Detecting structures in linguistic maps--Fuzzy clustering for pattern recognition in geostatistical dialectometry

Abstract

When geographic language variation is examined on a broad level, it is common practice to aggregate large quantities of data to identify dominating structures. This practice, however, may result in a neglection of valuable geolinguistic data. In this talk, new perspectives on non-aggregative methods of dealing with large corpora of dialect maps are presented that aim at preserving the distinct features of individual maps. In order to achieve this, methods derived from spatial statistics, stochastic image analysis and pattern recognition are applied and adapted to the analysis of dialect data. Firstly, we introduce a new method of clustering individual maps. We employ fuzzy clustering algorithms which provide a new angle on clustering techniques in linguistics by allowing the researcher to detect and measure gradual similarities between individual maps rather than form "hard" clusters as in conventional hierarchical clustering. This is achieved by automated comparison of statistical properties of the maps. The method can be used for grouping maps based on their spatial similarities while at the same time allowing for the investigation of semantic/phonetic/ontic etc. relationships between such spatially related maps. Secondly, we employ means of stochastic image analysis and pattern recognition in order to automatically and objectively detect geometric structures in dialect map corpora, a usually tedious and error-prone task when done manually. A quantitative method for the detection of ellipsoid patterns and their centres in dialect maps is outlined, along with the identification of parameters necessary for adapting the technique to any dataset. We argue that these methods offer promising enhancements for quantitative research techniques on geographical language variation.

Advances in Dialectometry

“…This visualization approach reveals distinct groupings, but Pröll et al emphasize that it is not the borders between the groups, but rather the dark centers in each area that should be most informative. Pröll (2013 elsewhere proposes an alternative approach. He starts from dialectometric intensity estimation (Rumpf et al 2009), which visualizes the distribution of a single linguistic variable (i.e., a single-dialect atlas map) on a so-called area class map.…”

“…Dialectometric intensity estimation is based on the idea that dialect atlas data also contain some random fluctuations due to having selected a specific speaker as being representative for the location. To better detect which variant is likely used in a location, the method takes into account the variants used in the locations within a certain radius around it [Pröll (2013 uses geographical distances as the basis for this radius, whereas Pickl et al (2014) argue that using linguistic distances might be better]. The proposed method then uses these surrounding locations to assign a probability representing how likely it is that the variant is used in the location.…”

“…In addition, numerical measures (i.e., complexity and homogeneity) can be derived from these area class maps, and these values appear to vary depending on the semantic category of the items (Pickl 2013). Grieve et al (2011) propose multivariate spatial analyses that have some similarities to Pröll's (2013 approach. As Grieve et al's approach builds on well-established spatial and geostatistical techniques (Chun & Griffith 2013), it is valuable for placing dialectometry within the larger field of geostatistics [also see Lameli (2013, chapter 5)].…”

Japanese Lexical Variation Explained by Spatial Contact Patterns

“…Since then, MDS has become a common tool for dialectometric visualisations [28,33,41,42], showing the association across localities with regards to a multitude of phenomena. Many contemporary dialectometric studies use principal components analysis (PCA), e.g., [43], or factor analysis, e.g., [44][45][46], to detect linguistic items showing similar geographical patterns. Besides, hierarchical cluster analysis is often used for finding linguistically similar locations [43,47,48].…”

Japanese Lexical Variation Explained by Historical Contact Patterns

“…Many contemporary dialectometric studies use principal components analysis (PCA), e.g. [43]), or factor analysis, e.g., [44][45][46], to detect linguistic items showing similar geographical patterns. Besides, hierarchical cluster analysis is often used for finding linguistically similar locations [43,47,48].…”