| With the explosion of big data information,the amount of scattered and redundant news text data on the Internet increases rapidly.News classification can integrate and efficiently present massive fragmented network information,so as to achieve news simplification,refinement and effective governance.As a brief summary of the news,the headline can quickly help readers understand the core content of the news.However,the title length is short and the text features are sparse,so it is difficult to obtain the best performance of classification only by extracting text word features.Therefore,how to effectively extract various deep features of news headlines,so as to improve the performance of text classification model for Chinese news title,is of great significance for related news research.This paper mainly uses deep learning knowledge to propose two kinds of news title classification methods,and constructs a UCL news title classification prototype system.Specifically,the main work of this paper is as follows:(1)Aiming at the problem that it is difficult to obtain comprehensive text features by simply modeling the syntactic information of the title,this paper proposed a news title classification method based on syntactic and semantic enhancement.The core idea is to enhance the dependency syntactic structure through semantic features,starting from the syntactic information which contains important clues of the title category.Firstly,the syntactic information of the title was obtained by dependency parsing.Secondly,the graph attention network was used to encode the syntactic information and semantic information of the title respectively.Finally,after enhancing the information interaction between the two sides through the dual representation interaction module,the adaptive attention operation is performed respectively to fuse them into a single sentence representation,which is sent to the fully connected layer and Softmax layer for classification.Experiments on two public data sets show that the classification effect of the proposed method is better than that of the comparison model.(2)Aiming at the problem that traditional multi-feature fusion methods such as concatenation or dot product are difficult to dynamically distinguish features that are more beneficial to the classification task,a news title classification method based on multi-feature selection mechanism was proposed.The core idea is to start with the deep feature extraction and dynamic screening of news headlines,and give higher weights to the feature categories that are helpful for current classification.Firstly,the cascade graph attention module was used to extract the syntactic features of the title.Secondly,the cascaded iterative dilated convolution module was used to extract the n-gram features of the title,and then the multi-kernel attention module was used to extract the context features of the title.Finally,the gated mechanism was used to screen and fuse the three types of features obtained,and the fused feature vector was sent to the fully connected layer and Softmax layer to obtain the probability distribution of each category.Experimental results on two data sets show that the proposed method further improves the effect of text classification after adding multi-feature categories and filtering mechanism.(3)This paper designs and implements a UCL news title classification prototype system based on Django framework,which includes three functional modules:system environment configuration module,news title classification module,background management module,and they are responsible for the configuration of classification model and server,the classification of news title text input by users,the management of background news data and account.The actual operation test shows that the system can realize the classification of news title text and the visual display of results. |
