Research And Implementation Of Text Sentiment Analysis Technology Based On Deep Learning

As the development and popularization of the Internet,online life has been taken for granted.Online social media provides an excellent platform for the public to easily and quickly participate and share their views and opinions.These opinions often contain a lot of emotional information,such as consumer evaluations of products and services and public perceptions of new government policy making.Therefore,using sentiment analysis technology to analyze this information is of great value and significance to the commercial sales of products and publicfacing public opinion analysis.Conventional sentiment analysis methods can be divided into basic dictionary methods and machine learning-based methods.The conventional machine learning methods require a large amount of labeled data in some fields to train the model,so as to achieve better classification accuracy.And the classification model trained on data in one field cannot be applied to other fields.Although the sentiment dictionary method can be applied to the classification of sentiment in different fields,it cannot accurately identify the sentiment expression in a specific field.However,the method based on deep learning does not require a large amount of labeled data to train the model,and the classification model trained from one field of data can also be applied to other fields.This article summarizes the existing sentiment analysis algorithms which based on deep learning and analyzes their shortcomings.The main work is the following four aspects:(1)This paper proposes a sentiment classification model based on the attention mechanism(AM)of a hierarchical gated recurrent unit(GRU).The model is divided into word layer and sentence layer when extracting text feature representations.Each layer uses a GRU model,which has a lighter weight structure than other recurrent neural network models without complicated model parameters.(2)Based on the convolutional neural network(CNN),a fusion model of CNN and bidirectional GRU is proposed.The model first extracts and combines text features of different dimensions through CNN,and then uses bidirectional GRU to capture feature information between long-distance words to better extract the semantic information of the text.(3)Add a layer of GRU or CNN after the embedding layer in the Transformer encoder,and remove the position encoding embedding link.Such a model fusion method not only helps the Transformer to more accurately extract the semantic information of the context,but also allows GRU and CNN to obtain the high-performance component of the Transformer—SelfAttention Mechanism(SM).(4)This paper uses transfer learning to apply a pre-trained Transformer-based Bidirectional Encoder Representation from Transformers(BERT)to cross-domain sentiment analysis tasks.And by adding a layer of GRU to the original model to improve the model,it can extract fine-grained context information and not be more suitable for transfer learning without affecting the calculation efficiency.Experiments prove that the above improved models have achieved better classification performance than the original model.