Research On Neural Network Based Abstract Meaning Representation Parsing

meaning representation(AMR)has become one of the most popular topics in English natural language processing recently.However,due to the limited development time of the task and the complexity of the overall procedure,there is still room for improvement in all parts of the system.The transition-based parser has achieved great results in semantic analysis tasks.In this paper,we propose a system for AMR parsing by investigating the existing transition-based parsers,several aspects have been improved: 1)A new rich resource enhanced AMR aligner has been proposed for aligning concepts in AMR graph to the word or span in natural language sentences;2)A new list-based AMR transition system along with its oracle parser is proposed;3)Embeddings from language models are introduced to our parser.Our aligner produces multiple alignments which are further tuned by the oracle parser to pick the best alignment for downstream parser training.We evaluate our aligner on previous open-sourced AMR system and it outperforms the widely used rule-based aligner in previous work.Our system ensembles multiple models to predict AMR graphs and achieves comparable results to the state-of-the-art AMR parser with only word information.Experimental results show that after we further introduce embeddings from language models,the performance of our final ensembled parser is better than the state-of-the-art AMR parser.