Research On Sentiment Analysis Method Of Multi-feature Fusion Based On Attention

With the advent of the information age,the Internet has changed the way people express their opinions,people have started to actively publish their opinions and comments.Under this trend,more and more text resources have appeared on the Internet.By mining and analyzing the emotional information in these text resources,people can understand people's feelings about other people and things,which has very important practical application significance.The methods currently used in sentiment analysis include sentiment word-based method,machine learning-based method,and deep learning-based method.Among them,the deep learning-based method has become the most popular method of sentiment analysis at present because of its autonomous learning ability and its advantages on large-scale data.Common deep learning methods include recurrent neural networks,convolutional neural networks,and long short-term memory models,but these models are based on supervised learning.The effect of model training is closely related to the scale and quality of labeled data.How to obtain more information in limited labeled data has become an urgent problem.This paper uses bidirectional long short-term memory models and convolutional neural networks to extract temporal and local features of data.Aiming at the problem of limited labeled data in the field of sentiment analysis,this paper proposes an attention-based multi-feature fusion sentiment analysis method.The self-attention mechanism is added to the feature extraction process of the long short-term memory model,calculating the weight based on the distribution of attention to determine the amount of information in the current state in the area,thereby giving different attention and extracting the information in the data deeper.In the process of bidirectional long short-term memory model and convolutional neural network model fusion,the attention mechanism is also added to adjust the model's focus,and the local feature vector is compared with the time-series feature vector at different times to give similar attention and give different attention,mining the connection between the two to better integrate the two features,and strive to maximize the use of limited labeled data to obtain better feature extraction results.In addition,a priori knowledge is introduced into the model,and the a priori emotion vector generated by the emotion dictionary is spliced on the input vector to make up for the problem that the deep learning model only focuses on the semantic information of the data and ignores the emotional information.In this paper,we perform comparative experiments on different combinations of convolutional neural networks and bidirectional long short-term memory models,and present the effect on the commonly used text binary classification evaluation indicators.The results show that the attention-based multi-feature fusion sentiment analysis method proposed in this paper improves the accuracy,precision,recall and F1 score to a certain extent,proving the validity and practicability of the method.