TL;DR
This paper investigates how two measures of morphological complexity—i-complexity (predictive structure) and e-complexity (size and diversity of forms)—affect language learning. Using both neural networks and human participants, they found that e-complexity has a stronger impact on learnability than i-complexity, challenging previous hypotheses that prioritized predictive structure.
What Problem Does This Paper Address?
Languages exhibit vast differences in their morphological systems, from those with minimal inflection to those with extensive marking of grammatical information. A key question in linguistics is how these differences in complexity influence the ease with which humans acquire and process these systems. Previous work by Ackerman and Malouf (2013) proposed that while languages vary widely in "surface" complexity (e-complexity), they tend to have low "predictive" complexity (i-complexity), suggesting a pressure for low i-complexity in language evolution. This paper aims to test this hypothesis by directly examining how i-complexity and e-complexity independently affect learnability, and whether learners are more sensitive to one over the other.
Methods
The researchers employed a multi-pronged approach:
- Artificial Language Learning: They created four artificial nominal inflectional paradigms, systematically manipulating i-complexity and e-complexity while keeping other factors constant (number of markers, inflectional classes).
- Neural Network Simulations: Long Short-Term Memory (LSTM) networks were trained on these paradigms. LSTMs are a type of recurrent neural network well-suited for sequence processing and are used here as a model of an "ideal learner."
- Human Behavioral Experiments: Human participants were recruited and trained on the same artificial paradigms using a trial-and-error learning task.
- Statistical Modeling: Linear mixed-effects regression models were used to analyze the learning data from both LSTMs and humans, assessing the significance and effect size of i-complexity, e-complexity, and other factors like syncretism.
- Random Paradigm Generation: To explore the relationship between i- and e-complexity more broadly, 1000 random paradigms were generated and analyzed for their complexity values and learnability by LSTMs.
Datasets
The study did not use pre-existing linguistic datasets. Instead, they created four artificial nominal inflectional paradigms for the experiments with LSTMs and human learners. These paradigms were designed to isolate the effects of i- and e-complexity. Additionally, 1000 randomly generated paradigms were created to explore the relationship between complexity measures across a wider range.
Key Insights and Ideas
- E-complexity Reigns Supreme: Contrary to the hypothesis that low i-complexity is the primary driver of learnability, the study found that e-complexity (related to the size and diversity of forms) had a significantly larger impact on both neural network and human learning.
- I-complexity's Limited Role: While i-complexity did show some effect in neural networks, its impact was weaker than e-complexity. For human learners, the effect of i-complexity was largely absent.
- The Interplay of Complexity: The study revealed a strong negative correlation between i-complexity and e-complexity in randomly generated paradigms. This suggests that high e-complexity often co-occurs with low i-complexity, and vice versa, which might explain some of the patterns observed in natural languages without necessarily invoking specific evolutionary pressures.
- Syncretism Matters (for LSTMs): The type of syncretism (how different grammatical functions share the same form) affected learning in LSTMs, with cross-class syncretism leading to slower learning. However, this effect was not observed in human learners.
- Frequency is Key: The strong effect of e-complexity suggests that the frequency and distribution of forms within a paradigm play a crucial role in learnability.
Key Results
- LSTM Learning: Networks trained on low-i/low-e paradigms learned fastest, followed by high-i/low-e. Low-i/high-e paradigms showed the slowest learning. The effect size of e-complexity was approximately four times larger than that of i-complexity.
- Human Learning: Human participants showed a significant effect of e-complexity, with low e-complexity paradigms being learned better. No significant effect of i-complexity was found.
- Random Paradigms: A strong negative correlation (r = -0.92) was found between i-complexity and e-complexity across 1000 randomly generated paradigms. Paradigms with high e-complexity tended to have low i-complexity, and vice versa.
- Syncretism Effects: In LSTMs, paradigms with across-class syncretism were learned slower than those with within-class syncretism. This effect was not observed in human learners.
Limitations and Future Work
- Limited i-complexity Range: The study notes that the differences in i-complexity between their low- and high-complexity paradigms were relatively small (0.222 bits). They suggest that larger differences might reveal a more pronounced effect, but this would require training participants on much larger, potentially multi-day, experiments.
- Adult Learners: The human experiments focused on adult learners, making the findings most relevant to adult second language acquisition. The authors acknowledge that child learners or first language acquisition might differ, and suggest this as a future research direction.
- Syncretism Nuance: While syncretism type affected LSTM learning, it did not in human learners, calling for more systematic investigation into syncretism's role.
- Generalization vs. Learning: The study's findings suggest that i-complexity's primary role might be in facilitating generalization to novel forms, rather than general learning. Future work could further disentangle these effects.
- Natural Language Complexity: The study used simplified artificial paradigms. Future work could explore these complexity measures in more complex, natural language paradigms.
Summary updated