Research On Leveraging Graph Structure For AMR Parsing

Abstract Meaning Representation(AMR)abstracts the semantic features of a given text into a single-root directed acyclic graph.AMR parsing aims to generate the corresponding AMR graph for a given text.Due to the availability of sequence-to-sequence(Seq2Seq)pretrained models,recent studies on AMR parsing often regard this task as a seq2seq problem,in which the AMR graph is serialized in the reprocessing,then an input sentence is converted into an AMR token sequence,and finally the generated AMR token sequence is recovered to the corresponding AMR graph.However,serialization of an AMR graph into an AMR token sequence tends to lose the structured information.Therefore,this paper aims to improve the performance of AMR parsing by integrating the structured information of the target AMR graph into the seq2seq decoder.In particular,the main work of this paper includes:(1)Modeling history graph structure information for AMR parsing.On the one hand,in vanilla decoder of seq2seq model,the structural relationships between the AMR tokens are not explicitly leveraged.On the other hand,the prediction history of any decoding step can be viewed as an AMR subgraph,which contains the structural information between the current AMR token and others.Therefore,this paper proposes to model history graph structure for AMR parsing.Specifically,this paper firstly defines AMR tokens in the prediction history that have structural relationship with the current input AMR token.Then it uses the graph attention network(GAT)to encode structured information.Finally,the GAT output is integrated into the AMR parsing decoder.Experimental results show that the approach can effectively capture the structured information of AMR graph and improve the performance of AMR parsing.(2)Predicting future graph structure information for AMR parsing.In decoding,the decoder only predicts an AMR token at each decoding time step,which makes the prediction independent from relevant AMR tokens in future.Given that the structured information in prediction history is useful,this paper further propose to predict future graph structure for AMR parsing.Simultaneously learning future structure when predicting an AMR token can enhance the capability of foreseeing future,and thus can help the model to make better prediction.Specifically,at the(t-1)-th time step to predict yt in the training stage,for example,this paper firstly obtain future structural token set St that have structural relationship with yt-1.Then,we learn to predict the future structural token set St under a multi-tasking scheme when predicting yt at the(t-1)-th time step.Experimental results show predicting future structured token set can enhance the semantic representation of the target side and improve the performance of AMR parsing.(3)Leveraging graph structure for cross-lingual AMR parsing.With the progress of English AMR parsing,recently cross-lingual AMR parsing has attracted more and more attention.This paper explores to leverage graph structure for cross-lingual AMR parsing.Specifically,due to the lack of annotated AMR corpora for target languages(such as German),this paper proposes cross-lingual AMR parsing based on unsupervised neural machine translation.Then,on this basis,the structured information of AMR graph is integrated into the AMR parsing task.Experimental results show that the graph-structured information proposed in this paper can significantly improve the performance of cross-lingual AMR parsing.