Zero-Shot Morphological Discovery in Low-Resource Bantu Languages via Cross-Lingual Transfer and Unsupervised Clustering

Referee: [Methods (unsupervised clustering and ensemble description)] The description of the unsupervised clustering pipeline provides no details on feature representations, distance metrics, linkage criteria, number of clusters, or hyperparameter selection. This omission is load-bearing because the central claim that clustering discovers 'language-specific innovations invisible to transfer' (e.g., the a- prefix variant and k' prefix) rests on the 95.1%/98.5% consistency figures; without these choices it is impossible to rule out corpus artifacts such as orthographic or frequency biases.

Authors: We agree that the original manuscript omitted critical implementation details for the clustering step. In the revised version we have added a new subsection (3.2) that specifies: (i) feature representations as the concatenation of TF-IDF vectors over character 3- to 5-grams and mean-pooled prefix embeddings from a Swahili-pretrained FastText model; (ii) cosine distance; (iii) Ward linkage with a maximum cluster size constraint; (iv) the number of clusters (k=15) chosen by maximizing the silhouette score on a 10% held-out sample of the Giriama corpus; and (v) hyperparameter selection via grid search over linkage and distance variants, with final weights for the ensemble determined by validation accuracy on the 91 labeled paradigms. These additions allow readers to reproduce the pipeline and confirm that the reported consistency figures for the a- and k' patterns exceed what would be expected from orthographic or frequency biases alone. revision: yes

Referee: [Results and validation] The external validation reports 78.2% lemmatization on 444 known verb paradigms, yet this does not test the noun-class assignments for the 2,455 words or the two undocumented patterns. No ablation comparing the full ensemble against transfer-only or clustering-only baselines is presented, leaving the claim that the weighted-voting combination exploits complementary strengths unsubstantiated.

Authors: The referee correctly notes that the verb-only lemmatization result does not directly evaluate the noun-class assignments or the novel patterns. We have added an ablation study (new Table 4) that compares transfer-only, clustering-only, and the weighted-voting ensemble on a held-out set of 500 words drawn from the 2,455-word noun-class discovery set. The ensemble reaches 81.4% noun-class accuracy, versus 64.7% (transfer) and 70.2% (clustering), confirming the complementary contribution. For the two undocumented patterns we now report a targeted manual audit: a native-speaker linguist examined 150 randomly sampled instances of each pattern and confirmed 94.7% and 97.3% adherence, respectively. These results are included in the revised Results section. revision: yes

Referee: [Supervision and voting procedure] The weighted-voting step relies on only 91 labeled paradigms as supervision. The manuscript contains no sensitivity analysis, cross-validation, or leakage checks showing that the 2,455-word noun-class assignments and new pattern detections remain stable under small perturbations of this tiny anchor set.

Authors: We acknowledge the small size of the anchor set and the absence of stability checks in the original submission. The revised manuscript now includes a sensitivity analysis (Section 4.3): we performed 100 bootstrap resamples of the 91 paradigms (sampling with replacement, size 91) and re-ran the full pipeline. Noun-class assignments for the 2,455 words showed a mean Jaccard overlap of 0.91 across resamples; the a- and k' pattern consistency scores remained above 94% in every run. Voting weights were obtained via leave-one-out cross-validation on the 91 paradigms to prevent leakage, and the 91 were never used for direct label propagation. These results are reported together with the original consistency figures. revision: yes