Sentiment Analysis Of Short Text Based On Improved Bidirectional LSTM Neural Network

With online shopping experience,a large number of emotional comments,microblogs and other short texts are accompanied.Whether it is from the perspective of consumers or businessmen to formulate consumption and sales strategies,or from the perspective of monitoring public opinion by government departments,or from the perspective of more effective dissemination of information by news and other media,the study of methods and models for quickly learning the emotional orientation of these co the updating and development of computer and Internet technology,more and more social and commercial platforms have emerged.Faced with the hotspots of public opinion andrpuses not only has practical value,but also has theoretical significance.The related research field is called emotional analysis.The theoretical methods of emotional analysis can be divided into rule dictionary,machine learning and neural network model.In order to make the research more representative,this paper mainly uses improved emotional score(modified emotional score),support vector machine,bidirectional BLSTM neural network and its improved model for more than 60,000 commodities.Short text reviews are used for emotional analysis.Through the study of relevant academic literature,we find that there are several problems in this field:(1)traditional emotional score symbols can discriminate emotional polarity,but the value changes randomly with the length of the text,so basic statistical learning can not discover the meaning of emotional score numerical part;(2)word vector as the input of most machine learning and neural network models,most studies use correlation level.On the platform,a word vector library based on certain corpus training is used as input,but the generalization ability of this method needs to be studied when it is used for specific research;(3)Most of the self-trained word vectors are calculated by equal weight method,ignoring the influence of word frequency and inverse text frequency index on the weight of each word vector,and the experimental input is not perfect.In view of the above problems in the field of emotional analysis,the following work has been done in this paper:(1)To study and sort out the relevant literature,put forward some ideas for improvement,and elaborate the model theory used in this paper;(2)Considering the emotional score of traditional rule dictionary,the relationship between the emotional polarity of text and the length of text,and the text of commodity review have been neglected.This paper improves the relationship between the length of data and emotional intensity by two steps,defining the average length and introducing modified emotional score to replace the traditional emotional score;(3)Pre-processing more than 60,000 corpus by adding positive and negative emotional dictionaries to the standard lexicon of word-cutting package,removing stop words,and using Word2 in the Python language package.Vec function carries out equal weight word vector training and TF-IDF weighted word vector training as input of support vector machine and bidirectional long-term and short-term memory neural network model;(4)Introduces the evaluation system of the empirical part of this paper,and introduces the key functions in each model programming,and carries out modified emotional score,support vector machine,bidirectional long-term and short-term memory neural network and its improved model together.The F1 values of the four models are 0.80,0.85,0.87 and 0.89 respectively.The experimental results show that the improved two-way long-term and short-term memory neural network model has certain advantages in the function of affective analysis.Finally,the research results,problems and imperfections in the research process are summarized and prospected.