Research On Deep Learning Methods For Text Classification Tasks

Text classification is one of the most important tasks in the field of Natural Language Processing.In the era of big data with information explosion,massive text information is constantly generated and pushed to people's hands,but people usually only need the information they are interested in.Therefore,it is very important to classify and manage the massive text information.Based on deep learning method,this thesis focuses on the problem that incomplete consideration of semantic information,label information and Chinese text representation in current popular text classification models,and constructs efficient and applicable classification models for different types of text classification tasks in order to improve the accuracy of text classification.The specific innovation and work content of this thesis are as follows:(1)A text classification model based on recurrent neural network is proposed,which combines entity information with context semantic information.It improves the defects of the existing single label classification model,which can not fully consider the entity information and context semantic information,and can not further enrich the text semantic representation and highlight the text content features.The model first extracts the entity information of the text,calculates the attention of the pre-trained entity representation and text representation to get the most relevant and clear entity representation of the text,and then fully extracts the context semantic features of the text by using the bidirectional recurrent neural network,and obtains the context representation after weight allocation through the attention mechanism,and then stitches the two outputs to strengthen the attention to the entity information and highlight the content characteristics of the text,finally send it to the Highway network to optimize the features,which greatly enriches the semantic representation of the text.The experimental results on the public data set show that the proposed model has improved accuracy compared with other advanced models,and the ablation experiment also shows the rationality and effectiveness of the model.(2)A text classification model based on Attention mechanism and Convolutional Neural Networks is proposed,which improves the existing multi label classification models that fail to fully integrate the relationship between text and tags and insufficient extraction of semantic information.In the model,positional encoding is introduced to emphasize the global order relationship of text words,Convolutional Neural Networks is used to extract the local information of text semantics,Multi-Head Attention mechanism is used to fully learn the global dependence relationship between text words,interactive attention calculation is used to combine with label representation of text content,and adaptive fusion strategy is used to output comprehensive text representation for classification.The experimental results on the public data set show that the proposed model has improved P@K and nDCG@K compared to other advanced models,and the ablation experiment also shows the rationality and effectiveness of the model.(3)A text representation method based on granularity fusion is used to improve the semantic ambiguity of the existing Chinese word embedding representation,and help the two models to improve the classification results in Chinese text classification.It combines word embedding and character embedding to form a new word embedding,which enriches the semantic information of word embedding,and then takes it as the input of the model.The experimental results on Chinese dataset show that the text representation after granularity fusion can effectively help the model improve the classification effect.