Research And Application Of High Quality Content Recognition Algorithm Based On Deep Learning

With the development of the Internet,more and more creators publish articles on social media.How to automatically filter high quality content from a large number of multimedia articles is one of the core functions of information recommendation,search engine,and other systems.However,existing approaches typically suffer from three limitations:(1)They usually model content as word sequences,which ignores the semantics provided by long-distance word dependency,non-consecutive phrases.(2)Since most of the existing methods only focus on text content,they ignore that the content on social media platforms has multi-modal information(such as text,images).(2)They rely on a large amount of manually annotated data to train a quality assessment model while users may only provide labels of interest in a single class for a small number of samples in reality.To address these limitations,we propose a Multi-modal Graph Convolutional Network(MGCN)and a Cross-modality Attention Network(CMAN).In order to capture long-distance word dependencies and non-consecutive phrases in text,we propose a multimodal graph convolution network.Unlike other methods that transform text information into a sequence of words,we model the text information as a graph.Meanwhile,we transform the visual information into corresponding words,and model these words into the graph,so as to realize the semantic complementarity between different modes.In addition,we also use a non-negative risk estimator for identify high-quality content,and use loss back propagation to learn the model.Thus,high-quality content recognition is achieved with few labeled samples.In order to better fuse multi-modal features and capture the inter-modal and intramodal relationships,we propose a cross-modality attention network.First,we use the pre-trained models BERT and Res Net50 to extract coding information for word content and image regions respectively,and then use a cross-modal attention network to extract the correlation of word features and word features,image region features and image region features,and word features and image regions.Thus,a more accurate multimodal feature representation is obtained to achieve high-quality content recognition.In order to better integrate multi-modal features and capture the inter-modal and intra-modal relationships,we propose a cross-modal attention network.First,we use the pre-trained models BERT and Res Net50 to extract coding information for word content and image regions respectively,and then use a cross-modal attention network to capture the relationship between word features,image region features,as well as word features and image regions.Thus,a more accurate representation of multi-modal features is obtained.A large number of experiments on real high-quality datasets verify that our model is better than the most state-of-the-art approaches methods.In addition,we also conducted some experiments on the liar data set(a public dataset of true and false news)to evaluate the effectiveness of our model.