Automatic Generation of English Respellings

What it is about

A respelling is an alternative spelling that tells a reader how to pronounce a word, as in "wah-tool-koh". The task is to generate one from a phoneme sequence. The system overgenerates syllabifications, filtered by sonority and lax-vowel constraints that are relaxed in a fixed order when nothing survives. Candidate syllables come from a phoneme-to-letter model trained on 4,215 monosyllabic words and from a hand-written context-sensitive respeller. A letter-to-phoneme model rejects candidates that do not read back correctly. A vowel counter and an SVM over letter n-grams (1 to 5) scored the ambiguity of each syllable, and a dictionary reviser can swap in a real word. On 231 words from the Web, the eSpeak synthesiser reads the generated respellings with 93.0% phoneme accuracy, against 76.2% for the original spelling. In a 20-judge study, 70.0% of the system's respellings were judged unambiguous, and 41.3% unambiguous and correct, against 32.6% for the Web respellings and 25.5% for the baseline. An ablation shows that removing the letter-to-phoneme filter almost doubles the phoneme error rate.

Ideas

• I1. 05. The writing system: a closed table of respelling rules per phoneme (the context-free mapping, plus the context-sensitive variants) as explicit rules with their reason, so each respelling can say which rule produced each syllable.
• I2. The ambiguity score as a count, not an SVM: for each letter n-gram in a respelling, count how many pronunciations it has in a dictionary, and reject a syllable whose n-grams have more than one reading. Deterministic and explainable; measure its agreement with the judges' decisions on the paper's examples.
• I3. The round-trip check as a negative control for spelling output: a generated spelling must read back to the same sound through a letter-to-phoneme rule set, and the rate of failures must be reported with the set of failing rules.
• I4. Constraint relaxation in a fixed order (sonority, lax vowels in open syllables) as an explicit schedule in code, with the relaxation step recorded per output.

For Mova

The chapters are 05. The writing system and 06. Spelling rules: inflection, derivation, compounds and national standards. The code points are en/src/tok.rs (syllable boundaries if they are added) and en/src/morph.rs (rule tables). Nothing here is a phonology module yet, so the first step is a data question: a pronunciation dictionary whose licence allows use. The measure: round-trip accuracy by rule on a held-out word list, plus the ablation of the reading filter, run twice for determinism.

Cite

The paper itself, not this card. BibTeX from the library's record of the work.

@inproceedings{hauer2013automatic,
  title = {{Automatic Generation of English Respellings}},
  author = {Hauer, Bradley and Kondrak, Grzegorz},
  booktitle = {Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies},
  year = {2013},
  url = {https://aclanthology.org/N13-1072/}
}

References

19 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Susan Bartlett, Grzegorz Kondrak, and Colin Cherry. 2009. On the syllabification of phonemes. In Proc. of HLT-NAACL, pages 308–316. find
2. Maximilian Bisani and Hermann Ney. 2008. Joint-sequence models for grapheme-to-phoneme conversion. Speech Communication, 50(5):434–451. find
3. Doğan Can, Erica Cooper, Arnab Ghoshal, Martin Jansche, Sanjeev Khudanpur, Bhuvana Ramabhadran, Michael Riley, Murat Saraçlar, Abhinav Sethy, Morgan Ulinski, and Christopher White. 2009. Web derived pronunciations for spoken term detection. In Proc. of ACM SIGIR, pages 83–90. find
4. Edward Carney. 1994. A Survey of English Spelling. Routledge. find
5. Noam Chomsky and Morris Halle. 1968. The Sound Pattern of English. New York: Harper & Row. find
6. Helen Fraser. 1997. Dictionary pronunciation guides for English. International Journal of Lexicography, 10(3):181–208. find
7. Lucian Galescu. 2009. Extending pronunciation lexicons via non-phonemic respellings. In Proc. of HLT-NAACL: Short Papers, pages 129–132. find
8. Arnab Ghoshal, Martin Jansche, Sanjeev Khudanpur, Michael Riley, and Morgan Ulinski. 2009. Web-derived pronunciations. In Proc. of ICASSP, pages 4289–4292. find
9. Sittichai Jiampojamarn and Grzegorz Kondrak. 2010. Letter-phoneme alignment: An exploration. In Proc. of ACL, pages 780–788. find
10. Sittichai Jiampojamarn, Grzegorz Kondrak, and Tarek Sherif. 2007. Applying many-to-many alignments and hidden Markov models to letter-to-phoneme conversion. In Proc. of HLT-NAACL, pages 372–379. find
11. Sittichai Jiampojamarn, Colin Cherry, and Grzegorz Kondrak. 2008. Joint processing and discriminative training for letter-to-phoneme conversion. In Proc. of ACL, pages 905–913. find
12. Thorsten Joachims. 1999. Making large-scale SVM learning practical. In B. Schalkopf, C. Burges, and A. Smola, editors, Advances in Kernel Methods - Support Vector Learning. MIT Press. find
13. Michael Kenstowicz. 1994. Phonology in Generative Grammar. Blackwell. find
14. John Kominek and Alan W Black. 2006. Learning pronunciation dictionaries: Language complexity and word selection strategies. In Proc. of HLT-NAACL, pages 232–239. find
15. Grzegorz Kondrak. 2000. A new algorithm for the alignment of phonetic sequences. In Proc. of NAACL, pages 288–295. find
16. John Noble. 2012. Mexico. Lonely Planet, 13th edition. find
17. Korin Richmond, Robert Clark, and Sue Fitt. 2009. Robust LTS rules with the Combilex speech technology lexicon. In Proc. of Interspeech, pages 1295–1298. find
18. Henry Rogers. 2000. The Sounds of Language. Pearson. find
19. Briony Williams and Rhys James Jones. 2008. Acquiring pronunciation data for a placenames lexicon in a less-resourced language. In Proc. of LREC. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/hauer-2013-automatic-generation-english-respellings/