Training a model with incomplete or imprecise labels, like event captions without exact timing information.