Modeling Markedness with a Split-and-Merger Model of Sound Change

Andrea Ceolin, Ollie SayeedACL 2019

TL;DR

This paper proposes a novel computational model of sound change based on random "splits" and "mergers" of phonemic categories. By simulating these processes, the authors demonstrate that the concept of "markedness" in phonology can be explained as an emergent property of phonetically grounded sound change, rather than an inherent feature of sounds.

What Problem Does This Paper Address?

The concept of "markedness" has been a cornerstone of phonological theory for decades, used to categorize sounds or linguistic features as more or less "natural" or "common." However, the precise origin and nature of markedness have remained a subject of debate. This paper aims to bridge the gap between theoretical phonology and computational modeling by investigating whether the observed patterns of markedness can be explained as an epiphenomenon of stochastic sound change. Specifically, they question if a simple, abstract model of sound change can naturally produce the correlations between language-internal and language-external properties that are often attributed to markedness.

Methods

The authors introduce a "split-and-merger" model, a discrete-time stochastic process that simulates sound change. The model operates on a set of phonemic segments, each with a specific frequency. At each time step, with equal probability, either a "split" or a "merger" occurs:

  • Split: A random pair of segments ($x_i, x_j$) is chosen. Half of $x_i$'s probability mass is transferred to $x_j$.
  • Merger: A random pair of segments ($x_i, x_j$) is chosen. All of $x_i$'s probability mass is transferred to $x_j$.

Crucially, the selection of segments for splitting and merging is not entirely random. The model incorporates two probability distributions:

  • $P_s(x_j)$: The probability that segment $x_j$ will have its frequency increased after a split. Segments with low $P_s(x_j)$ are termed "splitwise marked."
  • $P_m(x_i)$: The probability that segment $x_i$ will be lost in a merger. Segments with high $P_m(x_i)$ are termed "mergerwise marked."

The authors then run simulations of this model, varying parameters related to these markedness probabilities, to observe the emergent frequency distributions of segments.

Datasets

Not discussed in the paper. The simulations are run on abstract sets of 20 segments, with initial frequencies set to be uniform, and some segments assigned an initial frequency of zero.

Key Insights and Ideas

  • Markedness as an Epiphenomenon: The most significant insight is that markedness, a concept traditionally treated as an intrinsic property of sounds, can emerge naturally from a simple, phonetically grounded model of sound change.
  • Power-Law Distributions: The split-and-merger model, like other random fragmentation and aggregation processes, naturally generates power-law frequency distributions for phonemes, which have been observed in natural languages.
  • Splitwise vs. Mergerwise Markedness: The model distinguishes between two types of markedness:
  • Splitwise markedness (low probability of being created after a split) primarily affects the across-language frequencies of "unmarked" segments, making them more common in the linguistic landscape.
  • Mergerwise markedness (high probability of being lost in a merger) is responsible for higher within-language and across-language frequencies for "unmarked" segments and lower frequencies for "marked" segments.

Key Results

The simulations demonstrate that:

  • When splitwise markedness is implemented, "unmarked" segments (like 'a' in the paper's examples) tend to have higher frequencies both within and across simulated languages compared to "neutral" or "marked" segments. The effect on "marked" segments was less pronounced.
  • When mergerwise markedness is implemented, "unmarked" segments consistently exhibit higher within- and across-language frequencies, while "marked" segments show lower frequencies. This type of markedness appears to be more influential in shaping segment frequencies.
  • The model successfully reproduces the correlation observed in real languages between within-language and across-language phoneme frequencies.

Limitations and Future Work

  • Abstract Model: The model uses abstract segments and simplified rules for splits and mergers. Future work could incorporate more linguistically realistic phonetic features and constraints.
  • Parameter Sensitivity: The paper explores different values of a parameter 'r' that quantifies the degree of markedness. Further investigation into the sensitivity of the model to various parameter choices would be beneficial.
  • Phonetic Grounding: While the authors suggest the model is "phonetically grounded," the specific phonetic basis for the probabilities $P_s$ and $P_m$ is not detailed. Future research could explore how specific phonetic properties might translate into these probabilities.
  • Cross-Linguistic Variation: The model currently focuses on emergent frequency distributions. Extending it to account for the diversity of phoneme inventories across languages would be a valuable next step.

Summary updated

BibTeX
@inproceedings{ceolin2019modeling,
  author = {Andrea Ceolin and Ollie Sayeed},
  title = {{Modeling Markedness with a Split-and-Merger Model of Sound Change}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
}
Figure