Computational Phonology: Merged, Not Mixed

Egon Berendsen, Simone Langeweg, Hugo van LeeuwenCOLING 1986

TL;DR

This paper presents a novel grapheme-to-phoneme conversion system for Dutch, designed for flexibility and efficiency in text-to-speech applications. It introduces a two-level linguistic and computational approach, allowing linguists to define rules independently of implementation, and a powerful rule format that extends standard SPE conventions.

What Problem Does This Paper Address?

The paper addresses the challenge of accurately converting written text (graphemes) into spoken sounds (phonemes) for Dutch text-to-speech (TTS) systems. While existing systems exist, the authors highlight the need for a system that is not only accurate but also flexible enough to serve multiple purposes beyond basic synthesis, such as generating phonemized lists and providing insights into grapheme-phoneme relationships. They also emphasize the importance of a user-friendly system for linguists without extensive programming knowledge.

Methods

The core of the system is a two-level approach: a linguistic component where linguists define conversion rules, and a computational component that implements these rules. They propose a grapheme-to-phoneme conversion process that involves several modules:

  • Preprocessing: Handling of numbers, acronyms, and abbreviations.
  • Word Recovery: Identifying units like compounds that influence pronunciation.
  • Grapheme-to-Phoneme Assignment: The central module, assigning phonemes to graphemes using a rule-based system.
  • Phonological Processes: Modules for stress assignment and segmental phenomena.

The authors extend the standard Chomsky and Halle (1968) SPE rule format with several key features:

  • Negation: Ability to negate elements or groups of elements in the rule environment.
  • Global Rules: Rules that can refer back to information earlier in the derivation, leveraging the two-level approach.
  • Definitions: Allowing the use of defined sequences (e.g., syllable boundaries) in rule environments.
  • Coordinations: The ability to specify multiple conditions that must all be met for a rule to apply.

For implementation, they describe a dynamic data structure composed of linked units, each containing fields for grapheme, phoneme, features, and pointers to other units. This structure supports alternative, simultaneous, and optional validity of rule components.

Datasets

Not discussed in the paper. The paper focuses on the linguistic rules and computational implementation rather than specific datasets for training or evaluation. They mention testing the system on a lexicon of about 4000 monomorphemic words.

Key Insights and Ideas

  • Separation of Concerns: The clear distinction between linguistic rule definition and computational implementation is crucial for user-friendliness and maintainability. Linguists can focus on phonological rules, while engineers handle the technical aspects.
  • Extended Rule Format: The addition of negation, global rules, definitions, and especially coordinations significantly enhances the expressive power of the rule system, allowing for more nuanced and accurate phonological descriptions.
  • Two-Level Representation: Maintaining separate levels for graphemes and phonemes is key to capturing the complex relationships and allowing for flexible processing.
  • Dynamic Data Structure: The use of a dynamic, linked data structure for representing linguistic rules is essential for efficiently handling the complexity and variable lengths introduced by the extended rule format.
  • Efficiency is Paramount: The system is designed with speed in mind, both for constructing the rule database and for on-line transcription, aiming for practical real-time performance.

Key Results

  • Performance: The system achieves approximately 25 grapheme-to-phoneme conversions per second.
  • Database Construction: A new database for testing rules can be constructed in about 20 seconds.
  • Flexibility: The system is designed to be adaptable for various applications beyond basic TTS, including generating phonemized lists and providing linguistic insights.
  • Linguistic Accuracy: The extended rule format allows for the expression of complex phonological phenomena, such as alternations in pronunciation based on context (e.g., the 'ch' sound in Dutch).

Limitations and Future Work

  • Handling of Ambiguity: The paper acknowledges that the system might encounter unexpected problems with complex negated structures, requiring careful consideration of all paths. Further discussion on this is referred to another publication (Van Leeuwen et al., 1986).
  • Exception Lexicon: For cases that cannot be captured by rules, an exception lexicon is mentioned as a necessary component.
  • Scope: The paper explicitly states that abbreviations, acronyms, numbers, and phonological processes above the word level are disregarded in this discussion.
  • Future Work: The system is being linked to a diphone speech synthesis system and is currently being tested on a lexicon of 4000 monomorphemic words, implying ongoing development and evaluation.

Summary updated

BibTeX
@inproceedings{berendsen1986computational,
  author = {Egon Berendsen and Simone Langeweg and Hugo van Leeuwen},
  title = {{Computational Phonology: Merged, Not Mixed}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1986},
}
Figure