Research On Text Sentiment Analysis Based On Pre-trained Language Mode

Natural language processing is an important direction in the field of computer science and artificial intelligence,and sentiment analysis is an important branch and key task in the field of natural processing.It is widely used in social media,question and answer service and public opinion analysis.It helps people solve various difficult problems and becomes a hot research direction at present.However,this research is faced with difficulties such as flexible text content,diverse expression ways and ambiguous sentences,and traditional sentiment analysis methods require dictionary construction or data annotation,which is time-consuming and laborious and also extremely dependent on domain language knowledge,which greatly hinders the further development of sentiment analysis.Pre-trained language model can complete feature extraction from massive text data with little human intervention,which provides great impetus for the development of text sentiment analysis.In view of the difficulties and challenges faced by current text sentiment analysis tasks at different levels and at different granularity,this paper explores the internal principles of pre-training language model,and summarizes and improves it.The main work and contributions are as follows:(1)In order to solve the problem that most networks can not make full use of the semantic information and correlation information between different channels in sentence level sentiment analysis task,and that traditional networks ignore the word location information of text,a text sentiment analysis model ALBERT-AFSFN combining ALBERT(A Lite BERT For Self Supervised Learning Of Language Representations)and Attention Feature Segmentation and Fusion network was proposed.By introducing ALBERT’s excellent structure design,this network extracts sentence semantic information and location information dynamically.It also integrates the excellent channel feature extraction ability of Split Attention network to effectively mine the semantic information in each sentence channel.At the same time,the excellent feature fusion ability of Attention Feature Fusion network is utilized to effectively fuse semantic information and correlation information between channels.Validation experiments were performed on public datasets Chn Senti Corp,Waimai-10 K and Weibo-100 k,respectively,and the accuracy reached 93.33%,88.98% and 97.81%,indicating that the proposed model has strong applicability and competitiveness in sentence level sentiment analysis tasks.(2)In order to solve the problem that most aspect-level sentiment analysis networks do not make full use of local and global feature information in text,a lightweight ALBERT convolution cascade network ALBERTC-CNN is proposed in this paper.This model can make use of the ability of convolutional network to effectively capture local features and effectively combine with the ability of ALBERT model to capture long-dependent information by self-attention mechanism to extract more fine-grained and comprehensive feature information from text and obtain more accurate sentiment analysis results.It is verified on the laptop and Restaurant review data sets of semeval-2014 open task,and the experimental results show that the classification accuracy of this model is improved by 1.72% and 0.67% respectively on the basis of ALBERT network,which verifies that the proposed model can effectively improve the analysis effect of text emotional orientation.(3)In order to intuitively observe the effect of the proposed model and realize the visualization of text sentiment analysis results,a set of short text sentiment analysis system is developed in this paper.The system can quickly and accurately analyze the user’s needs,and render the analysis results to the front end,realizing the closed-loop analysis and display process of text emotion orientation.At the same time,the system can also carry out self-upgrading and optimization,that is,when the results of model analysis do not meet users’ expectations,the system will automatically bring the correct results of users’ feedback into a new round of model training,so as to update the network parameters,so as to increase its robustness and ensure its service cycle.