I'm reading this paper for sub-character decomposition for logographic languages and the authors mention decomposition at inference-time. They're using Transformer architecture.
More specifically, the authors write:
We propose a flexible inference-time sub-character decomposition procedure which targets unseen characters, and show that it aids adequacy and reduces misleading overtranslation in unseen character translation.
What do inference-time and inference-only decomposition mean in this context? My best guess would be that inference-time would be at some point during the decoding process, but I'm not 100% clear on whether that's the case and, if so, when exactly.
I'm going to keep digging and update if I find something helpful. In the meantime, if anyone needs more context just let me know.