A provable SVD-based algorithm for learning topics in dominant admixture corpus