---
title: "Pronunciation Modeling in Spelling Correction for Writers of"
description: "Boyd studies English learners whose native language is Japanese (abbreviated JWEFL) and builds on… 9 references, cited by 9 papers here."
authors: ["Adriane Boyd"]
year: 2009
venue: "Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, Companion Volume: Student Research Workshop and Doctoral Consortium"
source: "https://aclanthology.org/N09-3006/"
licence: "CC BY-NC-SA 3.0 (ACL Anthology, before 2016)"
licence_url: "https://creativecommons.org/licenses/by-nc-sa/3.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/boyd-2009-pronunciation-modeling-spelling-correction-writers/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/boyd-2009-pronunciation-modeling-spelling-correction-writers/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2009](https://md.research.elamur.ai/free/en/papers/2009/page/2/) / Pronunciation Modeling in Spelling Correction for Writers of English as a Foreign Language

# Pronunciation Modeling in Spelling Correction for Writers of English as a Foreign Language

**Authors:** [Adriane Boyd](https://md.research.elamur.ai/free/en/authors/adriane-boyd/) · **Year:** 2009 · **Venue:** [Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, Companion Volume: Student Research Workshop and Doctoral Consortium](https://md.research.elamur.ai/free/en/venues/human-language-technologies-the-annual-conference-of-the-north-american-chapter/) ([Human Language Technologies: The Annual Conference of the North American Chapter of the Association for Computational Linguistics, Companion Volume: Student Research Workshop and Doctoral Consortium 2009](https://md.research.elamur.ai/free/en/venues/human-language-technologies-the-annual-conference-of-the-north-american-chapter/2009/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY-NC-SA 3.0 (ACL Anthology, before 2016)](https://creativecommons.org/licenses/by-nc-sa/3.0/)

**Links:** [Publisher's page](https://aclanthology.org/N09-3006/) · [PDF](https://aclanthology.org/N09-3006.pdf) · [OpenAlex](https://openalex.org/W2026584864)

**Citations:** 12 citations (OpenAlex, 2026-10-14) · [9](https://md.research.elamur.ai/free/en/papers/boyd-2009-pronunciation-modeling-spelling-correction-writers/cited-by/) in our library, of the 149,439 papers whose bibliographies we read

## Summary
Standard spell checkers are built for native writers. They often miss the right fix for learners whose errors come from their first-language phonology. Boyd studies English learners whose native language is Japanese (abbreviated JWEFL) and builds on the noisy-channel corrector of Toutanova and Moore (2002), which combines a letter-based error model with a pronunciation-based one.

The new step models how Japanese speakers actually pronounce English. A monophone recognizer trained on native speech (TIMIT) is adapted to the English Read by Japanese corpus (ERJ), which has no phonetic transcription. The observed phone variation is used to enlarge the CMUDICT pronouncing dictionary: each word gets its n-best variant pronunciations, generated with OpenFst.

The models are trained and tested on misspellings from the Atsuo-Henry Corpus. The letter model and the combined model clearly beat GNU Aspell. Adding the pronunciation variants significantly improves the pronunciation-only model (PHL), but the combined model does not gain. The work shows that non-native phonology can be learned from untranscribed speech and fed into error correction.

## Key points
- **Data.** The Atsuo-Henry Corpus was cleaned to 4,769 unique misspellings of 1,046 target words and split 80/10/10. The word list is SCOWL size 50, filtered to CMUDICT entries, giving 54,001 words. With the top five variants per word, it grows to 255,827 unique pronunciations.
- **Letter-to-phone model.** This is the n-gram LTP model of Fisher (1999), extended to use the top M most specific contexts. With M = 3 it reaches 95.5% phone accuracy and 74.9% word accuracy on held-out CMUDICT data.
- **1-Best accuracy.** Aspell scores 44.1%. The letter model (L) scores 64.7%. PHL rises from 47.9% to 50.6% with variants. The combined model (CMB) moves from 64.9% to 65.5%, which is not significant.
- **Significance.** PHL with variants is significantly better (p\<0.05) at 1-Best, 3-Best and 4-Best.
- **Tuning.** The tuned settings are N = 3 for letters, N = 4 for phones, minimum identity probability m = 80%, and combination weight λ = 0.15. The word prior P(w) is held uniform.

## Interesting ideas
- Instead of changing the matcher, enlarge a canonical lexicon with population-specific variants: the n-best paths through per-phone probability FSTs are added as extra entries. This keeps the downstream algorithm unchanged and makes each variant inspectable.
- When training data contains only error pairs, identity edits look unrealistically rare, so a floor probability m is reserved for "no change". This corrects the bias that comes from sampling errors only.
- Two independent views of the same error, spelling and sound, are merged by a weighted sum of log scores, with a small weight tuned on dev data. This yields an interpretable two-part score instead of one opaque model.
- A component that improves significantly on its own may add nothing once combined with a stronger model. This argues for always measuring gains inside the full pipeline.
- Fixing the word prior P(w) as uniform isolates the error model being studied. The paper also reports n-best accuracy and shows ranked candidate lists for one misspelling.

## For Mova

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

- **I1** (tokens): Learner or dialect spellings could be mapped to lemmas through an explicit table of generated variants, with each variant's origin recorded for explanation.
- **I2** (learn): Self-induced repair rules learned only from failures could get a similar floor, so the system does not over-edit tokens that were already correct.
- **I3** (learn): Witness parsers or repair signals could be combined with explicit log-weights, so every decision can be split into each witness's contribution.
- **I4** (learn): Evaluation should report each module's improvement both in isolation and end-to-end, so that locally nice but globally redundant additions are flagged.
- **I5** (answer): QA evaluations could neutralize frequency priors as a control, so that improvements can be credited to the reasoning step rather than to answer popularity.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

## Institutions

- [The Ohio State University](https://md.research.elamur.ai/free/en/universities/the-ohio-state-university/) — University, Columbus, United States · 90 works · 19,112 citations

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 17,450 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 17,296 works

🧭 Topics (2): [English](https://md.research.elamur.ai/free/en/topics/english/) · 2,379 works · [🖍️ Grammatical error correction & learner language](https://md.research.elamur.ai/free/en/topics/grammatical-error-correction/) · 162 works

## Cite

Cite as: Adriane Boyd. 2009. Pronunciation Modeling in Spelling Correction for Writers of English as a Foreign Language. Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, Companion Volume: Student Research Workshop and Doctoral Consortium.

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

```bibtex
@inproceedings{boyd2009pronunciation,
  title = {{Pronunciation Modeling in Spelling Correction for Writers of English as a Foreign Language}},
  author = {Adriane Boyd},
  booktitle = {Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, Companion Volume: Student Research Workshop and Doctoral Consortium},
  year = {2009},
  doi = {10.3115/1620932.1620938},
  url = {https://doi.org/10.3115/1620932.1620938}
}
```

- [← previous Aspect and the Biblical Hebrew Niphal and Hitpael](https://md.research.elamur.ai/free/en/papers/benton-2009-aspect-biblical-hebrew-niphal-hitpael/)
- [next → A Probabilistic Model for Associative Anaphora Resolution](https://md.research.elamur.ai/free/en/papers/sasano-2009-probabilistic-model-associative-anaphora-resolution/)
- [↑ back to papers of 2009](https://md.research.elamur.ai/free/en/papers/2009/page/2/)

## References

9 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Brill, Eric and Robert C. Moore (2000). [An Improved Error Model for Noisy Channel Spelling Correction](https://md.research.elamur.ai/free/en/papers/brill-2000-improved-error-model-noisy-channel/). In Proceedings of ACL 2000.
2. CMUDICT (1998). CMU Pronouncing Dictionary version 0.6. [http://www.speech.cs.cmu.edu/cgi-bin/cmudict](http://www.speech.cs.cmu.edu/cgi-bin/cmudict). [↗](http://www.speech.cs.cmu.edu/cgi-bin/cmudict)
3. Fisher, Willam (1999). A statistical text-to-phone function using ngrams and rules. In Proceedings of ICASSP 1999. [find](https://scholar.google.com/scholar?q=Fisher%2C+Willam+%281999%29%2E+A+statistical+text%2Dto%2Dphone+function+using+ngrams+and+rules%2E+In+Proceedings+of+ICASSP+1999%2E)
4. Kukich, Karen (1992). Technique for automatically correcting words in text. ACM Computing Surveys 24(4). [find](https://scholar.google.com/scholar?q=Kukich%2C+Karen+%281992%29%2E+Technique+for+automatically+correcting+words+in+text%2E+ACM+Computing+Surveys+24%284%29%2E)
5. Minematsu, N., Y. Tomiyama, K. Yoshimoto, K. Shimizu, S. Nakagawa, M. Dantsuji, and S. Makino (2002). English Speech Database Read by Japanese Learners for CALL System Development. In Proceedings of LREC 2002. [find](https://scholar.google.com/scholar?q=Minematsu%2C+N%2E%2C+Y%2E+Tomiyama%2C+K%2E+Yoshimoto%2C+K%2E+Shimizu%2C+S%2E+Nakagawa%2C+M%2E+Dantsuji%2C+and+S%2E+Makino+%282002%29%2E+English+Speech+Database+Read+by+Japanese+Learners+for+CALL+System+Development%2E+In+Proceedings+of+L)
6. Mitton, Roger and Takeshi Okada (2007). The adaptation of an English spellchecker for Japanese writers. In Symposium on Second Language Writing. [find](https://scholar.google.com/scholar?q=Mitton%2C+Roger+and+Takeshi+Okada+%282007%29%2E+The+adaptation+of+an+English+spellchecker+for+Japanese+writers%2E+In+Symposium+on+Second+Language+Writing%2E)
7. Okada, Takeshi (2004). A Corpus Analysis of Spelling Errors Made by Japanese EFL Writers. Yamagata English Studies 9. [find](https://scholar.google.com/scholar?q=Okada%2C+Takeshi+%282004%29%2E+A+Corpus+Analysis+of+Spelling+Errors+Made+by+Japanese+EFL+Writers%2E+Yamagata+English+Studies+9%2E)
8. TIMIT (1991). TIMIT Acoustic-Phonetic Continuous Speech Corpus. NIST Speech Disc CD1-1.1. [find](https://scholar.google.com/scholar?q=TIMIT+%281991%29%2E+TIMIT+Acoustic%2DPhonetic+Continuous+Speech+Corpus%2E+NIST+Speech+Disc+CD1%2D1%2E1%2E)
9. Toutanova, Kristina and Robert Moore (2002). [Pronunciation Modeling for Improved Spelling Correction](https://md.research.elamur.ai/free/en/papers/toutanova-2002-pronunciation-modeling-improved-spelling-correction/). In Proceedings of ACL 2002.

*Updated 31 October 2026*