| With the rapid development of "Internet plus" and the integration of the three networks,new network technologies have been constantly upgraded,injecting new energy into the current video platform.Bullet screen text has gradually developed into a new form of emotional expression and communication for video users.With its unique advantages,it has gradually become an indispensable way of communication in video platforms and video users.Therefore,it is of great significance to mine text features with emotional color based on emotion analysis model.In this paper,natural language processing is used to integrate and refine the effective information with a certain market value.Based on the web crawler technology,462,000 pieces of bullet screen information from the third season of the cultural and museological variety National Treasure were selected as text data,and 2,500 bullet screen information were randomly selected from each of the 10 issues,a total of 25,000 bullet screen information were selected as data set.Based on Chinese word segmentation and information extraction techniques,this paper explores the quantitative law and distribution characteristics of bullet screen text by using sequence diagram and word cloud diagram from qualitative and quantitative perspectives.To keep the applicability and superiority of traditional machine learning model and deep neural network model,a text sentiment analysis model based on TIL-LSTM is proposed.The new model has the advantages of condensed topic extraction and deep learning classification,providing an effective theoretical framework for sentiment analysis of bullet screen text.The specific work of this paper includes the following three aspects: First,extracting feature words from bullet screen text based on TF-IDF algorithm.The unstructured text feature words are transformed into TF-IDF values and represented by VSM model mapping as vectors.Second,subject extraction was carried out according to LDA atlas.Through the judgment of the evaluation index of confusion and consistency,five optimal topic numbers were finally selected according to the atlas.After model training,the high-frequency words and corresponding weights of each theme were obtained,and the expression of various emotional themes was analyzed based on video content.Thirdly,sentiment analysis of bullet screen text is carried out by TIL-LSTM deep neural network model according to the results of theme extraction.Py Torch’s learning framework was used to build the model,which was divided into test set,validation set and training set to obtain the model classification results.At the same time,the traditional machine learning model based on Bayes,LR and SVM and the recurrent neural network model based on RNN were set up as the control experimental group to compare the results of subject-specific emotion classification.The results show that the three traditional machine learning models have little difference in effect and low efficiency.The accuracy of emotion classification based on RNN is between the traditional machine learning model and til-LSTM model,and the accuracy of TIL-LSTM model is obviously superior to other control models.The sentiment analysis model based on topic extraction shows better performance and has better classification effect. |
