Content uploaded by Martijn Wieling
Author content
All content in this area was uploaded by Martijn Wieling on Aug 19, 2015
Content may be subject to copyright.
Tracking linguistic features underlying lexical
variation patterns: A case study on Tuscan
dialects
Simonetta Montemagni
Istituto di Linguistica Computazionale
“Antonio Zampolli”, ILC-CNR
Martijn Wieling
University of Groningen, CLCG
In this paper, we illustrate the application of hierarchical spectral partitioning of
bipartite graphs in the study of lexical variation in Tuscany based on the data from a
regional linguistic atlas. This method makes it possible not only to identify existing
patterns of lexical variation in Tuscany, but also to uncover the underlying lexical
features in terms of the most characteristic concept-lexicalization pairs. The results
are promising, demonstrating the potential of the method for tracking the linguistic
features underlying identified patterns of lexical variation and change across
generations.
Introduction
In dialectometry (Séguy 1971) the focus lies on the aggregate analysis of dialect variation. In
contrast to “cherry-picking” a few linguistic items confirming the analysis one wishes to settle on
(Nerbonne 2009), the advantage of the aggregate approach is that it offers a more objective view of
dialect variation. Unfortunately, many studies focusing on the aggregate pattern of dialect variation
have disregarded the underlying linguistic basis. As a consequence, linguists have remained critical
of the dialectometric approach (Schneider 1988, Woolhiser 2005, Loporcaro 2009).
To counter this criticism, various new dialectometric methods have been developed aimed at
identifying the linguistic basis of dialectal variation (as reviewed in Wieling and Nerbonne 2015).
For example, Nerbonne (2006) and Pröll et al. (in press) use an approach based on factor analysis,
whereas Shackleton (2005) uses principal component analysis. Grieve et al. (2011) follow the
workflow of traditional dialectology (i.e. identifying isoglosses, bundling isoglosses and cluster
analysis) by using multivariate spatial analysis.
The method we will apply here, hierarchical bipartite spectral graph partitioning (HBSGP), has
been developed by Wieling and Nerbonne (2009, 2010, 2011), who adopted it from information
retrieval (Dhillon 2001) and applied it to dialectology. HBSGP results in a clustering of
geographical varieties while simultaneously providing a linguistic basis for each of the identified
clusters. The approach of Wieling and Nerbonne (2011) has been successfully applied to study
phonetic variation in Dutch dialects (Wieling and Nerbonne 2011), English dialects (Wieling et al.
2013) and Tuscan dialects (Montemagni et al. 2012, 2013). More recently, the method has also been
applied to investigate lexical variation in contemporary English dialects on the basis of the BBC
Voices data (Wieling et al. 2014a).
In this study, we focus on lexical variation. Our dataset, a regional lexical atlas of Tuscan dialects
whose data have a diatopic and diachronic characterization, allows us to explore the potential of the
HBSGP method in the study of lexical variation. In particular, it enables us to identify lexical
features and their relationships on the one hand and to reconstruct the dynamics of lexical change
across generations on the other hand. Technically, a new measure is proposed for determining the
most important lexical features associated with the identified dialectal areas.
Data
We investigate Tuscan lexical variation on the basis of a linguistic atlas of Tuscany, the Atlante
Lessicale Toscano (ALT, Giacomelli et al. 2000), now available as an online resource
(http://serverdbt.ilc.cnr.it/ALTWEB). ALT is a regional Italian lexical atlas focusing on dialectal
variation throughout Tuscany, where both Tuscan and non-Tuscan dialects are spoken. In this paper
we focus on Tuscan dialects only, recorded in 213 localities by a total of 2060 informants who were
selected with respect to various socio-demographic parameters (such as age, education and gender).
ALT interviews were carried out on the basis of a questionnaire of 745 target items, designed to
elicit mainly lexical, but also semantic and phonetic variation. This study is based on the results of
onomasiological questions, i.e. starting from concepts and looking for their lexicalizations. A
typical onomasiological question asks how a given concept is designated or named, e.g. “what is the
name for flat and crispy bread, seasoned with salt and oil?”. To avoid interference with non-
lexicalized answers, we excluded questions prompting 50 or more distinct lexical items.
Furthermore, we only considered nouns (the large majority of items of ALT questionnaire) in this
study. The resulting subset consists of 170 questionnaire items for which a total of 5,174 distinct
normalized answers were given (on average 30 lexical variants per concept) distributed into 61,496
geo-referenced responses (i.e. associated with locations). The total number of speaker-responses
was 384,454.
To abstract away from phonetic variation, we used the most abstract representation level present in
ALT (Cucurullo et al. 2006). This normalized representation was meant to abstract away from
phonetic variation (caused by productive phonetic processes), but did not remove morphological
variation or variation caused by unproductive phonetic processes. In this study we used the
normalized lexical answers to the selected subset of 170 onomasiological questions. The same set of
questions has also been used by Wieling et al. (2014b) in a study of lexical differences between
Tuscan dialects and standard Italian.
The representativeness of the selected sample with respect to the whole set of ALT onomasiological
questions (i.e. a total of 460 questionnaire items) was assayed using the correlation between overall
lexical distances and lexical distances obtained from the selected sample (Wieling et al. 2014b). The
Pearson’s correlation coefficient was r = 0.94, showing the representativeness of the selected
sample with respect to the whole set of onomasiological questions.
Methods
In this study, we use hierarchical bipartite spectral graph partitioning as our method of choice
(Wieling and Nerbonne 2011). As mentioned before, this approach simultaneously clusters the
geographic locations together with the linguistic features characterizing them. In this case, a cluster
of locations is characterized by a linguistic basis expressed in terms of the most salient lexical
features. These lexical features can be seen as a proxy of the traditional notion of lexical isoglosses,
establishing the boundaries of dialectal areas.
Every variety attested in a given location is described in terms of concept-lexicalization (CL) pairs
linking each of the 170 selected concepts with its lexicalization(s) (reported in the normalized form)
in the specific location. CL frequencies are normalized by dividing the number of recorded answers
by the number of informants in a given location, with their value ranging between 0 and 1. Since for
each location there was a socio-demographically differentiated group of informants potentially
giving rise to multiple responses to denote the same concept, the sum of normalized frequencies of
lexical variants associated with the same concept in a certain location can be greater than 1.
The input for the HBSGP method is a bipartite graph which contains two sets of vertices, locations
and CL pairs, connected by lines. There exists a line between a location and a CL pair whenever at
least one of the speakers in the location uses the lexical variant. The lines are weighted between 0
and 1. A value of 0 indicates that no speakers in the location use the lexical variant (and thus equals
the absence of a line), whereas a value of 1 indicates that all speakers in the location use the lexical
variant to denote the concept being investigated. Table 1 gives an example of (a tabular
representation of) the bipartite graph, with the rows corresponding to the locations and the columns
to the CL pairs. About 80% of the speakers in Caprese Michelangelo use the form aràncio to denote
an ORANGE. A similar number of speakers also uses melàngola to denote the same (speakers
frequently provided multiple lexicalizations to denote a certain concept).
The input matrix is then subjected to singular value decomposition (SVD), and the k-means
clustering algorithm (with k equals 2) is applied to the results of the SVD resulting in a two-way
clustering. The k-means clustering was repeated 1000 times for robustness. As the output of the
SVD combines the locations with the CL pairs, the clustering likewise groups locations and CL
pairs. Consequently, lexical variants grouped with locations can be seen as characteristic elements
of those locations. For more mathematical details, we refer the interested reader to Wieling and
Nerbonne (2011).
Location
ORANGE-arància
ORANGE-aràncio
ORANGE-melàngola
Caprese Michelangelo
0.1379
0.7931
0.7931
Pieve Santo Stefano
0.4000
0.7333
0.2000
Anghiari
0.0000
0.7059
1.0000
Sansepolcro
0.0000
1.0000
1.0000
Table 1: Tabular representation of a bipartite graph. The numbers represent the normalized
frequency (obtained by dividing by the number of speakers) of the lexical variant associated with a
given concept in the different locations which ranges between 0 and 1. As the speakers may use
multiple variants to denote a concept, the normalized frequencies associated with a concept in a
certain location do not have to sum to 1.
In order to identify the most characteristic linguistic features for a group of locations, Wieling and
Nerbonne (2011) combined two different criteria which were implemented in two different and
complementary measures: representativeness and distinctiveness. Representativeness measures the
relative frequency of the lexicalization of a given concept in the locations in the cluster. For
example, if the cluster contains ten locations and all speakers in seven locations use the lexical
variant, the representativeness is 0.7. Distinctiveness measures how frequently the lexical variant
occurs within as opposed to outside of the cluster (corrected for the relative size of the cluster,
which is calculated by dividing the number of locations in the cluster by the total number of
locations in the dataset). A distinctiveness of 1 indicates that the lexical variant is only used inside
the cluster. The distinctiveness equals 0 when the relative frequency of the lexical variant in the
cluster is equal to the relative size of the cluster (i.e. it is not distinctive). Interestingly, the measures
of representativeness and distinctiveness are reminiscent of the Consistency and Homogeneity
measures introduced by Labov and colleagues for the construction of isoglosses in the Atlas of
North American English (Labov et al. 2006). Homogeneity measures how much variation exists
within the region defined by the isogloss (i.e. corresponding to a non-chance corrected variant of
distinctiveness) and Consistency (i.e. corresponding to representativeness) measures how strongly
the variable is concentrated within a given region.
The two measures capture two different equally important desiderata of isoglosses: to put it in the
words of Labov et al. (2006), “First, we want the area defined to be as uniform as possible […].
Second, we want as high a proportion of hits as possible to be located within the isogloss”. For this
reason they need to be combined. Wieling and Nerbonne (2011) combined representativeness and
distinctiveness measures by averaging them, yielding the importance score. Here, we propose that
to determine the relevance of CL pairs in the characterization of identified lexical areas it is better
to multiply the two values. The advantage of this approach is that it is not possible to assign high
importance values to lexical variants which score high on a single measure only. For example,
lexical variants occurring in all locations are highly representative, but not distinctive. Similarly, a
lexical variant only occurring in a single location is highly distinctive, but not representative (unless
the cluster contains a single location). Note that constraints on isogloss construction were also
foreseen by Labov et al. (2006) by enforcing frequency thresholds. However, the advantage of the
approach proposed by Wieling and Nerbonne (2011) and its evolution presented here consists in the
fact that no a priori constraints on the values of individual measures are defined.
Results
In this section, we report the results of applying the HBSGP method to the selected ALT dataset.
The results obtained are based on 5,174 CL pairs and 213 locations, which correspond to all lexical
data gathered through fieldwork (as opposed to a dataset in which infrequent lexical variants are
filtered out) for the 170 selected concepts. See Wieling and Montemagni (2015) for a discussion of
the advantages connected with this dataset.
The map in Figure 1 shows the geographic visualization of the clustering of Tuscan varieties into
seven groups designated as follows: the Florence area (A), the western Tuscan area (C) and the
dialects from Arezzo, Siena, Grosseto and Mount Amiata (E) which represent the three main
groupings, together with the dialects from Elba island (D), Chiana Valley (F), Capraia Island (G)
and Apuan Alps (B) which are minor but clearly distinct dialectal areas.
Figure 1: Geographic visualization of the clustering of Tuscan varieties into seven groups.
It is interesting to note that this result is in line with the classifications of Tuscan dialects proposed
by Giacomelli (1975) for what concerns the lexicon, and by Giannelli (1976, 2000) which is based
instead on phonetic, phonemic, morpho-syntactic and lexical features. It is also in line with the
subdivision of Tuscan dialects by Pellegrini (1977), in spite of it being mainly based on the
distribution of phonetic phenomena.
Linguistic features underlying identified lexical areas
For what concerns the underlying lexical features, we first focus on the three main dialectal clusters
(A, C and E). Table 2 reports for each cluster the five most important CL pairs with associated
values of representativeness, distinctiveness and importance.
The relevance of the lexical features with respect to the dialectal subdivision emerges clearly from
the value maps in Figure 2, which show the geographic distribution of the first and second topmost
lexical features of each of the three main identified clusters (A, C and E). The topmost lexical
features associated with each identified cluster can be assimilated with the traditional notion of
bundle of isoglosses, which have long been considered a major criterion for the definition of dialect
areas: as Chambers and Trudgill (1998) put it, “the significance of a dialect area increases as more
and more isoglosses are found which separate it from adjoining areas”.
Cluster
Concept-Lexicalization pair
Representativeness
Distinctiveness
Importance
E
TURKEY-bìllo
0.863
0.700
0.604
CORNER OF TISSUE-pìnzo
0.724
0.795
0.576
EYE GUM-cipìcchia
0.624
0.920
0.574
OIL JAR-zìro
0.879
0.609
0.535
VAT-bigónzo
0.649
0.821
0.533
A
ORANGE-arància
0.779
0.675
0.526
LADLE-romaiòlo
0.788
0.536
0.423
OIL JAR-órcio
0.671
0.590
0.396
TURKEY-tàcco
0.390
1.000
0.390
BRAWN-capofréddo
0.432
0.900
0.389
C
OIL JAR-cóppo
0.749
0.696
0.522
EYE GUM-cìspia
0.702
0.676
0.474
BREAST-pùppa
0.649
0.717
0.466
FLEA-pùce
0.602
0.686
0.413
CLUSTER OF GRAPES-pìgna
0.570
0.701
0.400
Table 2: The five topmost lexical variants for the three main clusters of Tuscan dialects.
By comparing the maps of Figure 2, we can observe that the geographic distribution of the topmost
CL pairs of the E, A and C clusters does not cover all and only the locations in the cluster. Each of
them can be seen as a quantitative visualization of individual isoglosses, where darkness of color
denotes the frequency of occurrence of the represented lexical variant (dark colors denote a greater
frequency, lighter colors lower frequency, and no coloring indicates the absence of the variant). As
can be observed, lexical variants shown in Table 2 may occur beyond the border of the cluster area,
thus lowering the distinctiveness value of the CL pair, or they may not occur in the whole cluster
area resulting in a lower representativeness. For instance, in cluster A comparable
representativeness values are observed for the two topmost CL pairs (0.77 - 0.78), whereas the CL
ranked in second place, i.e. LADLE-romaiòlo, has a lower distinctiveness value (0.53) than the
topmost CL (i.e. whose distinctiveness value is 0.67). Different patterns can be observed in clusters
E and C, with decreasing representativeness and increasing distinctiveness in the former case, and
with both of them decreasing in the latter case. Despite these slight differences, in all cases
representativeness and distinctiveness show relatively high values which never reach the value of 1
(with the only exception of the CL pair TURKEY-tàcco in cluster A whose distinctiveness is equal to
1). The average values of the five topmost lexical features for representativeness and distinctiveness
range between 0.61 and 0.74, and 0.69 and 0.77 respectively, demonstrating that the corresponding
dialect areas are not marked by very clear and strong dialect borders.
Figure 2: Value maps of the first (row 1) and second (row 2) topmost CL pairs for the A, C and E dialectal clusters.
Areas with darker (blue) color denote a greater frequency of occurrence of the selected lexical variant; lighter colors
denote a lower frequency, while no coloring (white) denotes the absence of the variant.
Different distinctiveness-representativeness patterns are observed in the case of the smaller
peripheral areas B, D, F and G (see Table 3). Here, the most salient CL pairs are highly distinctive
(their average values ranges from 0.84 to 1), with the average representativeness ranging from 0.49
to 1. Thus smaller dialect areas are characterized by much more distinctive features than the larger
areas.
Besides the strength of dialectal borders, granularity of the identified dialectal areas is another open
issue in the study of dialectal variation. Consider, for instance, the traditional dialectal subdivision
of Tuscan dialects by Pellegrini (1977) and Giannelli (1976, 2000). In his Carta dei Dialetti
d’Italia, Pellegrini (1977) identifies a western variety of Tuscan which is further subdivided into
Pisano-Livornese-Elbano, and Pistoiese and Lucchese. On the other hand, Giannelli (1976, 2000)
identifies Pisano-Livornese, Lucchese, Elbano and Pistoiese as independent dialectal varieties in his
seminal work Toscana. The two subdivisions are compatible with each other but adopt different
levels of granularity, i.e. they are seen through lenses differing in their magnifying power.
Depending on the specific goals of a study, different levels of granularity of the dialectal landscape
may be appropriate. By exploiting the hierarchical clustering results, the HBSGP method can also
be used to identify increasingly smaller dialectal areas associated with progressively more specific
lexical features. These nested dialect areas are characterized by nested isoglosses (i.e. the spatial
distribution of one feature is entirely contained within that of another). To assess these nested
isoglosses, we compare the geographical and linguistic results obtained by clustering the selected
dataset into two, four and seven groups (with the latter representing the clustering discussed so far).
Cluster
Concept-Lexicalization pair
Representativeness
Distinctiveness
Importance
F
FINCH-frenguéllo
1.000
1.000
1.000
CUCUMBER-citróne
1.000
0.973
0.973
HAIL-granìschia
0.667
1.000
0.667
GOOSE-ciucióne
0.667
1.000
0.667
LIZARD-racanàccio
0.667
1.000
0.667
B
SNOW-gnéva
0.429
1.000
0.429
ROLLING PIN-canèlla
0.429
1.000
0.429
STYE-orzaiolo
0.653
0.633
0.414
GARBAGE-rùsco
0.531
0.734
0.389
LIZARD-ciortellóne
0.430
0.853
0.367
D
HORNET-buffóne
0.950
1.000
0.950
KHAKIS-cicàchi
0.500
1.000
0.500
KHAKIS-cicàco
0.500
1.000
0.500
PINE CONE-pignòcca
0.500
1.000
0.500
TROUGH-tròlego
0.500
1.000
0.500
G
WATERMELON-patècca
1.000
1.000
1.000
MELON-melòne
1.000
1.000
1.000
CLUSTER-raspòllo
1.000
1.000
1.000
SQUIRREL-miseràngolo
1.000
1.000
1.000
LIZARD-bìscia
1.000
1.000
1.000
Table 3: The five topmost lexical variants for the smaller peripheral areas B, D, F and G.
Figure 3 reports the geographic visualization of clustering the Tuscan varieties into two, four and
seven groups. In the map with two clusters (Figure 3, left), the large red cluster corresponds to the
composite set of Tuscan dialects, excluding only the Chiana Valley dialects (cyan cluster). The map
with four clusters (Figure 3, middle) shows the main subdivision of Tuscan dialects between
Northern dialects (cyan and green clusters), covering (from east to west) Fiorentino, Pistoiese,
Lucchese and Pisano-Livornese, and Southern dialects (violet and red clusters), i.e. (from east to
west) the dialect from Arezzo, Siena and Grosseto (violet cluster) and from the Chiana valley (red
cluster). The map containing seven clusters (Figure 3, right) has already been discussed above.
Figure 3: Geographic visualization of the clustering of Tuscan varieties into two, four and seven groups.
Table 4 shows the lexical features characterizing the red, cyan and blue clusters in the first, second
and third map, respectively. These clusters cover a progressively restricted area including the
province of Florence (in the map with seven clusters it almost coincides with the province). Table 4
reports, for each of these clusters, the five topmost lexical variants with their associated scores. The
most salient CL pairs characterizing the red cluster of the two-clusters map coincide with pan-
Tuscan words well known from the literature (Giacomelli and Poggi Salani 1984): they show a
distinctiveness value equal to 1 and very high representativeness values (≥ 0.79). Similar
observations hold for the cluster corresponding to the set of Northern Tuscan dialects (the cyan
cluster in Figure 3, middle) with one main difference: all values are considerably lower, with a
general reduction observed at the level of representativeness. This illustrates that the cyan cluster is
a heterogeneous area. However, by comparing the CL pairs underlying the cyan cluster in the
second map and the blue cluster in the third map, we can also see there are two shared lexical
variants, namely OIL JAR-cóppo and EYE GUM-cìspia, which appear among the topmost features
whose importance values in the smaller blue cluster are higher (determining a higher ranking),
despite their unavoidably lower distinctiveness. In this case, these CL pairs are more characteristic
of the smaller cluster, whereas a word such as THIMBLE-anèllo is more characteristic of the larger
cluster. This suggests that whenever the same features appear to qualify nested clusters, they should
be taken as relevant features for the cluster in which they play a more prominent role (i.e. having a
higher importance value). Consequently, OIL JAR-cóppo and EYE GUM-cìspia should be removed
from the most salient features of the cyan cluster due to the lower importance (0.424 against 0.522
for the former, and 0.397 against 0.474 for the latter) with respect to the nested blue cluster.
In sum, these results show that hierarchical spectral partitioning can be usefully exploited to
identify dialectal areas at different levels of granularity with their associated lexical features. In
particular, the method may help in the selection of the most appropriate isoglosses for each dialectal
area and in the reconstruction of nested isoglosses.
Cluster
Concept-Lexicalization pair
Representativeness
Distinctiveness
Importance
Two-
cluster
map:
Red
SINK-acquàio
0.909
1.000
0.909
CELERY-sèdano
0.853
1.000
0.853
MELON-popóne
0.844
1.000
0.844
LAUREL-allòro
0.801
1.000
0.801
WATERMELON-cocómero
0.794
1.000
0.794
Four-
cluster
map:
Cyan
THIMBLE-anèllo
0.495
0.857
0.424
OIL JAR-cóppo
0.525
0.808
0.424
CATERPILLAR-brùcio
0.448
0.928
0.416
EYE GUM-cìspia
0.498
0.798
0.397
TURKEY-lùcio
0.445
0.872
0.388
Seven-
cluster
map:
Blue
OIL JAR-cóppo
0.749
0.696
0.522
EYE GUM-cìspia
0.702
0.676
0.474
BREAST-pùppa
0.649
0.717
0.466
FLEA-pùce
0.602
0.686
0.413
CLUSTER OF GRAPES-pìgna
0.570
0.701
0.400
Table 4: The five topmost lexical variants of the red, cyan and blue areas in the two, four and seven-
cluster maps of Tuscan dialects.
Reconstructing the dynamics of lexical change
The hierarchical spectral partitioning method can also be used for studying the dynamics of lexical
change across generations. For this purpose, ALT speakers were grouped in an old age group (born
in 1930 or earlier – 1930 was the median year of birth) and a young age group (born after 1930). To
guarantee comparability of results, we focused on two maps each having four clusters. As Figure 4
shows, the analysis of the two datasets results in slightly different, partially overlapping lexical
areas, with the area corresponding to the southeastern (cyan) cluster being more restricted for the
older speakers. Major differences, however, are explicitly clear at the level of the underlying lexical
features. In particular, the central blue area is more restricted (and also linked with fewer CL pairs:
881 vs. 1193) in the map built on the basis of the answers by the young speakers.
Besides the different size of the set of associated linguistic features (i.e. more reduced in the case of
young speakers), it is interesting to note that 424 salient lexical features underlying the old speakers
map do not appear among the features underlying the young speakers map. These CL pairs
emerging from old speakers correspond typically to old-fashioned and traditional notions as well as
less common plants and animals. Examples include STRUCTURE FOR BED WARMER-prète, POPPY-
ròsolo, MUTTON-bìrro, SET OF POPLARS-alborellàia. These CL pairs can be seen as lexical variants
which are no longer being used by younger speakers, and these are likely to disappear altogether.
The number of CL pairs restricted to young speakers is much lower (112) than the number of CL
pairs restricted to the old speakers. In this case, the CL pairs correspond to standard Italian words
(e.g., CLOSET-ripostìglio, WEEPING WILLOW-sàlice piangènte, HARVEST-mietitùra), generic terms
(e.g., AFTERNOON-dòpo mangiàto, SLUG-lumàca ignùda) or “distorted” (i.e. deviant with respect to
traditional pronunciation) variants of dialectal terms (e.g., TUSCAN COLD CUT FROM PORK
SHOULDER-capricòllo). The typology of these lexical variants shows the dynamics of lexical change
ongoing in younger Tuscan generations, characterized by the loss of local features in favor of
generic or standard terms, and by the creative distortion of dialectal words.
In both cases, however, these CL pairs are not highly ranked (i.e. not the most important) for the
associated old and young clusters. Instead, the CL pairs underlying both maps (a total of 769) show
clear differences with respect to their ranking. For example, the 1st, 10th, 20th and 50th lexical
variants in the ranked list of CL pairs underlying the old speakers map correspond to the 60th,
809th, 59th and 818th position in the young CL pairs list, respectively. Similarly, the 1st, 10th, 20th
and 50th ranked lexical variants of the young speakers are ranked (respectively) in the 100th, 13th,
17th and 69th position in the old speakers list. The asymmetry between the old-young vs. young-old
correspondences can be seen as the result of a dialect leveling process, causing the lower
importance of old-fashioned lexical variants for the young speakers (which are top-ranked for the
old speaker). Seen from the perspective of young speakers, the disalignment of the ranking is more
reduced, reflecting an additional shared set of dialectal lexical items.
Table 5 reports the five topmost CL pairs underlying the blue cluster in the two maps. Clearly, the
importance values associated with the blue cluster of the old speakers are higher than those
associated with the blue cluster of the young speakers. This pattern is confirmed by comparing the
average importance scores of the top-10 and top-100 CL pairs in the two lists, which are much
higher for the old speakers (0.42 vs. 0.34 for the top-10 and 0.26 vs. 0.17 for the top-100). This may
also be seen as evidence in support of dialect leveling: lexical areas inferred from young speakers
data are characterized by less distinctive and/or representative features.
Figure 4: Geographic visualization of a four-way clustering of Tuscan varieties on the basis of data from young vs. old
speakers.
Cluster
Concept-Lexicalization pair
Representativeness
Distinctiveness
Importance
Old
speakers:
Blue
cluster
GRAPE-chìcco
0.721
0.828
0.597
CHESTNUT HUSK-rìccio
0.706
0.661
0.467
EMBERS-bràce
0.673
0.632
0.425
BRAZIER-bracière
0.596
0.680
0.405
HAZELNUT-nocciòla
0.794
0.507
0.403
Young
speakers:
Blue
cluster
BAT-pipistrèllo
0.736
0.538
0.396
BREAST-pùppa
0.428
0.900
0.385
THIMBLE-anèllo
0.394
0.893
0.352
OIL JAR-cóppo
0.437
0.772
0.337
EYE GUM-cìspia
0.431
0.779
0.335
Table 5: The five topmost lexical variants of the blue cluster in the young vs. old speakers maps of
Tuscan dialects.
Conclusion
In this paper, we illustrated the application of hierarchical spectral partitioning of bipartite graphs in
the study of lexical variation in Tuscany based on the dialectal corpus of the Atlante Lessicale
Toscano. Our results demonstrate the potential of the method in bridging the gap between models of
linguistic variation based on aggregate analyses and more traditional analyses based on individual
linguistic features.
By using the HBSGP method, we not only identified existing patterns of lexical variation in
Tuscany on the basis of the whole dialectal corpus, but also uncovered the underlying lexical
features in terms of the characterizing concept-lexicalization pairs. The most relevant CL pairs
represent the features used to classify and define each identified lexical area. To put it in more
traditional terms, they can be seen as a proxy of lexical isoglosses marking both the qualitative and
quantitative distribution of the lexical variants identified as discriminating features of a given
lexical dialect area. This entails that the set of the topmost CL pairs associated with each identified
lexical dialect area acts as a proxy of bundles of isoglosses, where the grading of individual
isoglosses within the bundle is determined on the basis of the combination of representativeness and
distinctiveness. If the representativeness score associated with identified isoglosses (CL pairs) can
help to shed light on how much variation exists within the area defined by a given isogloss, the
distinctiveness score reflects how strongly the lexical variant is concentrated within that area. By
comparing the results obtained for different dialect areas, we have seen that different stages of the
process of dialect differentiation can be inferred from the different values of these two measures:
dialectal subdivisions range from clearly defined areas to areas characterized by fuzzy borders.
We also investigated whether and to what extent patterns of lexical variation and their associated
features varied with respect to the granularity of the identified dialectal areas and with the age of
informants, revealing interesting results. The possibility of exploring linguistic variation at different
levels of granularity makes it possible to customize the analysis with respect to the user’s needs.
The linguistic features associated with increasingly smaller areas can be seen as nested isoglosses,
occurring when the spatial distribution of one feature is contained entirely within that of another
and establishing an implicational relationship between the two.
The analysis and comparison of lexical variation patterns and associated features across generations
showed that the method can also be usefully exploited to track the change in the typology of
features in young vs. old informants and to monitor the vitality of a dialect in a given area. In
particular, the HBSGP method turned out to effectively capture the dynamics of lexical change in
Tuscany, by highlighting the emergence of lexical innovations and the obsolescence of old-
fashioned traditional dialectal words.
Current directions of research include testing the robustness of these results by noisy clustering and
the analysis of lexical variation patterns across semantic domains.
References
Chambers, J.K. & Peter Trudgill. 1998. Dialectology. Second edition. Cambridge University Press,
Cambridge.
Cucurullo, Sebastiana, Simonetta Montemagni, Matilde Paoli, Eugenio Picchi & Eva Sassolini.
2006. Dialectal resources on-line: the ALT-Web experience. In Proceedings of the 5th International
Conference on Language Resources and Evaluation (LREC-2006), Genova, Italy, 24-26 May 2006,
pp. 1846-1851.
Dhillon, Inderjit S. 2001. Co-clustering documents and words using bipartite spectral graph
partitioning. In Proceedings of the 7th ACM SIGKDD International Conference on Knowledge
Discovery. ACM, New York, pp. 269-274.
Giacomelli Gabriella. 1975. Aree lessicali toscane. La ricerca dialettale, I, Pacini Editore, Pisa, pp.
115-152.
Giacomelli, Gabriella & Teresa Poggi Salani. 1984. Parole toscane. Quaderni dell’Atlante Lessicale
Toscano, 2(3), Leo S. Olschki Editore, Firenze, pp. 123-229.
Giacomelli, Gabriella, Luciano Agostiniani, Patrizia Bellucci, Luciano Giannelli, Simonetta
Montemagni, Annalisa Nesi, Matilde Paoli, Eugenio Picchi & Teresa Poggi Salani. 2000. Atlante
Lessicale Toscano. Lexis Progetti, Roma.
Giannelli Luciano. 2000. Toscana. Pacini Editore, Pisa (1976, first edition).
Grieve Jack, Dirk Speelman & Dirk Geeraerts. 2011. A statistical method for the identification and
aggregation of regional linguistic variation. Language Variation and Change, 23, Cambridge
University Press, pp. 193-221.
Labov, William, Sharon Ash & Charles Boberg. 2006. The Atlas of North American English.
Phonetics, Phonology and Sound Change. Mouton de Gruyter, Berlin-New York.
Loporcaro, Michele. 2009. Profilo linguistico dei dialetti italiani, Laterza, Roma-Bari.
Montemagni, Simonetta, Martijn Wieling, Bob de Jonge & John Nerbonne. 2012. Patterns of
language variation and underlying linguistic features: A new dialectometric approach. In Patricia
Bianchi, Nicola De Blasi, Chiara De Caprio and Francesco Montuori (Eds.), La Variazione
nell’italiano e nella sua Storia. Varietà e Varianti Linguistiche e Testuali. Atti dell’XI Congresso
SILFI (Società Internazionale di Linguistica e Filologia Italiana). Franco Cesati Editore, Firenze.
Vol. II, pp. 879-889.
Montemagni, Simonetta, Martijn Wieling, Bob de Jonge & John Nerbonne. 2013. Synchronic
patterns of Tuscan phonetic variation and diachronic change: Evidence from a dialectometric study.
Literary and Linguistic Computing, 28(1), Oxford University Press, pp. 157-172.
Nerbonne, John. 2006. Identifying linguistic structure in aggregate comparison. Literary and
Linguistic Computing, 21, Oxford University Press, pp. 463-476.
Nerbonne, John. 2009. Data-driven dialectology. Language and Linguistics Compass 3(1), pp. 175-
198.
Pellegrini, Giovan Battista. 1977. Carta dei Dialetti d'Italia, Pacini Editore, Pisa.
Pröll, Simon, Simon Pickl, Aaron Spettl. In press. Latente Strukturen in geolinguistischen Korpora.
In Michael Elmentaler, Markus Hundt, Jürgen E. Schmidt (Eds.), Deutsche Dialekte - Konzepte,
Probleme, Handlungsfelder. Kongress der Internationalen Gesellschaft für Dialektologie des
Deutschen (IGDD), Kiel vom 13.–15. September 2012. Stuttgart: Steiner.
Séguy, Jean. 1971. La relation entre la distance spatiale et la distance lexicale. Revue de
Linguistique Romane, 35(138), pp. 335-357.
Schneider, Edgar. 1988. Qualitative vs. quantitative methods of area delimitation in dialectology: A
comparison based on lexical data from Georgia and Alabama. Journal of English Linguistics, 21,
pp. 175-212.
Shackleton, Robert G., Jr. 2005. English-American speech relationships: A quantitative approach.
Journal of English Linguistics 33, pp. 99-159.
Wieling Martijn & John Nerbonne. 2009. Bipartite spectral graph partitioning to co-cluster varieties
and sound correspondences in dialectology. In Proceedings of the 2009 Workshop on Graph-Based
Methods for Natural Language Processes. ACL, Stroudsburg, PA, pp. 14-22.
Wieling Martijn & John Nerbonne. 2010. Hierarchical spectral partitioning of bipartite graphs to
cluster dialects and identify distinguishing features. In Proceedings of the 2010 Workshop on
Graph-Based Methods for Natural Language Processing, ACL, Stroudsburg, PA, pp. 33-41.
Wieling, Martijn & John Nerbonne. 2011. Bipartite spectral graph partitioning for clustering dialect
varieties and detecting their linguistic features. Computer Speech and Language 25, pp. 700-715.
Wieling, Wieling, Robert G. Shackleton, Jr. & John Nerbonne. 2013. Analyzing phonetic variation
in the traditional English dialects: Simultaneously clustering dialects and phonetic features. LLC
28(1). 31-41.
Wieling, Martijn, Clive Upton & Ann Thompson. 2014a. Analyzing the BBC Voices data:
Contemporary English dialect areas and their characteristic lexical variants. Literary and Linguistic
Computing, 29(1), Oxford University Press, pp. 107-117.
Wieling, Martijn, Simonetta Montemagni, John Nerbonne & R. Harald Baayen. 2014b. Lexical
differences between Tuscan dialects and standard Italian: Accounting for geographic and
sociodemographic variation using generalized additive mixed modeling. Language 90(3), pp. 669-
692.
Wieling, Martijn & John Nerbonne. 2015. Advances in dialectometry. Annual Review of Linguistics
1.
Wieling, Martijn & Simonetta Montemagni. 2015. Infrequent forms: noise or not? In this volume.
Woolhiser, Curt. 2005. Political borders and dialect divergence/convergence in Europe. In Peter
Auer, Frans Hinskens & Paul Kerswill (Eds.), Dialect Change. Convergence and Divergence in
European Languages, Cambridge University Press, New York, pp. 236-262.
