Datastore refresh dynamics and zero-shot retrieval extensions

Investigate the dynamics of periodically refreshing the retrieval datastore for continuously emerging non-canonical vocabulary and develop zero-shot extensions of the retrieval mechanism for user-generated-text phonemization.

Background

The retrieval datastore is constructed from training-set surface-to-canonical pairs, with an optional LLM-based expansion. Because user-generated-text vocabulary changes continuously, the current approach may lose coverage of newly emerging non-canonical forms. The paper explicitly leaves both the dynamics of datastore refresh and zero-shot extensions of the retrieval mechanism unresolved.

References

UGT vocabulary drifts continuously, so maintaining coverage on emerging non-canonical forms will require periodic datastore refresh; we leave the dynamics of such refresh, and zero-shot extensions of the retrieval mechanism, to future work.

— Phonemizing User-Generated Text: A Benchmark, Taxonomy, and Compositional Approach  (2609.27205 - Jeon et al., 23 Sep 2026) in Limitations section