Towards robust cross-linguistic comparisons of phonological networks

Philippa Shoemark, Sharon Goldwater, James Kirby, Rik SarkarSIGMORPHON 2016 · DOI

TL;DR

This paper tackles the challenge of comparing phonological networks across languages, which are often confounded by factors like lexicon size and word length. They propose a novel method using matched pseudolexicons to isolate true cross-linguistic differences and find that while many "small-world" properties of these networks are artifacts of their construction, some interesting variations in how network statistics change with lexicon size hint at deeper linguistic differences.

What Problem Does This Paper Address?

The mental lexicon, our internal dictionary of words, is often modeled as a network where words are nodes and edges connect words that are phonologically similar (differing by one phoneme). Researchers have used network science to study these phonological networks, looking for universal properties and cross-linguistic differences that might explain variations in language processing. However, previous studies have often compared networks of vastly different sizes and with different word length distributions, making it unclear whether observed differences are due to genuine linguistic properties or simply artifacts of the data. This paper aims to address this by developing a more robust method for cross-linguistic comparison that accounts for these confounding factors.

Methods

The core of the paper's methodological contribution lies in its approach to controlling for confounding variables. Instead of directly comparing real lexicons, the authors propose:

  • Matched Pseudolexicons: For each real lexicon, they generate multiple random "pseudolexicons." These pseudolexicons are designed to match the real lexicon in terms of size, phoneme inventory, and crucially, word-length distribution. This allows them to isolate the effect of genuine linguistic properties from the effects of these structural factors.
  • Range-Based Comparison: Rather than relying on single point estimates for network statistics, they analyze how these statistics change across a range of lexicon sizes. This helps to understand the dynamics of network properties as lexicons grow, rather than just their static values.
  • Network Statistics: They employ standard network science measures:
  • Average Shortest Path Length (ASPL): The average minimum number of steps to get between any two nodes in the network.
  • Average Clustering Coefficient: A measure of how "cliquish" a node's neighborhood is.
  • Degree Assortativity Coefficient: Measures whether nodes tend to connect to other nodes with similar degrees (number of neighbors).
  • Fraction of Nodes in the Giant Component: The proportion of nodes in the largest connected part of the network.

Datasets

The study analyzes phonological networks derived from lexicons of eight languages: English, Dutch, German, French, Spanish, Portuguese, Polish, and Basque. For each language, they used both lemma-only lexicons and lexicons including inflectional variants (all wordforms), where available. The sources for these lexicons and their pronunciations are detailed in Table 1.

Key Insights and Ideas

  • "Small-World" Properties are Often Artifacts: The paper strongly supports the idea that the "small-world" characteristics (high clustering, relatively short path lengths) and high degree assortativity observed in phonological networks are largely a consequence of how these networks are constructed (connecting similar strings of phonemes), rather than evidence of deeper cognitive organization or specific growth processes. This is demonstrated by the fact that their matched pseudolexicons exhibit similar properties.
  • Lexicon Size Matters (and Varies): Network statistics are demonstrably affected by lexicon size. Crucially, the way these statistics change with size can differ across languages, suggesting that simple comparisons of networks of different sizes are misleading.
  • Word-Length Distribution is a Major Confounder: The distribution of word lengths significantly impacts network statistics, particularly the fraction of nodes in the giant component. This highlights the importance of matching this property when comparing languages.
  • Cross-Linguistic Differences Emerge in Dynamic Behavior: While many static properties are similar or explained by structural factors, interesting cross-linguistic differences appear in how network statistics (especially the fraction of nodes in the giant component) evolve with lexicon size. This dynamic behavior, particularly how less frequent words behave, could be linked to differences in language processing.

Key Results

  • Pseudolexicons Mimic Real Networks: The phonological networks derived from random pseudolexicons (matched for size and word-length distribution) exhibit many of the same "small-world" properties and high degree assortativity as real language lexicons. This weakens claims that these properties indicate unique linguistic organization.
  • Giant Component Size Varies Significantly: The fraction of nodes in the giant component differs substantially between languages, even in pseudolexicons, largely due to word-length distributions. However, real lexicons often have larger giant components than their matched pseudolexicons, suggesting that phonotactic constraints in real languages make words more similar than random sequences.
  • Dynamic Behavior of Giant Component: In most languages, the fraction of nodes in the giant component decreases with increasing lexicon size (as more infrequent words are added), implying infrequent words are phonotactically unusual. However, in Spanish, Portuguese, and Basque, this trend is reversed or less pronounced, suggesting less frequent words are more similar to others in these languages. This finding directly illustrates the hypothetical scenario presented in Figure 2, showing how comparing at a single lexicon size could lead to opposite conclusions.
  • ASPL Differences: While ASPLs are generally similar between real lexicons and their pseudolexicons, some languages (Spanish, Polish) show consistently higher ASPLs in their real lexicons. The paper also notes that the difference in ASPL between real and random lexicons can be influenced by morphological variants, particularly in Polish.

Limitations and Future Work

  • Methodological Complexity: The proposed method of comparing differences across a range of lexicon sizes using pseudolexicons is more complex than direct comparisons but is necessary for robust analysis.
  • Incomplete Data: The authors note limitations in obtaining phonemic transcriptions and frequencies for certain languages (e.g., Portuguese inflected wordforms, Spanish lemmas).
  • Future Work: The authors hope their findings will inspire further work to investigate the connections between these network properties and language processing differences, and to extend the analysis to additional languages. They also suggest exploring the impact of morphological variants more deeply.

Summary updated

BibTeX
@inproceedings{shoemark2016towards,
  author = {Philippa Shoemark and Sharon Goldwater and James Kirby and Rik Sarkar},
  title = {{Towards robust cross-linguistic comparisons of phonological networks}},
  booktitle = {SIGMORPHON},
  year = {2016},
  doi = {10.18653/v1/w16-2018},
}
Figure