Constructing meaningful negative examples for source–grounding consistency

Identify a principled procedure to construct meaningful negative examples for a binary classification formulation of translation that labels source–grounding pairs (S, z) as consistent or inconsistent, specifically defining what grounding data should constitute inconsistency with a given source communication so that the reduction is well-posed for learning and evaluation.

Background

The paper discusses whether active learning insights from binary classification can be transferred to translation by classifying pairs of source communications and grounding data (e.g., observations) as consistent or inconsistent. While positive examples arise naturally from observed pairs, the construction of negative examples is non-trivial.

Without a principled way to generate inconsistent grounding for a given source, the binary reduction lacks a meaningful supervision signal for the negative class, undermining both theoretical analysis and practical learning setups. Resolving this would establish the feasibility of binary-consistency formulations for translation tasks.