Automatic Generation of English Respellings

Bradley Hauer, Grzegorz KondrakNAACL 2013

TL;DR

This paper tackles the challenge of automatically generating "respellings" – intuitive, alphabet-based spellings that convey word pronunciation. They introduce a novel system combining machine learning with linguistic rules, achieving respellings that are more understandable and correct than those found online, and approaching expert quality.

What Problem Does This Paper Address?

Pronouncing unfamiliar words, especially foreign names or technical terms, can be a hurdle. While phonetic transcriptions (like IPA) exist, they require specialized knowledge. "Respellings" offer a user-friendly alternative, using standard English letters to guide pronunciation. However, creating good respellings is difficult; those found online are often ambiguous or incorrect. This paper aims to automate the generation of high-quality respellings from a word's phonemic representation, bridging the gap between technical phonetic notation and everyday understanding.

Methods

The authors propose a multi-stage system for generating respellings:

  • Syllabification: They first break down the phonemic representation into syllables, a crucial step for managing pronunciation. They employ an "overgenerate-and-rank" approach, considering multiple syllabification possibilities and then ranking them based on linguistic constraints (like sonority principles and vowel rules).
  • Candidate Generation: Two main modules generate respelling candidates for each syllable:
  • Trained Phoneme-to-Letter (P2L) Model: A machine learning model trained on a pronunciation dictionary to learn common phoneme-to-letter mappings.
  • Context-Sensitive Respeller: A hand-crafted module that uses linguistic rules and considers neighboring phonemes and syllable structure to propose more contextually appropriate respellings.
  • Candidate Selection (Filtering and Scoring):
    • L2P Correctness Filter: A letter-to-phoneme (L2P) model is used to "translate" the generated respelling back into phonemes. If the resulting phonemes don't match the original intended pronunciation, the respelling is rejected.
  • Vowel Counter: A simple filter to remove potentially confusing syllables with multiple vowel groups.
  • SVM Ambiguity Classifier: A Support Vector Machine (SVM) is trained to predict the ambiguity of a respelling syllable, helping to select the clearest options.
  • Lexical Reviser: As a final step, this module attempts to improve the top-ranked respelling by replacing parts with actual English words where appropriate (e.g., changing "surr-sin-uss" to "sir-sin-us" for "circinus").

Datasets

  • Training Data:
    • A pronunciation dictionary (Combilex) was used to extract monosyllabic words and their pronunciations for training the P2L model and the L2P correctness filter.
  • Web-derived corpora (from Ghoshal et al., 2009) were used to train the SVM ambiguity classifier, extracting word-respelling pairs.
  • Evaluation Data:
    • Fraser's Set: 27 words from Helen Fraser's (1997) work, considered "expert" respellings and serving as an upper bound for quality.
  • Web Set: 231 words extracted from the Ghoshal et al. (2009) corpus, filtered for phonetic transcriptions from Wikipedia and matching Soundex codes, and containing at least eight phonemes.

Key Insights and Ideas

  • Respellings as a Bridge: The paper strongly advocates for respellings as a practical, accessible way to convey pronunciation, overcoming the barrier of technical phonetic alphabets.
  • Syllable-Level Processing is Key: Breaking down words into syllables is fundamental for both generating and evaluating respellings, simplifying the task and improving clarity.
  • Combining ML and Rules Works: The success of the system hinges on the synergy between data-driven machine learning (P2L model, SVM) and linguistically informed, hand-crafted rules (context-sensitive respeller, syllabification constraints).
  • Ambiguity is Inherent, but Manageable: English orthography is inherently ambiguous. The system acknowledges this by explicitly modeling and filtering for ambiguity, rather than trying to eliminate it entirely.
  • Automated Evaluation is Powerful: Using an L2P converter (like eSpeak) as an automated evaluator provides a scalable and objective way to assess respelling quality, complementing human judgment.

Key Results

  • Human Judgment: The system's generated respellings significantly outperformed both baseline (phonemic respelling) and Web-sourced respellings in terms of unambiguity and correctness. They were also more likely to be preferred by human evaluators. On the "expert" Fraser's set, the system's quality was much closer to the expert baseline than the simple baseline.
  • Automated Appraisal (using eSpeak): The system dramatically boosted phoneme accuracy (to over 90%) when evaluated by eSpeak, significantly outperforming baseline and Web respellings, and even surpassing expert respellings.
  • Ablation Study: Removing the P2L generator or the L2P correctness filter led to substantial drops in accuracy, highlighting their importance. Removing syllable breaks had a particularly detrimental effect.

Limitations and Future Work

  • Limitations:
    • The system's P2L model is trained on a relatively small set of syllables, which might limit its coverage.
  • The L2P correctness filter is described as being more conservative, potentially rejecting some valid respellings.
  • The paper acknowledges that even expert respellings can be sub-optimal or require labor-intensive creation.
  • Future Work:
    • Developing methods to improve existing respellings.
  • Assisting writers in creating respellings without direct access to phonemic representations.
  • Exploring the impact of stress indicators on respelling quality.
  • Potentially incorporating more sophisticated linguistic knowledge into the generation and evaluation processes.

Summary updated

BibTeX
@inproceedings{hauer2013automatic,
  author = {Bradley Hauer and Grzegorz Kondrak},
  title = {{Automatic Generation of English Respellings}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2013},
}
Figure