Text Sentiment Analysis Based On BiSRU Fusion Attention Mechanism

With the increasing development of the Internet and mobile technology,social media and e-commerce platforms are becoming more and more common.While providing people with channels to publicly express their personal opinions and views,they also produce a large number of comment texts with personal emotional colors.As a typical network news text,it has the characteristics of large information capacity,strong timeliness,clear views,unrestricted and strong subjectivity.As an important branch of machine learning,sentiment analysis helps people better understand users ’ attitudes,emotions and opinions by analyzing the emotional information hidden in text data,and then is used in product promotion,public opinion monitoring,marketing and other fields.With the rise of deep neural network,the sentiment analysis model based on deep neural network has achieved good results.However,the existing methods are inefficient in processing largescale text review data,and cannot quickly extract features and identify text sentiment polarity.At the same time,the existing models only focus on the representation of each position in the input sequence,without considering the interaction between different positions.In view of the above problems,this paper studies the pre-training model,recurrent neural network and attention mechanism and proposes two text sentiment analysis models under different granularity.The main contributions of this paper include :(1)Aiming at the problem that the feature extraction ability of basic modules such as CNN and BiLSTM is insufficient,and the key features that have great influence on the classification results cannot be identified,and the recent excellent deep learning model lacks the learning of text context semantic features at different scales,a Chinese sentence-level text sentiment analysis model composed of Chinese BERT model,multi-channel fusion feature network secondary semantic learning layer and emotional category linear output layer is proposed.Among them,the Chinese BERT model adds unique font and pinyin information to the pre-training process,and enhances the understanding of semantics and grammar by integrating character semantics,font and pinyin feature information;the secondary semantic learning layer of multi-channel fusion feature network is composed of multi-channel convolutional neural network,bi-directional built-in attention simple loop unit and soft attention mechanism,and uses residual connection architecture to enhance model expression ability.The experimental results show that each module has a significant positive effect on the performance improvement of the model,making the feature extraction more comprehensive.The highest F1 scores of 76.58 % and 97.59 % were obtained on the SMP2020 Weibo epidemic sentiment classification data set and the shopping comment data set,respectively,higher than the recent excellent experimental comparison deep learning model.(2)In order to comprehensively consider the emotional level of words in different aspects of a sentence from a more fine-grained perspective,and strengthen the learning of the relevance between the target vocabulary and the context,an interactive attention neural network model based on bidirectional simple recurrent units is proposed.Through interactive learning of the semantic representation of context and target words,the information fusion of context and target words is realized to highlight the key information.The experimental results show that the proposed model can effectively identify the sentiment tendency of comment text,and achieve the accuracy of 80.3 %,73.3 % and 73.6 % on the three data sets,respectively.