📄 Classifying Graphemes in English Words Through the Application of a Fuzzy Inference System

Summary

The paper looks at how to segment English words into graphemes, the written units that stand for phonemes. The authors see this as a level between character-based and word-based analysis. English spelling is messy: one letter sequence can be a grapheme by itself or part of a longer one, and the right split can shift with accent. So the authors treat a word's grapheme count as a fuzzy quantity rather than a crisp label. They annotated the Google 10,000 word list by hand, using England's National Curriculum phonics rules, and kept the EnglishGrammarHere 10,000 list for validation. From the annotated data they computed Gaussian statistics for word length, vowel count and consonant count at each grapheme count. These statistics feed a Mamdani Fuzzy Inference System that predicts how many graphemes a word has. A greedy longest-first matcher over a 284-grapheme inventory then produces the actual split. The predicted count is exact for about half the training words and within one for over 93%. Compared with a baseline that maps IPA from a pronunciation dictionary, the fuzzy pipeline counts less accurately but more often produces the correct split. It also needs no dictionary entry, so it can handle misspelled words.

Key points

• The grapheme inventory has 284 units taken from Brooks' Dictionary of the British English Spelling System: 89 in the "main system" plus 195 rarer ones. English has 44 phonemes.
• The FIS has three inputs (word length, vowels, consonants) with 13 sets per input. Sets for counts 2–12 are Gaussian, and the edge cases (1 and 14) use Z- and S-shaped curves. Outputs are triangular sets aligned to rounding boundaries, there are 13 one-to-one rules, and defuzzification uses the centroid method.
• Count prediction on Google 10,000 gave:
  • 50.18% exact;
  • 26.73% one too high;
  • 16.6% one too low;
  • 6.49% off by more than one (93.51% within ±1).
• On count accuracy, the IPA mapping baseline scored 69.27% (training) and 71.20% (validation), versus 50.18% and 55.68% for the fuzzy pipeline.
• On correct split given a correct count, the fuzzy pipeline scored 62.16% and 59.93%, versus 36.41% and 38.94% for IPA mapping.

Interesting ideas

• I1. Predict the number of units first and then constrain a greedy longest-match segmenter to hit that count, retrying when the match fails; the count acts as a cheap global prior on a local search.
• I2. Score predictions both exactly and with a ±1 tolerance, since for inherently variable targets (here, dialect-dependent graphemes) near-misses carry information that exact match hides.
• I3. Build interpretable membership functions straight from per-class means and standard deviations of simple features, so each rule can be read and checked by hand.
• I4. Keep a dictionary-free method next to a dictionary-based one, because the lookup approach fails on typos and unseen words while the rule-based one still produces output.
• I5. Report two separate metrics for the two pipeline stages (getting the count right, then getting the split right), which showed that the method better on stage one was worse on stage two.

For Mova

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

• I1 (tokens): a predicted morpheme or subpart count could guide splitting of unknown compounds and derived words in crate en, with the count guess logged so the split stays explainable.
• I2 (learn): evaluations of tasks with fuzzy gold, such as syllable counts or event-order distance, could report exact and tolerance-band accuracy side by side to separate real errors from annotation variance.
• I3 (tokens): transparent Gaussian-feature scores could serve as an explainable fallback signal for out-of-vocabulary handling in the tagger.
• I4 (tokens): pair lexicon lookup with a dictionary-independent analyzer, so misspelled or novel words in school texts still get a morphology/lemma guess rather than a dead end.
• I5 (learn): score each step of text → tree → frames separately against negative controls, so a component that wins upstream but hurts downstream is caught.