Research On Long Text Semantic Matching Based On Graph Convolution Neural Network

Long text semantic matching task has become a research hotspot in academia and industry in the field of natural language processing due to its complex structure and variable semantics,which makes it difficult to capture the hidden dependencies between entities.When dealing with actual scenes such as breaking news mining,information flow recommendation,etc.,new requirements are also imposed on the real-time and accuracy of the matching algorithm.At present,the industry still widely uses sequence-based recurrent neural network architecture.When applied to long text tasks,there are shortcomings such as a large increase in training time and difficulty in capturing non-linear contextual relationships.In response to this phenomenon,based on the Attention mechanism and graph convolutional neural network,this paper considers the long text semantic space structure and multi-granular semantic representation,etc.,and make optimizations for the input layer,presentation layer and matching layer of the deep semantic matching model architecture.Optimization includes constructing entity association graphs of text,combining convolution feature extraction and Attention mechanism to optimize semantic representation,and fusion matching based on graph convolution introducing dynamic update mechanism.The main work and innovations of the paper are as follows:1.We proposed a method for constructing entity association graph based on keywords and word order features.Compared with serialized text input,using a graph structure to represent long text can store semantic structure information,which helps to mine deep semantic interactions and avoid time-consuming explosive growth.Therefore,based on named entities and ideographic real words,combined with word order and writing features,an entity association graph is constructed as the basis of multi-granular semantic representation.It is found through experiments that compared with the GCN benchmark model,the S-GCN model based on writing structure and word order information is evaluated.The indicators(accuracy and F1 score)have improved by an average of about 2.4%.2.We proposed a graph convolution fusion method that introduces dynamic updates.Since the initial entity association graph weights and connection methods are only determined by the shallow word order and semantics,it cannot well reflect the deep semantic interaction based on context.Therefore,in the process of GCN aggregation,this paper dynamically updates the node and edge information based on the structure of the association graph,and at the same time updates the structure of the association graph with potential edge connections,to mine and improve the interaction between concepts.Through experiments,it was found that the D-SGCN model with dynamic update mechanism improved the average evaluation index by about3.0% compared with the GCN benchmark model.3.We proposed a conceptual semantic representation method based on Attention and convolutional neural network.Due to the characteristics of long text,such as complex semantic structure,the use of convolutional neural networks that support parallel computing and good feature extraction capabilities is the first choice for comprehensive performance and efficiency.However,the convolution feature extraction is limited by the size of the convolution kernel.In order to make up for this shortcoming,we use the Attention mechanism to obtain the semantic interaction of cross-segment sentences,and constructs a concept-oriented global feature.Through experiments,it was found that the Att-D-S-GCN model,which introduced the Attention mechanism,improved the average evaluation index by about 4.4% compared to the GCN benchmark model.