Identify the mechanism underlying palatalization-error reduction

Identify the channel through which Tag-Based Replacement reduces palatalization errors when combined with SpecAugment despite leaving category exposure essentially unchanged, because token frequency does not account for this repair.

Background

The paper finds that Random Replacement reduces tone and palatalization errors introduced by SpecAugment, apparently because uniformly sampling word types increases the training exposure of rare categories. Tag-Based Replacement, however, preserves category exposure while still reducing palatalization errors. This result contradicts the proposed exposure-based account, leaving the mechanism responsible for the repair unidentified.

The authors explicitly defer identification of the responsible channel to future work. The unresolved issue concerns the causal explanation of the palatalization improvement, not whether the empirical reduction occurs.

References

One result falls outside the exposure account: Tag-Based Replacement also reduces palatalization errors ($-24\%$, $p = .041$) despite changing exposure by nothing. Whatever drives this repair, it is not token frequency; identifying the channel is left to future work.

— Error Rate is Not Enough: Costs of Augmentation in Low-Resource ASR  (2610.13807 - Ibaraki et al., 9 Oct 2026) in Section 5.3, “Method Complementarity”