Research On Document Level Text Sentiment Analysis Based On Deep Learning Model

Nowadays,with the rapid development of network technology,the Internet has become the main information carrier,and people are accustomed to expressing their sentiments and opinions in the Internet.There is a large amount of text data in the Internet,and it is of great scientific value to analyze the sentiment contained in these text data.Document level sentiment analysis is one of the fundamental tasks in the field of sentiment analysis,which aims to summarize and classify the sentiment expressed in the text by computation.In recent years,the rapid development of deep learning has promoted the development of the research of sentiment analysis,and more and more researchers have devoted themselves to solve the problems of sentiment analysis with deep learning.Based on these backgrounds,with the goal of improving classification performance,this paper aims to further explore the application of deep learning in document level sentiment analysis.The main contributions of this paper are as follows.(1)A sentiment analysis model based on neural bag-of-words attention(NBOW-Att)has been proposed.The model extracts the contextual semantic features of the text by bidirectional long short-term memory network(Bi LSTM)at first,then uses the neural bag-of-words of text as the query vector to calculate the attention distribution,and finally aggregates these features according to the attention distribution.The neural bag-of-words vector is a shallow representation,which can provide the information about the input text compared to the fixed query vector.Guided by this information,the attention mechanism can identify the key information in the input more accurately,and thus the sentimental semantic features contained in the text can be extracted more effectively.The experimental results show that the NBOWAtt can focus attention on the keywords in the text and obtain better classification performance.(2)A global information based hierarchical attention network(G-HAN)has been proposed.This model encodes each sentence in the text by the sentence encoder to obtain the sentence representation at first,and then uses the document encoder to encode all the sentence representations and obtain the text representation.Especially,the sentence encoder and the document encoder are both consist of the bidirectional gated recurrent network(Bi GRU)with attention mechanism.Since the process of sentence encoding is independent of each other,in order to enable the sentence encoder to establish the connection between the content in the current sentence and the overall content of the text,the G-HAN uses the abstract representation of the neural bag-of-words vector of text as the initial hidden state of Bi GRU when encoding sentences,and the global information of the text is passed into Bi GRU by this way.The experimental results demonstrate both the effectiveness of incorporating global information when encoding sentences and the effectiveness of using bag-of-words vector to provide global information.