[1]ZENG Biqing,HAN Xuli,WANG Shengyu,et al.Hierarchical double-attention neural networks for sentiment classification[J].CAAI Transactions on Intelligent Systems,2020,15(3):460-467.[doi:10.11992/tis.201812017]
Copy
CAAI Transactions on Intelligent Systems[ISSN 1673-4785/CN 23-1538/TP] Volume:
15
Number of periods:
2020 3
Page number:
460-467
Column:
学术论文—自然语言处理与理解
Public date:
2020-05-05
- Title:
-
Hierarchical double-attention neural networks for sentiment classification
- Author(s):
-
ZENG Biqing1; HAN Xuli2; WANG Shengyu2; ZHOU Wu2; YANG Heng2
-
1. College of Software, South China Normal University, Foshan 528225, China;
2. College of Computer, South China Normal University, Guangzhou 510631, China
-
- Keywords:
-
sentiment analysis; attention mechanism; convolutional neural network (CNN); sentiment classification; recurrent neural network (RNN); word vector; deep learning; feature selection
- CLC:
-
TP391
- DOI:
-
10.11992/tis.201812017
- Abstract:
-
In sentiment classification, feature extraction in the document level is more difficult than the analysis in the common sentence level because of the length of the text. Most methods apply a hierarchical model to the sentiment analysis of text in the document level. However, most existing hierarchical methods mainly focus on a recurrent neural network (RNN) and attention mechanism, and the feature extracted by a single RNN is unclear. To solve the sentiment classification problem in the document level, we propose a hierarchical double-attention neural network model. In the first step, we improve a convolutional neural network (CNN), construct a word attention CNN, and then extract the features of the chapter from two levels. In the first level, the attention CNN can identify important words and phrases from every sentence, extract the word feature of the sentence, and construct the feature vector of the sentence. In the second level, the semantic meaning of the document is derived by the RNN. The global attention mechanism can find the importance of every sentence in the document, attribute different weights to them, and construct the whole semantic representation of the document. The experiment results on IMDB, YELP 2013, and YELP 2014 datasets show that our model achieves a more significant improvement than all state-of-the-art methods.