Methods Based On Combined Deep Neural Networks For Text Sentiment Analysis

With the advancement of the computer and information technology revolution,the frequent activities of global Internet users in e-commerce and social news platforms have produced massive amounts of imbalanced text data.People are eager to obtain useful information from these text data,and text sentiment analysis method is an efficient solution.Text sentiment analysis methods are currently mature and used in many fields,such as ecommerce review classification,news classification,and public opinion control.However,there are still the following challenges that need to be resolved: First,for the characteristics of document-level data,its constituent semantic characteristics need to be considered,and the semantics between sentences and the dependence between sentence representations are very important for text classification;Second,in view of the characteristics of text data sets,the imbalanced distribution of data sample categories should be considered.In order to solve the above problems,this paper studies text sentiment analysis methods based on combined deep neural networks.The main innovations of this paper are as follows:This paper proposes a document-level sentiment classification(Att DR-2DCNN)method based on attention-based neural network for the structural characteristics of document-level text data.This method first constructs a document matrix representation with a compositional semantic relationship,and at this step,the attention mechanism is used to distinguish the importance between different sentences and words;then a two-dimensional convolutional neural network is applied to capture the important dependencies in the sentence dimension and feature dimension of the document representation,and at this step,the convolution attention module is used to improve the representation of features generated by convolution operation.The experimental results prove the effectiveness of Att DR-2DCNN on the document dataset.Second,according to the characteristics of imbalanced distribution of text data categories in real production environments,an adaptive imbalanced learning method based on memory module for text sentiment classification(AILM)is proposed.This method uses a memory module to remember difficult samples and improves the frequency of difficult samples in later training to adaptively adjust the distribution of text samples belonging to minor class without manual data resampling and sensitive function design,thereby improving the accuracy of text classification.Third,based on the above research,the Chinese text dataset is used to practice the universality of the adaptive balanced text classification learning method based on the memory module,and an optimal Chinese news classification model is constructed by fine-tuning the parameters.