Analyzing the BBC Voices data: Contemporary English dialect areas and their characteristic lexical variants

This study investigates data from the BBC Voices project which contains a large amount of vernacular data collected by the BBC between 2004 and 2005. This project was designed primarily to collect information on vernacular speech around the United Kingdom for broadcasting purposes. As part of the project, a web-based questionnaire was created, to which tens of thousands of people supplied their way of denoting thirty-eight concepts which were known to exhibit marked lexical variation. Along with their variants, those responding to the online prompts provided information on their age, gender, and —significantly for this study— their location, this being recorded by means of their postcode. In this study we focus on the relative frequency of the top-ten variants for all concepts in every postcode area. By using hierarchical spectral partitioning of bipartite graphs, we are able to identify four contemporary geographical dialect areas together with their characteristic lexical variants. Even though these variants can be said to characterize their respective geographical area, they also occur in other areas, and not all people in a certain region use the characteristic variant. This supports the view that dialect regions are not clearly defined by strict borders, but are fuzzy at best. Introduction In 2004 and 2005, the British Broadcasting Corporation conducted a large-scale survey in order to obtain a contemporary view of English dialectal variation. People visiting a speciallyconstructed website were invited to offer their variants for thirty-eight concepts that were known to exhibit marked lexical variation. Along with their lexical use, informants were asked to provide details of their age, gender, and geographical (post-coded) location. Upwards of 29,000 people participated in this project (“BBC Voices”) to a greater or lesser degree, resulting in a substantial electronic dataset as a consequence. As dialectologists we are interested in investigating geographical structure which might be present in our data. Given the enormous size of the Voices lexical dataset (containing more than 700,000 responses in total), we use quantitative methods from dialectometry to provide an aggregate view of the contemporary English dialectal landscape. Dialectometry originated in the 1970’s (Séguy, 1973) to provide a more objective method of identifying dialect differences than by “cherry-picking” the features which support the analysis one wishes to settle on (Nerbonne, 2009). Unfortunately, dialectometry has not been received very favorably by some traditional dialectologists, as aggregate analyses obscure the importance of individual linguistic features, on which they are required to focus for their often philologically-directed purposes. Consequently, there have been a number of attempts to develop quantitative methods which enable the identification of characteristic linguistic variables. For example, Shackleton (2007) uses cluster analysis and principal component analysis (PCA) to identify linguistic variables which show a specific geographic distribution, while Grieve et al. (2011) uses spatial autocorrelation to detect significant geographical patterns in forty individual lexical alternation variables. Prokić et al. (2012) examine each item in a dataset, seeking those that differ minimally with a candidate area and maximally with respect to sites outside the area. In this study, however, we use hierarchical bipartite spectral graph partitioning (BiSGP), which allows a simultaneous identification of geographical areas together with their characteristic linguistic features. This approach has been successfully used to obtain the linguistic basis (in terms of sound correspondences) with respect to a certain reference pronunciation for Dutch (Wieling et al., 2010), English (Wieling et al., submitted) and Tuscan (Montemagni et al., forthcoming) dialect datasets. In contrast to analyzing pronunciation data, however, we investigate the use of specific lexical variants (per concept variable) from Voices data. Dataset The BBC Voices data contains a total of 38 concepts which are shown in Table 1 below. 1. Hot 2. Cold 3. Tired 4. Unwell 5. Pleased 6. Annoyed 7. Play a game 8. Play truant 9. Throw 10. Hit hard 11. Sleep 12. Drunk 13. Pregnant 14. Left-handed 15. Lacking money 16. Rich 17. Insane 18. Attractive 19. Unattractive 20. Moody 21. Baby 22. Mother 23. Grandmother 24. Grandfather 25. Friend 26. Male partner 27. Female partner 28. Young person in cheap trendy clothes and jewelry 29. Clothes 30. Trousers 31. Child’s soft shoes worn for PE 32. Main room of house (with TV) 33. Long, soft seat in the main room 34. Toilet 35. Narrow walkway alongside buildings 36. To rain lightly 37. To rain heavily 38. Running water smaller than a river Table 1: List of all 38 concepts in the BBC Voices dataset The complete dataset contains (on average) 19,326 responses per concept. We only include responses from the online questionnaire as the responses on the (identical) paper questionnaire have not been digitized. As a consequence of paper copies not being included, the average age of the people is relatively low (about 33) and more than sixty percent of the people were aged below thirty. A total of 57.3 percent of the participants were female. The responses were lemmatized in order to abstract away from variation in spelling. For example, ‘skive’, ‘scaive’, ‘scive’ (for the concept PLAY TRUANT) were grouped together. To simplify the data somewhat, we only select the top-ten variants for every concept (on average containing 84 percent of all responses). We group the responses by postcode area (there are a total of 121 UK postcodes) and for every (lemmatized) variant we calculate the percentage of people in the postcode area using this variant. Our input data thus consists of a table with 121 rows (the postcode areas) and 380 columns (38 concepts having 10 variants each) containing these percentages.