TL;DR
This paper uses causal discovery to untangle the complex web of relationships between word properties like length, frequency, and complexity across 25 languages. They find that while some associations (like word length and frequency) are near-universal, their underlying causal mechanisms are often obscured by unmeasured factors, suggesting that simple correlations don't tell the whole story.
What Problem Does This Paper Address?
Languages, despite their diversity, exhibit surprising regularities in their lexicons. For instance, shorter words tend to be more frequent (Zipf's law of abbreviation), and frequent words are often phonetically simpler. These "lexical trade-offs" are thought to be driven by cognitive pressures for efficient communication. However, the exact nature of these relationships and how they interact remains unclear. Previous studies often look at one trade-off at a time, but this paper aims to understand the entire system of lexical biases and their causal underpinnings across languages. They question whether these trade-offs are universal or if they vary significantly, and crucially, whether the observed correlations imply direct causation or are influenced by hidden factors.
Methods
The core of this paper's methodology is causal discovery, a statistical approach that aims to infer causal relationships from observational data. Specifically, they employ the Fast Causal Inference (FCI) algorithm. Unlike algorithms that assume all relevant variables are measured (causal sufficiency), FCI is designed to handle situations where unmeasured confounding variables might exist. It outputs a Partial Ancestral Graph (PAG), which can represent direct causal links, unmeasured confounders, and uncertainty about relationships. The authors applied FCI to lexical data, treating word length, frequency, phonotactic complexity, and morphological irregularity as variables. They used a bootstrapping procedure to assess the robustness of the discovered relationships across multiple resamples of the data for each language.
Datasets
The study utilizes data from Doucette et al. (2024), which comprises lexical information for 25 languages. This dataset includes measures of:
- Word length: Number of phones.
- Frequency: Log count per million (from Wikipedia).
- Phonotactic complexity: Derived from a neural network model predicting word probability.
- Morphological irregularity: Derived from a neural network predicting inflected word form from its lemma.
It's important to note that the UniMorph data used in this dataset primarily consists of morphologically complex words, which might influence the findings for morphologically simple words.
Key Insights and Ideas
- Correlation is not causation, especially in language: The paper strongly suggests that observed correlations between lexical properties are often influenced by unmeasured factors. This means that simply observing that shorter words are more frequent doesn't automatically mean word length causes frequency, or vice-versa.
- The shadow of the unmeasured: For many fundamental lexical trade-offs (like word length and frequency, or word length and phonotactic complexity), the causal discovery algorithm consistently points to the presence of unmeasured confounding variables. This is a significant finding, implying that our current understanding of these relationships is incomplete.
- Surprisal as a potential culprit: The paper repeatedly suggests that "surprisal" (or predictability in context) might be a key unmeasured confounding variable, potentially explaining the observed associations between word length, frequency, and complexity.
- Universality is debatable: While some relationships (like word length and frequency) appear to hold across nearly all languages, the causal structure behind them, especially the role of confounders, is consistently identified. For other relationships (like frequency and morphological irregularity), the paper finds significant variation across languages, challenging claims of universal trade-offs.
- Lexical biases as constraints: Even without fully identifying direct causal links, the identified relationships and potential confounders act as constraints on what a possible lexicon can look like, regardless of the specific cognitive pressures driving them.
Key Results
- Word Length & Frequency: A strong association is found in all 25 languages, consistent with Zipf's law of abbreviation. However, the causal discovery consistently identifies an unmeasured confounder, preventing a clear causal direction.
- Word Length & Phonotactic Complexity: A universal association is found, but again, with strong evidence of an unmeasured confounder.
- Word Length & Morphological Irregularity: A near-universal association is identified, also confounded by an unmeasured variable.
- Other Relationships (Frequency-PC, Frequency-MI, PC-MI): These relationships show significant variation across languages. In about half the languages, no clear association is found. Where associations do exist, they are often confounded. This suggests these trade-offs might not be universal.
- Graph Diversity: The "most-likely graphs" (PAGs) for individual languages show considerable variation, with only a few languages sharing the same structure.
Limitations and Future Work
- Causal Sufficiency: The primary limitation is that the four variables studied (WL, FR, PC, MI) are likely not causally sufficient. The consistent identification of unmeasured confounders highlights the need for more variables to fully specify the causal models.
- Dataset Bias: The UniMorph dataset primarily contains morphologically complex words, potentially skewing results for morphologically simple words.
- Statistical Assumptions: The FCI algorithm relies on a Gaussian distribution assumption for its independence tests, which may not hold for all variables (e.g., morphological irregularity can be bimodal).
- Future Work: The authors plan to apply causal discovery to larger datasets with more languages and variables to identify more specified causal models. They also suggest exploring alternative independence tests and different causal discovery algorithms to assess their impact. Investigating the role of "surprisal" as a confounder is also a key future direction.
Summary updated