| AMR(Abstract Meaning Representation)is a sentence-level semantic representation formalism and is widely used in many downstream natural language processing(NLP)tasks.AMR parsing is the process of converting sentences of natural language into their corresponding AMR graphs,while its parsing performance has a critical effect on its application in downstream NLP tasks.With the recent success of neural networks in NLP,sequence-to-sequence(seq2seq)based AMR parsing has attracted more and more attention.On this basis,this paper explores modeling different kinds of structural information in AMR parsing,including the structural syntactic and semantic information,and the structural information of the AMR graph.The main contents are as follows:(1)Modeling source syntax and semantic role via sequence-aware linearization.Considering the fact that in the absence of linguistic knowledge,seq2seq models for AMR parsing tend to produce results that do not well respect syntax and semantics of the input sentence,this paper proposes an approach to modeling source syntax and semantics via sequence-aware linearization.Specifically,the approach first linearizes the structural syntactic and semantic information of source sentences into sequences,which consists of syntactic labels,semantic roles,and words of original sentences.Then,the new sequences will be fed into the encoder to predict their AMR graphs.(2)Modeling source syntax and semantic role via structure-aware encoding.The transformer uses multi-head self-attention,which enables the encoder to learn a sentence-wide context for every source word.Therefore,to make the encoder aware of sentence syntactic and semantic structure,we propose the structure-aware self-attention to encode structural information.The proposed self-attention model supports an arbitrary number of input features to represent the structural relationships between word pairs.Consequently,we focus on a few well known syntactic and semantic features extracted from parse tree augmented with semantic information,including syntactic path,syntactic distance,and semantic relation.Finally,this paper combines the above structural features to improve the performance further.(3)Modeling target structure information for AMR parsing.The Seq2seq model views the output as a sequence and ignores the inherent and useful structural information.This paper explores modeling target structural information to boost parsing performance.Specifically,in the decoding phase,this paper firstly recovers the structural information from partial results,then applies the structure-aware self-attention to encode the recovered structural information.On publicly available English AMR dataset LDC2017T10,the experimental results show that the proposed approaches in this paper significantly improve the performance of AMR parsing. |
