TL;DR
This paper introduces novel approaches to morphological decomposition for speech synthesis, aiming to improve word pronunciation accuracy by handling complex word structures. It proposes Aronoff-style truncation rules for level 1 morphology to better capture derivational relationships and suggests a compound stress theory for level 2 affixes as an alternative to the traditional stress-neutral model.
What Problem Does This Paper Address?
Speech synthesizers convert text to speech signals, and a crucial step is converting words into phonemes. Traditionally, this involves either dictionary lookups or letter-to-sound rules. Dictionary lookups fail for unknown words (like proper nouns), while letter-to-sound rules struggle with irregular words. A hybrid approach is common, but morphological decomposition offers a way to reduce dictionary size and increase coverage by breaking words into their constituent morphemes (roots and affixes). However, existing systems like MITalk's Decomp had limitations, particularly in handling complex morphological and phonological interactions, such as stress alternations and vowel shifts, often resorting to listing exceptions. This paper aims to address these limitations by proposing more sophisticated morphological decomposition techniques that can better account for these linguistic phenomena.
Methods
The paper proposes two main methodological advancements:
- Aronoff-style Truncation Rules for Level 1 Morphology: Instead of simple concatenation (e.g.,
form+-al+-ly), the authors advocate for truncation rules, inspired by Aronoff's work. This approach captures the asymmetry in "derived from" relationships. For example,presidencyis derived frompresidentby truncating-entand adding-cy, butpresidentis not derived frompresidencyby truncating-cyand adding-ent. This method is argued to be more linguistically accurate and can help avoid spurious derivations. - Compound Stress Theory for Level 2 Affixes: The paper challenges the "stress-neutral" assumption for level 2 affixes (like
#ness,#ly). It proposes that level 2 affixes behave more like compounds, with stress patterns dependent on whether the word is a noun (stress on the left) or a verb/adjective (stress on the right). This is presented as an alternative to the MITalk Decomp's approach, which treated these affixes as stress-neutral.
Datasets
Not discussed in the paper. The paper refers to using a dictionary for lookup and analysis, but no specific datasets are listed for training or evaluation.
Key Insights and Ideas
- Asymmetry in Derivation is Crucial: The "derived from" relationship is not symmetrical. Aronoff-style truncation rules are a more linguistically sound way to model this than simple concatenation, leading to more accurate morphological analysis.
- Level 2 Affixes Aren't Always Stress-Neutral: The traditional view of level 2 affixes being stress-neutral is questioned. The paper argues that they often behave like compounds, with stress patterns influenced by the word's lexical category (noun, verb, adjective). This offers a more unified explanation for stress assignment in words with these affixes.
- Lexical Marking for Syllable Weight: For accurate stress assignment, knowing syllable weight is essential. The paper suggests that stems should be marked in a lexicon for syllable weight, which can then be combined with morphological decomposition to determine overall word stress.
- Bracketting Paradoxes and Stress: The paper touches upon "bracketing paradoxes" (where syntactic structure and morphological structure seem to conflict) and suggests that stress assignment can provide evidence for the correct underlying structure, often favoring a left-branching structure that violates strict level ordering.
Key Results
- The MITalk Decomp program achieved 95% coverage of running text, but this came at the cost of not handling complex phonological rules and listing many exceptions.
- The proposed decomposition program, using Aronoff-style truncation rules, was able to construct forests of related words that "agree very closely with intuition."
- Statistical estimation of suffix relationships (e.g.,
-ency > -entwith 73% probability) was used to inform the decomposition process. - The paper provides extensive tables detailing the statistical relationships between various suffixes and their potential preceding elements, offering a data-driven approach to morphological analysis.
Limitations and Future Work
- Implementation of Stress Assignment: While the paper proposes how morphological decomposition can inform stress assignment (by using a lexicon with marked syllable weights), the full implementation of this stress assignment mechanism was not yet completed at the time of writing.
- Defining "Branching": The paper acknowledges a challenge in precisely defining "branching" for compounds, particularly in cases like
superconductorvs.superconductivity, to correctly predict stress. This is identified as a promising but not fully resolved area. - Pragmatic vs. Theoretical Payoff: The authors admit that the immediate payoff in terms of lexicon size reduction might be marginal, but they bet on the long-term value of pushing theoretical linguistic understanding through demanding applications like speech synthesis.
- Further Refinement of Heuristics: The statistical estimation of suffix relationships could be improved by refining heuristics, such as excluding short stems that might lead to spurious analyses.
Summary updated