Neural Semantic Parsing by Character-based Translation: Experiments with Abstract Meaning Representations

We evaluate the character-level translation method for neural semantic\nparsing on a large corpus of sentences annotated with Abstract Meaning\nRepresentations (AMRs). Using a sequence-to-sequence model, and some trivial\npreprocessing and postprocessing of AMRs, we obtain a baseline accuracy of 53.1\n(F-score on AMR-triples). We examine five different approaches to improve this\nbaseline result: (i) reordering AMR branches to match the word order of the\ninput sentence increases performance to 58.3; (ii) adding part-of-speech tags\n(automatically produced) to the input shows improvement as well (57.2); (iii)\nSo does the introduction of super characters (conflating frequent sequences of\ncharacters to a single character), reaching 57.4; (iv) optimizing the training\nprocess by using pre-training and averaging a set of models increases\nperformance to 58.7; (v) adding silver-standard training data obtained by an\noff-the-shelf parser yields the biggest improvement, resulting in an F-score of\n64.0. Combining all five techniques leads to an F-score of 71.0 on holdout\ndata, which is state-of-the-art in AMR parsing. This is remarkable because of\nthe relative simplicity of the approach.\n