Training regime using noisy, partial, or indirect labels like heuristics and distant supervision to scale datasets when full annotation is expensive.