The Dialectometry of Luxembourgish

In 1997, Guillaume Schiltz published the first dialectometric study of Luxembourgish, based on data from the Luxemburgischer Sprachatlas (LSA). Dialectometry refers to the quantitative analysis of linguistic variation across space using geo-statistical techniques (Scherrer & Stöckle 2016; Goebl 2022). In Schiltz’s study, the relative similarity of linguistic variants across all LSA maps was computed for each pair of localities, resulting in a similarity matrix. This matrix served as the basis for clustering analyses and spatial visualisations of dialect similarity.

A First Dialectometric Study

One example of this approach is shown in Fig. 1 (Map 3 from Schiltz 1997), which visualizes the similarity profile of the locality Keespelt (white polygon in the center) compared to all other LSA localities. Darker shades indicate greater linguistic similarity, while lighter shades represent linguistic distance. This map suggests that Keespelt shares high similarity with much of central Luxembourg, while the northern and eastern regions exhibit lower similarity, possibly indicating separate dialect zones.

Fig. 1 Map 3 from Schiltz (1997) displaying the relative similarity of the locality Keeschpelt (white polygon) to all other localities.

Similarly, Fig. 2 (Map 5 from Schiltz 1997) shows the similarity profile for Echternach in the east. The concentration of dark polygons in the eastern region reflects strong internal similarity, in contrast to the linguistic distance observed toward the center and southwest of the country. These early analyses already demonstrate how dialectometry can uncover regional linguistic structures.

Fig. 2 Map 5 from Schiltz (1997) with the similarity profile for the locality Echternach in the east

Comparison with Present-day Data: The Schnëssen Corpus

With the development of the Schnëssen corpus—created via crowdsourcing and published in the Atlas of Variation in Luxembourgish (Variatiounsatlas)—a new, extensive dataset is now available for applying similar dialectometric analyses to contemporary Luxembourgish. This corpus contains approximately 500,000 observations for over 800 linguistic variables (phonetics, lexicon, grammar, contact phenomena), sampled across 102 communes.

Using the most frequent variant per variable per commune, we constructed a 102 × 102 similarity matrix. Scripting was performed in R using standard libraries. Fig. 3 shows a fragment of this matrix, with similarity values ranging from 0 (no identity) to 1 (complete identity). In the Schnëssen data, the actual range lies between 0.43 and 1. By comparison, Schiltz’s original data exhibited a lower minimum value (~0.3), indicating greater linguistic diversity at that time.

Fig. 3 Fragment of the similarity matrix

The entire similarity matrix can be displayed as a heatmap, which already gives a first overview of the clustering and divergence of the similary of lcoations.

We can now compare the situation from the LSA (representing Luxembourgish from pre-war times) with today’s. Note that instead of the Voronoi tesselation used by Schiltz, we are mapping based on the polygons of the communes. This entails also that the number of locations is reduced from 236 villages/cities to 102 communes. The similarity values are distributed into six bins using Jenks breaks. Fig. 4 displays the updated similarity profile for Kehlen (which includes the village Keispelt). The former similarity with the central communes is still quite pronuncend persists, but the contrast toward the East and North is less marked than in earlier data. This suggests a convergence of dialects, although regional distinctions remain detectable.

Fig. 4 Similarity profile for reference locality Kehlen (based on the Schnëssen corpus)

For Echternach (Fig. 5) the similarity profile offers a somewhat mixed picture: There is quite some similarity distance to most of the localities in the surrounding making this locality standing out. Compared with the former situation (Fig. 2), the surrounding areas seem to have lost typical dialect features of the East (light-grey polygons) and are now more similar to central Luxembourg. Echternach itself seems to show more similarity with the centre of the country (dark shaded polygons). Generally speaking, this situation is hinting at a greater amount of mixture of variants. The dialectometric analysis does not, however, reveal which specific variants are involved—this requires a different analytical approach.

Fig. 5 Similarity profile for reference locality Echternach (based on the Schnëssen corpus)

For both maps it is clearly visible that the North remains a rather distinct dialect region also today (white or fairly light polygons). This is corroborated by the next map (Fig. 6) with the similarity profile for Troisvierges in the North, which clearly singles out this region and is presenting the strongest contrast to the largest part of the country southwards. The many white polygons demonstrate the least similar values compared to the reference.

Fig. 6 Similarity profile for reference locality Troisvierges (based on the Schnëssen corpus)

While these similarity profiles do not provide an instant insight into the regional distribution of dialect areas, due to the dependency on a certain reference point, they nevertheless allow – when compared to each other – important insights into the relative level of dialectality in the area. An online tool has been created to generate dynamically all similarity profiles.

Synopsis of Skewness

To move beyond the dependence on reference localities, Schiltz used also the Synopsis of Skewness, a method that skewness of the similarity scores to the right or the the left (see the histogramms above for examples). This offers a holistic view of regional cohesion and divergence. Fig. 7 and 8 contrast the skewness maps based on Schiltz’s LSA data and the new Schnëssen data. Light polyons indicate that – compared to all other polyons – their similarity integrates well with most of the other polygons, i.e. hinting at a rather homogenous dialect. Darker polyons on the other hand signal that here the dialect here is greatly diverging from most of the polygons. Compared to the former situation, the new map reveals a tremendous homogenisation of regional variation throughout the country, visible to an extend and granularity  They illustrate a general trend toward regional homogenization, though distinctive areas in the North and South remain.

Fig. 7 Synopsis of Skewness based on LSA data (map 6 from Schiltz 1997)

Fig. 8 Skewness of Similarity by Location (based on Schnëssen corpus)

Cluster Maps

By transforming the similarity matrix into a distance matrix, we generated hierarchical clusters which were then mapped (Fig. 9). With three clusters, a clear North-South divide emerges. Increasing the number of clusters to four or five further subdivides the North and reveals the distinct ‚Minette‘ dialect area in the South-West. Interestingly, no coherent cluster appears for the East—localities along the Moselle tend to cluster with central Luxembourg, and places like Vianden or Echternach either merge with the North or stand alone. This suggests the erosion of traditional eastern dialect identity.

Fig. 9 Maps of Linguistic Clusters

Outlook

This comparative study applies dialectometric tools to both historical (LSA) and contemporary (Schnëssen corpus) data to trace changes in Luxembourgish dialect geography over time. While this blog post has focused on methodological applications, many linguistic and sociolinguistic questions remain and require more fine-grained qualitative investigation.

Future analyses will expand the dialectometric toolkit to include Getis-Ord G* statistics, Multidimensional Scaling (MDS), and t-SNE visualisations. These will provide deeper insight into the structure and dynamics of regional variation.

All results will be integrated into the Variatiounsatlas to complement variable-based visualisations with dialectometric perspectives.

Bibliography