Content Order-Controllable MR-to-Text

Content order is critical in natural language generation (NLG) for emphasizing the focus of a generated text passage. In this paper, we propose a novel MR (meaning representation)-to-text method that controls the order of the MR values in a generated text passage based on the given order constraints. We use an MR-text dataset with additional value order annotations to train our order-controllable MR-to-text model. We also use it to train a text-to-MR model to check whether the generated text passage correctly reflects the original MR. Furthermore, we augment the dataset with synthetic MR-text pairs to mitigate the discrepancy in the number of non-empty attributes between the training and test conditions and use it to train another order-controllable MR-to-text model. Our proposed methods demonstrate better NLG performance than the baseline methods without order constraints in automatic and subjective evaluations. In particular, the augmented dataset effectively reduces the number of deletion, insertion, and substitution errors in the generated text passages.