Cross Linguistic Phoneme Correspondences

Lynne Cahill, Carole TiberiusCOLING 2002

TL;DR

This paper introduces "metaphonemes" to model systematic sound correspondences between closely related languages, extending the concept of phonemes beyond a single language. This approach aims to improve multilingual lexicon sharing for applications like speech recognition and synthesis, and offers insights into language evolution.

What Problem Does This Paper Address?

Traditional phonology treats sounds within a single language, defining phonemes and their allophones. However, closely related languages often share systematic sound similarities that aren't captured by treating each language in isolation. For instance, the vowel in English "cat," Dutch "kat," and German "Katze" are subtly different but clearly related. The challenge is to formally represent these cross-linguistic sound relationships to enable more efficient data sharing and understanding, particularly for computational linguistics applications. Existing methods like archiphonemes (within-language generalizations) and morphophonemes (morphological generalizations) don't address this cross-linguistic aspect.

Methods

The authors propose "metaphonemes" as abstract units that represent sets of distinct phonemes appearing in different but related languages. The core idea is to group phonemes that correspond systematically across languages, similar to how allophones are grouped within a single language.

Their methodology involved:

  1. Database Construction: Compiling a database of approximately 800 cognate words from Dutch, English, and German. Phonological forms were sourced from the CELEX database.
  2. Data Preprocessing: Simplifying word forms by ignoring leading/trailing schwa syllables and focusing on core root structures.
  3. Syllabic Analysis: Analyzing word forms into syllabic structures, then collating onsets, peaks (vowels), and codas for each language.
  4. Correspondence Analysis:
    • Cluster Analysis: Examining absolute correspondences for consonant clusters and single consonants, noting their frequencies to identify likely patterns and potential phonological conditioning.
  5. Individual Consonant Analysis: Extracting individual consonant correspondences semi-manually to handle cases where a phoneme might correspond to zero in another language (e.g., English /x/ to German zero).
  6. Metaphoneme Definition: Deriving tables that map consonants from one language to their potential counterparts in others, and then defining metaphonemes based on these observed correspondences, sometimes incorporating phonological context.

Datasets

  • CELEX Lexical Database: Used to automatically obtain phonological forms for cognate words.
  • Custom Cognate Database: A newly compiled database of around 800 cognate words from Dutch, English, and German, specifically prepared for this analysis.

Key Insights and Ideas

  • Metaphonemes as a Bridge: The concept of metaphonemes offers a way to bridge the gap between individual language phonologies, allowing for generalizations across related languages.
  • Consonants vs. Vowels: Consonants present different challenges than vowels for metaphoneme definition. The consonant space is more discrete, inventories overlap significantly across these languages, consonants can form clusters, and they can be lost entirely (resulting in zero correspondences).
  • Phonological Conditioning is Crucial: While simple correspondences exist, phonological context (e.g., onset vs. coda position, preceding consonants) plays a significant role in defining metaphonemes, especially for consonants. For example, English /s/ might correspond to Dutch /z/ in onset but /s/ in coda.
  • Language Change Clues: The analysis of zero correspondences provides potential insights into historical language change, suggesting which segments are more likely to be lost. Approximants, liquids, glides, and velar stops (/k/, /g/) appear to be more prone to loss.
  • Dutch vs. German vs. English Loss: Dutch appears less likely to have zero correspondences than German, which in turn has fewer than English.

Key Results

  • The paper successfully defines metaphonemes for consonants in Dutch, English, and German, building on prior work on vowels.
  • Analysis of the /s/, /S/, /z/ sibilants in English, Dutch, and German revealed specific metaphonemes, such as |szz| for the correspondence of English /s/ to Dutch /z/ and German /z/ in onset position, while English /s/ corresponds to /s/ in coda.
  • The authors estimate that using consonant metaphonemes, combined with vowel metaphonemes, can increase shared phonological information in multilingual lexicons by around 40%.
  • The study identified that approximants, liquids, glides, and velar stops are more likely to be lost in language change across these languages.

Limitations and Future Work

  • Limitations: The authors acknowledge that the initial metaphoneme definitions are simplified and that more complex definitions incorporating phonological context are necessary. The analysis of zero correspondences is based on a subset of data and provides preliminary insights.
  • Future Work: The authors plan to fully integrate metaphonemes into a multilingual lexicon for testing with a speech synthesis system. They also suggest that this framework could be extended to model differences between dialects or accents within a single language, and that the analysis of segment loss could be further explored for historical linguistics.

Summary updated

BibTeX
@inproceedings{cahill2002cross,
  author = {Lynne Cahill and Carole Tiberius},
  title = {{Cross Linguistic Phoneme Correspondences}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2002},
}
Figure