Simple MAP Inference via Low-Rank Relaxations