Content uploaded by Steven Moran
Author content
All content in this area was uploaded by Steven Moran on Mar 25, 2017
Content may be subject to copyright.
Zurich Open Repository and
Archive
University of Zurich
Main Library
Strickhofstrasse 39
CH-8057 Zurich
www.zora.uzh.ch
Year: 2013
Investigating the Relatedness of the Endangered Dogon Languages
Moran, Steven; Prokić, Jelena
Abstract: In this article we apply up-to-date methods of quantitative language comparison, inspired by
algorithms successfully applied in bioinformatics to decode DNA and determine the genetic relatedness
of humans, to language data in an attempt to shed light on the current situation of a family of languages
called Dogon, which are spoken in Mali, West Africa. Our aim is to determine the linguistic subgroupings
of these languages, which we believe will shed light on their prehistory, highlight the linguistic diversity
of these groups and which may ultimately inform studies on the cultural boundaries of these languages.
DOI: 10.1093/llc/fqt061
Posted at the Zurich Open Repository and Archive, University of Zurich
ZORA URL: http://doi.org/10.5167/uzh-84673
Originally published at:
Moran, Steven; Prokić, Jelena (2013). Investigating the Relatedness of the Endangered Dogon Languages.
Literary and Linguistic Computing, 28(4):676-691. DOI: 10.1093/llc/fqt061
Investigating the relatedness of the
endangered Dogon languages
............................................................................................................................................................
Steven Moran
University of Marburg and University of Zurich
Jelena Prokic
´
University of Marburg
.......................................................................................................................................
Abstract
In this article we apply up-to-date methods of quantitative language comparison,
inspired by algorithms successfully applied in bioinformatics to decode DNA and
determine the genetic relatedness of humans, to language data in an attempt to
shed light on the current situation of a family of languages called Dogon, which
are spoken in Mali, West Africa. Our aim is to determine the linguistic sub-
groupings of these languages, which we believe will shed light on their prehistory,
highlight the linguistic diversity of these groups and which may ultimately inform
studies on the cultural boundaries of these languages.
.................................................................................................................................................................................
1Introduction
This year’s theme of the Digital Humanities confer-
ence is ‘digital diversity: cultures, languages and
methods’. Our research fits in naturally with this
motif because human cultures and languages are
intimately intertwined with each other. What is
less obvious is that culture and language both dis-
seminate in the same ways: through genealogical
descent and through areal contact. Although, it is
incredibly difficult to track the dissemination of cul-
tures among societies, there are tried and tested
methods for determining the genealogical related-
ness of languages.
In this article we apply up-to-date methods of
quantitative language comparison, inspired by algo-
rithms successfully applied in bioinformatics to
decode DNA and determine the genetic relatedness
of humans, to language data in an attempt to shed
light on the current situation of a family of lan-
guages called Dogon, which are spoken in Mali,
West Africa. Our aim is to determine the linguistic
subgroupings of these languages, which we believe
will shed light on their prehistory, highlight the lin-
guistic diversity of these groups and which may ul-
timately inform studies on the cultural boundaries
of these languages.
Some of the Dogon languages are spoken by rela-
tively small groups and thus can be considered
endangered languages. The study of endangered
languages and elaborate cultures is particularly
urgent at this time in history, so we start by describ-
ing in Section 2 factors of language endangerment
and we provide an overview of Dogon. In Section 3
we describe the data that we use in this work and in
Section 4 we discuss the theory of how languages
can be compared to establish genealogical related-
ness. In Section 5 we discuss modern methods of
quantitative language comparison and how they
have been adapted from biology to linguistics. In
Section 6 we give the analyses and results of apply-
ing two quantitative methods, the normalized edit
distance (NED) and LexStat approaches, to the
Dogon data. In Section 7 we provide an evaluation
of these approaches and discuss their shortcomings.
Lastly, in Section 8 we give our conclusion.
Correspondence:
Steven Moran, Seminar fu
¨r
Allgemeine
Sprachwissenschaft,
Plattenstrasse 54, Raum 206,
University of Zurich, 8033
Zurich, Switzerland.
Email:
steven.moran@uzh.ch
Literary and Linguistic Computing ßThe Author 2013. Published by Oxford University Press on
behalf of ALLC. All rights reserved. For Permissions, please email: journals.permissions@oup.com
1of 16
doi:10.1093/llc/fqt061
Literary and Linguistic Computing Advance Access published September 17, 2013
by guest on September 18, 2013http://llc.oxfordjournals.org/Downloaded from
2Background
2.1 Language endangerment
Current estimates are that there are around 7,000
languages spoken in the world today (Lewis et al.,
2013). Since the early 1990s, linguists have estimated
that half of the world’s languages will no longer be
spoken by the end of this century (cf. Hale et al.,
1992). Recent findings based on precise data from
the Catalogue of Endangered Languages show that
43% of the world’s languages are currently endan-
gered and that the rate at which languages are dis-
appearing, roughly —three to four per year, has
highly accelerated in the last forty years (Campbell
et al., 2013).
1
These figures represent a dramatic
loss of the number of languages and the range of
linguistic diversity of humans. When a language is
lost, humanity loses a piece of the human know-
ledge base, which includes an irreplaceable encoding
of culture-specific ways of thinking (Harrison,
2007).
Most people are familiar with the idea of endan-
gered species and the extinction of animals, that is,
there are very few remaining animals, or none remain-
ing at all, of a particular species due to factors such as
the destruction of the species’ habitat, over hunting or
pollution. A lesser known fact is that languages are
becoming extinct at a rapid rate. But why is it that
languages become endangered and/or extinct?
There are several causes of language endanger-
ment (cf. Austin and Sallabank, 2011), including
natural catastrophes, famine, and disease. For ex-
ample, the small number of speakers of languages
of the Andaman Islands were seriously affected by
the Indian Ocean earthquake tsunami in 2004. War
and genocide are also a factor that can lead to lan-
guage endangerment and extinction; for example
the atrocities of genocide by colonists in
Tasmania, where all indigenous Tasmanian lan-
guages have been lost. Another factor, repression
and forced linguistic and cultural assimilation, has
lead to a decrease in the number of Native American
languages in both North and South America. There
are also influences from culturally, politically, and
economically dominant cultures and countries.
Often these causes of language endangerment
overlap with each other, but they typically share
common influences that lead to language shift. The
degree of language shift in a community of speakers
determines the intergenerational language transmis-
sion rate, which is one way of measuring language
vitality and endangerment. For example, UNESCO
has a language vitality and endangerment frame-
work, given in Table 1.
Language endangerment is happening all over
the world and it is of the utmost urgency that lin-
guists document and describe these languages before
they disappear.
2
In many places in West Africa, for
example, there is language shift towards colonial
languages like English and French or towards large
lingua-francas like Arabic and Swahili. Basically, as
an increasing number of people become economic-
ally mobile and educated, those people may try to
provide better future opportunities to their children
by teaching them the more ‘powerful’ (or presti-
gious) language as their mother tongue, instead of
passing on their own smaller, lesser known lan-
guage.
3
One area where this shift is occurring is in
Dogon country, a region in Mali renowned for its
secluded and geographically divided village commu-
nities, some of which are embedded in an escarp-
ment, that is, a sandstone cliff that rises 500 m
(1500 feet) above the plains below it.
4
2.2 Dogon languages
Dogon languages are spoken predominantly in east-
ern Mali in West Africa. The Dogon people were
made famous by Marcel Griaule, a French anthro-
pologist who pioneered ethnography in France, and
who worked with the Dogon between 1931 and
1956. Reportedly the Dogon had advanced astro-
nomical knowledge of the Sirius binary star
system, knowledge that is not possible without tele-
scopes. Since then, the Dogon have been shrouded
in controversy and mystery.
As late as 1989, Dogon appeared in reference
books as if it were a single language, for example.,
Bendor-Samuel (1989). In 2004, an extensive socio-
linguistic survey by Hochstetler et al. (2004) esti-
mated no less than seventeen distinct languages
and described the language family as highly intern-
ally divided. The standard encyclopedic reference on
the world’s languages, the Ethnologue,
5
recently
increased the number of Dogon languages that it
S. Moran and J. Prokic
´
2of 16 Literary and Linguistic Computing, 2013
by guest on September 18, 2013http://llc.oxfordjournals.org/Downloaded from
lists from 14 to 19 (Lewis, 2009;Lewis et al., 2013),
but this figure is still too low. Since 2004, much
initial survey work on Dogon has been undertaken
by Professor Jeffrey Heath’s Dogon Languages
Project (DLP),
6
which has lead to the ‘discovery’
of a web of divergent dialects, some of which have
been raised to the status of distinct languages based
on standard linguistic criteria. The team has identi-
fied between 80 and 100 locally named varieties that
have been tentatively group into approximately 20–
25 languages. Nevertheless, the current Dogon lin-
guistic situation is not at all transparent and much
work is needed to document and describe these lan-
guages. Dogon languages are very under-described,
some are highly endangered, and all are genealogic-
ally not well established (Blench, 2005;Heath,
2008). And although, Dogon is generally considered
to be a division of the vast Niger–Congo language
family, much work remains to establish this firmly.
The DLP provides a tentative yet detailed inventory
of known Dogon languages. There are currently
twenty distinct languages grouped (crudely) into
eight geographical regions, with no implications for
genealogical subgrouping. The internal structure of
the Dogon language family is unknown, as is the
number of mutually unintelligible languages it con-
tains. In fact, the Ethnologue gives a flat genealogical
tree. The position of the Dogon languages relative to
other African language families is also unclear because
of Dogon’s typological characteristics. Its lineage has
long been disputed, as summarized in Table 2.
7
Thus, the current Dogon linguistic situation is
not at all clear. Table 3 provides an approximate
number of Dogon speakers, based on recent field-
work by members of the DLP.
One important factor regarding the need to
understand the linguistic situation in Dogon has
to do with education. Over the decades there have
been several initiatives by the Malian government
that involved sociolinguistic studies to determine
which particular Dogon should be the ‘official’ lan-
guage for all Dogon people. Or in other words, the
studies were aimed at determining which Dogon
language was the most mutually intelligible by the
most number of speakers of other Dogon languages.
The ‘official’ Dogon language, currently determined
to be Toro So by crude lexicostatical methods
(Hochstetler et al., 2004), is now the language in
which all literacy and written materials for schools,
Table 1 UNESCO’s language vitality and endangerment framework
Degree of endangerment Intergenerational language transmission
Extinct There are no speakers left
Critically endangered The youngest speakers are grandparents and older, and they speak the language partially and
infrequently
Severely endangered Language is spoken by grandparents and older generations; while the parent generation
may understand it, they do not speak it to children or among themselves
Definitely endangered Children no longer learn the language as L1
Vulnerable Most children speak the language, but it may be restricted to certain domains (e.g. home)
Safe Language is spoken by all generations; intergenerational transmission is uninterrupted
Table 2 Historical classification of Dogon
Year Classification
(language family)
Author
1924 Nige
´ro-Se
´ne
´galais Delafosse
1941 Voltaic (Eng. Gur) Homburger
1948 Voltaic; Gurunsi Baumann and
Westermann
1951 Mande
´Holas
1952 Mande
´Delafosse
1952 Gur (Fr. Voltaic) Westermann and Bryan
1953 Voltaic Bertho
1953 Non-classified de Tressan
1950/60 Gur Calame-Griaule
1963 Gur Greenberg
1971 Gur Bender-Samuel
1981 Voltaic Manessy
1981 Volta-Congo Bendor-Samuel
1993 Unresolved; non-classified Galtier
1994 Unresolved; non-classified Plungian and Tembine
´
2000 Ijo-Congo Williamson and Blench
2009 Volta-Congo Lewis
Relatedness of the endangered Dogon languages
Literary and Linguistic Computing, 2013 3of 16
by guest on September 18, 2013http://llc.oxfordjournals.org/Downloaded from
etc., are to be developed. However, as linguists
know, determining how closely a pair of languages
are related is no trivial task and forcing groups of
different language speakers to learn the same lan-
guage can lead to dire consequences for minority
languages.
3Dogon Languages Data
There is much language and cultural data about the
Dogon people that was collected in the field and is
made available online through the DLP website.
8
These data include photos of flora and fauna,
photos of lexical senses that illustrate areas other
than flora and fauna, and photos of villages and
topographic features of Dogon country. The website
also provides videos of festivals, events, music, food
and drink preparation, agriculture, pottery, hunting
and gathering, and much more. The project has an
even stronger focus on disseminating language data
via the website. Grammars, dictionaries, and typo-
logical discussions are made openly available, as is
the raw lexical data in an Excel spreadsheet, an
example of which is shown in Table 4.
The Dogon comparative spreadsheet, which con-
tains nearly 9,000 rows (meanings) and 20 columns
(languages), aligns the words across different Dogon
languages by their meanings. As the example in
Table 4 shows, there is an English concept, such as
‘sheep’ or ‘grass, herbs’, and in each cell the particu-
lar Dogon word(s) are given in phonetic transcrip-
tion. These data are collected by the DLP team in
the field from speakers of these languages through
common elicitation methods practiced by field lin-
guists. The transcriptions of these words follow
principles of the International Phonetic Alphabet.
Even without linguistic training and without
knowing exactly what sound each letter stands for,
the astute reader can glean some relations among
the different words in the different languages in
Table 4. For example the word ‘sheep’ in Jamsay
and Togo Kan are the same, and in Togo Kan and
Tomo Kan, it looks like their words for ‘sheep’ are
not so different. On the other hand, compare the
seemingly many different words in these different
languages for ‘grass’ or ‘herbs’. This is the type of
reasoning that is behind historical and comparative
linguistics.
4Language Comparison and the
Historical-comparative Method
Similar to species in biology, languages also evolve.
Words are lost, new words are gained, and the pro-
nunciation of all words changes slightly from day to
day. During its history, a language may split into
two or more descendant languages when the speak-
ers separate and their languages keep on changing
independently. To uncover how languages have
Table 3 Approximate number of Dogon speakers by
language
Tomo Kan 132,800 Tebul Ure 3000
Jamsay (?) 130,000 Toro Tegu 2900
Togo Kan (4) 127,000 Walo (2) 2000
Toro So (?) 50,000 Yanda Dom 2000
Donno So 45,300 Bangeri Me
a
1200
Tommo So 40,000 Ampari Pa 1000
Najamba-Kindige (?) 24,700 Bunoge 500
Mombo 24,000 Ana <500
Dogul Dom 12,200 Nanga Unknown
Ambaleenge 6000 Tiranige Unknown
Nyambeenge 5000 Pena Unknown
Language names that are followed by a number or a question
mark in parentheses indicate that there are one or more distinct
languages under the same language name.
a
Bangeri Me is a language isolate (along the lines of Basque) that
is spoken in Dogon country, but that does not seem to belong to
the Dogon language family. It may be an even more ancient
language that no longer has any surviving relative languages.
Table 4 Example of Dogon comparative wordlist
English Toro Tegu Nanga Jamsay Tommo So Togo Kan Tomo Kan
sheep bKlu
´pMrgKpe
´ju
´pe
´du
´pe
´ju
´pe
`jı
´
grass, herbs s[lYbe
`rı
´,sa
`wa
ˆ[yYke
`ru
´,be
`lu
´gu
´-gu
´ru
´gwı
`nı
´
S. Moran and J. Prokic
´
4of 16 Literary and Linguistic Computing, 2013
by guest on September 18, 2013http://llc.oxfordjournals.org/Downloaded from
evolved into their current shape is one of the major
tasks of historical linguistics.
Uncovering language history is an incredibly dif-
ficult task for many reasons. Foremost is the fact
that language, specifically the speech stream, is a
changing and variable acoustic and articulatory
signal. In fact, many acoustic and articulatory phon-
eticians believe that one cannot characterize speech
sounds with discrete and invariant symbolic repre-
sentations (such as letters to represent sounds).
Nevertheless, researchers need some type of tangible
object, even if a necessarily abstract one, to under-
take any type of comparison and analysis of lan-
guages. Consider the fact that no two persons’
pronunciation is identical, nor at a very fine-grained
level does anyone say the same sound in exactly the
same way twice in his or her lifetime. It is quite
obvious, however, that Spanish and Italian are
more closely related than Spanish and Mandarin
Chinese. One might ask then, at what level then
can we compare languages?
When a language is documented and described
for the first time (most languages do not have writ-
ing systems), someone has to identify the different
sounds in the language. That someone, perhaps a
linguist or a trained missionary, determines the spe-
cific sounds in that language which trigger contrasts
in the ear of the listeners. For example, in English
the sounds /l/ and /r/ contrast so that we recognize a
tangible difference between the words ‘lake’ and
‘rake’. However, these sounds do not contrast in
every language, for example, Korean speakers do
not readily distinguish between them.
What we know about uncovering language his-
tory is that it is very difficult. For example, there are
only a few languages whose history is directly re-
flected in written sources. For the majority of the
7,000 or so languages spoken today, we would not
know anything about their past if we did not have
methods to infer their history. In order to uncover
language history, the languages spoken today are
typically manually searched for traces of common
origin. Finding these traces, however, is an ex-
tremely complicated task.
Constructing historical scenarios involves com-
paring words from different languages and identify-
ing cognates. Cognates are words from different
languages that go back to a common ancestor
word (compare German Hund ‘dog’ and English
hound). Cognate words exhibit a specific kind of
similarity which does not necessarily show up in
the form of surface resemblances of the sounds
that the words are made of, but rather in structural
similarities of cognate words. This kind of similarity
is also not easy to detect: German Zahn ‘tooth’ and
English tooth, for example, are cognate, while Greek
mati and Malay mata are not.
The ultimate goal of historical linguistics is to
construct historical scenarios. How such a construc-
tion (or reconstruction) is usually carried out can be
easily illustrated by comparing the following four
words in German, English, Italian, and French,
which all mean ‘tooth’.
9
Each is presented in
Figure 1 in the same phonetic alphabet. In the
first step, the words are all compared with each
other, and common sound segments are analyzed,
such as /d/ to /d/ and /t/ to /ts/.
In a second step, the sound sequences are
‘aligned’, that is, they are arranged in such a way
that all corresponding segments occur in the same
column, as illustrated in Figure 2.
Based on these identified correspondences,
proto-stages of the languages are reconstructed.
An ancestor word for both the German and
English and the Italian and French word pairs is
selected by certain principles that basically follow
Occam’s razor. This is shown in Figure 3.
Fig. 1 Identify sound correspondences
Relatedness of the endangered Dogon languages
Literary and Linguistic Computing, 2013 5of 16
by guest on September 18, 2013http://llc.oxfordjournals.org/Downloaded from
These reconstructed proto-stages can then further
be compared. Finally, a historical scenario which
shows how the words evolved into their current
shape is reconstructed, as illustrated in Figure 4.
Historical reconstruction along these lines has
typically been done by manual inspection by scholars
with deep knowledge of different languages and their
words’ etymologies. Today, scientists are increasingly
using quantitative approaches to automate and speed
up this laborious and time-consuming task. Some of
these quantitative approaches have been co-opted
from biology, where inspiration has been taken
from comparing sequences of DNA.
5Quantitative Language
Comparison
In both biology and historical linguistics, there are
several parallels. First, sequences are defined as
ordered collections drawn from a fixed set of char-
acters; in molecular biology sequences constitute the
basic unit of replication, in languages sequences
constitute words. Second, whether DNA strings or
words in different languages, sequence comparison
is of paramount importance because it is the main
method for determining relatedness between two
objects. Lastly, both evolutionary biology and
historical linguistics use phylogenetic trees (family
trees) as their basic classification scheme. The
former describe the evolution of species, and the
latter, the evolution of languages.
There are several problems with these parallels
between evolutionary biology and historical linguis-
tics. Although the comparison problem is similar,
the amino acid alphabet for proteins has only 20
characters. Compare this with the languages of the
world, which have more than 2,000 different sounds
(Moran, 2012). Additionally, biological sequences
(such as proteins) are very long, while linguistic se-
quences (words) are relatively very short in com-
parison. And lastly and most importantly, in
biology the alphabet remains stable during evolu-
tion, while it changes constantly in language history.
Thus algorithms in biology were designed for long
Fig. 2 Align the sounds
Fig. 3 Reconstruction of language proto-forms
Fig. 4 Reconstructing a historical scenario
S. Moran and J. Prokic
´
6of 16 Literary and Linguistic Computing, 2013
by guest on September 18, 2013http://llc.oxfordjournals.org/Downloaded from
sequences drawn from small alphabets. On the other
hand, in quantitative historical linguistics we need
algorithms for short sequences drawn from large
alphabets.
10
To address these differences in biolo-
gical comparison and language comparison, several
different approaches are taken by linguists. In this
article we will use two approaches to attempt to
decode the relatedness of the Dogon languages:
the normalized edit distance (NED) method and
the LexStat method.
5.1 Normalized edit distance
Edit distance, also known as Levenshtein distance, is
a metric used to measure the distances between two
strings (Levenshtein, 1965). It represents the smal-
lest number of edit operations (insertions, deletions,
or substitutions) needed to transform one string
into the other. At the same time, it aligns the two
strings, as illustrated in Figure 5, which presents the
alignment of two words for ‘liver’ taken from the
Tebul Ure and Yorno So languages (two Dogon
languages from our data). These two words differ
in position 2 and position 4, where [K] corresponds
to [ı
´], and [d] has no corresponding sound (i.e. it
corresponds to insertion or deletion of a sound),
which means that the aggregate distance between
these two strings is 2. In order to discard the influ-
ence of the lengths of the strings being compared,
we normalize edit distance by dividing it by the
length of the longer string.
The Levenshtein method has been extensively
used in dialectometry to measure the distances be-
tween various dialects (Kessler, 1995;Nerbonne
et al., 1996;Heeringa, 2004). It has also been used
to analyze the relatedness between languages, such
as Indo-European (Serva and Petroni, 2008;
Blanchard et al., 2010), Austronesian (Petroni and
Serva, 2008), and a very large sample of 3,002 lan-
guages (Holman, 2010). While dialect comparison
based on the edit distance can give a pretty accurate
picture of the aggregate distances between the dia-
lectal varieties, some recent studies have shown the
limits of this method when applied to languages at
larger phylogenetic distances. For example, Prokic
´
and Moran (forthcoming) compare three methods
(the Levenshtein algorithm, n-gram approaches,
and a very simple zipping technique) used to
measure the distances between languages and show
that these algorithms are not suitable for revealing
deep genealogical relations on a set of sixty-nine
indigenous South American languages. And
Greenhill (2011) has shown that the accuracy of
the Levenshtein method in classification of the
Austronesian languages reaches only up to 65%.
Furthermore, he has observed that the accuracy of
Levenshtein classification decreases rapidly with
phylogenetic distance.
In this article, we rely on the NED method to
shed more light on the synchronic relatedness
among Dogon languages, without any reference to
deep genealogical relatedness. We compare words of
the same meaning by calculating their edit distance,
and identify those groups of words that show 70%
and higher similarity. We refer to the words that
have the same meaning and a very similar form as
‘homologies’. They could be real cognates or the
result of internal or external borrowing. However,
for the synchronic comparison of languages this is
not relevant. To estimate the relation between each
two language varieties in the data set, we calculate
the number of shared ‘homologies’ between each
two languages. The higher the number of shared
homologies, the more similar two language varieties
are and the higher mutual intelligibility between
them. Since the current Dogon linguistic situation
is not at all clear, we find this an important first step
in comparing languages at the synchronic level, that
is, for comparing the differences in languages as they
are currently spoken. In order to automatically
detect cognates and estimate genealogical related-
ness between Dogon languages, we apply the
LexStat approach described in the next section.
5.2 LexStat method
In recent years there have been several approaches
to automatic cognate detection (Covington, 1996;
Kondrak, 2002;Steiner et al., 2011;Wettig et al.,
Fig. 5 Illustration of two aligned strings
Relatedness of the endangered Dogon languages
Literary and Linguistic Computing, 2013 7of 16
by guest on September 18, 2013http://llc.oxfordjournals.org/Downloaded from
2012). In this research we rely on the LexStat
method proposed by List (forthcoming, 2012b).
11
In the LexStat method each word is represented as
a tuple of sound classes and prosodic strings. Sound
classes are used to guess initial correspondences of
sounds. The main idea of a sound class approach is
that sounds which often occur in correspondence
relations in genealogically related languages can be
clustered into classes (or types). It is assumed ‘that
phonetic correspondences inside a ‘‘type’’ are more
regular than those between different ‘‘types’’ ’
(Dolgopolsky 1986: 35). Figure 6 provides an illus-
tration of how a linguist might group sounds into
classes.
12
Given a list of sounds, illustrated by /p, b,
t, d, .../, sets of sounds can be grouped together
into classes based on their regular sound corres-
pondences between languages (List, forthcoming,
chap. 4).
This process helps to reduce the number of
sounds into a manageable size of twenty-eight
classes (List, forthcoming, chap. 4.2.1), which is
comparable to the number of characters used in
biological algorithms. At the same time, for each
word, a sonority profile is calculated. The sonority
profile is used to derive a prosodic string, which is a
vector of integer weights representing the relative
sonority of segments in a sonority hierarchy (List,
forthcoming, chap. 4.2.2). The prosodic strings
combined with a scoring function helps account
for the well-known fact that certain types of sound
changes are more likely to occur in specific prosodic
contexts. There are seven different prosodic envir-
onments distinguished in the LexStat method and
each sound is assigned one of the seven values. In
the next step, a permutation test is used to derive
language-specific similarity scores between the
sounds, and consequently between the words. This
method compares the observed and expected distri-
butions of the aligned sounds. While the observed
distributions are calculated from the aligned words
that show high phonetic and semantic similarity,
that is, they can be considered potential cognates,
the expected distributions are calculated from the
randomly aligned words regardless of their phonetic
or semantic similarity.
In the following section we use the NED and
LexStat methods to investigate the relatedness of
the Dogon languages. The Dogon data are taken
from the comparative lexical spreadsheet that is
available online (as discussed in Section 3). From
this spreadsheet we take a small subset of the data,
based on the Swadesh word list, which includes
words that are thought to be the most commonly
found across languages, for example, ‘sun’, ‘moon’,
‘man’, and ‘woman’ (Swadesh, 1971).
13
We then
parsed this list into a format that allowed us to
run the NED and LexSat methods.
6Analyses and Results
Our analysis is based on the 100 Swadesh list col-
lected at nineteen Dogon villages and provided by
the DLP. The distribution of the villages can be seen
in Figure 7.
We apply the NED and LexStat approaches in
order to align words and search for similar sets of
words, both from the synchronic and diachronic
perspectives. In both methods, once the similarities
between the strings (words) are calculated, a certain
threshold is required that determines which words
are homologies/potential cognates. We determine
this threshold for both approaches by empirically
examining the aligned strings. The relatedness be-
tween languages is then determined by calculating
the percentage of shared homologies/cognates.
What we discovered is that while the LexStat
Fig. 6 Illustration of collapsing sounds into sound classes
S. Moran and J. Prokic
´
8of 16 Literary and Linguistic Computing, 2013
by guest on September 18, 2013http://llc.oxfordjournals.org/Downloaded from
method is designed specifically to discover cognates,
the two approaches give almost identical results. In
Table 5 we show the homologies identified by the
NED method and the cognates identified by the
LexStat method by inserting space between each
group of identified homologies/cognates. The two
columns list the lexical realizations of the same con-
cept according to the NED method, on the left, and
the LexStat method on the right.
The similarity matrices that express the percent-
age of shared homologies/cognates between each
pair of languages are separately calculated using
the NED and LexStat approaches. Each is first ana-
lyzed by means of noisy clustering (Nerbonne et al.,
2008). In noisy or composite clustering, small
amounts of random noise are added to the matrices
during repeated clustering in order to obtain stable
clustering results.
14
The results of noisy clustering
are probability dendrograms that show grouping of
the data, as well as the confidence of the obtained
groups (Figures 8 and 9).
With very high confidence, both analyses show
separation of the western villages at the highest level
(the biggest split in the data), where Mombo,
Penange, Bunoge, Ampari, and Tiranige are
spoken. The difference between the two methods
is that in the LexStat method, unlike in the NED
method, the village where Dogul Dom is spoken is
not classified together with the other western vil-
lages. In order to see the geographical spread of
the identified groups, we annotate with different
shades the areas around each village in the data as
determined by GPS coordinates for these villages.
The areas around each village are determined by
using Voronoi triangulation. The two way classifi-
cation obtained by the two methods can be seen in
Figures 10 and 11, where we use two different
shades to show the division, dark gray for western
and light gray for the eastern villages.
In the east, two groups of language clusters are
identified by both methods (Figures 12 and 13).
ThedifferenceisthatintheNEDanalysis,thearea
Fig. 7 Geographical distribution of villages where Dogon languages are spoken
Relatedness of the endangered Dogon languages
Literary and Linguistic Computing, 2013 9of 16
by guest on September 18, 2013http://llc.oxfordjournals.org/Downloaded from
where Tomo Kan is spoken (9) is not classified in any
of the two eastern clusters and we leave it shaded the
same gray as the outlying area (Figure 12).
15
In the
LexStat method, the area where Dogul Dom is spoken
is not classified in any of the two eastern groups (also
shaded the same as the outlying area in Figure 13).
The maps presented in Figures 12 and 13 reveal
that the geographic spread of the identified language
families is almost identical regardless of the method
used. In the west, a homogeneous group of villages
is clearly identified on both maps. The other two
groups seem to be geographically mixed.
7Evaluation
We evaluated classifications produced by our two
approaches against a language family tree from the
MultiTree project, based on Heath (2012) which
represents genealogical classification of the Dogon
languages.
16
We discarded the languages not present
in our data set in order to make the comparison
easier (Figure 14).
17
The main division of the languages in Figure 14
separates the western and eastern groups. The western
group includes five languages identified by our two
methods, but also languages Tebul Ure, Najamba,
and Yanda Dom, classified as eastern languages by
both of our methods. The tree in Figure 14 is not
well resolved and the only two subgroups of languages
present are the Bunoge, Mombo, Ampari, and
Penange cluster in the west, and Nanga, Ben Tey,
and Bankan Tey in the east. These two smaller
groups of languages can also be identified in the prob-
ability clusters in Figures 8 and 9. Geographic
Fig. 8 Probabilistic dendrogram based on the NED approach
Table 5 Homologies identified by the NED
method (left) and cognates identified by the
LexStat method (right)
NED LexStat
1. ku
ˆl1.ku
ˆl
1. kwe
´:
2. kwe
´:
2. pı
`re
´
2. pı
`re
´3. be
`ra
ˇ:
3. bMrK
3. do
´le
`3. bMr[
3. bMrK
4. be
`ra
ˇ:3. bMrK
4. bMdM3. pı
`re
´
4. bMrK3. bMrK
4. bMr[
4. bMrK4. do
´le
`
4. bMrK
4. bMrK5. bMdM
4. pı
´dZı
´5. pı
`ndı
`
5. bı
`ndı
`
5. pı
`ndı
`5. be
`te
´
5. bı
`ndı
`5. be
`nde
´
6. bı
ˇn6.pı
´dZı
´
6. be
`te
´
6. be
`nde
´7. dZa
`˛ga
`
7. dZa
`˛ga
`8. bı
ˇn
S. Moran and J. Prokic
´
10 of 16 Literary and Linguistic Computing, 2013
by guest on September 18, 2013http://llc.oxfordjournals.org/Downloaded from
Fig. 9 Probabilistic dendrogram based on the LexStat approach
Fig. 11 Two groups of Dogon-speaking areas based on
LexStat approach Fig. 13 Three groups of Dogon-speaking areas based on
LexStat approach
Fig. 10 Two groups of Dogon-speaking areas based on
NED approach
Fig. 12 Three groups of Dogon-speaking areas based on
NED approach
Relatedness of the endangered Dogon languages
Literary and Linguistic Computing, 2013 11 of 16
by guest on September 18, 2013http://llc.oxfordjournals.org/Downloaded from
distribution of the main west-east division proposed
by Heath (2012) can be seen in Figure 15.
The classifications suggested by the automatic
methods offer a somewhat different view on the
relatedness of the Dogon languages when compared
to the genealogical classification. The difference is
not only in the languages classified in the two main
groups (western and eastern), but also in the finer
subgroupings which are not present in the tree in
Figure 14. The reason for some of the differences
may be due to language contact, and subsequently
language shift. As a result of physical movement of
the whole villages to the east, Tebul Ure and Yanda
Dom, genetically western languages, are nowadays
deeply embedded in the area where eastern Dogon
languages are spoken. Our analyses reveal that,
probably due to contact with neighbouring lan-
guages, these genetically western varieties are shift-
ing towards neighbouring languages at the lexical
level. Najamba seems to be closely related to both
Tebul Ure and Yanda Dom and lexically rather
divergent from the immediately adjoining Dogon
languages such as Jamsay and Tommo So. Our
two automatic methods gave inconsistent classifica-
tions of Dogul Dom.
Fig. 15 Two-way classification of languages based on
Heath (2012)
Fig. 14 Classification of languages based on Heath (2012)
S. Moran and J. Prokic
´
12 of 16 Literary and Linguistic Computing, 2013
by guest on September 18, 2013http://llc.oxfordjournals.org/Downloaded from
We also evaluated the classifications produced by
the two automatic methods against a language
family tree that was created by Prokhorov, an
expert on Dogon languages (Prokhorov et al.,
2012). These trees are based on the number of
shared ‘lookalikes’, that is, the set of true cognates
and words that look alike due to either the bor-
rowing of words between languages or pure
chance. The lookalikes were manually selected by
Prokhorov and his genealogical tree of Dogon is
given in Figure 16.
Both automatic classifications that we used show
an almost identical grouping of languages as the
analysis done by Prokhorov. Two main groups of
languages are clearly identified: western and eastern
that is further divided into eastern and northern.
Both computational methods gave almost identical
results when compared to the opinion of an expert,
with the difference that our method is fully auto-
mated and it can easily handle large amounts of
data.
The results of both automatic methods suggest
that (1) the major classification of Dogon languages
follows the east-west division and that (2) Yanda
Dom, Tebul Ure, and Najamba are lexically closely
related to the neighboring eastern varieties probably
due to language shift. Groupings obtained by the
NED and LexStat methods nicely correspond to
the classification suggested by Prokhorov and sup-
port it even at the level of dialect variation.
Considering that research on the Dogon languages
is still in its early stages, both when it comes to the
traditional scholarship and quantitative methods,
we find our work is an important first step in resol-
ving the relatedness of the Dogon languages from
both synchronic and diachronic perspectives.
8Conclusion
In this article we have given a brief overview of
language endangerment and the Dogon languages.
Our goal has been to use quantitative methods
repurposed from biology and applied to linguistics
to automate methods used in traditional historical
linguistics to compare and classify the relatedness of
languages. Our goal fits well within digital huma-
nities scholarship; we use digital methods to study
and uncover knowledge about languages, and per-
haps ultimately, their speakers’ cultures. Our goal
has been to try and untangle the mystery of how
the different Dogon languages are related and we
have identified some of the limitations of the cur-
rent quantitative methods for language comparison.
These quantitative methods can only be seen as
the first step in producing an automatic analysis
Fig. 16 Prokhorov’s classification of Dogon languages
Relatedness of the endangered Dogon languages
Literary and Linguistic Computing, 2013 13 of 16
by guest on September 18, 2013http://llc.oxfordjournals.org/Downloaded from
towards solving the genealogical relatedness of lan-
guages in general. Closer inspection of the potential
cognates identified by the LexStat method reveals
that the method is successful in cases where the
two words that are compared show similar forms.
However, this method alone cannot deal with the
potential problem of borrowed words and add-
itional analyses are needed. While for the synchronic
comparison of languages this is not an important
issue, for in-depth reconstruction of the historical
relatedness of languages, we need to refine these
digital methods even further in order to detect bor-
rowings. However, the results of the two tested
automatic methods are encouraging and we hope
that they will be an important contribution in sol-
ving the problem of the diversity of Dogon
languages.
Software
NED and LexStat analyses were performed using the
LingPy software.
18
We use the GabMap software to
produce Figures 7-13 and 15 (Nerbonne et al.,
2011).
19
Funding
This work was supported by the ERC starting grant
240816: ‘Quantitative modeling of historical-com-
parative linguistics’.
References
Austin, P. K. and Sallabank, J. (eds), (2011). The
Cambridge Handbook of Endangered Languages.
Cambridge University Press.
Bendor-Samuel, J., Olsen, E. J., and White, A. R. (1989).
Dogon. In Bendor-Samuel, J. (ed.), The Niger-Congo
Languages: A Classification and Description of Africa’s
Largest Language Family. Lanham, Maryland:
University Press of America, pp. 169–77.
Blanchard, P., Petroni, F., Serva, M., and
Volchenkov, D. (2010). Geometric Representations of
Language Taxonomies. Computer Speech and Language,
25(3): 679–99.
Blench, R. (2005). A Survey of Dogon Languages in Mali:
Overview. OGMIOS,26: 14–15.
Campbell, L. (1998). Historical Linguistics: An
Introduction. Edinburgh: Edinburgh University Press.
Campbell, L. et al.(2013). New Knowledge: Findings
from the Catalogue of Endangered Languages. Paper
presented at the 3rd International Conference on
Language Documentation and Conservation (ICLDC).
University of Hawai’i at Ma
¯noa. February to March
2013.
Covington, M. A. (1996). An Algorithm to Align Words
for Historical Comparison. Computational Linguistics,
22: 481–96.
Greenhill, S. (2011). Levenshtein Distances Fail to
Identify Language Relationships Accurately. Computa-
tional Linguistics,37(4): 689–98.
Hale, K. et al.(1992). Endangered Languages: On Endan-
gered Languages and the Safeguarding of Diversity.
Language,68(1): 1–42.
Harrison, K. D. (2007). When Languages Die: The
Extinction of the World’s Languages and the Erosion of
Human Knowledge. New York and London: Oxford
University Press.
Heath, J. (2008). A Grammar of Jamsay. Berlin: Mouton
de Gruyter.
Heath, J. (2012). Dogon and Bangime linguistics. http://
dogonlanguages.org/ (accessed 1 June 2013).
Heeringa, W. (2004). Measuring Dialect Pronunciation
Differences Using Levenshtein Distance. Ph.D. thesis,
University of Groningen.
Hochstetler, J. L., Durieux, J., and Durieux-Boon, E.
(2004). Sociolinguistic Survey of the Dogon Language
Area. SIL Electronic Survey Reports 2004, SIL
International.
Holman, W. E. (2010). Do Languages Originate and
Become Extinct at Constant Rates? Diachronica,
27(2): 214–25.
Kessler, B. (1995). Computational Dialectology in Irish
Gaelic. In Proceedings of the European chapter of the
Association for Computational Linguistics (EACL),
pp. 60–6.
Kondrak, G. (2002). Algorithms for Language
Reconstruction. Ph.D. thesis, University of Toronto.
Ladefoged, P. (1992). Another View of Endangered
Languages. Language,68(4): 809–11.
Levenshtein, V. (1965). Binary Codes Capable of
Correcting Deletions, Insertions and Reversals.
Doklady Akademii Nauk SSSR,163: 845–8.
S. Moran and J. Prokic
´
14 of 16 Literary and Linguistic Computing, 2013
by guest on September 18, 2013http://llc.oxfordjournals.org/Downloaded from
Lewis, M. P. (2009). Ethnologue: Languages of the World,
Sixteenth Edition. Dallas, TX: SIL International.
Lewis, M. P., Simons, G. F., and Fennig, C. D. (eds),
(2013). Ethnologue: Languages of the World,
Seventeenth Edition. Dallas, Texas: SIL International.
List, J.-M. (2011). Multiple Sequence Alignments in
Historical Linguistics. Paper presented at the
ConSOLE XIX: The 19th Conference of Student
Organization of Linguistics in Europe. Groningen,
Netherlands, January 2011.
List, J.-M. (2012a). Multiple Sequence Alignment in
Historical Linguistics: A Sound Class Based Approach.
In Boone, E., Linke, K., and Schulpen, M. (eds),
Proceedings of ConSOLE XIX 2011. Leiden, The
Netherlands: University of Leiden, pp. 241–60.
List, J.-M. (2012b). LexStat: Automatic Detection of
Cognates in Multilingual Wordlists. In Butt, M.,
Prokic
´, J., Mayer, T., and Cysouw, M. (eds),
Proceedings of the European Chapter of the Association
for Computational Linguistics (EACL) 2012: Joint
Workshop of LINGVIS & UNCLH, pp. 117–25.
List, J.-M. (Forthcoming). Sequence Comparison in His-
torical Linguistics.Du
¨sseldorf: Du
¨sseldorf University
Press.
List, J.-M. and Moran, S. (2013). An Open Source Toolkit
for Quantitative Historical Linguistics. Proceedings of
the Association for Computational Linguistics (ACL)
2013. Sofia, Bulgaria.
Moran, S. (2012). Phonetics Information Base and Lexicon.
Ph.D. thesis, University of Washington.
Needleman, S. B. and Wunsch, C. D. (1970). A General
Method Applicable to the Search for Similarities in the
Amino Acid Sequence of Two Proteins. Journal of
Molecular Biology,48(3): 443–53.
Nerbonne, J., Heeringa, W., van den Hout, E., van de
Kooi, P., Otten, S., and van de Vis, W. (1996).
Phonetic Distances between Dutch Dialects. In
Durieux, G., Daelemans, W., and Gillis, S. (eds), CLIN
VI: Proceedings of the Sixth CLIN Meeting. Antwerp:
Centre for Dutch Language and Speech, pp. 185–202.
Nerbonne, J., Kleiweg, P., Heeringa, W., and Manni, F.
(2008). Projecting Dialect Differences to Geography:
Bootstrap Clustering vs. Noisy Clustering. In
Preisach, C., Schmidt-Thieme, L., Burkhardt, H., and
Decker, R. (eds), Data Analysis, Machine Learning, and
Applications. Berlin, Heidelberg: Springer, pp. 647–54.
Nerbonne, J., Colen, R., Gooskens, C., Kleiweg, P., and
Leinonen, T. (2011). Gabmap: A Web Application for
Dialectology. Dialectologia: revista electro
`nica,2: 65–89.
Petroni, F. and Serva, M. (2008). Language Distance and
Tree Reconstruction. Journal of Statistical Mechanics:
Theory and Experiment,2008: P08012.
Prokhorov, K., Heath, J. and Moran, S. (2012). Dogon
Classification. Paper Presented at the Niger-Congo
Congress, Paris, September 2012.
Prokic
´, J. and Moran, S. (Forthcoming). Black Box
Approaches to Genealogical Classification and Their
Shortcomings. In Borin, L. and Saxena, A. (eds),
Approaches to Measuring Linguistic Differences. Berlin,
Boston: De Gruyter Mouton, pp. 429–45.
Serva, M. and Petroni, F. (2008). Indo-European
Languages Tree by Levenshtein Distance. Europhysics
Letters,81(6): 68005.
Steiner, L., Stadler, P. F., and Cysouw, M. (2011). A
Pipeline for Computational Historical Linguistics.
Language Dynamics and Change,1(1): 89–127.
Swadesh, M., Sherzer, J., and Hymes, D. (1971). The
Origin and Diversification of Language. Chicago:
Aldine De Gruyter.
Wettig, H., Reshetnikov, K., and Yangarber, R. (2012).
Using Context and Phonetic Features in Models of
Etymological Sound Change. In Butt, M., Prokic
´, J.,
Mayer, T., and Cysouw, M. (eds), Proceedings of the
European Chapter of the Association for Computational
Linguistics (EACL) 2012: Joint Workshop of LINGVIS &
75 UNCLH, pp. 108–16.
Notes
1 http://www.endangeredlanguages.com/.
2 See the UNESCO Atlas of the World’s Languages in
Danger at: http://www.unesco.org/culture/languages-
atlas/.
3Ladefoged (1992) challenges the assumption that lan-
guages and cultures should be preserved, arguing that it
is ‘paternalistic of linguists to assume they know what is
best’ for a given speech community. Indeed, the adop-
tion of major languages by certain small-language
speech communities affords speakers better socio-eco-
nomic opportunities. This issue is a fiercely debated in
linguistics.
4 Dogon country has been recognized as a UNESCO
World Heritage Site since 1989.
5 http://www.ethnologue.com/.
6 http://dogonlanguages.org/.
7 See Hochstetler et al. (2004) and Hantgan’s Dogon
bibliography for references at:http://dogonlanguages.
org/bibliography.cfm.
Relatedness of the endangered Dogon languages
Literary and Linguistic Computing, 2013 15 of 16
by guest on September 18, 2013http://llc.oxfordjournals.org/Downloaded from
8 http://dogonlanguages.org.
9Figures 1–4 are taken, with permission, from List,
forthcoming, and List 2012a.
10 Note also that while biologists can model their se-
quences in ASCII characters, linguists cannot easily
do without the Unicode encoding due to the large
number of different characters needed for phonetic
transcription.
11 Here we give a high-level overview of the SCA pro-
cedure. A step-by-step tutorial with explanation is
given with the Dogon data in the LingPy tutorial:
http://lingpy.org/tutorial/. See also List & Moran,
forthcoming.
12 Figure 6 is adapted, with permission, from List 2011.
13 An anonymous reviewer points out that Swadesh
wordlists, which come in several versions that differ
in length and semantic content, have been challenged
as appropriate vehicles for lexicostatistics and glotto-
chronology (cf. Campbell 1998). In general we agree
with this criticism. For our study, we could have in
principle used any set of words from the Dogon com-
parative wordlist because we know in general that
Dogon languages are very closely related, both genea-
logically and geographically, and we have not at-
tempted to do any dating of their genealogical
relatedness. The Swadesh 100 list, as adapted slightly
by the Dogon language experts to Dogon culture and
geography (e.g. original ‘earth’ to ‘earth (as brick mix
or to repair walls)’ or ‘cold (of weather)’ to ‘cold
(e.g. water)’), is nicely present in all villages in our
survey.
14 In biology, in order to obtain stable clustering results a
bootstrap procedure is often employed by randomly
resampling the observed data (Felsenstein, 2004).
Tested on dialect data, bootstrapping and noisy clus-
tering produce distance matrices that correlate nearly
perfectly (r¼0.997). Unlike bootstrapping, noisy clus-
tering can be applied on a single distance matrix.
15 In Figures 12,13, and 15 we replace language areas
with numbers in order to make the grouping of east-
ern areas visible: Toro_Tegu (1), Ben_Tey (2),
Bankan_Tey (3), Nanga (4), Jamsay (5), Perge_Tegu
(6), Togo_Kan (7), Yorno_So (8), Tomo_Kan (9),
Tommo_So (10), Dogul_Dom (11), Tebul_Ure (12),
Yanda_Dom (13), Najamba (14), Mombo (15),
Ampari (16), Tiranige (17), Bunoge (18) and
Penange (19).
16 http://multitree.org/trees/
Dogon%3A%20Heath%202012.
17 Classification of the Dogon language in the
Ethnologue is flat and cannot be used to compare
the groups obtained by out two methods.
18 http://lingpy.org/.
19 http://www.gabmap.nl/.
S. Moran and J. Prokic
´
16 of 16 Literary and Linguistic Computing, 2013
by guest on September 18, 2013http://llc.oxfordjournals.org/Downloaded from
