Research On Textual Knowledge Transfer Of Cross-domain Social Events

In current society,with the popularization of computers and the development of the network technology,more and more multimedia platforms have emerged.The vast number of internet citizens can more and more conveniently use the network to publish their life experience,interesting events and social events around themselves on these platforms.Therefore,when a hot social event occurs in the world,it can spread rapidly in various media platforms through internet channels.At the same time,a large number of relevant text data can also be produced.The text data upload by these network users are all related to the theme of an certain social event in our society,and the information contained in the text data of these social events plays an important role in economic development of society and government management.In order to extract these important information effectively,we have to collate and analyze these texts of different structure on many different platforms.The classification of these social events text data is one of the most important steps.It would be time-consuming and laborious to classify these event texts manually.Besides,there are a large number of existing multimedia platforms,and the data organization and structure of each platform are different.So it is challenging to classify and process these text data on the platforms.But at the same time,an effective cross-domain text classification model is very important.It can not only reduce the workload of manual text categorization,but also reduce the cost of retraining the traditional text classification model when dealing with classification tasks in different domains.In order to construct the common textualdeep feature space among different domains,we use unsupervised topic model to construct the shared topic feature space among domains,and use the text similarity information in the topic feature space to assist the construction of textualdeep feature space,so as to realize the knowledge transfer of different feature spaces and make the classification model trained in one domain achieve better performance in another domain.This paper can be divided into three sub-tasks.We design and implement specific methods for each sub-task.The corresponding contributions of this paper are as follows:(1)Acquisition and processing of text data about social events: Due to the particularity of the model proposed in this paper,the data available for research on related tasks is very scarce.So in order to obtain suitable training data for the model,we use the crawler technology to capture the text data about social events from Wangyi and Sougou platforms.Besides,some related pre-processing works are performed on the obtained text data,so as to facilitate feature extraction and computation of topic model and cross-domain text classification model;(2)The realization of topic model: The function of topic model is to extract textual topic features from a large number of texts,which is convenient to construct a textual topic feature space.On this basis,we can use the similarity of each text in the topic feature space to restrict the deep feature representation that the text classification model learns,so that the deep feature of the text learned by the text classification model can contain the similarity information of the text in the topic feature space,and ultimately realize the cross-domain transfer of unsupervised textual topic feature knowledge;(3)The realization of cross-domain social event text classification model: The purpose of the cross-domain social event text classification model is to realize the feature learning of event text in different domains and the mapping of textual deep features to text categories.Based on the traditional text classification model,the proposed model is optimized by adding the text topic model.The feature learning of the traditional model is constrained by the topic features learned from the topic model,and the knowledge transfer of the topic features is realized,which improves the performance of the cross-domain event text classification model.