Phonological distances for linguistic typology and the origin of Indo-European languages

This paper demonstrates that an information-theoretic analysis of short-range phoneme dependencies in 67 modern languages can effectively reconstruct major language families, detect contact-induced convergence, and support the Steppe hypothesis for the Indo-European homeland by revealing a strong correlation between phonological and geographic distances.

The Gist

Imagine you are a detective trying to solve a mystery that is thousands of years old: Where did the Indo-European language family (which includes English, Spanish, Hindi, Russian, and many others) originally come from?

Usually, linguists solve this by looking at ancient words, like comparing the word for "mother" or "fire" across different languages. But this new paper takes a different approach. Instead of looking at what words mean, they looked at how the sounds of the languages feel in your mouth.

Here is the story of their investigation, explained simply.

1. The "Sound Fingerprint"

Imagine every language has a unique "sound fingerprint." This isn't just about which letters are used, but how the sounds flow together.

The researchers took the Bible (because it's available in almost every language and has a similar length) and turned the text into a string of pure sounds (phonemes). They didn't just look at single sounds; they looked at groups of three sounds (like "b-u-k" in "book").

Think of it like listening to a song. If you only listen to one note, you don't know the melody. But if you listen to three notes in a row, you start to hear the rhythm and the style. The researchers found that these "three-sound groups" capture the unique musical style of a language's sound system.

2. The "Mouth Gym" Map

To measure how different two languages are, they didn't just count how many sounds were different. They looked at how hard your mouth has to work to switch from one sound to another.

• Analogy: Imagine your mouth is a gym.
  • Moving from a sound made with your lips (like "b") to another lip sound (like "p") is a tiny, easy step.
  • Moving from a lip sound to a sound made deep in your throat (like "k") is a giant, exhausting leap.

They built a map where languages that use similar "mouth movements" are close together, and languages that require very different gymnastics are far apart.

3. The Big Discovery: Distance Equals Time

When they plotted all 67 languages on this map, something magical happened. They found a strong rule: The farther two languages are geographically, the more different their sound fingerprints are.

• The Logic: Imagine a group of people leaving a home village. As they walk further away, they drift apart. Over centuries, their accents change. The people who stayed home sound most like the original group. The people who walked the furthest developed the most unique accents.
• The Result: The researchers found that the "sound distance" between languages perfectly matched their "geographic distance."

4. Solving the Mystery of the "Homeland"

Now, they applied this rule to the Indo-European family. They asked: "If we treat the average sound of all these languages as a 'center of gravity,' where on the map would that center be?"

• They calculated the "sound distance" from every modern Indo-European language (like English, Hindi, Greek) back to a theoretical "average ancestor."
• Then, they translated that sound distance back into miles.
• The Answer: The spot that minimized the error was north of the Black Sea, in the grassy plains known as the Pontic-Caspian Steppe.

Why This Matters

This supports the "Steppe Hypothesis," which suggests that the ancestors of Indo-European speakers were nomads on horseback who started in the steppes (modern-day Ukraine/Russia) and spread out, carrying their language with them.

This contradicts the older "Anatolian Hypothesis," which suggested the origin was in modern-day Turkey (where farming began earlier). The "sound math" of this paper leans heavily toward the Steppe.

The Takeaway

The authors didn't need to dig up ancient bones or translate old tablets. They simply used math and the physics of how our mouths move to prove that languages, like people, leave a trail of "sound footprints" that reveal where they started and how far they traveled.

It's like finding a family's origin story just by listening to how they all laugh.

Technical Summary

1. Problem Statement

The paper addresses the challenge of quantitatively measuring linguistic relatedness and evolutionary history using phonological data. While language distance metrics have traditionally relied on lexical (vocabulary) or morphosyntactic features, there is a need for robust methods to analyze phonological systems independently. Specifically, the authors aim to:

• Determine if short-range phoneme dependencies encode large-scale patterns of linguistic relatedness.
• Quantify distances between languages to recover known language families and identify contact-induced convergence.
• Investigate the correlation between phonological distance and geographic distance to infer the homeland (Urheimat) of the Indo-European (IE) language family, testing competing hypotheses (e.g., Steppe vs. Anatolian).

2. Methodology

A. Dataset and Preprocessing

• Corpus: The study utilizes a parallel corpus of 129 Bible translations to ensure consistent genre and sample length across languages, minimizing biases from different text types.
• Languages: The analysis covers 67 modern languages from diverse families (Indo-European, Altaic, Dravidian, etc.).
• Transcription: Texts are converted to the International Phonetic Alphabet (IPA) using Phonemizer (eSpeak backend) and Epitran.
  • Filtering: Suprasegmental features (tones) and diacritics are removed. Vowel length is normalized, but distinctive features like aspiration, palatalization, and nasalization are retained.
  • Validation: Transcriptions were cross-referenced with the WikiPron database. Discrepancies were analyzed and found to be mostly due to allophonic variations or regional dialect differences rather than systematic errors.

B. Information-Theoretic Modeling

The authors model phoneme sequences as stochastic processes to capture statistical correlations:

• Markov Chain Approach: Phoneme sequences are treated as higher-order Markov chains. The authors analyze overlapping blocks of phonemes ($r$-phones).
• Entropy and Predictability: They calculate block entropy ($H_r$) and the predictability gain ($G_u$), defined as the negative second discrete derivative of entropy.
  • $G_u$ measures the information gain when increasing the memory order from $u$ to $u+1$.
  • Key Finding: The analysis reveals that $G_u$ drops significantly between $u=0$ and $u=1$, and stabilizes for $u \ge 3$. This indicates that second-order Markov chains (memory $m=2$) are sufficient to capture the essential statistical properties of the phonological systems. Consequently, 3-phone (triphone) models are used for the distance metric.
• Coarse-Graining: To ensure statistical reliability with limited data, phonemes are sometimes grouped by articulatory features (e.g., voiced/voiceless, vowel openness) to reduce the state space size ($L$).

C. Distance Metric

To quantify the divergence between two languages ($L$ and $L'$), the authors define a phonological distance based on the Wasserstein distance (Earth Mover's Distance):

1. Feature Embedding: Each phoneme is represented as a vector of 24 articulatory features (e.g., syllabic, nasal, voice, anterior). A 3-phone sequence is mapped to a 72-dimensional vector (reduced to 60 dimensions after removing constant features).
2. Ground Metric: The distance between two 3-phone vectors is calculated using the feature edit distance, which assigns substitution costs based on the similarity of their articulatory feature vectors.
3. Wasserstein Distance ($W$): The distance between two languages is the minimum cost to transform the probability distribution of 3-phones in one language into the distribution of the other, weighted by the articulatory feature distance. This accounts for the fact that phonologically similar sounds (e.g., /b/ and /p/) are "closer" than dissimilar ones.

3. Key Contributions

• Novel Metric: Introduction of a phonological distance metric that combines information-theoretic modeling (Markov chains) with articulatory feature-based embeddings and the Wasserstein distance.
• Validation of 3-Phone Models: Demonstration that second-order Markov chains (triphones) effectively capture the statistical structure of phonological systems across diverse languages.
• Phylogenetic Recovery: Successful reconstruction of major language families (Germanic, Slavic, Romance, Indo-Aryan) based solely on phonological statistics, without prior knowledge of genetic classification.
• Geographic-Phonological Correlation: Establishment of a robust positive correlation between phonological distance and geographic distance, particularly within the Indo-European family.

4. Results

A. Language Clustering

• Hierarchical Clustering: Applying Ward linkage to the phonological distance matrix successfully grouped languages into known families:
  • Altaic: Grouped together (Turkic languages).
  • Indo-European: Clear separation into Germanic, Slavic, Balto-Slavic, and Indo-Aryan clusters.
  • Contact Effects: The model captured non-genetic similarities, such as the clustering of Spanish and Basque (due to long-term contact) and the proximity of Armenian to Persian (geographic influence).
• Indo-European Specifics: Within the IE family, the model correctly identified sub-families (Balto-Slavic, Germanic, Italic) and placed Albanian and Greek within the Romance cluster due to historical contact and phonological convergence.

B. Geographic Correlation

• A significant positive correlation was found between phonological distance and geographic distance ($d_{geo}$).
• The relationship follows a logarithmic fit: $W \propto \ln(d_{geo})$.
• The correlation is stronger for Indo-European languages ($R_d = 0.496$) than for the full dataset ($R_d = 0.428$), suggesting that phylogenetic inheritance plays a dominant role in phonological divergence within a single family.

C. Indo-European Homeland Inference

• Method: The authors calculated the average 3-phone distribution ($P_{av}$) of the 39 IE languages. They assumed that phonological distance from this average increases with geographic distance from the origin.
• Optimization: They minimized the sum of squared residuals ($\chi^2$) between the predicted geographic distances (derived from phonological divergence) and the actual geographic coordinates of the languages.
• Result: The location minimizing $\chi^2$ (the most probable homeland) is located north of the Black Sea.
• Uncertainty: The 95% uncertainty region (calculated via bootstrapping with Dirichlet-distributed weights) also lies north of the Black Sea.
• Conclusion: This result strongly supports the Steppe Hypothesis (Kurgan hypothesis) regarding the origin of Indo-European languages and is less compatible with, though not entirely ruling out, the Anatolian hypothesis.

5. Significance

• Quantitative Typology: The study provides a rigorous, data-driven framework for linguistic typology that moves beyond simple lexical comparisons to deep structural phonological analysis.
• Evolutionary Linguistics: It demonstrates that phonological systems retain a "fossil record" of migration and divergence, allowing for the reconstruction of language origins even without ancient texts.
• Interdisciplinary Impact: The work bridges statistical physics (Markov chains, entropy, optimal transport) with computational linguistics and historical linguistics, offering new tools to resolve long-standing debates about language origins.
• Limitations & Future Work: The authors note limitations regarding dataset size (only 67 languages) and the synchronic nature of the analysis. Future work could incorporate diachronic corpora and Bayesian phylogenetic methods to estimate divergence dates.

In summary, the paper successfully argues that phoneme sequences modeled as second-order Markov chains encode sufficient statistical information to reconstruct language families and infer the geographic origins of the Indo-European family, providing strong quantitative evidence for the Steppe hypothesis.