Sequence Models For Constituent Parsing

Constituent parsing is an important task in natural language processing,aiming to parse the phrase structure tree of a given sentence.At present,the mainstream constituent parsing methods can be divided into three categories.One is transitionbased methods.The second is based on dynamic programming,and the third is based on linearizations of constituent trees.These three methods have their own advantages.Method one and method three are very efficient,but the effect is bad,and method two is the opposite.Therefore,the main content of this paper is to find a better balance between efficiency and effect based on sequence representation.The first work of this paper proposes a constituent tree decoding method based on in-order traversal.This work performs greedy decoding in the order of in-order traversal,which reduces the(!)time complexity of the original dynamic programming method to(log),where is the length of the input sentence.At the same time,the decoding algorithm based on in-order traversal combines the advantages of the algorithms based on pre-order and post-order traversal,which can further improve the effect of the model to a certain extent.In addition,by predicting the in-order traversal sequence,it can also make full use of the historical decision information that has been predicted to improve the effect of subsequent sequence prediction.The final experimental results show that this method outperforms dynamic programming-based methods in efficiency and slightly increases the effect.The second work proposes a span-based sequence representation method of constituent trees.It proposes that only need to predict all the left child spans in the constituent tree to restore the entire tree.In this way,all the left child spans can be encoded and represented as a series of integer sequences equal to the sentence length.Through the attention mechanism and the fast local normalization method,the final constituent tree can be predicted with a high degree of parallelism,and the time complexity is only(log).The final experimental results show that this method can use the high efficiency of the sequence tagging methods to achieve the same best effect as the dynamic programming decoding methods.The third work proposes a structural representation method of constituent tree based on graph neural network.The node count and structure of constituent tree are unknown before prediction,so it is difficult to model with graph neural network.However,using the sequence representation method proposed in the second work,the non-leaf nodes of the constituent tree can be ignored,and an approximate fully connected graph can be constructed for all leaf nodes.Then the graph attention network is used to encode this graph to enhance the representation of nodes.The final experimental results prove that the use of graph neural network can well encode the connection between adjacent phrases,further improve the model effect,and finally surpass the traditional dynamic programming decoding methods.