Statistical Modeling of Creole Genesis

Yugo MurawakiNAACL 2016

TL;DR

This paper tackles the complex question of how creole languages form, challenging traditional tree-like evolutionary models. By applying statistical mixture models inspired by population genetics, the authors find that a "restructurer" component, representing innate linguistic universals, plays a dominant role in creole genesis, rather than just the lexifier or substrate languages.

What Problem Does This Paper Address?

The formation of creole languages is a fascinating linguistic phenomenon that doesn't fit neatly into traditional evolutionary models, like the "language tree" used to depict how languages diverge. These tree models assume that once languages split, they evolve independently. However, creole genesis involves intense contact between multiple languages, making it a prime example of non-tree-like evolution. Linguists have debated whether creoles are typologically distinct from non-creoles and how their unique grammars emerge. Previous work, like that by Daval-Markussen and colleagues, used clustering methods (NeighborNet) to suggest creoles form a distinct cluster, supporting a "universalist" view. However, this paper argues that clustering methods might not fully capture the complex, non-tree-like nature of creole formation and that the synchronic question of distinctiveness is conflated with the diachronic question of genesis.

Methods

The authors propose a novel approach using mixture models, drawing inspiration from population genetics and admixture analysis. Instead of a tree, they model a creole as a probabilistic mixture of three sources: a lexifier (the dominant language providing vocabulary), substrate(s) (languages of the less dominant groups), and a crucial "global restructurer". This restructurer is posited to capture innate human linguistic abilities or universals that shape creole grammar, independent of the specific source languages.

They introduce two Bayesian generative models:

  • MONO Model: A simpler model where each creole is a mixture of the three sources, with a single mixing proportion for the entire language.
  • FACT Model: An extension that allows for more nuanced mixing by decomposing proportions into per-feature and per-creole factors. This means different features within a creole can be influenced by different sources, and the influence can vary across creoles.

For inference, they use techniques like Gibbs sampling and Metropolis-Hastings algorithms. They also employ Principal Component Analysis (PCA) for exploratory data visualization to understand the distribution of creoles and non-creoles in the feature space.

Datasets

The primary dataset used is the Atlas of Pidgin and Creole Language Structures (APICS). This database contains sociolinguistic features, phonological inventories, and example texts for pidgin and creole languages. To compare creoles with non-creoles, they also utilized features from the World Atlas of Language Structures (WALS). They selected 48 typological features common to both datasets. For the analysis, they identified 64 creoles and 541 non-creoles, ensuring at least 30% of features were present for each language. Missing data was imputed using multiple correspondence analysis.

Key Insights and Ideas

  • The Restructurer Dominates: The most striking finding is the overwhelming influence of the "restructurer" component in creole genesis. This suggests that innate linguistic predispositions or universals play a more significant role than the specific lexifier or substrate languages in shaping creole grammars.
  • Non-Tree-like Evolution is Key: The paper argues that mixture models are more appropriate than tree-based models for understanding creole genesis, as they can explicitly account for the complex interplay of multiple source languages and the emergence of novel structures.
  • Creoles Aren't Necessarily "Simple": While creoles are often described as simpler, the "restructurer" component implies a complex underlying mechanism driven by innate linguistic principles, rather than just a reduction of complexity.
  • Heterogeneous Feature Influence: The FACT model reveals that different linguistic features can have varying origins within a creole, and this influence can differ from one creole to another, highlighting the dynamic nature of creole formation.
  • Challenging Traditional Theories: The results provide counter-evidence to purely superstratist (lexifier-dominant) or substratist (substrate-dominant) theories, and also question the "feature pool" theory, suggesting a more complex, internally driven process.

Key Results

  • Binary Classification: A linear SVM classifier, while achieving high accuracy (97.2%), still struggled with borderline cases, indicating that creoles and non-creoles are not perfectly separable based on the typological features used. PCA also showed overlap between the two groups, despite distinct distributions.
  • Mixture Model Proportions:
    • In the MONO model, the restructurer accounted for a substantial majority of features (around 74.1%), with the lexifier and substrate playing smaller roles (16.6% and 9.3% respectively).
  • The FACT model showed similar patterns, with the restructurer being dominant. When decomposed into per-feature and per-creole factors, the per-creole factors showed less uneven distributions, suggesting that the influence of sources can vary across creoles.
  • Top Restructurer Features: Features like "Numeral Classifiers," "Intensifiers and Reflexive Pronouns," and "Expression of Pronominal Subjects" were strongly associated with the restructurer, suggesting these might be universal tendencies in creole formation.

Limitations and Future Work

  • Data Limitations: The authors acknowledge limitations in the APICS dataset, including the need for better proxies for substrates, handling of missing values, and the representation of multi-valued features.
  • Simplifications: The models currently use only one substrate per creole, which is a simplification. Extending this to multiple substrates is a clear future direction.
  • Staged Development: The current models mix sources in a single step. Future work could model the staged development of pidgin formation and creole formation, and also investigate decreolization (the process by which creoles become more like their lexifiers over time).
  • Defining Universals: While the "restructurer" suggests universals, the authors refrain from definitively calling them "restructuring universals" without further breakdown. Future work could aim to identify specific linguistic universals within the restructurer.
  • Historical Depth: The dominant role of the restructurer has implications for tracing deep historical relationships, as many features might not be directly inherited from known ancestors. Further research could explore how long creole-like traits persist.

Summary updated

BibTeX
@inproceedings{murawaki2016statistical,
  author = {Yugo Murawaki},
  title = {{Statistical Modeling of Creole Genesis}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2016},
}
Figure