A Method Of Text Classification Based On Feature Fusion

Text classification tasks have received widespread attention in today’s society.Analyzing the categories and sentiments of texts can provide businesses with highquality user attributes and data,improving the practical effects of downstream models and services,and increasing enterprise revenue.With the development of the internet society,various text forms have also brought challenges to text research.Most current research uses deep learning methods to mine key text information,but existing pretrained and neural network models require improvement in analyzing fine-grained semantics,weakening the relationship between opposition and unity among finegrained semantics in the text,and misjudging the sentiment and meaning of the sentence.In order to enhance the model’s ability to analyze the relationship between fine-grained semantics,this thesis designs a dual-stream network to associate contradictory semantics and core sentence segments,and designs feature projection and convolution layers to process the semantic information of contradictory words.The main research content of this thesis can be summarized as follows:(1)Considering the conflict between core words expressing semantics and contradictory words in a sentence,this thesis proposes a Feature Augmentation Network(FA-Net)based on gradient reversal layers and feature projection layers.This network includes general feature extraction modules and backward feature extraction modules.Specifically,the general feature extraction modules include Text CNN,BERT,Ro BERTa;the backward feature extraction module internally stacks a gradient reversal layer to reverse the gradient of the feature vector,causing the auxiliary network parameters to update in the opposite direction of the gradient.The purpose of this is to extract the backward information from the text.The feature projection layer eliminates harmful information in the backward feature vector using feature projection while retaining semantic information related to the general feature vector.Finally,the filtered backward features and normal features are combined to correct the model’s excessive focus on contradictory words and enrich the semantic information of the feature vector.(2)Considering the contradictory relationships between fine-grained semantics in sarcasm detection constitute sarcastic meaning,this thesis proposes a Multi-CLS Fusion Network(MCF-Net)that uses convolutional neural networks as a fusion mechanism.This network includes multi-CLS feature extraction modules and convolutional fusion modules,utilizing different convolutional kernel receptive fields to merge several classification features,thereby enhancing the model’s ability to capture global semantics.Firstly,several CLS features are extracted using the multi-CLS feature extraction module.Secondly,the convolutional fusion module integrates feature vectors to combine the semantics expressed by different feature vectors.Finally,the fused feature vector is used as the network’s ultimate expression feature,allowing the model to absorb more semantic information,improve its understanding of fine-grained semantics,and identify sarcastic meaning by analyzing the relationships between finegrained semantics.This thesis applies the proposed method to three base models and conducts experiments on six datasets.The results show that the method proposed in this thesis achieves higher results on the six datasets compared to the base models.Furthermore,this thesis demonstrates from the perspective of attention visualization that the model can analyze the relationship between fine-grained semantics according to the task.Therefore,this thesis has certain practical value and provides a new approach for text analysis with contradictory semantics.