Research On Text Sentiment Classification Technology Based On Deep Learning

Currently,the expression of emotions is reckoned essential.News and short commentaries appeal to middle and older aged viewers who post their own attitudes and opinions occasionally;younger people enjoy sharing their experiences on platforms such as Weibo and Zhihu.The emergence of increasing video-sharing social networking services such as Tik Tok and Vlogs has boosted people's desire to express their emotions,which demonstrates the significance of emotions in life.Positive emotions are beneficial at work;negative emotions can impair the physical functions to an extent and even the health of the personality.The normalization of epidemic prevention and control has raised people attention on mental health.To this end,this thesis aims to develop a mental-health-based sentiment classification system that analyzes users' mental health through text,and helps doctors to screen patients with depression or users understand their health.Using existing researches as groundwork,this thesis proposes three sentiment classification models which are applied in the implementation of a sentiment classification system.At present,there is less research on depression emotional classification,and the technique used by most models is model fusion or multi-modal,and the importance of text characteristics is ignored,resulting in the accuracy of emotional classification.In order to solve these problems,this thesis proposes three sentiment classification models:the first model is based on Bi-directional Gated Recurrent Unit which is the improvement of Char SCNN model,it fuses word and sentence level features,using the attention mechanism to extract word vector features and sentence vector features at the word level and sentence level respectively,achieving more emotional semantics,and it makes up for the deficiencies that CNN networks cannot be well extracted;the second model employs the Bi-directional Long Short-Term Memory unit based on knowledge distillation.With BERT serving as a teacher model to supervise the learning of the student model(Bi-directional Long Short-Term Memory),distillation techniques compress the student model to a certain extent,improving the portability of the model;finally,the second model is further improved when the Attention Bi-directional Long Short-Term Memory unit is trained with the focal loss function to tackle the imbalance between positive and negative samples in the emotion dataset.Secondly,comparative experiments on the three sentiment classification models above are conducted using the distress analysis interview corpus.The experimental results show that they can all successfully classify emotions.By comparing the experimental results on this dataset,among them,the Att-Bi LSTM model based on focus loss outperforms others,achieving 0.78,0.81,and 0.80 in accuracy,recall,and F1 value,respectively.In addition,the dataset used in this thesis is obtained to the real population,which can reflect the proportion of depression and unsatisfactory populations in the actual life,Att-Bi LSTM model based on focus loss whose precision rate can reach 78% in mental health dataset,indicating that it has certain availability and effectiveness,and can be used in the mental health screening system to complete the emotional classification task.Finally,this thesis develops and implements a mental health sentiment classification system,which includes three main functional modules: psychological evaluation forms,open Q&A,and report display.The Psychological evaluation forms provides multiple types of measurements for users to choose from and the questions are answered by filling in the questionnaire;in the Q&A module,the user answers questions by having conversations with the robot,and the classification model at the back of the system outputs the results according to the user's answering texts;at last,users can view their own reports.