Boostrapping phonetic lexicons for new languages

Sameer Maskey, Alan W. Black, Laura Tomokiya2004 · DOI

TL;DR

This paper introduces a novel "bootstrapping" algorithm to automatically build phonetic lexicons for new languages, significantly reducing the manual effort typically required. By leveraging large amounts of unlabeled text, a small set of "seed words," and an active learning approach with native speaker feedback, the method efficiently creates accurate phonetic dictionaries.

What Problem Does This Paper Address?

Speech applications like synthesis and recognition heavily rely on phonetic lexicons, which map words to their pronunciations. However, creating these lexicons for new or less-resourced languages is a time-consuming and labor-intensive process. Existing methods often require extensive manual annotation or assume the availability of a pre-existing phonetic lexicon, creating a barrier for broader language coverage. This paper tackles the challenge of efficiently constructing these essential resources for languages lacking them.

Methods

The core of the paper is a bootstrapping algorithm that iteratively builds and refines a phonetic lexicon. It starts with a small set of "seed words" with known pronunciations and uses these to train an initial Letter-to-Sound (LTS) rule system. The algorithm then enters an active learning loop:

  1. Word Selection: It identifies the most frequent words in a large corpus of unlabeled text that are not yet in the lexicon.
  2. Pronunciation Prediction: It uses the current LTS rules to predict pronunciations for these new words.
  3. Confidence Scoring: Two confidence scores are computed:
  4. Orthographic Similarity (Co): Measures how similar the new word's spelling is to words already in the lexicon using Longest Common Subsequence (LCS).
  5. Phonetic Similarity (Cp): Assesses how similar the predicted pronunciation is to the pronunciations of orthographically similar words, using a variant of the DICE phonetic distance measure.
  6. Automatic Lexicon Expansion: If both confidence scores exceed predefined thresholds, the word and its predicted pronunciation are automatically added to the lexicon.
  7. Manual Verification: For words that don't meet the confidence thresholds, a native speaker is asked to verify and correct the predicted pronunciation.
  8. LTS Rule Retraining: The lexicon and LTS rules are updated with the newly added (and corrected) words, and the process repeats until a desired accuracy is achieved.

Datasets

The authors experimented with three languages:

  • Nepali: Over 1 million words of text from daily newspapers. No pre-existing phonetic lexicon was available.
  • German: A large volume of text from online newspaper sources. The CELEX lexicon was used as an oracle for evaluation and correction.
  • English: 50 million words of Wall Street Journal (WSJ) text. The Carnegie Mellon Pronouncing Dictionary (CMUDICT) was used as an oracle.

For evaluation, they created three types of test sets: TOKENS (held-out text), TYPES (unique words from TOKENS), and TYPES_PLUS (unique words occurring more than once).

Key Insights and Ideas

  • Active Learning for Lexicon Building: The paper demonstrates that active learning principles are highly effective for lexicon construction. By strategically selecting words that are most likely to improve the LTS rules (i.e., frequent words), the algorithm efficiently expands the lexicon.
  • Dual Confidence Scoring: Combining orthographic and phonetic similarity measures provides a robust way to automatically accept new pronunciations. This two-tiered approach helps filter out potentially incorrect predictions before they are added to the lexicon.
  • Bootstrapping Efficiency: The iterative bootstrapping process allows the system to leverage its own growing knowledge. As the lexicon and LTS rules improve, they become better at predicting pronunciations for new words, creating a positive feedback loop.
  • Language Opacity Matters: The paper highlights that languages with a more direct relationship between orthography and pronunciation (like Nepali) are easier to build lexicons for, but the method is still effective for more opaque languages like English and German.
  • Smaller Lexicons, Higher Accuracy: The bootstrapped English lexicon achieved high accuracy with significantly fewer entries than an existing lexicon (CMUDICT), suggesting that the incremental, targeted approach is more efficient in terms of lexicon size.

Key Results

  • English: Achieved over 90% accuracy on the TOKENS test set within 18 iterations and approximately 4000 lexicon entries. A final lexicon of 16,000 entries reached 94.31% accuracy on TOKENS and 87.23% on TYPES, being about one-seventh the size of CMUDICT.
  • Nepali: Achieved 91.47% precision for confidence metrics (Co and Cp). After 16 iterations, pronunciation accuracy reached 92.9% on new data, and for tokens not in the lexicon, accuracy improved from 69.8% to 94.6%.
  • German: With a lexicon of only 350 words, achieved 35% accuracy on the TOKENS test set. After 13 iterations, accuracy reached 90% on TOKENS, 80.67% on TYPES, and 87.04% on TYPES_PLUS.
  • Confidence Score Impact: On average, confidence scores automatically added 9.92% of test words in English. The average precision for the confidence metric was 87.79%.

Limitations and Future Work

  • Native Speaker Dependency: While the algorithm reduces manual effort, it still relies on native speakers for verification, especially in early iterations or for words that don't meet confidence thresholds.
  • Language-Specific Tuning: The thresholds for confidence scores (Co and Cp) may need to be tuned for different languages.
  • Future Work: The authors suggest that the method can be extended to build lexicons for languages with even more complex orthographic-phonetic relationships. They also imply that further refinement of confidence metrics and active learning strategies could further reduce manual annotation.

Summary updated

Abstract

Although phonetic lexicons are critical for many speech applications, the process of building one for a new language can take a significant amount of time and effort. We present a bootstrapping algorithm to build phonetic lexicons for new languages. Our method relies on a large amount of unlabeled text, a small set of ’seed words’ with their phonetic transcription, and the proficiency of a native speaker in correctly inspecting the generated pronunciations of the words. The method proceeds by automatically building Letter-to-Sound (LTS) rules from a small set of the most commonly occurring words in a large corpus of a given language. These LTS rules are retrained as new words are added to the lexicon in an Active Learning step. This procedure is repeated until we have a lexicon that can predict the pronunciation of any word in the target language with the accuracy desired. We tested our approach for three languages: English,
BibTeX
@misc{maskey2004boostrapping,
  author = {Sameer Maskey and Alan W. Black and Laura Tomokiya},
  title = {{Boostrapping phonetic lexicons for new languages}},
  year = {2004},
  doi = {10.21437/interspeech.2004-62},
}
Figure