Research On Video Mining Based On Bullet Text

With the further development of science and technology,video is fully integrated into people’s daily life.Video has become an important way of entertainment in people’s daily life,and constantly changes people’s contact with the world,communication mode and integration trend.At the same time,the prevention and control of the new coronavirus have accelerated the development of network and video at all levels of society and in all fields,making the technology,content and platform of video more socialized.Moreover,with the massive growth of video files,as an important function and main interactive mode of video,the bullet screen is deeply loved by users.Massive video files produce massive bullet screen text.The bullet screen has gradually become the main source for users to obtain video information,and also provides important text data for researchers to analyze video.Under the guidance of the overall national security concept,cyberspace security has become an important issue of public concern.To maintain cyberspace security,it is necessary to maintain the security of network content and network communication.With the vigorous development of online video and live broadcasting,while enriching the way of mass entertainment,it is also accompanied by security risks,and the spread of bad information is becoming more and more intense.There are a lot of bad information in the massive bullet screen data,and bullet screen has become one of the important ways to spread bad information.Bad bullet screen mainly adopts the way of keyword variant to avoid system detection,and different from the traditional video review text,the bullet screen text has the characteristics of colloquialism and contains many network terms and expressions,which contains rich semantic information and emotional information.Therefore,the analysis of bullet screen data focuses on mining its semantic information and emotional information combined with video.In order to ensure the reliability and authenticity of the analysis results,it is necessary to preprocess the bullet screen data,filter the bad bullet screen and ensure the quality of the bullet screen data.Based on the above background,this paper explores the video shot text processing method by grasping the video shot text,and excavates and analyzes the video shot text.The main work and contributions are as follows:(1)Select the typical bullet screen video website Bilibili as the empirical analysis sample source,grab the bullet screen text data.In order to solve the noise problem of bullet screen text,the bullet screen text is preprocessed,and all kinds of bullet screen text variants are processed through eight preprocessing steps,including Chinese character processing,English character processing,emoticons processing,and construction of Chinese-Pinyin dictionary.In order to eliminate the bad bullet screen,this paper constructs the bullet screen filtering model based on the text convolution neural network model,and generates the word vector of the preprocessed bullet screen,and enters the text convolution neural network model for filtering.(2)Based on the neural network,the bullet screen text is mined,and the key bullet screen recognition model is constructed.Based on the TOPSIS model,by setting the key bullet screen recognition index,the key bullet screen is identified from the bullet screen text(pre-similarity and post-similarity)and the bullet screen time(natural time and video time),so as to further simplify the bullet screen data and improve the quality of the bullet screen data.At the same time,the key frames of the video are extracted based on the inter-frame difference,and then the SIFT algorithm is used to extract the image features.The key frame of the video is compared with the image frame corresponding to the key shot.It is found that the image frame corresponding to the key shot is highly similar to the key frame of the video,indicating that the key shot can be used as a reference for the extraction of the key frame.(3)The bullet-screen sentiment analysis model based on ALBERT-Text CNN is constructed by mining the bullet-screen sentiment tendency based on neural network.The sentiment analysis effect of this model is compared with that of the other four models(SVM,CNN,Text CNN and ALBERT model)on the Chinese dialogue sentiment analysis dataset.The results show that the text features identified by ALBERT model not only better utilize the context information of words in sentences,but also better distinguish the different meanings of the same words in different contexts.Compared with the other four models,the sentiment analysis effect of bullet-screen text can be improved.Finally,combined with the video,the emotional tendency of bullet screen is analyzed.