Regularizing translation models for better automatic image annotation