Research On Sentiment Analysis Of Shooting Mobile Games Reviews Based On Deep Learning

With the development of the big data era,sentiment classification of review texts has become a research hotspot.The quality of a sentiment dictionary affects the accuracy of sentiment classification,and there are different sentiment dictionaries in different fields.Note that new word discovery is helpful for the construction of a sentiment dictionary.Shooting mobile games have many users and their review texts have many new words,so this thesis attempts to design a new word discovery algorithm to build a dictionary of shooting mobile games review domain,and then filter out the sentiment dictionary for the shooting mobile games review domain from it,and perform sentiment classification based on the dictionaries and rules.In addition,this thesis also tries to build a deep learning model to classify the sentiment of shooting mobile game review texts and further explores the classification effect of the deep learning model integrated with the domain dictionary.This thesis works on these two aspects as follows:In terms of building domain dictionaries and sentiment dictionaries,after mainly extracting candidate new words based on the N-Gram method and filtering out candidate new words with statistical features such as PMI that are less than the threshold.this thesis constructs a domain dictionary for shooting mobile game reviews.Next,based on word vectors.this thesis uses the modified sum of similarity method and the maximum similarity method to filter out the domain sentiment dictionary from the domain dictionary.The results show that the new word discovery algorithm designed in this article can effectively identify new words in the field of shooting mobile games.Meanwhile,by comparing the results of sentiment classification,this thesis found that the domain dictionary and the domain sentiment dictionary can improve the accuracy and other performance indicators of sentiment classification,and the impact of the domain sentiment dictionary constructed by the two methods on sentiment classification performance is not significantly different.In terms of deep learning modeling,this thesis focuses on the influence of two factors,the text vectorization method and the neural network algorithm,on the effect of sentiment classification.Considering two text vectorization methods,Word2vec and Bert,and two neural network algorithms,TexCNN and BiLSTM,this thesis designs a two-factor factorial design with replications.The experimental results show that the text vectorization method,the neural network algorithm and their interaction all have a significant impact on the accuracy,precision and F1 value of sentiment classification,but the recall rate of sentiment classification only is affected by the text vectorization method.In addition,replicated experimental results show that the Bert-TextCNN model has the best sentiment classification effect,and the deep learning model integrated with the shooting mobile game domain dictionary cannot improve the classification effect.