Ambiguity-Aware Inference for Multi-Valid Structured Generation

This research reformulates structured generation from deterministic mapping to an ambiguity-aware inference problem. By adopting a multi-valid learning framework, the method preserves multiple semantically valid interpretations instead of collapsing outputs to a single reference during training.