Spatio-Temporal Prediction of Dialectal Variant Usage

Péter Jeszenszky, Panote Siriaraya, Philipp Stoeckle, Adam JatowtACL 2019

TL;DR

This paper explores how age and geographic proximity influence the usage of dialectal variants, aiming to predict language change. By modeling contact possibility through these factors, the authors demonstrate that while age is a significant predictor globally, its local influence is more nuanced, suggesting that understanding dialectal variation requires a fine-grained, spatially aware approach.

What Problem Does This Paper Address?

Language is constantly evolving, and understanding how and why it changes is a central question in linguistics. Dialectal variants, specific forms of words or grammar used in particular regions, are key indicators of this change. Traditionally, linguistic geography has focused on spatial patterns, while sociolinguistics has examined social factors like age. This paper bridges these fields by proposing a computational model to predict the usage of dialectal variants, thereby offering a way to understand and potentially forecast language change. The core challenge is to move beyond static descriptions of dialectal distribution and develop a dynamic model that accounts for both the spatial relationships between communities and the temporal aspect of language acquisition and change, particularly as influenced by age.

Methods

The authors employ a logistic regression approach to model the probability of a specific dialectal variant being used. They test two main hypotheses:

  1. Global Scale: They first assess the predictive power of age across all survey sites combined. This involves training a logistic regression model using age as the sole predictor for variant usage and evaluating its performance using AUC (Area Under the ROC Curve) as a measure of separability.
  2. Local Scale (Regionalisation): Recognizing that global patterns might mask local variations, they introduce a regional approach. For each survey site, they consider its k nearest neighbors (using Euclidean distance) and build a logistic regression model for that site based on the age and variant preferences of respondents in its local neighborhood. This allows them to assess the predictive power of age at a more granular, local level. They test this approach with varying values of k (from 5 to 50).

Datasets

The study utilizes data from the Syntactic Atlas of German-speaking Switzerland (SADS). This dataset comprises:

  • Survey Sites: 383 sites across German-speaking Switzerland.
  • Respondents: 3,174 respondents across various age groups (12-94 years old), with a median age of 57.
  • Linguistic Phenomena: Data on 54 different morphosyntactic phenomena, with the study focusing on approximately half of these (60 phenomena in total, as detailed in the appendices).
  • Data Type: The data includes responses to multiple-choice questions where respondents specified their "preferred" variants, which is crucial for analyzing variant usage.

Key Insights and Ideas

  • Age as a Global Predictor: Age is indeed a significant factor in predicting dialectal variant usage on a global scale for many linguistic phenomena. However, the predictive power (measured by AUC) is often relatively low, suggesting that age alone doesn't fully capture the complexity of variant distribution.
  • Local Context Matters: The study highlights that age's predictive power can vary significantly from one location to another. What might be a strong predictor in one region might be weak in another, indicating that local contact patterns and social dynamics are crucial.
  • Syntax Changes Slowly: The paper acknowledges the common observation that syntactic changes tend to be slower than lexical changes. This might explain why age, a strong driver of lexical change, shows a less pronounced effect on syntax at a global level.
  • The "Apparent Time Hypothesis" in Action: The research implicitly tests the "apparent time hypothesis," which suggests that an individual's language reflects the linguistic environment of their youth. By correlating current variant usage with age, they are indirectly assessing past contact patterns.
  • Limitations of Global Models: The low AUC values at the global scale suggest that a one-size-fits-all approach is insufficient. Local variations and interactions are likely masked by averaging across the entire dataset.

Key Results

  • Global Prediction: For over half of the analyzed variants, age was found to be a significant predictor. However, the AUC values were generally low (around 0.5-0.6), indicating limited discriminative power. Variants with fewer users tended to show higher AUCs, but this is often a statistical artifact.
  • Local Prediction: The study mapped the significance of age as a predictor at the local level using k=13 nearest neighbors. While age proved to be a significant predictor in a large number of survey sites for only a few variants, this local analysis revealed patterns not evident in the global analysis. Figure 3, for instance, shows a detailed map of significance and accuracy for a specific variant, illustrating how age's predictive power is spatially distributed.
  • Accuracy vs. Significance: The authors note that logistic regression can sometimes produce statistically significant results with low accuracy, especially with imbalanced datasets (many instances of one variant, few of another). This is a crucial consideration when interpreting the results.

Limitations and Future Work

  • Limitations:
    • Age Alone: The study acknowledges that age alone is not an exceptionally good predictor of syntactic variation, partly due to the slower nature of syntactic change compared to lexical change.
  • Class Imbalance: Logistic regression can be sensitive to class imbalance, potentially leading to misleading accuracy measures.
  • Distance Metric: The use of Euclidean distance for nearest neighbors might not fully capture complex spatial relationships.
  • Future Work:
    • Refining Spatial Models: The authors plan to explore different values of k, incorporate a distance decay approach, and use weights based on parameters like age.
  • Contact Potential: They intend to investigate contact potential metrics beyond Euclidean distance, such as travel time, linguistic gravity (settlement population as weights), and linguistic distance (assuming closer varieties result from closer contact).
  • Advanced Algorithms: Future research will explore more sophisticated algorithms like random forests, SVM, and XGBoost for prediction.
  • Broader Scope: Testing the methodology across different geographic areas and with a larger number of respondents is also planned.

Summary updated

BibTeX
@inproceedings{jeszenszky2019spatio,
  author = {Péter Jeszenszky and Panote Siriaraya and Philipp Stoeckle and Adam Jatowt},
  title = {{Spatio-Temporal Prediction of Dialectal Variant Usage}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
}
Figure