Attention-based NMT Models as Feature Functions in Phrase-based SMT