Accenting and Deaccenting: a Declarative Approach

Arthur DirksenCOLING 1992

TL;DR

This paper introduces PROS-3, a novel language-independent system for deriving pitch accent in text-to-speech (TTS) systems. It tackles the challenge of marking "given" vs. "new" information by employing a declarative approach based on focus-accent theory, offering a more maintainable and extensible alternative to traditional rule-based systems.

What Problem Does This Paper Address?

Accurate prosody, particularly pitch accent, is crucial for natural-sounding speech synthesis. A key challenge in TTS is determining which words in a sentence should be accented to reflect the distinction between "given" (already known or contextually inferable) and "new" (novel or emphasized) information. Existing systems often rely on complex, derivational rules that are difficult to manage and extend. This paper addresses this by proposing a declarative approach to pitch accent derivation, aiming for a more robust and adaptable TTS system.

Methods

The PROS-3 system utilizes a declarative approach to model the syntax-to-prosody interface. The core methodology involves:

  • Syntactic Parsing: Input text is parsed using context-free phrase-structure rules augmented with "argument" structure information.
  • Metrical Tree Mapping: The syntactic representation is then mapped onto a metrical tree, which represents the relative prominence of nodes.
  • Focus-Accent Theory: The system is grounded in focus-accent theory, where accentuation is determined by the focus/non-focus distinction.
  • Declarative Rules: The system employs a strictly declarative rule set. Once information is added by a rule, it is never removed, contrasting with traditional derivational systems.
  • Key Rules:
    • Default Accent: This rule models the deaccenting of "given" information, such as pronouns, by shifting accent to more prominent elements.
  • Rhythmic Deaccenting: This rule accounts for rhythmic alternations in accentuation within specific contexts, ensuring a more natural flow.
  • Implementation: The system uses a binary-branching metrical tree as its central data structure, with accenting defined using shared variables that can be instantiated to "true" (accented) or "false" (deaccented).

Datasets

Not discussed in the paper. The paper mentions that PROS-3 is currently being evaluated for Dutch using a grammar of about 125 rules and a lexicon of approximately 80,000 word forms derived from the CELEX lexical database. Grammars and lexicons of comparable size are also being developed for English and German.

Key Insights and Ideas

  • Declarative Approach for Prosody: The paper demonstrates that a declarative, non-derivational approach can effectively model complex prosodic phenomena like pitch accent, offering significant advantages in terms of maintainability and extensibility.
  • Language Independence: By modeling language-specific differences solely within the syntactic rules, the core prosodic derivation mechanism remains language-independent, a crucial feature for multilingual TTS systems.
  • Focus-Accent Theory as a Foundation: The work highlights the power of focus-accent theory as a theoretical basis for deriving pitch accent, showing how its principles can be directly translated into computational rules.
  • Unified Treatment of Deaccenting: The system provides a unified framework for handling both "given" information deaccenting (Default Accent) and rhythmically motivated deaccenting (Rhythmic Deaccenting).

Key Results

The paper presents a system (PROS-3) that is designed to derive pitch accent in a language-independent manner. While specific quantitative results are not detailed, the paper emphasizes the system's ability to:

  • Model the distinction between "given" and "new" information through focus-accent theory.
  • Handle deaccenting of pronouns and other "given" elements.
  • Incorporate rhythmic factors for more natural prosody.
  • Offer a declarative and extensible framework for prosody generation.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Crude Heuristics for Focus/Non-Focus: The current modeling of the focus/non-focus distinction relies on simple heuristics (e.g., considering major phrases as candidates for focus, lexical specification for pronoun deaccenting).
  • Future Work: Incorporating a more flexible and "discourse-aware" model, potentially along the lines of Hirschberg (1990), is desired.
  • Simple Syntactic Categories: The PROS-3 grammar formalism is kept deliberately simple (atomic syntactic categories) to ensure efficient processing.
  • Limitation: This simplicity means that simple context-free rules may not disambiguate very well.
  • Future Work: Exploring more sophisticated syntactic analysis is needed.
  • Verb Subcategorization: Simple rules cannot fully account for verb subcategorization.
  • Limitation: This can make it difficult to distinguish between arguments and non-arguments, which is crucial for metrical rules.
  • Future Work: Finding an optimal compromise between syntactic analysis sophistication and processing efficiency is a key goal.
  • Phonological Parsing: Preliminary work has been done on phonological parsing to derive word stress and temporal structure of words.
  • Future Work: Further development in this area is planned.

Summary updated

BibTeX
@inproceedings{dirksen1992accenting,
  author = {Arthur Dirksen},
  title = {{Accenting and Deaccenting: a Declarative Approach}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1992},
}
Figure