| Semantic composition is a process of combining word semantics into sentence semantics,and is one of the main techniques to obtain sentence semantic representation.It is widely used in natural language processing tasks such as emotion analysis,machine translation,sentence matching,etc.The principle of compositionality shows that sentence semantics are combined by lexical semantics according to syntactic structure,so semantic composition is closely related to word meaning and syntactic.At present,a few researches begin to focus on the use of syntactic structure information,while most researches just focus on the improvement of lexical semantic representation.However,these structured methods still have the following problems:(1)At present,the analysis accuracy of Chinese dependency parsing on long-distance dependency needs to be improved to support semantic composition;(2)Using external parser,only the information of syntax structure tree provided by it can be used;At the same time,because the training data used by syntactic parsing model and semantic composition model are usually different,the accuracy of syntactic parsing will decline when it applied to semantic composition.In order to solve these problems,we propose a Chinese semantic composition method joint with dependency parsing.Firstly,this paper proposes a top-down Chinese joint dependency parsing model to strengthen the use of global information,so as to improve the accuracy of Chinese dependency parsing.Then,on this basis,this paper constructs a Chinese semantic composition model joint with dependency parsing.A method of utilizing the intermediate information of dependency parsing in the joint model is proposed to improve the robustness of the semantic composition model.At the same time,the dependency parsing model and semantic composition model are optimized cooperatively to reduce the influence of data inconsistency.The main work and contributions of this paper are summarized as follows:(1)Propose a top-down Chinese joint dependency parsing model.The existing joint dependency models mainly use Shift-Reduce algorithm,which makes decisions only based on the transfer actions of adjacent nodes in each step.Thus,it is difficult to make full use of global information,so there is still space for improving long-distance dependency parsing.To solve this problem,this paper proposes a top-down Chinese joint dependency parsing model,which uses the Encoder-Decoder model as a framework.We use the attention mechanism to predict the dependency nodes,by select the optimal node among all candidate nodes to construct the dependency relationship,to improve efficiency of utilizing global information.The experimental results on the public data set CTB5(Penn Chinese Treebank 5.0)show that the performance of the proposed method in Chinese word segmentation,part-of-speech tagging and dependency parsing tasks are significantly better than those of the existing models using Shift-Reduce algorithm,with F1 values reaching 98.25%(+0.01%),95.13%(+0.64%),and 85.44%(+3.79%),respectively.The design experiments verified the effectiveness of the proposed method in long-distance dependency parsing.(2)Propose a Chinese semantic composition model joint with dependency parsing.Aiming at the problems in the existing structured semantic composition methods,we propose a Chinese semantic composition model joint with dependency parsing on the basis of(1).Firstly,this paper proposes a semantic composition model based on attention.The dependency syntax tree is used as the graph in graph attention calculation,and the semantic of each node is composed according to its child nodes.Secondly,a method of utilizing the intermediate information in dependency parsing is proposed.The semantic information as the head node in the dependency relationship is introduced into the semantic composition model,to reduce the influence of prediction errors of the dependency parsing model and improve the robustness of the semantic composition model.Finally,through the joint learning of dependency parsing and semantic composition,the dependency parsing model is subject to domain adaptation to improve its robustness.We apply the semantic composition model to the paraphrase recognition task.The experimental results on the Chinese paraphrase recognition dataset LCQMC show that the prediction accuracy and F1 value of the proposed method in the paraphrase recognition task reach 76.81% and 78.03% respectively.We further design experiments to verify the effectiveness of joint learning and intermediate information utilization,and compare our model with representative work.This paper proposes a top-down Chinese joint dependency parsing algorithm,and constructs a Chinese semantic composition model joint with dependency analysis,to solve the problems existing in Chinese semantic composition.The effectiveness proposed models are verified through the comparative experiments on public data sets. |
