Research On Text-based Tag Recommendation Methods Based On Deep Semantics

With the development of the information age,the amount of data and information on the Internet is increasing,and the searching process is usually inefficient and time-consuming when users search the information they want.Tags are a simple description of information,and recommending appropriate tags for text is an effective means of organizing and using text content.In recent years,tag recommendation has gained much attention from the research communities.At present,the existing studies about tag recommendation are mainly divided into two categories: tag recommendation methods based on traditional semantics and tag recommendation methods based on deep semantics.Tag recommendation methods based on traditional semantics generally conduct text semantic mining based on artificially extracted features,and the performance of the recommendation is not satisfactory.In recent years,with the rapid development of deep neural networks,which can be used to extract text features for mining the deep semantics of text.Furthermore,it can address the limitations existing in traditional text tag recommendation methods.Therefore,it has received more and more attention and has become the hot research topic.At present,most of the tag recommendation methods based on deep semantics mainly utilize the deep semantics in the text to recommend tags for this text.However,the data information does not exist independently.For example,the co-occurrence of words between texts in the corpus can construct a complex structure network.Since this type of methods only considers the text content information and ignores the structure information between the texts,these affects the accuracy of tag recommendation.At the same time,the most frequently used deep neural network models for tag recommendation mainly are convolutional neural networks and recurrent neural networks,and these methods ignore other suitable models,such as language models and pre-trained models.To address the problems mentioned before,based on the existing research work,this paper proposes two methods for tag recommendation.1.We propose a tag recommendation method combining network structure information and text content information.This method firstly uses text and vocabulary as nodes to construct a text heterogeneous graph.Then it utilizes the graph convolutional neural network to extract network structural features between texts and leverages a recurrent neural network model to encode text features to obtain the text sequential semantic features.Finally,the network structure features between texts and the semantic features of text order are combined together for recommending tags by using the attention mechanism.Experimental results show that this method can effectively improve the effect of tag recommendation.2.We propose a tag recommendation method based on Transformer structure.This method leverages the texts and each candidate tags to construct text-tag pairs and encodes these text-tag pairs to capture the basic matching relationships.Then it uses a self-attention mask matrix and a pre-trained large-scale language model to treat the tag recommendation task as a downstream task.Experimental results show that this method is significantly better than most current tag recommendation methods based on deep semantics.