Generating from Abstract Meaning Representation (AMR) is a non-trivial problem, as many syntactic decisions are not constrained by the semantic graph. Current deep learning approaches in AMR generation almost depend on a large amount of "silver data" in general domains. While the text in the legal domain is often structurally complicated, and contain specific terminologies that are rarely seen in training data, making text generated from those deep learning models usually become awkward with lots of "out of vocabulary" tokens. In our paper, we propose some modifications in the training and decoding phase of the state of the art AMR generation model to have a better text realization. Our model is tested using a human-annotated legal dataset, showing an improvement compared to the baseline model.
展开▼
