“…With the development of NLP, Sentiment Analysis has been paid more attention by researchers and many efforts have been made in word embeddings. Jiang et al [5] proposed Bag-of-words text representation method based on sentiment topic words, which is composed of deep neural network, sentiment topic words and context information, and performed well in Sentiment Analysis; Rezaeinia et al [6] proposed refined word embeddings method based on Part-of-Speech(POS) tagging technology and sentiment lexicons, which improved the performance of pre-trained word embeddings in Sentiment Analysis; Pham et al [7] proposed a joint model of multiple Convolutional Neural Networks (CNNs), which is focused on word embeddings from Word2Vec, GloVe and the one-hot character vectors, and it achieved good performance in aspect sentiment classification tasks; Zhou et al [8] constructed a text representation model containing TF-IDF and topic features based on LDA for Sentiment Analysis, which reduced the dimension of word vector space in the traditional representation model; Han et al [9] built a hybrid neural network model using convolutional neural networks and long short-term memory(LSTM) for document representation, and it incorporated user's and product's information; Devlin J. et al [10] proposed the BERT model to represent text, which can better reflect the modifying relationship between words in texts, and it had good performance in Sentiment Analysis tasks; Liu et al [11] proposed latent topic information of the text that used Neural topic model into word-level semantics representations to deal with the problem of data sparsity, and presented a new topic-word attention mechanism to explore the semantics of words from the perspective of topic word association; Li et al [12] proposed a framework that combined different levels of prior knowledge into word embeddings for Sentiment Analysis, which improved the performance of Sentiment Analysis; Xu et al [13] proposed an improved word representation method, which integrated the contribution of sentiment information into the traditional TF-IDF algorithm and generated weighted word vectors, and the method had higher F1 score; Peters M. E. et al [14] proposed a text representation model based on deep learning framework, and it constructed an English text representation model which contained grammar feature, semantics feature and sentiment feature by training a large number of sentiment text corpus; Hao et al [15] proposed a method for cross domain sentiment classification using random embeddings, which retained similar structure in embedding space and achieved well results in the task of Sentiment Analysis; Usama et al [16] merged multilevel features which are from different layers of the same network and different network architectures to improve the accuracy of Sentiment Analysis; Majumder et al [17] demonstrated the correlation between sarcasm detection and sentiment classification, and proposed a multitasking learning framework to improve the performance of two tasks; Ma et al [18] proposed Sentic...…”