Arapaho historical phonology, which, like the phonology of the insular Celtic languages, underwent some radical changes.

The aim of such an approach, enhanced by using a computer, is to employ an algorithm that reads the given input data (proto-forms), makes the appropriate changes based on the changes that are usually assumed to have taken place in the development of the particular languages (regular sound changes) and generates the output which can be either manually or automatically compared with the actually attested forms. The purpose of this approach is to test the regularity of sound changes and their relative chronology. If the generated form is exactly the same as the attested one, then the relative chronology is assumed to be correct.

However, only two programs known to the author have tried to replicate sound changes comprehensively by taking into account most of the changes (Reference ManietManiet 1985; Reference Hartman, Miyares, Álvarez Moreno and SilvaHartman 2003). Most of the other programs have only used fragmentary data and applied only the main sound changes. I will list and discuss the most important ones below.Footnote ⁹

The first application of historical derivation was a program by Raoul Reference SmithSmith (1969) which applied twenty-one regular rules to 650 Proto-Indo-European reconstructed lexemes as taken from an etymological dictionary and, through the application of those rules, derived the modern Russian forms from the Proto-Indo-European ones. However, only nine lexemes were derived regularly because too few sound-change rules were applied.

Further attempts were made by Reference Burton-HunterBurton-Hunter (1976) to generate Old French forms from their Latin sources and by Reference EastlackEastlack (1977), to produce Ibero-Romance from its Latin source. Both programs yielded a high percentage of expected forms since they took into account a larger number of rules within a shorter period of time (Latin to Old French or Old Spanish).

Probably one of the most comprehensive computational sound-change replication programs was the one devised by Albert Reference ManietManiet (1985). He simulated the changes from Proto-Indo-European to Latin (252 rules) on a corpus of approximately 15,000 words from Plautus. However, his research went largely unnoticed by both linguists and computer scientists, and the program is, unfortunately, no longer accessible today.

In more recent times, in his thesis on cognate alignment and reconstruction, Reference KondrakKondrak (2002: 141–3) included a short appendix on a Perl program which generated Polish forms from their Proto-Slavic sources. From the 626 lexemes taken into account, 72.5 per cent were regular.

Reference Hartman, Miyares, Álvarez Moreno and SilvaHartman (2003) has been developing a program since the 1980s which simulates the sound changes from Latin to Spanish (with approximately 122 rules and 1,806 coded vocabulary items) using sets of distinctive features (from Reference Chomsky and HalleChomsky & Halle 1968) coded as binary strings rather than the usual string substitution as in most programs developed so far. As input, the etymon is “fed” into the computer via the keyboard or from a file. Then the individual letters are translated into sets of features, and changes are applied to the features in accordance with the programmed rules and their relative chronology. Finally, the features are translated back into characters and displayed using the International Phonetic Alphabet.

Hartman also enabled the program itself to be used to model sound changes in other languages. Recently, the hard-coded version with the Spanish model was made available on-line, and a new working version of the program was presented (Reference HartmanHartman 2018) but with no significant modifications. The earlier version of Hartman’s program was used by Towhid bin Muzzafar to simulate the changes from Proto-Algonquian to Shawnee (Reference bin Muzzafarbin Muzzafar 1997). Apart from some additional examples that he was able to find for Shawnee and minor improvements in the relative chronology of changes, he pointed out an important aspect of the computational replication of sound change which probably constitutes the main obstacle to this kind of work:

Some more recent programs, apart from the ones developed by people to create fictitious languages (so-called “Conlangers”), include the simulation of Spanish from Latin by Marcel Schmuki (“ETYMO”, 2001), Proto-Germanic from Proto-Indo-European by Brett Kessler (“Derive”, 2004), Gothic from Proto-Indo-European by Roland Reference Mittmann, Chiarcos, de Castillo and StedeMittmann (2009) and the “Sound change transducers” by Reference ZeldesAmir Zeldes (2008).

Sims-Williams tried some computational replication of sound change in Celtic historical phonology just by using the “Find & Replace” function in the word processor (Reference Sims-WilliamsSims-Williams 2018: 562). By applying forty-three sound changes on the material of 159 selected Common Celtic forms, he was able to find amendments in the relative chronology of changes and identify usually overlooked Celtic cognates from a Proto-Indo-European root. He further argued that with modern programs, this kind of research could be pursued and expanded significantly so that it might include complete Proto-Indo-European reconstructions as input to find new etymologies, correct existing ones and check the relative chronology of sound changes (Reference Sims-WilliamsSims-Williams 2018: 564). I agree with the author in principle, although I think implementing such an expanded approach is quite complicated, because of several key factors:

1. 1. circularity of the method – there are competing hypotheses on what the Proto-Indo-European reconstructed forms should look like, and the applied sound changes can be used to fit the proto-form

2. 2. problems in “translating” the sound changes into computational notation

3. 3. problems in coding the forms and characters as the input for the program

4. 4. lack of high-quality computational databases for Indo-European languages and Proto-Indo-European which would also include verbal and nominal paradigms

5. 5. lack of monographs and works that comprehensively present the complete relative chronology of changes from Proto-Indo-European to the attested languages

6. 6. not applying morphological changes which are usually assumed to have been very numerous in paradigms and which will inevitably blur the regular outcomes from simulations, especially if the simulation has a very large time span (e.g. from Proto-Indo-European to Latin etc.).

To give just one example, if we take the Proto-Indo-European reconstructed form for ‘foot’ (cf. (Reference Wodtko, Irslinger and SchneiderWodtko et al. 2008: 530 n. 2) and apply the sound changes that occurred in Attic Greek (cf. Reference LiesnerLiesner 2015: 110–15), we will get the following result:

Apart from the fact that some forms do not match the attested ones because of not being included in the morphological changes, we will encounter problems with the reconstruction of the proto-form, since different scholars propose competing forms: nom.sg. *pṓds, gen.sg. *pedés (Reference RingeRinge 2017a: 59) or nom.sg. *pṓds, gen.sg. *pedós (Reference ClacksonClackson 2007: 72). There are also problems with the assumed sound changes (cf. Reference SzemerényiSzemerényi 1996: 116 for the view that *póds developed into *póss and Greek *pṓs).

3.6 Potential Ways to Enhance the Computational Replication of Sound Change

In this section, I will try to address some of the problems involved in computational replication of sound change.

3.7 The Potential Use of Relative Chronology of Sound Changes in Subgrouping

In an article devoted to the position of West Germanic, Don Ringe observed that

If that is so, then it would be possible to use the computational replication of sound change in this area as well, depending on the quality and the availability of the data as discussed below.Footnote ¹²

3.8 Perspectives on the Inclusion of Morphological Changes

It may be possible to expand the scope of the computational replication of sound change in such a way as to apply computationally generated sound changes along with morphological changes to the complete lexicon of the Proto-Indo-European language as it is reconstructed today in order to generate the main Indo-European languages. The program would apply the sound changes and the analogical changes in their relative chronology to the lexicon of Proto-Indo-European and generate output which in turn would be compared with the actual data relating to those languages. With this approach, the amount of regular sound change from a more or less complete lexicon would be uncovered along with the exact interferences causing irregularities – errors in the formulation, chronology or translation into computational terms of the programmed sound changes, borrowing and, especially, analogy.

This approach could potentially address a very direct and practical question of interest to every practising historical linguist: whether one analogical solution is more probable than another. The usual answer to this question depends on one’s view of the system of the language in which the change occurred. However, different scholars might view a certain language system (in its earlier or even reconstructed phases) differently and so pose different analogical explanations with their own models and motives. They will look for parallel developments and typologically similar changes in the material they are working on and, most importantly, in their previous experiences. We can deem one solution of analogical remodelling as more plausible than another by providing other analogical solutions of exactly the same type, with similar models and motives along with an in-depth analysis of the synchronic situation. Warren Cowgill, in his work on universals in Indo-European diachronic morphology, noted that with regard to small-scale innovations “[a] sufficiently large collection of such individual changes, appropriately classified, should give linguists measure of the relative plausibility of different solutions for problems in historical grammar” (1966: 115). He continues:

Using a computational algorithm and an electronic database with word forms and the phonological and morphological changes which occurred in their development, it would be possible to create a virtually complete picture of the phonological and morphological system of the language at every stage of its development and to investigate any possible phonetic and analogical changes. Because the reconstruction of the proto-language and each of its stages of development remains hypothetical, its validity and accuracy can only be checked against the general typology of both synchronic language systems and types of diachronic changes along with the internal coherence of the system. Most notably, the compatibility of every single sound change can be checked against the hypothesis by applying it to the lexicon of the language.

Such an approach would allow us to formulate hypotheses concerning the relative chronology and tendencies of sound change and analogical levelling based on fairly complete empirical data. The results would confirm or challenge the existing theories on sound change and analogical remodelling and could form the basis for comprehensive historical grammars in the future which, with the expansion of integrated corpus linguistics, could encompass all corpora of texts from all periods of the documented language development. Such a large database would enable scholars to pursue further research in the area, allow the explicit discussion of competing hypotheses and serve as an educational tool. Additionally, the method itself could be applied to other language families, thus forming the basis for research on universal tendencies in language change. Moreover, it would break the so-called “handbook tradition” mentioned by Reference Eichner, Panagl and KrischEichner (1992: 61), whereby a sound change is illustrated only by a handful of examples (usually the same in various historical grammars) and in order to find more of these, one has to consult an etymological dictionary.

In fact, a recent project carried out by Jouna Reference Pyysalo, Tiedemann and TahmasebiPyysalo (2017) has managed to achieve most of what is described above. With the use of computational simulation, the project aims to generate all the forms of the Indo-European languages from their reconstructed proto-forms. However, the author uses an idiosyncratic reconstruction of Proto-Indo-European which deviates significantly from the current communis opinio, i.e. with only one laryngeal, at the same time basing his argument against the classical laryngeal theory only on Anatolian data and completely ignoring the data relating to Ancient Greek (cf. Reference Janhunen and PyysaloJanhunen & Pyysalo 2018). If the input proto-forms are hard-coded so that they cannot be modified, this project will only serve to present the author’s own views on the subject.

3.9 Perspectives on the Computational Methods

Czekanowski, Kroeber and Chrétien pursued an interesting way of handling language classification back in the days when such methods were far from popular or even acceptable. They were often very harshly criticised by linguists, to the extent that virtually nobody followed their lead to improve the method. Just as Czekanowski, Kroeber and Chrétien were pioneers in the use of statistics for language classification applied to Indo-European, so were Ringe et al. and Gray and Atkinson pioneers in the use of computational methods in the same area. Even though their work is relatively recent, a large amount of new research has been done in the field, which has become so popular that some scholars argue that historical linguistics appears to have taken something of a quantitative turn. Indeed, new methods are being implemented in an attempt to meet the standards of traditional historical linguistics in paying careful attention to the annotation of the data and to detail in general, while at the same time taking advantage of the replicability, robustness and formality of the computational approach. Progress is being made in modelling language characteristics and change using algorithms, and more attention is being given to making the programs and the data openly accessible, in a more or less standard format, easy to use by non-computational scientists and, importantly, annotated jointly by experts in the respective fields. It seems that only through combining qualitative and quantitative methods can further progress be made in the field of Indo-European linguistics, and the current thinking is that such progress is only possible if scholars from different disciplines contribute collectively.

The advantages of the computational approach are its speed, error-free processing of data and ability to handle large amounts of data. I would further argue that, thanks to the computational approach and the explicit presentation of the material, it would be much easier to compare different linguists’ competing hypotheses, even for people from outside the exact field of specialisation, thus making the whole enterprise of Indo-European linguistics easily accessible for interdisciplinary studies.