Short Text Classification Algorithm Based On Deep Learning And Graph Learning

With the development of the Internet,especially with the rapid development of various social media such as We Chat,QQ and so on,the number of short Internet texts has been growing rapidly with an explosive trend.Short text classification is an important research topic in natural language processing.Short text classification is widely used in text subject classification,sentiment analysis of user comments,as well as spam detection of website comments and email headers.Short text classification is a fundamental problem in natural language processing,unlike long text,this essay has its own obvious characteristics,mainly including the following: on the one hand,short text,is generally within 150 words and texts,this essay mainly appear in social media,web site comments,post title,and messages in weibo.In addition,the short text in the Internet now has the characteristics of loud noise,fast update,and massive generation.For short text classification,it is mainly divided into rule-based,traditional machine learning,deep learning and graph network learning methods.With the development of deep learning,more and more deep learning algorithms have been successfully applied in short text classification.Deep learning can effectively model and classify short texts.However,due to the sparsity of short texts and the polysemy of keywords,there are variants or deformed words in short texts,so how to classify short texts more effectively is still a big challenge.This paper carries out an in-depth study on the sparsity of short text and the polysemy of short text keywords.The main research work is as follows:(1)To solve the problem of polysemy and sparsity of keywords in short texts,a context-sensitive topic memory network(CS-TMN)is creatively proposed.This network uses context-sensitive word vector representation and global topic knowledge to classify short texts.The0020context-sensitive word vector is mainly composed of local context representation and global topic representation,which can effectively solve the problem of word polysemy in short text.Then,the context-sensitive word vector is used to match the topic knowledge related to the content to solve the problem of short text coefficient.(2)With the development and combination of graph neural network and deep learning,short text classification can also be classified and analyzed by graph neural network.In this paper,a conductive graph convolutional neural network short text classification algorithm(STGCN)is proposed.At the same time,aiming at the sparsity problem of short text,the topic model is used in the construction of the graph,and the relevant topic information is also taken as part of the graph,which can effectively alleviate the sparsity problem of short text.At the same time,the word vector and document vector obtained by STGCN are combined with the word vector generated by the latest large-scale language model BERT.Input the classifier together to get the final classification result.(3)Finally,To further effectively solve the sparsity issue,this paper proposes a text classification method based on gating graph neural network,which takes external knowledge as feature supplement.Specifically,the algorithm first retrieves knowledge from external knowledge sources such as Microsoft Concept Net,Freebase,and Sentic Net.Then,multi-head knowledge attention is used to match the relevant knowledge and reduce the noise of the retrieved knowledge.Finally,the previous researchers' work simply incorporated knowledge nodes into a text graph with dense connections,which would assign equal weights to each connection,inevitably resulting in performance degradation.In contrast,this paper proposes a variant of the Gated Graph Neural Network,which adds an attention aggregation function,which can effectively aggregate neighbor nodes of different importance degrees.In conclusion,this paper mainly studies the short text classification algorithm based on deep learning and graph neural network.Aiming at the problems of sparse features and polysemy of keywords in short text,this paper proposes three network frameworks to effectively classify short text.