TL;DR
This paper introduces a novel method for automatically discovering sound correspondences between related languages, a crucial step in historical linguistics. By adapting techniques from statistical machine translation, the authors show it's possible to identify these patterns even in noisy wordlists with many unrelated words, leading to more accurate cognate identification.
What Problem Does This Paper Address?
Historical linguists often rely on the comparative method to understand language evolution. A key step in this method is identifying "sound correspondences" – regular patterns of sound changes between related languages (e.g., English 't' often corresponds to Latin 'd'). These correspondences are vital for distinguishing true cognates (words with a common ancestor) from chance resemblances or loanwords. However, manually identifying these correspondences is incredibly time-consuming. Existing computational approaches often require a pre-existing set of cognates, creating a circular dependency. This paper aims to automate the discovery of sound correspondences directly from bilingual wordlists, even when a large portion of the word pairs are not cognates.
Methods
The core innovation is adapting models from statistical machine translation (SMT) to the problem of sound correspondences. In SMT, translation models learn word-to-word alignment probabilities in parallel text (bitexts). This paper frames sound correspondence discovery as a similar alignment problem, but between phonemes in cognate words.
Key aspects of their approach include:
- Translating SMT Concepts: The probability of a word pair being a translation in SMT is analogous to the probability of a phoneme pair being a sound correspondence.
- One-to-One Alignment: The method assumes a one-to-one correspondence between phonemes, similar to how SMT often models word alignments.
- Dynamic Programming for Alignment: Instead of Melamed's competitive linking, they use a dynamic programming approach (similar to the Wagner-Fischer algorithm) to find optimal phoneme alignments, incorporating a "no-crossing-links" constraint which is more appropriate for phonological changes than for sentence translation.
- Indel Penalty: Insertion and deletion of sounds are modeled using a constant penalty, reflecting that these changes tend to happen at the segment level rather than across entire words.
- Likelihood Scores: Phoneme pairs with high likelihood scores (indicating they frequently co-occur in aligned cognate pairs) are considered probable correspondences.
- Iterative Refinement: Similar to SMT training, the model iteratively refines likelihood scores and alignments until convergence.
- Multiple Methods (A, B, C, D): They explore variations, including methods that explicitly model noise (Method B), classify phonemes (Method C), or focus only on consonants (Method D).
Datasets
- Development Set: Three 200-word list pairs from the Comparative Indo-European Data Corpus (Dyen et al., 1992). These lists included cognation judgments from a historical linguist. The pairs were Polish-Russian, Spanish-Romanian, and Italian–Serbocroatian. These were transcribed into phonetic notation.
- Test Set: Five 200-word lists representing English, German, French, Latin, and Albanian, compiled by Kessler (2001). These lists were automatically converted from XML to phonetic notation. Word pairs classified as "doubtful cognates" were treated as unrelated.
Key Insights and Ideas
- SMT as a Powerful Analogy: The paper demonstrates that the underlying principles of learning alignments in SMT can be effectively repurposed for discovering sound correspondences in historical linguistics. This highlights the transferability of techniques across seemingly disparate fields.
- Robustness to Noise: The method is surprisingly effective even when a large majority (up to 71%) of word pairs in the input list are unrelated. This is a significant advantage over methods that require clean cognate sets.
- Explicit Correspondence Positing: Unlike some statistical approaches, this method explicitly posits correspondences, allowing linguists to examine and verify them directly.
- Cognate Identification as a Downstream Task: The discovered correspondences can then be used to improve the accuracy of identifying cognates, creating a more robust pipeline for linguistic analysis.
Key Results
- Accurate Correspondence Discovery: In noisy English-Latin data (only 29% cognates), the method discovered the top ten correspondences, with all but one being valid. This outperformed a baseline using the $\chi^2$ statistic.
- Improved Cognate Identification: When used for cognate identification, the proposed methods (especially B, C, and D) achieved higher average precision than established programs like COGNATE and JAKARTA on the test set.
- Method Performance: Method B (with an explicit noise model) performed best on closely related languages, while Method D (focusing on consonants) was superior for more distantly related languages.
Limitations and Future Work
- One-to-One Model Limitation: The current model assumes one-to-one phoneme correspondences. The authors acknowledge this limitation and suggest future work to incorporate "complex correspondences" and handle non-compositional compounds, potentially by extending Melamed's algorithm.
- Phonological Context: The current approach does not explicitly take phonological context into account, which can be crucial for understanding sound changes. Future work could combine this approach with phonetic-based methods.
- Dictionary-Type Data: The authors suggest that the techniques could be applied directly to dictionary-type data, rather than just wordlists.
- Transferability: The paper suggests that the techniques developed for diachronic phonology could also be applied to SMT, indicating a potential for cross-pollination of ideas.
Summary updated