TL;DR
This paper introduces Gaussian Processes (GPs) as a novel way to model sound change in Indo-Aryan languages, capturing complex correlations between changes that simpler models miss. By using GPs, the authors demonstrate improved predictive accuracy, offering a more nuanced understanding of dialectal relationships and linguistic evolution.
What Problem Does This Paper Address?
Understanding how languages change over time, particularly sound changes, is crucial for reconstructing linguistic history and mapping dialectal relationships. Indo-Aryan languages, due to extensive contact and borrowing, present a particular challenge. It's difficult to disentangle genuine sound change from lexical borrowing, leading to many "residual forms" that deviate from expected patterns. Traditional models often treat sound changes as independent events, failing to capture how certain changes might be correlated (e.g., if one sound changes, another similar sound might also change in a predictable way). This paper aims to address this by developing a probabilistic model that explicitly accounts for these correlations between sound changes.
Methods
The core of this paper is the application of Gaussian Processes (GPs) to model sound change. GPs are a flexible statistical tool that can model covariance (correlation) between data points. In this context, they are used to model the covariance between different sound changes.
The authors propose a generative Bayesian model. At a high level, each word in a language is assumed to be generated by one of several latent dialect components. These components, in turn, are characterized by collections of distributions over sound changes. The key innovation is how these sound change distributions are modeled:
- Diagonal Model (Baseline): This model treats sound changes as independent, using a simple softmax-transformed multivariate Gaussian prior.
- GP Models (BGP and GGP): These models use GPs to generate the unnormalized sound change probabilities. The GP's covariance function, parameterized by a kernel, allows it to learn how sound changes are related.
- Binary GP (BGP): Uses a simpler kernel that considers segmental identity (input/output sounds) and environmental identity.
- Granular GP (GGP): Employs a more sophisticated Automatic Relevance Determination (ARD) kernel that incorporates detailed phonological features (like place and manner of articulation, voicing, etc.) for both segments and their environments. This allows the model to learn which features are most influential in determining correlations between sound changes.
Inference is performed using Stochastic Gradient Variational Bayes (SGVB) to learn the model parameters.
Datasets
The primary dataset used is derived from Ralph L. Turner's "A Comparative Dictionary of Indo-Aryan Languages" (1962-1966). This dataset includes Old Indo-Aryan (OIA) headwords and their reflexes in modern Indo-Aryan languages.
- Data Preprocessing: The data was normalized to IPA, morphological endings were stripped, and languages were mapped to geographic metadata. Languages with fewer than 100 forms were excluded, resulting in 50 languages. Sound changes appearing fewer than 7 times were also filtered out, leaving approximately 38,479 modern Indo-Aryan words.
- Alignment: OIA/NIA string pairs were aligned using the Needleman-Wunsch algorithm.
- Feature Engineering: Sound changes affecting specific OIA sounds (∫, v, n, n, s, ṛ, h, i, i, j, kș, l, n, r, s, u, u:) were selected as they are considered diagnostically important for Indo-Aryan dialectology.
No new benchmarks were introduced; the focus was on applying the proposed modeling framework to existing linguistic data.
Key Insights and Ideas
- Sound Changes Aren't Independent: The paper strongly suggests that sound changes are often correlated, and failing to model this correlation leads to a less accurate understanding of linguistic evolution.
- Phonological Features Matter for Correlation: The GGP model, by incorporating detailed phonological features, shows promise in capturing nuanced relationships between sound changes. This moves beyond simple phonetic similarity to a more feature-based understanding of linguistic evolution.
- GPs as a Flexible Prior: Gaussian Processes offer a powerful and flexible way to define priors over sound change distributions, allowing them to adapt to the data and learn complex covariance structures.
- Geographic Signal: The BGP model, which considers segmental identity and environment, shows the strongest correlation between inferred dialect component makeup and geographic location, suggesting that spatial proximity plays a role in linguistic similarity.
Key Results
- Improved Predictive Accuracy: The GP models (BGP and GGP) generally outperformed the simpler Diagonal model in terms of posterior predictive checks, specifically in accurately regenerating the observed data (accuracy). The BGP model showed slightly better accuracy than the GGP model.
- Geographic Correlation: The BGP model exhibited the strongest geographic signal, with a Spearman's ρ of .34 for the correlation between inter-language JS divergence and geographic distance.
- Feature Relevance: Analysis of the GP kernels (squared inverse length scales) indicated that "change" (input-output segments) was generally more relevant than "environment" for the BGP model. For the GGP model, while most features showed low relevance, "vowel height" in the environment was noted as potentially influential.
- Entropy: The Diagonal model performed better in terms of entropy (uncertainty in dialect component assignment), suggesting a potential trade-off between model complexity and certainty in assigning labels.
Limitations and Future Work
- Interpretability of Relevance: Interpreting the relevance of featural dimensions in the GP kernels can be challenging, especially when features have different scales.
- Kernel Function Choice: The paper used a standard squared exponential kernel. Future work could explore other kernel functions, including those that can model negative covariance.
- Prior Specification: The choice of priors for kernel parameters was relatively uninformative. Further empirical work is needed to determine optimal priors.
- Model Complexity vs. Certainty: The Diagonal model's better performance on entropy suggests that simpler models might assign labels with more certainty, even if they are less accurate overall. This trade-off warrants further investigation.
- Future Directions: The authors suggest exploring more sophisticated GP kernels, investigating the Gaussian Stick-Breaking construction for language-level priors, and refining the featural representation of environments. They also highlight the potential for this methodology to be applied to other language families and to leverage large, unexploited high-definition linguistic datasets.
Summary updated