Typological Word Order Correlations with Logistic Brownian Motion

Kai Hartung, Gerhard Jäger, Sören Gröttrup, Munir GeorgesNAACL 2022

TL;DR

This paper investigates whether word order patterns in languages are universally correlated or if these correlations are specific to language families. Using a logistic Brownian motion model on a large dataset of languages, they found that correlations are largely lineage-specific, contradicting some previous findings and suggesting that cultural evolution plays a significant role in shaping language structure.

What Problem Does This Paper Address?

For decades, linguists have been fascinated by the similarities observed across human languages. One area of intense study is word order, the arrangement of subject, verb, and object, among other elements. While some word order patterns appear frequently across diverse languages, the underlying reasons for these patterns are debated. Are they driven by universal cognitive constraints, or do they arise from historical and cultural interactions within specific language groups? Previous studies have yielded conflicting results, with some suggesting universal correlations and others pointing to lineage-specific ones. This paper aims to resolve this by applying a more sophisticated phylogenetic modeling approach to a broad dataset of languages.

Methods

The core of this research lies in the application of a logistic Brownian motion model. This approach models the evolutionary change of linguistic traits over time, treating languages as evolving entities akin to biological species.

  • Brownian Motion: This statistical model describes how traits change randomly over time. In this context, it's used to model the gradual evolution of word order traits.
  • Logistic Linking: Since word order traits are categorical (e.g., adjective-noun or noun-adjective), a logistic function is used to link the continuous Brownian motion process to the observed binary trait values. This allows the model to predict the probability of a specific word order.
  • Bayesian Framework: The model is implemented within a Bayesian framework, allowing for the estimation of parameters (like trait correlations) and the quantification of uncertainty.
  • Phylogenetic Trees: The model explicitly incorporates the historical relationships between languages, represented by phylogenetic trees. The lengths of the branches in these trees represent the time since the last common ancestor, influencing how traits are expected to co-evolve.
  • Model Comparison: The authors compare different model configurations:
  • Single-family models: Analyzing correlations within each of the 34 language families separately.
  • Universal models: Analyzing correlations across all families simultaneously.
  • Independence models: Testing for correlations while assuming traits evolve independently.
  • Metrics: They use Bayes Factors, WAIC, and LOOIC to compare the models and determine which best explains the data.

Datasets

The study utilizes two main types of data:

  • Language Features: Data on eight specific word-order traits for 768 languages. These traits are derived from the World Atlas of Languages (WALS) and are summarized into binary values (present or absent).
  • Phylogenetic Data: Phylogenetic trees for 34 language families, estimated by Gerhard Jäger (2018), representing the historical relationships between the languages.

Key Insights and Ideas

  • Lineage-Specific Correlations Dominate: The most striking finding is that word order correlations are predominantly specific to individual language families, rather than being universal across all languages. This suggests that historical contact and cultural transmission play a more significant role than universal cognitive biases in shaping these linguistic patterns.
  • Contradiction with Previous Universal Findings: The results challenge some earlier studies that suggested universal correlations for certain word order pairs. This highlights the importance of the modeling approach and the breadth of the dataset used.
  • Bayes Factors vs. Information Criteria Disagreement: A notable point is the discrepancy between Bayes Factors and information criteria (WAIC/LOOIC) when comparing universal models. While Bayes Factors favored lineage-specific correlations, information criteria leaned towards universal models. This suggests a need for further investigation into the interpretation of these metrics in this context.
  • Adposition-Noun Order as a Potential Universal: Among the few potential universal correlations identified, the Adposition-Noun order (e.g., "to the house" vs. "the house to") emerged as a candidate, though its universality was not definitively established across all comparison methods.

Key Results

  • Single-family models: For 15 out of 34 families, no strong preference for correlated models over independent models was found using Bayes Factors. When correlations were found, they appeared in only a minority of families, not indicating a universal pattern.
  • Universal models: Bayes Factors strongly favored lineage-specific correlations over universal ones for each trait pair.
  • Information Criteria: WAIC and LOOIC, however, strongly favored universal correlation models, indicating a complex picture where different statistical measures offer different perspectives.
  • Figure 2: Illustrates that for most trait pairs, correlations were found in only a small number of families, reinforcing the idea of lineage specificity.
  • Figure 3: Visually demonstrates that Bayes Factors consistently favor lineage-specific correlations over universal ones for most trait pairs.

Limitations and Future Work

  • Information Criteria Discrepancy: The authors acknowledge the contradiction between Bayes Factors and information criteria regarding universal models. They suggest this might be due to the unreliability of information criteria with the specific data and model used, and call for further investigation.
  • Model Complexity: While the logistic Brownian motion model is sophisticated, the paper notes that representing all families in one model did not add significant information compared to single-family models, suggesting that the complexity of a single universal model might not be fully justified by the data.
  • Future Directions: The paper implicitly suggests further exploration into the reasons behind lineage-specific correlations, potentially involving more detailed analyses of cultural transmission and contact. Investigating the nuances of the information criteria disagreement is also a clear avenue for future research.

Summary updated

BibTeX
@inproceedings{hartung2022typological,
  author = {Kai Hartung and Gerhard Jäger and Sören Gröttrup and Munir Georges},
  title = {{Typological Word Order Correlations with Logistic Brownian Motion}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2022},
}
Figure