Research On Dependency Parsing Based Semantic Computing Method For Chinese Sentence

Sentence semantic computing aims to transform the word sequence of input sentence into continuous low-dimensional vector.It is the core technique for many tasks of natural language processing,such as information retrieval and question answering system,and therefore becomes the focus of research.Syntactic structure based methods show better performance on downstream tasks.However,there are two main problems.One is the low accuracy of dependency parsing for Chinese and the other is that parallel computing is almost impossible for existing methods to implement.This thesis makes an in-depth study of Chinese dependency parsing model and the semantic computing model integrating dependency parsing.The research results are summarized as follows:(1)Propose a joint model of Chinese word segmentation,part-of-speech tagging and dependency parsing under the graph framework.First,we design a conversion strategy,which converts the part-of-speech tagging into the prediction of dependencies between Chinese characters,to realize the joint learning of the three tasks.We introduce the features of n-grams to strengthen the semantic representation of Chinese characters.Then we design a dependency scoring method based on biaffine attention to obtain the dependency score matrix between Chinese characters.Finally,we designs a decoding algorithm to obtain the annotations of the three tasks.The experimental results on the public dataset CTB7 show that the introduction of part-of-speech tagging can simultaneously improve the accuracy of word segmentation and dependency parsing.(2)Propose a semantic computing graph network model based on the dependency parsing.We consider taking advantages of parallel computing of graph networks.First,we construct graph structure for semantic computing based on the dependency score matrix obtained in work(1),where nodes and edges correspond to the semantic representations of Chinese characters and the scores of dependencies between characters.Then we design a semantic computing method that aggregates the semantic representation of other nodes according to the dependency score and integrates it into the representation of the current node.Finally,the sentence representation is computed by averaging the semantic representations of all nodes.The experimental results on the public paraphrase identification dataset LCQMC show that our model with 81.8% accuracy.is superior to the baseline models and the computing speed is increased by up to 7 times.