Modeling Letter-to-Phoneme Conversion as a Phrase Based Statistical Machine Translation Problem with Minimum Error Rate Training

Taraka Rama, Anil Kumar Singh, Sudheer KolachinaNAACL 2009

TL;DR

This paper tackles the challenging task of converting written text to its phonetic pronunciation by framing it as a Statistical Machine Translation (SMT) problem. By leveraging phrase-based SMT techniques and optimizing with Minimum Error Rate Training (MERT), the authors achieve state-of-the-art results on several languages, demonstrating a novel and effective approach to this fundamental NLP problem.

What Problem Does This Paper Address?

The conversion of written text (orthography) to its phonetic representation (phonemes), known as Letter-to-Phoneme (L2P) conversion, is crucial for many speech and language technologies like spell checkers, speech synthesis, and transliteration. However, this task is inherently complex due to the irregular and often many-to-many mapping between letters and phonemes, especially in languages like English. While rule-based systems are impractical due to the sheer number of rules required, previous machine learning approaches often relied on simpler one-to-one mappings or complex, handcrafted features. This paper aims to bridge this gap by proposing a more robust and language-independent L2P system using established SMT frameworks.

Methods

The core innovation of this paper lies in modeling L2P conversion as a phrase-based SMT problem. This means treating sequences of letters as "source language" phrases and sequences of phonemes as "target language" phrases. The authors utilize the noisy channel model, where the goal is to find the most probable phoneme sequence ($f$) given a letter sequence ($l$), expressed as $P(f|l) \propto P(l|f)P(f)$.

Key methodological components include:

  • Phrase-Based SMT Framework: Leveraging existing SMT toolkits and concepts, where the system learns to translate letter "phrases" to phoneme "phrases."
  • Letter-to-Phoneme Alignment: Using GIZA++ (a standard SMT alignment tool) to align letters with phonemes in training data. Crucially, they treat each letter as a "word" in the source side and each phoneme as a "word" in the target side, enabling the learning of many-to-many alignments.
  • Minimum Error Rate Training (MERT): Instead of relying on standard SMT training objectives, the authors employ MERT to directly optimize for word accuracy on a development set. This is a novel application of MERT to the L2P task.
  • Log-Linear Model: The posterior probability $P(f|l)$ is modeled using a log-linear framework with feature functions, allowing for flexible incorporation of various linguistic cues.
  • A* Beam Search Decoder: Used for efficient decoding to find the best phoneme sequence.

Datasets

The authors evaluated their models on publicly available speech dictionaries for four languages:

  • English: CMUDict
  • French: Brulex
  • German: Celex
  • Dutch: Celex

The datasets were divided into 10 folds for cross-validation, with one fold for testing, one for parameter development, and the remaining eight for training.

Key Insights and Ideas

  • SMT as a Powerful L2P Framework: The paper convincingly demonstrates that established SMT techniques, when adapted, can effectively handle the complexities of L2P conversion, outperforming previous methods.
  • MERT for L2P Optimization: Applying MERT directly to optimize for word accuracy is a significant contribution, as it directly targets the desired evaluation metric, leading to better performance than standard SMT training objectives.
  • Language-Independent Approach: The SMT framework inherently allows for a language-independent approach, as the learning process is data-driven and relies on alignments rather than language-specific rules.
  • Difficulty Level Estimation: The introduction of "Weighted Symmetric Cross Entropy" (WSCE) as a language-independent measure to estimate L2P difficulty is a valuable theoretical contribution, showing a correlation with actual accuracy.

Key Results

The proposed approach achieved state-of-the-art results on several languages:

  • German and Dutch: The system achieved the highest word accuracy among all compared methods. For German, the accuracy reached 90.20±0.25%, and for Dutch, it was 91.63±0.24%.
  • English and French: While the baseline was met, the accuracy fell slightly short of being the absolute best, though the English dataset used was larger.
  • Overall Improvement: The paper reports an overall improvement of 5.8% over the baseline.
  • Phoneme vs. Word Accuracy: The study highlights the known correlation: approximately 90% phoneme accuracy corresponds to about 60% word accuracy, underscoring the difficulty of achieving high word-level precision.

Limitations and Future Work

  • Limitations:
    • The authors acknowledge that for English and French, their approach, while strong, did not surpass all previous methods.
  • The error analysis points to specific challenges, particularly with vowel and diphthong transcriptions, and the handling of allophonic variations.
  • The WSCE measure shows a "rough correlation" with accuracy, suggesting it's a useful indicator but not a perfect predictor.
  • Future Work:
    • Improved MERT: Incorporating phonetic feature-based edit distance or string similarity as the loss function in MERT could further enhance performance.
  • Decoding Enhancements: Exploring decoding schemes similar to substring-based transducers could improve the usage of lower-order language models.
  • More Model Parameters: Investigating additional feature functions and extensive parameter testing is suggested.
  • Further Language Independence: While the approach is language-independent, further exploration of its performance across a wider range of languages and linguistic typologies would be beneficial.

Summary updated

BibTeX
@inproceedings{rama2009modeling,
  author = {Taraka Rama and Anil Kumar Singh and Sudheer Kolachina},
  title = {{Modeling Letter-to-Phoneme Conversion as a Phrase Based Statistical Machine Translation Problem with Minimum Error Rate Training}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2009},
}
Figure