# LSTM language model for cloze tasks

I am looking for any papers or suggestions of how one may perform training and inference for cloze tasks (fill in the blanks) with multiple blanks in the sentence. In the case of just 1 blank, using an LSTM from both directions would be a good idea, but what about cases with multiple blanks?

After a brief pursuit, the ________ was caught and arrested by the ______.


Update: Some things I have considered are using Seq2Seq, but there is no easy way for a selection of one of the blank words to affect the selection of the other. For example, a input like _____ and ______ should be either answered with cats and dogs or apples and oranges but not cats and oranges.