Content uploaded by Peter Dekker
Author content
All content in this area was uploaded by Peter Dekker on Apr 01, 2017
Content may be subject to copyright.
Determining Dutch dialect phylogeny using
bayesian inference
Peter Dekker
Bachelor’s thesis (7.5 ECTS)
BSc Artificial Intelligence, Universiteit Utrecht
Supervisors: Alexis Dimitriadis and Martin Everaert
July 31, 2014
Abstract
In this thesis, bayesian inference is used to determine the phylogeny of Dutch
dialects. Bayesian inference is a computational method that can be used to
calculate which phylogenetic tree has the highest probability, given the data.
Dialect data from the Reeks Nederlandse Dialectatlassen, a corpus of words in
several Dutch dialects, serves as input for the bayesian algorithm. The data was
aligned and converted to phonological features. The trees generated by bayesian
inference were evaluated by comparing them with an existing dialect map by
Daan and Blok.
Contents
1 Introduction 2
1.1 Bayesianinference .......................... 2
1.2 Earlierresearch............................ 3
1.3 Applying bayesian inference to the Dutch dialects . . . . . . . . . 3
2 Method 4
2.1 Data.................................. 4
2.2 Alignment............................... 4
2.3 Phonological mapping . . . . . . . . . . . . . . . . . . . . . . . . 6
2.4 Bayesianinference .......................... 6
2.5 Evaluation............................... 7
3 Results 7
3.1 Netherlandic dialects . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.1.1 Equal rate variation model . . . . . . . . . . . . . . . . . 8
3.1.2 Gamma-distributed rate variation model . . . . . . . . . . 14
3.2 Comparison: Belgian dialects . . . . . . . . . . . . . . . . . . . . 15
4 Discussion 18
5 Conclusion 20
6 Literature 20
7 Appendix 22
7.1 Tree of Netherlandic dialects – equal rate variation . . . . . . . . 22
7.2 Tree of Netherlandic dialects – gamma rate variation . . . . . . . 25
7.3 Tree of Belgian dialects – equal rate variation . . . . . . . . . . . 28
7.4 Tree of Belgian dialects – gamma rate variation . . . . . . . . . . 30
7.5 Dialect map by Daan and Blok (1969) . . . . . . . . . . . . . . . 32
1
1 Introduction
How are languages related? Languages are genetically related if they share a
single ancestor from which they derive (Campbell, 1998). To prove a common
ancestor, an array of methods can be applied. The phylogeny, or evolutionary
relationship, of languages can be viewed as a tree, where a branching shows that
two languages Band Cderived from an ancestor A.
1.1 Bayesian inference
In recent years, computational methods have seen their advent in historical
linguistics. One of them is bayesian phylogenetic inference. This method is
inspired by Bayes’ law: the probability of a hypothesis Hxfor a certain phe-
nomenon Ecan be given using the probability of the phenomenon given the
hypothesis.
P(Hx|E) = P(Hx)·P(E|H x)
Pn
k=1 P(Hk)·P(E|Hk)
In our case, a phylogenetic tree is the hypothesis. A tree is a tuple ω= (τ , υ, φ)
with (Larget and Simon, 1999):
•a tree topology τ
•a vector of branch lengths υassociated with topology τ. Each branch
length in υrepresents the distance between two adjacent nodes in τ.
•a substitution model φ, which determines the probability that a certain
element of a language changes into a certain other element.
The tree topology τis defined as:
•a set of vertices V
•a set of edges E∈V×V
•the graph that E describes on V is strongly connected
•there are no cycles.
Branch lengths show how much distance there is between languages: a longer
branch means that more substitutions have been made between the input strings.
The substitution model determines how likely the change from a character in
the input string to a certain other character is.
The question is: what is the most probable tree to describe the linguistic data
X? Our application of Bayes’ law becomes (Ronquist and Huelsenbeck, 2003):
f(ω|X) = f(ω)·f(X|ω)
f(X)
f(ω) is the prior distribution, containing the a priori probabilities of the different
trees. f(X|ω) is the likelihood function, which returns the probability that the
2
data has been generated by a tree. f(X) is the total probability of the data.
f(ω|X) is the posterior distribution, containing the probabilities of all the trees.
Assumptions have to be made about the prior probabilities of trees, because they
are generally unknown. Calculating the posterior probability means summing
over all of the trees, whose posterior probabilities have not been calculated yet,
and integrating over all the possible combinations of τ,υand φ. The posterior
probability distribution cannot be calculated directly (Huelsenbeck et al., 2001).
To address this issue, bayesian algorithms use a technique called Markov Chain
Monte Carlo (MCMC) sampling. MCMC is an approximation of Bayes’ law,
with a number of simplifications. The prior probability distribution is deter-
mined by a Dirichlet distribution, inferring the prior probability of a tree from
the data itself (Ronquist et al., 2011). The algorithm starts off with a ran-
dom tree. Every generation, a small random change of the parameter values is
proposed. It is accepted or rejected with a probability given by the Metropolis-
Hastings algorithm (Larget and Simon, 1999). Every sgenerations (where s
is the sample frequency), the accepted tree of the current generation is saved
to the posterior sample. After a number of generations, the posterior sam-
ple should approximate the real posterior probability distribution (Huelsenbeck
et al., 2001). From this distribution, a best tree can be drawn, according to the
desired criteria.
In principle, the input for the bayesian method could be any kind of linguistic
data. When applied to languages, it is common to use Swadesh lists, a list of
words which are unlikely to be borrowed. A linguist manually classifies each
word in a dialect to be in a certain cognate class. Crucial in the cognate classi-
fication is the Neogrammarian hypothesis, which states that sound changes are
regular. Regularity means that if a sound in a word changed into another sound,
it will do so in every other word. Two words can only be cognate, if the common
ancestor can be reached from both cognate candidates by a number of known
sound changes (Campbell, 1998). Generally, in phylogenetic linguistics, a string
of the cognate classifications of every word in a language serves as the input for
the bayesian algorithm. For dialects, words in the different dialects are likely to
be cognates (Dunn, 2008). If cognacy is assumed, manual classification can be
omitted and an objective measure of distance between the phonological forms
in different dialects can be used.
1.2 Earlier research
In earlier research, the bayesian method has been applied to Bulgarian dialects
(Proki´c et al., 2011). The focus in the research was on vowel change. The
consonants were dropped and the vowels were classified into a limited number
of classes to reduce computational cost. A broader approach will be used in this
thesis, because consonants may also amount to important distinctions between
dialect groups.
1.3 Applying bayesian inference to the Dutch dialects
I would like to evaluate the bayesian method when used on the Dutch dialects.
My research question is: how well is bayesian inference suited to determine the
3
phylogeny of Dutch dialects?
2 Method
The input for the bayesian inference is a string for every dialect, which uniquely
describes that dialect. A corpus of words and their translations into different
dialects were used as the basis for the input string. Each word was aligned
with its counterparts in different dialects. All the aligned words for a dialect
were concatenated. The concatenated strings were converted to phonological
features. The resulting feature strings were used as input for the bayesian
inference.
2.1 Data
The dialect data was taken from the Reeks Nederlandse Dialectatlassen (RND).
This is a corpus of transcribed speech in Dutch and Frisian dialects in the
whole Dutch language area: The Netherlands, a neighbouring area in Germany,
Flanders and the north of France. The corpus was recorded between 1925 and
1982. A selection of 166 words and 363 dialects has been made from this corpus
and digititalized (Heeringa, 2001). An interesting addition to the digital version
is the Plautdietsch dialect from Protasovo, Siberia. This dialect descended
from 16th century Mennonites who migrated via Eastern Europe to Siberia. It
maintained its Dutch character in Slavic surroundings (Nieuweboer, 1998).
The data was written in X-SAMPA, an ASCII version of the International Pho-
netic Alphabet (IPA). The data was converted to IPA, represented in Unicode,
using the cxs2ipa script (Theiling, 2008)1. The dialect data was split into two
subsets: one set of 269 Netherlandic (and neighbouring German) dialects and
one set of 94 Belgian (and neighbouring French) dialects. It is interesting to
use the parameters from the Netherlandic data set on the Belgian data, to see
whether the setting of the parameters is generally applicable to different data
sets.
2.2 Alignment
Figure 1: Alignment of translations of the lemma are. The sound classes (colors
of the phones) enable comparable sounds to be matched.
In order to compare which sounds differ in the translation of a lemma in different
dialects, the words need to be aligned. Comparable sounds are put in the same
1The script was modified in order to convert the æ properly as well.
4
column (Figure 1). Making an alignment assumes that the words in different
dialects are cognates. For most, but not all, words in the RND, this is the
case. For example, the lemma chickens has entries which look like Dutch kippen
and entries that look like German H¨uhner. Aligning these with each other
is less informative (Figure 2). For some dialects, there was more than one
Figure 2: Alignment of translations of the lemma chickens. The entries are
not cognate, there are entries which look like kippen and entries that look like
H¨uhner. Still, they are aligned based on their phonological characteristics.
translation for a certain lemma. In these cases, the first one was chosen as the
only translation. There are also lemmas where the alignment may have been
distracted by morphological rather than phonological differences. For example,
the lemma sore throat has items which look like keelpijn (throat-sore) and other
entries which look like pijnindekeel (sore-in-the-throat). The sounds are roughly
the same, only the order of stems is different. This is not fully reflected in the
alignment (Figure 3).
Figure 3: Alignment of translations of the lemma sore throat. The phonological
similarity is not fully reflected in the alignment, because of the morphological
difference in order. The part pin in both words is however aligned.
The alignment is done using LingPy (List and Moran, 2013). This is a program
for multiple sequence alignment, which means that all words are aligned with
each other at the same time. LingPy matches phones, by classifying them into
a number of sound classes (List, 2012). Phones in the same sound class have
the highest probability of matching.
Before the alignment, the data was tokenized: phones were grouped with dia-
critic signs to form one token. The list of possible tokens was based on Hoppen-
brouwers and Hoppenbrouwers (1988). It supplies a list of IPA tokens and maps
those to phonological features. The list is based on the RND data. Still, some
combinations of vowels/consonants and diacritics that were used in my RND
data, were missing in the list. These tokens were omitted, because this means
there is no phonological mapping for these tokens as well. The result is that the
omitted diacritic signs are shown as a ?in the alignment, which means it can
be aligned with a random phone. It seems this has not decreased the quality of
the alignment heavily.
5
Standard Dutch k Ip - @mEi n b l u m @-
Standard German h y - n @mAi n b l u: m@n
Midsland h E-n: - m i- - b l u m @n
Figure 4: Concatenation of the lemmas chickens,my and flowers for three
dialects. The real concatenated strings are far longer, they contain 166 lemmas.
Figure 5: The feature strings which serve as input for the bayesian inference.
This is the result of converting the concatenations from figure 4 to phonological
features.
After the alignment, all aligned words for a dialect were concatenated with each
other, resulting in a long string of all words for that dialect (Figure 4).
2.3 Phonological mapping
A possiblity would be to directly use the concatenated string of all aligned words
as the input string for a dialect. The positions in the alignment, the phones,
would then be the features, on the basis of which a dialect can be compared with
other dialects. However, the symbol alphabet of all phones used in the RND
is too big to be computationally feasible. Furthermore, it would be nice if the
algorithm also takes into account that phones that are phonologically close to
each other can change more easily than phones that are further from each other.
For these two reasons, the aligned phone strings were converted to an array of
phonological features. Hoppenbrouwers and Hoppenbrouwers (1988) provide
a mapping from each character to 21 binary phonological features, which was
used. The result is a long string of 0’s and 1’s for every dialect (Figure 5).
2.4 Bayesian inference
The program used to execute the bayesian inference was MrBayes (Ronquist and
Huelsenbeck, 2003). As described in the introduction, an MCMC analysis starts
from a randomly chosen tree and proposes small random changes to this tree.
MrBayes runs two different MCMC analyses at the same time, starting from
6
two different randomly chosen trees. By calculating the convergence between
the two analyses, it is possible to get an indication whether a stable posterior
probability distribution has been reached. The algorithm was run for 1,000,000
generations. The sample frequency was set to 20, which means that every 20
generations, the most probable tree is saved.
The likelihood is determined by two parameters: the substitution model and the
rate variation model. Together they provide the probability of the data, given
a certain tree. The substitution model determines the chance that a certain
character changes into another character. We have only two characters (0 and
1), so the only state changes are 0 →1 and 1 →0. A substitution model with
equal probability for every state change was used. The rate variation model
determines the chance that a certain feature changes state. For example, a
rate variation model could state that letters at the end of a word have a higher
chance of changing than letters in the middle (Proki´c et al., 2011).
Two rate variation models were tried: an equal rate variation model and a
gamma-distributed rate variation model. In an equal rate variation model, every
feature has the same chance of changing. In a gamma-distributed rate variation
model, the bayesian algorithm infers from the data which features change more
often than others. It categorizes the features in rate classes of higher or lower
probability of change, according to a gamma distribution.
2.5 Evaluation
A dialect map by Daan and Blok (1969) is used as the gold standard to evaluate
the results of the bayesian analysis. The map is based on the perception of
speakers. In a questionnaire, people from villages in the Dutch language area
were asked which dialects from other villages were (almost) the same. Arrows
could be drawn between villages with roughly the same dialect. Daan and Blok’s
map is based on this arrow method, combined with some linguistic knowledge,
in cases where the arrow method did not match the known insights (Daan and
Blok, 1969).
Finally, the bayesian algorithm with the same settings was applied to the Belgian
dialects. The results were also compared with Daan and Blok’s map.
3 Results
The output of the bayesian inference is a set of trees, each with their own
probability. The consensus tree is a tree that tries to reconcile all trees. If the
branching is contradictory between trees, the consensus tree places the branch at
a lower level (Dunn, 2008). The trees were shown graphically using the FigTree
(Rambaut, 2013) program.
7
3.1 Netherlandic dialects
3.1.1 Equal rate variation model
The consensus tree that was outputted correctly shows groups of dialects that
are connected locally, but does not generally show higher-order grouping be-
tween the local groups. This probably happens because the consensus tree could
not decide between two specific branchings and places dialects at a lower level.
The MCMC analysis has also not converged optimally, even after 1,000,000
generations.
Hardly any false groupings are made. Dialects that are grouped in the tree, are
generally also in one group on Daan’s map. Sometimes dialects are linked with
a dialect that is just across the border of a different group on the map, but still
geographically close. This is visible in the grouping of the dialects of Zeeuws-
Vlaanderen with some neighbouring dialects from Noord-Brabant (Figure 6).
No strange groupings between different parts of the country are made. The
price for this accuracy is that a lot of dialects remain ungrouped or are only
connected with their direct neighbours (Figure 7).
Figure 6: Equal model. The dialects Clinge, Lamswaarde and Groenendijk
from Zeeuws-Vlaanderen have been grouped together. They all belong to the
Zeeuws group on the map. The geographically close Zundert, Roosendaal and
Ossendrecht dialects have been grouped together in the tree, although they
belong to the different Noord-Brabant group on the map.
Some distinguishing groups can be seen, which correspond with groups on Daan
and Blok’s map. The dialects of Groningen (Figure 8) and southern Dutch
Limburg (Figure 9) form groups which correspond with Daan and Blok’s map.
The dialects of northern Noord-Holland and the islands of Texel and Vlieland
form one group, as the map would predict (Figure 10). There is a branch of the
tree that splits into Frisian dialects and Frisian city dialects (Figure 11). It is
good to see this clear division but still close connection between Frisian dialects
and Frisian city dialects. The Frisian city dialects are dialects which originate
from Frisian, but have been influenced by the dialects from Holland in the 16th
century (Jansen, 2002).
It is clear that Daan’s Utrecht-Alblasserwaard group is not well-visible in the
tree. Many dialects are clustered with other groups. Utrecht and Amersfoort
are unresolved (Figure 7).
Dialects from eastern Noord-Brabant have been connected, but are not con-
nected with dialects from the west of Noord-Brabant, which from one group on
Daan’s map. The dialects are however closely connected with two dialects from
Zuid-Gelderland, a related, but different group on the map (Figure 12).
8
Figure 7: Equal model. A lot of dialects have not been grouped: dialects
from the Utrecht-Alblasserwaard group like Utrecht and Amersfoort are on the
same level as eastern dialects like Beilen and Emmen. Other dialects have been
clustered into small groups: for example Goirle, Oirschot and Loon op Zand.
9
Figure 8: Equal model. The dialects of Groningen have been grouped according
to Daan’s map.
Figure 9: Equal model. The dialects of southern Dutch Limburg form a well-
divided group that is coherent with the map.
10
Figure 10: Equal model. The dialects of northern Noord-Holland are grouped
with the dialects from the islands of Texel and Vlieland, as Daan predicts. The
group is on the same level with a totally different, but also coherent group, that
of Zeeland. The long branch length of Protasovo is remarkable and signifies a
large distance compared to the other dialects.
11
Figure 11: Equal model. The dialects of Friesland. The Frisian dialects (red)
and Frisian city dialects (green) are related, but it is clear that there is a division.
Figure 12: Equal model. Dialects from the eastern side of the Noord-Brabant
group (red) have been grouped with dialects from the river region (green). Al-
though these groups are related, it is remarkable that the Noord-Brabant di-
alects match with dialects from a different group, whereas they do not match
with dialects from the western part of the same Noord-Brabant group.
12
The tree lacks some higher-order grouping. The dialects of the southern Nether-
lands are shown as a family in Daan’s map using red shades. The Low-Saxon
dialects of the eastern and northern Netherlands are shown as a family using
green shades. These higher-order groupings are however not visible in the tree
(Figure 13).
Figure 13: Equal model. The red group is a mix of dialects from the Utrecht-
Alblasserwaard group (Oudewater, Soest, Driebergen, Polsbroek) and Zuid-
Holland (Berkel, Wateringen, Nieuwveen, Langeraar, Warmond, Zoetermeer).
Maybe the border between these groups is not really clear-cut, as Daan and
Blok (1969) state. A second observation is that there is no sufficient higher-
order grouping. The red group of western-central dialects is at the same level
as the two blue groups of northeastern (Low-Saxon) dialects. These two blue
groups would be expected to be on a different level, together with other Low-
Saxon dialects.
The Protasovo (Plautdietsch) dialect has a very long branch, which shows that
it differs a lot from the other dialects (Figure 10). This seems reasonable, given
that it is a form of Dutch that has not been in contact with other Dutch dialects
for centuries.
Concludingly, the dialects that have been grouped together form groups that
are coherent with Daan and Blok. The groups have however not been grouped
in higher-order groups that show relations between dialect regions. This makes
13
the explanatory power of the tree smaller.
3.1.2 Gamma-distributed rate variation model
The consensus tree of the gamma-distributed rate variation model shows the
same pattern as the consensus tree of the equal rate variation model. There are
some differences in the groupings, sometimes these are improvements, sometimes
these are degradations. It is not really clear whether these small differences are
caused by different rate variation models. Differenes across different executions
of the same rate variation model also occurred.
The groups of southern Dutch Limburg and Groningen are also salient in this
consensus tree. Some groupings from Daan’s map are better under the gamma
model. The dialects of Twente have been grouped together under this model,
whereas they were spread across different groups in the equal model (Figure
14). The dialects groups of Noord-Holland and Zuid-Holland are related, this
is shown to a greater extent in this tree (Figure 15).
Figure 14: Gamma model. The dialects of Twente (and the directly neighbour-
ing places in Germany) form one group under the gamma model, whereas they
were spread across several groups in the equal model.
The distinction between Frisian dialects and Frisian city dialects is still shown,
but this time the Frisian dialects are shown as a subgroup of the Frisian city
dialects (Figure 16). This is not correct from a historical point of view, since
the Frisian city dialects split off the Frisian dialects. However, from a distance
point of view, it is less remarkable. The Frisian city dialects are closer to the
14
Figure 15: Gamma model. Dialects of Noord-Holland and Zuid-Holland have
been combined as one group under the gamma model.
other Dutch dialects at the root of the tree, because they have been influenced
by the dialects from Holland.
An interesting result is that the dialect of Katwijk aan Zee, a coastal place in
Zuid-Holland, is grouped with the dialects of Zeeland, a different group further
to the south (Figure 17). Apparently there are some shared characteristics
between these coastal areas.
There are also dialects that were grouped in the consensus tree of the equal
model, but are unresolved under the gamma model. Examples are the places
Oldemarkt and Steenwijk.
It is hard to say whether the gamma or the equal model is better. The gamma
model shows a few interesting groups that the equal model does not show, but
it also leaves dialects ungrouped which the equal model grouped. Furthermore,
some differences can occur across different executions of the same model and
are not caused by the model choice.
3.2 Comparison: Belgian dialects
The Belgian data was kept apart to see whether the method works for a different
data set as well. The Belgian data was processed in the same way as the
Netherlandic data and the bayesian algorithm was run with the same parameters
(1,000,000 generations, sample frequency 20).
Again, an equal rate variation model and a gamma-distributed rate variation
were tried.
In the equal rate variation model, three important groups are seen. Only a few
small groups are available and few dialects remain unresolved. This could mean
15
Figure 16: Gamma model. The Frisian dialects are shown as a subgroup of the
Frisian dialects.
Figure 17: Gamma model. Katwijk aan Zee, a coastal place in the Zuid Holland
dialect region, is grouped with dialects from the Zeeland group, further to the
south.
16
there is less contradiction between the different trees than in the results of the
Netherlandic data set. Also, the convergence between the runs is better.
The first group contains places from the east of Belgium (Figure 18). Most
dialects belong to the Limburg group on Daan’s map, two belong to the Brabant
group and one belongs to the group of dialects between Brabant and Limburg.
This group in the tree seems to be a coherent group of Limburg dialects with
some other dialects which are geographically very close.
Figure 18: Equal model. This subtree contains dialects from the east of Belgium.
The red dialects belong to the Limburg group, the green dialects belong to the
Brabant group, the yellow dialect belongs to the group of dialects between
Brabant and Limburg.
The second big group consists solely of Brabant dialects (Figure 19). There is
a division into subgroups that are geographically close to each other. There
are only two small groups of Brabant dialects that are not included in this big
group and are connected separately in the tree. The grouping is stronger than
in the Netherlandic tree.
Figure 19: Equal model. This subtree consists of the Brabant dialects.
The third group consists of dialects from the west of Belgium and northern
17
France. The dialects are roughly in three groups from Daan’s map: Western
Flemish, Eastern Flemish and dialects between Western and Eastern Flemish.
As can be seen in figure 20 the tree is nicely subdivided into these three groups.
Figure 20: Equal model. This subtree contains dialects from the west of Bel-
gium. The red dialects belong to the Western Flemish group, the green dialects
belong to the Eastern Flemish group, the yellow dialects belong to the group of
dialects between the Western and Eastern Flemish dialects.
The consensus tree under the gamma model shows the same three main groups.
The only difference is that some dialects have split from the bigger groups and
formed a smaller group.
4 Discussion
The data from the RND which was used seems to have given a reliable set of
basic words. However, there is no guarantee that the words are used as often in
one area as in another area. The closest approximation to a list of words that
is used in every area would be a Swadesh list.
The alignment has been done using a system of sound classes, which gives good
results. The quality of the alignments could possibliy become even higher. To
focus on phonology and filter out morphological effects, the stems of composed
words could all be put in the same order. Furthermore, lemmas which contain
words from different cognate sets, could be split in several lemmas: one for every
cognate sets. Finally, more combinations of phones and their diacritics could be
added to the token list. The diacritics that were not listed in (Hoppenbrouwers
and Hoppenbrouwers, 1988) were not taken into account in the alignment now.
18
As a summary of the tree sample, consensus trees were used. A characteristic
of the consensus tree is that it is not guaranteed to be a real tree from the
sample, but a reconciliation of the trees. Contradicting branchings are solved
by placing a branch at a lower level. Other tree summaries are the the maximum
probability tree (Nichols and Warnow, 2008) and the maximum clade credibility
tree (Dunn, 2008). Both methods pick a tree which exists in the tree sample: the
tree with the highest probability or the tree with the highest sum of probabilities
of the branchings respectively. The phylogenetic program (MrBayes) that was
used, was not accustomed to the creation of these trees. It was possible to fetch a
maximum probability tree topology, but without branch lengths. Furthermore,
it was possible to create a maximum clade credibility tree with an external
program, but this did not succeed. For these reasons, only consensus trees were
used in my analysis.
Although bayesian inference is a quantitative method, which draws conclusions
from large amounts of data, the evaluation of the method in this thesis was
done qualitatively. Ideally, an objective measure of distance between a bayesian
inference tree and Daan’s dialect map would be used. Both representations
would then have to be converted to the same format. Zhang and Shasha (1989)
proposes an algorithm for edit distance between trees. Implementing this al-
gorithm and processing the tree data in such a way that it could be read by
the algorithm could be a direction for future research. It would also have to
be assessed whether the edit distance between language trees coincides with a
linguistic feeling of similarity between language trees.
In earlier quantative dialect research (Heeringa and Nerbonne, 2006), the eval-
uation of the tree was also done qualitatively. However, new methods are being
applied to visualize the data in such a way that it is easier to do a human com-
parison with the gold standard. Nerbonne et al. (2011) present the Gabmap
package, which has, among other features, the possibility to project a dialect
tree onto a map.
Gamma-distributed and equal rate variation models were evaluated. The dif-
ferences in the resulting trees of the models were not very large. It seems that
for the current input format, long strings of phonological features, the choice of
the rate variation model is not of utmost importance.
The results for the bayesian inference on the Belgian dialects were better than
the results on the Netherlandic dialects. The bayesian analyses for the Belgian
dialects had better convergence rates than the analyses for the Netherlandic
dialects. It must be noted that the number of Belgian dialects was smaller than
the number of Netherlandic dialects (94 vs. 269 dialects), but they ran the same
number of generations. It may be that the Netherlandic analyses should have
run for more generations, to compensate for the large number of dialects. This
is however made inattractive by the long running times of the algorithm. There
could also be other reasons for the better performance on the Belgian data set.
For example, it could be the case that the Belgian data set had clearer divisions
between the dialects, making it easier to generate a tree.
19
5 Conclusion
The application of bayesian inference on the Netherlandic dialects performed
well on local groups. Distinctive groups from common linguistic theory were
visible. The grouping was very accurate, hardly any groupings were made that
were not coherent with the dialect map by Daan and Blok. However, many
dialects remained ungrouped. Also, local groups were not grouped with other
groups in order to get higher-order families. This limited the explanatory power
of the results.
The bayesian inference for the Belgian dialects gave suprisingly good results.
Almost all dialects were grouped and there was higher-order grouping appar-
ent. There were a few large groups, which contained dialects from a bounded
geographical area, eg. the east of Belgium. These large groups were divided
into smaller groups, which mostly followed the dialect groups from the dialect
map by Daan and Blok.
All in all, bayesian inference seems to be a good addition to the tools used
to determine the phylogeny of dialects. The performance of the method is
not constant enough to use it as the only method to create a dialect tree. In
this thesis, the method performed better on the Belgian dialects than on the
Netherlandic dialects. However, once the results have been validated using a
dialect map for the researched area, insights from bayesian inference can be
used to get a full image of dialect kinship. For example, even if no higher-order
groupings are returned in a bayesian inference tree, local groupings (as in Figure
17) can give interesting clues about relationships between dialects.
6 Literature
Campbell, L. (1998). Historical linguistics: An introduction. MIT press.
Daan, J. and Blok, D. (1969). Van Randstad tot Landrand: Toelichting
bij de kaart: Dialecten en Naamkunde. Bijdragen en mededelingen der Di-
alectenkommissie van de Koninklijke Nederlandse Akademie van Wetenschap-
pen. Noord-Hollandsche Uitgevers Maatschappij.
Dunn, M. (2008). Language phylogenies (in press). Routledge hand-
book of historical linguistics. http://pubman.mpdl.mpg.de/pubman/
item/escidoc:1851319:5/component/escidoc:1851318/dunn-phylogenetic-
approaches.pdf.
Heeringa, W. (2001). De selectie en digitalisatie van dialecten en woorden uit
de reeks nederlandse dialectatlassen. TABU, Bulletin voor Taalwetenschap,
31, number 1/2:61–103.
Heeringa, W. and Nerbonne, J. (2006). De analyse van taalvariatie in het
nederlandse dialectgebied: methoden en resultaten op basis van lexicon en
uitspraak. Nederlandse Taalkunde, 11(3):18–257.
Hoppenbrouwers, C. and Hoppenbrouwers, G. (1988). De featurefrequen-
tiemethode en de classificatie van nederlandse dialecten. TABU, Bul-
letin voor Taalwetenschap, Jaargang 18, nummer 2, 1988. Retrieved
20
from http://urd.let.rug.nl/nerbonne/papers/inferring-sound-changes-Prokic-
et-al-2011-Diachronica.pdf on 15-06-2014.
Huelsenbeck, J., Ronquist, F., Nielsen, R., and Bollback, J. (2001). Bayesian
inference of phylogeny and its impact on evolutionary biology. Science,
294(5550):2310–2314.
Jansen, M. (2002). De dialecten van ameland en midsland in vergelijking met
het stadsfries. Us Wurk. Tydskrift foar frisistyk, 51:128–152. Retrieved from
http://depot.knaw.nl/9683 on 25-06-2014.
Larget, B. and Simon, D. (1999). Markov chain monte carlo algorithms for
the bayesian analysis of phylogenetic trees. Molecular Biology and Evolution,
16:750–759.
List, J.-M. (2012). Multiple sequence alignment in historical linguistics. a sound
class based approach. In Proceedings of ConSOLE XIX, pages 241–260.
List, J.-M. and Moran, S. (2013). An open source toolkit for quantitative histor-
ical linguistics. In Proceedings of the 51st Annual Meeting of the Association
for Computational Linguistics: System Demonstrations, August 4-9, Sofia,
Bulgaria., pages 13–18.
Nerbonne, J., Colen, R., Gooskens, C., Kleiweg, P., and Leinonen, T. ((2011)).
Gabmap – a web application for dialectology. Dialectologia, Special Issue
II:65–89.
Nichols, J. and Warnow, T. (2008). Tutorial on computational linguistic phy-
logeny. Language and Linguistics Compass, 2.5:760–820.
Nieuweboer, R. (1998). The altai dialect of plautdiitsh (west-siberian mennonite
low german). Master’s thesis, University of Groningen.
Proki´c, J., Gray, R., and Nerbonne, J. (2011). Inferring sound changes
using bayesian mcmc. Submitted to Diachronica, 1/2011. Retrieved
from http://urd.let.rug.nl/nerbonne/papers/inferring-sound-changes-Prokic-
et-al-2011-Diachronica.pdf on 15-06-2014.
Rambaut, A. (2013). Figtree 1.4.1. tree figure drawing tool.
Ronquist, F. and Huelsenbeck, J. (2003). Mrbayes 3: Bayesian phylogenetic
inference under mixed models. Bioinformatics, 19:15721574.
Ronquist, F., Huelsenbeck, J., and Teslenko, M. (2011). Draft mrbayes
version 3.2 manual: Tutorials and model summaries. Retrieved from
http://mrbayes.sourceforge.net/mb3.2 manual.pdf on 25-06-2014.
Theiling, H. (2008). cxs2ipa. an x-sampa to ipa converter. Retrieved from
http://www.theiling.de/ipa/ on 16-06-2014.
Zhang, K. and Shasha, D. (1989). Simple fast algorithms for the editing distance
between trees and related problems. SIAM journal on computing, 18(6):1245–
1262.
21
7 Appendix
7.1 Tree of Netherlandic dialects – equal rate variation
The tree spans two pages.
22
7.2 Tree of Netherlandic dialects – gamma rate variation
The tree spans two pages.
25
7.3 Tree of Belgian dialects – equal rate variation
28
0.06
Tienen
Kalken
Kinrooi
Bekegem
Nazareth
Diest
Geraardsbergen
Overpelt
Gent
Kieldrecht
Kapelle-Broek
Gits
Kortrijk
Thisselt
Kampenhout
Itegem
Eupen
Oostende
Hingene
Hondegem
Overijse
Brugge
Warhem
Houthulst
Houthalen
Balen
Reninge
Zevendonk
Steenbeek
Geel
Wijnegem
Grimbergen
Assenede
Waregem
Moerbeke
Heldergem
Vreren
Ingooigem
Baelen
Lebbeke
Boom
Essen
Lochristi
Hekelgem
Werchter
Velm
Oostkamp
Arendonk
Lot
Raeren
Woesten
Diepenbeek
Wemmel
Zolder
Ronse
Gierle
Oelegem
Kalmthout
Blankenberge
Zwevegem
Aubel
Moerkerke
Alveringem
Humbeek
Roeselare
Oostkerke
Damme
Wingene
Zomergem
Bree
Boutersem
Zelzate
Meerhout
Lippelo
Nukerke
Beveren
Bellegem
Aarschot
s-Gravenvoeren
Zandvliet
Middelkerke
Herselt
Nieuwkerke
Gistel
Aalst
Buggenhout
Vertrijk
Moorslede
Veurne
Bollezeele
Lauw
Bottelare
Rijkevorsel
Mechelen
7.4 Tree of Belgian dialects – gamma rate variation
30
0.2
Mechelen
Zomergem
Grimbergen
Middelkerke
Damme
Bellegem
Gierle
Aarschot
Bekegem
Heldergem
Steenbeek
Oostkerke
Diest
Raeren
Oelegem
Hondegem
Kortrijk
Zelzate
Vreren
Wingene
Nieuwkerke
Reninge
Roeselare
Moerkerke
Houthalen
Nazareth
Thisselt
Boutersem
Itegem
Hekelgem
Balen
Ingooigem
Werchter
Houthulst
Bottelare
Wijnegem
Blankenberge
Velm
Woesten
Kalken
Kinrooi
Baelen
Hingene
Assenede
Wemmel
Boom
Overijse
Kieldrecht
Rijkevorsel
Geraardsbergen
Kalmthout
Lochristi
Lot
Lebbeke
Aalst
Humbeek
Diepenbeek
Warhem
Zolder
Kampenhout
Aubel
Kapelle-Broek
Essen
Arendonk
Herselt
Gent
Bollezeele
Buggenhout
Zwevegem
Lauw
Zandvliet
Oostende
Bree
Moerbeke
Meerhout
Geel
Eupen
Lippelo
Alveringem
Brugge
Oostkamp
Nukerke Ronse
Overpelt
Veurne
Tienen
Waregem
Beveren
Zevendonk
Gistel
s-Gravenvoeren
Vertrijk
Gits
Moorslede
7.5 Dialect map by Daan and Blok (1969)
The first page shows the map, the second page shows the legend.
32
34
