Abstract

How much capacity does Turkish inflection require? An empirical study of GRU encoder--decoder bottlenecks

---

Abstract

How much capacity do neural networks actually need to successfully learn morphophonological tasks? We address this for a simple case of Turkish inflection by training small GRU encoder-decoder models and sweeping an enforced bottleneck dimension. The no-attention design makes the bottleneck a hard information constraint, rendering capacity a measurable property of the task rather than the architecture.

Sequence accuracy follows a sigmoid shape in bottleneck dimension, transitioning between h=16 and h=48, performance at near-ceiling h=64 (roughly 70k parameters total). Logistic regression probes on the bottleneck recover the stem's harmony class [±back] at ≥95% accuracy across *all* bottleneck sizes --- including sizes at which inflection sequence accuracy is near zero --- with [±round] and [±high] lagging behind and converging as capacity grows. An attention ablation confirms the bottleneck's role is load-bearing: attention without context transmission collapses to ~1% accuracy, producing harmonically fluent but morphologically incorrect outputs.

Together, these results establish a tight empirical lower bound on the capacity required for this task and reveal a clear feature-acquisition order that mirrors the structural hierarchy of Turkish vowel harmony.

Keywords: encoder-decoder, computational phonology, RNNs, vowel harmony

How to Cite:

Mailhot, F. "., (2026) “How much capacity does Turkish inflection require? An empirical study of GRU encoder--decoder bottlenecks”, Society for Computation in Linguistics 9(1). doi: https://doi.org/10.7275/scil.4085