Research Of Sentiment Analysis Based On Deep Learning And Multi-feature Fusion

With the rapid progress of Internet,social media platforms,e.g.,Sina Weibo and Facebook,have become increasingly popular.Using above platforms,end-users send posts to express sentiments and share opinions in daily life,leading to the accumulation of textual data with emotional tendency.These sentiment information on social media are valuable for studies to track the shift of human emotion about hot topics and analysis the public opinion when an event takes place.As a major research direction of natural language processing,sentiment analysis is the task that extracts the attitudes and opinions from a given piece of textual data which contains the emotional information.Traditional sentiment analysis methods generally fall into two categories,one is dictionary-based models,the other is machine learning-based approaches.The former heavily relies on the quality of the sentiment dictionary,while the latter depends on a large amount of training data.Therefore,current methods have strong limitations.In recent years,the emergence of deep learning models has provided an efficient way for sentiment analysis.In this paper,we focus on Sina Weibo short text sentiment analysis.Our main work lies in: Sina Weibo contains diverse textual content,however,most existing methods pay less attention to the nonliteral features,leading to poor performance.To solve this limitation,we incorporate three features,i.e.,dictionary-based sentiment value,emoji usage and adapted semantic feature,into sentiment classification model.We crawl textual content from Sina Weibo and then accomplish data preprocessing: cleaning and labeling.Afterwards,the classification dataset is constructed to conduct experiments.The results show that our method achieves competitive performance compared with traditional neural networkbased approaches because of the multi-feature fusion mechanism.Moreover,we utilize a word2 vec model weighted by term frequency–inverse document frequency(TF-IDF)to represent words in low-dimensional vector,which surpasses the original word2 vec method.Furthermore,we combine convolutional neural network(CNN)and bi-directional long short-term memory network(Bi LSTM)to build a novel sentiment classification model,namely KCNN-Bi LSTM.CNN could capture the local semantic information of input sentence,while Bi LSTM excels at modeling long-term contextual dependency.To integrate merits of the CNN and Bi LSTM,we replace max-pooling operation with K-max-pooling,which can retain the frequency and partial location information to some extent.Also,self-attention mechanism is introduced to highlight the important words.Substantial experiments on the binary classification and three classification datasets show that our proposed method KCNN-Bi LSTM achieves better performance than other deep learning methods on accuracy.When self-attention mechanism is leveraged,the performance is further improved.At last,we implement a sentiment analysis system based on Flask framework,which has various functions,e.g.,single sentence text prediction,batch data prediction,data display and data download etc.