AMR Parsing Based On Linearization Strategy And Pointer Network

Abstract meaning representation(AMR)uses a single directed acyclic graph to represent the semantics of sentences.AMR graph is based on the predicate-argument structure,abstracts the predicate and content words in a sentence as nodes,and abstracts the relationship between predicate and argument as directed edge with the core semantic label.If multiple edges point to the same node,they are called reentrant edges.AMR parsing aims to parse natural language sentences into AMR graphs.At present,the mainstream AMR parsing adopts the seq2 seq method which uses neural networks to generate the linearized sequence of the AMR graph.There are the following problems: 1)This kind of method usually uses depth-first traversal to obtain the sequence of the AMR graph.There is a long-distance dependence between the reentrant edge and the pointing node in the sequence,resulting in low reentrancy.2)Predicates have different word meanings in different contexts.The relationship expressed by the core semantic labels between predicates and arguments changes with the change of predicate word’s meaning.Such methods are difficult to deal with the polysemy of labels,resulting in more label errors in the generated AMR graph.To solve the above problems,the main work of this thesis is as follows:(1)AMR parsing based on linearization strategy and BART modelTo solve the long-distance dependence between reentrant edge and pointing node in the linearized sequence,three linearization strategies,namely,canonical,reconfigured and randomized are proposed.These strategies make the original non-adjacent reentrant edges and pointing nodes in the sequence adjacent,and shorten the distance between reentrant edges and pointing nodes,and different strategies get different linearized sequences.The experimental results show that on the datasets AMR bank 2.0 and AMR bank 3.0,the reentrancy of the model with linearization strategy is improved by 3.2% and 2.6% respectively compared with the model without linearization strategy.(2)AMR parsing with pointer network and BART modelTo solve the challenge of label ambiguity,this thesis proposed an AMR parsing method integrating pointer networks to correct the label errors in the AMR graph generated by the baseline model.The pointer network can make the model copy the correct label from the input and correct the wrong label.The experimental results show that on the data sets AMR bank 2.0 and AMR bank3.0,the core semantic labels of the model with pointer network are improved by 0.3% and 0.6%respectively compared with the baseline model.In addition,an ablation study is conducted to analyze the contribution of the pointer network.The results show that the pointer network can improve the performance of the model.