Hard temporal-negative mining for TEMPS training

Develop and evaluate hard temporal-negative mining for TEMPS training, so that negative examples are near-misses in time rather than randomly sampled temporally unrelated expressions.

Background

The training data-generation pipeline creates four random negative samples for each temporal expression. The paper acknowledges that such negatives may be temporally easy because they can be far from the positive example in time, while the evaluation candidate pools contain fixed benchmark-specific distractors. Constructing near-miss temporal negatives is therefore identified as a concrete unresolved extension for improving the temporal supervision used by TEMPS.