📄 Computer-Produced Representation of Dialectal Variation: Initial Fricatives in Southern British English

Summary

The authors wanted to show that dialect geography is well suited to computers. They chose a test problem large enough to be informative but quick to prepare: the voicing of word-initial fricatives (f-, s-, th-, sh-) in the south of England. The data came from the published Survey of English Dialects (SED). They took 68 words across the 75 SED localities in the ten southernmost counties and used only the first recorded form at each locality. Each narrow IPA transcription went into a simplified phonetic code on punch cards, giving 5100 items. The cards were sorted in four ways to support diagnosis. The team also produced frequency tables, one map per word and proportional maps drawn on a CalComp plotter. The results show strong variation from word to word and place to place, which supports the maxim that every word has its own history. They also suggest that the traditional voiced-fricative area should include western Hampshire and most of Cornwall. The broader point is that automated sorting and mapping can replace slow, error-prone hand cartography.

Key points

• The corpus has 68 words: 27 f-words, 22 s-words, 16 th-words and 3 sh-words. These were recorded at 75 localities in 10 counties, from 31 Somersetshire to 40 Sussex, giving 5100 items. The 59 missing responses were coded XXX.
• Data entry took roughly a dozen hours of investigator time and a similar amount of key-punching. Each card carries a locality code, the SED item reference and the coded pronunciation.
• Four sorted lists were produced:
  • by keyword, then locality;
  • by locality, then keyword;
  • by locality, then citation form;
  • by the sounds after the initial consonant (in articulatory order), then by the initial consonant.
• Outputs included per-set tables ranking words by voiced initials, counts of following vowels and consonants per initial consonant, 68 single-word maps and four voiceless:voiced ratio maps.
• The map outlines were captured with a digitizer; the full 311-locality SED map needed about 4000 x/y measurements. A PL/I program prepared the data and a FORTRAN program drove the plotter using a simple instruction format: map number, county, locality, message.

Interesting ideas

• I1. The coding scheme kept the phonetically significant distinctions and dropped fine detail that seemed irrelevant, so the data became quick to sort without losing what mattered.
• I2. One dataset was re-sorted along several keys, and the sorted lists served as cheap diagnostic views before any counting or mapping.
• I3. Sorting by the sounds after the target segment exposes whether the surrounding context conditions a change, instead of assuming it does.
• I4. Taking only the first, citation-style response per locality and marking non-responses explicitly gave a uniform grid that could be compared cell by cell.
• I5. The plotter took a generic stream of "map, place, message" records, which kept the linguistic analysis separate from rendering.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (tokens): Use a similar reduced feature code for normalizing spelling or dialect variants in the en crate, while keeping the original form so the mapping stays explainable.
• I2 (learn): Dump parser disagreements sorted by lemma, by relation and by the following token to quickly see where errors cluster before writing repair rules.
• I3 (tree): Group attachment errors by their right-hand context (next POS or word) to test whether a parsing mistake depends on context or belongs to a particular lemma.
• I4 (learn): Build evaluation sets with one canonical item per cell and explicit missing markers, so coverage gaps are not silently counted as errors.
• I5 (back): Have the graph-to-text generator consume a neutral instruction stream from the meaning graph, so new output formats can be added without touching the semantics.