| Video barrage is a new cultural expression form that reflects social hot spots and the dynamic emotions of netizens on the internet,and it is also an important part of the emerging internet culture in today’s society.As a special form of communication,barrage comments possess authenticity and objectivity,providing richer emotional information.Furthermore,they offer a substantial corpus for sentiment analysis,which has significant research value.Addressing the issues such as non-linearity,informality,and polysemy in the structure of Chinese barrage comments,this thesis conducts research using deep learning technology to further explore the potential emotions expressed in barrage comments.The specific research contents are as follows:(1)To address the issue of polysemy that cannot be resolved by static word embedding representations,a Chinese pre-training model RoBERTa-wwm-ext with whole word masking was introduced in combination with the popular BERT pre-training language model to obtain dynamic word vectors that are contextually relevant.Furthermore,utilizing it as an upstream word embedding model in sentiment analysis tasks to provide input features for downstream tasks can result in higher quality dynamic word vectors.When the downstream classification model is also an LSTM,the RoBERTa model outperformed the BERT model by 1.28% and the static word vector model Word2 Vec by 7.32% in terms of accuracy.(2)Based on the upstream pre-training language model,this thesis presents the RoCBiLSA,a sentiment analysis model,which employs multiple deep learning techniques in its approach.Extracting local semantic features of text using CNN for optimization in the downstream sentiment classification model,afterward,the locally extracted features are subjected to max-pooling and combined with the word embedding matrix.This composite input is then fed into a BiLSTM layer,allowing for the integration of both the contextual and local semantic features of the text.Ultimately,a hybrid feature is generated that represents the global semantic information.Moreover,a Self-Attention mechanism was incorporated to capture significant local information within the sequence,which can improve network performance and achieve better classification results.The RoCBiLSA model achieves an accuracy rate and F1 score of 89.79% and 89.53%,respectively,on the Chinese video barrage dataset collected in this thesis.(3)On the basis of the trained RoCBiLSA sentiment analysis model,this thesis selects videos on the Bilibili platform as analysis objects and collects barrage comments,which are preprocessed and classified for sentiment polarity using the RoCBiLSA model.Finally,visual analysis of the sentiment distribution of video barrage text is conducted,discussing public opinion trends,exploring the video meaning,and proposing corresponding recommendations for the control of related events so as to better comprehend the development trends of current society. |
