Text Sentiment Analysis based on Fusion of Structural Information and Serialization Information

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 336–341, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Text Sentiment Analysis based on Fusion of Structural Information and Serialization Information Ling Gan and Houyu Gong College of Computer Science and Tecnology Chongqing University of Posts and Telecommunications Chongqing 400065, China ganling@cqupt.edu.cn,gonghouyu b103@163.com Abstract Tree-structured Long Short-Term Memo- ry (Tree-LSTM) has been proved to be an effective method in the sentiment analy- sis task. It extracts structural information on text, and uses Long Short-Term Memo- ry (LSTM) cell to prevent gradient vanish. However, though combining the LSTM cell, it is still a kind of model that extract- s the structural information and almost not extracts serialization information. In this paper, we propose three new mod- els in order to combine those two kind- s of information: the structural informa- tion generated by the Constituency Tree- LSTM and the serialization information generated by Long-Short Term Memory neural network. Our experiments show that combining those two kinds of infor- mation can give contributes to the perfor- mance of the sentiment analysis task com- pared with the single Constituency Tree- LSTM model and the LSTM model. 1 Introduction Text sentiment analysis, namely Opinion mining, is an important research direction in the field of Natural Language Processing (NLP). It aims to ex- tract the author’s subjective information from text and provide useful values for us. In recent years, there were more and more researchers paying at- tention to the study of text sentiment analysis. Up to date, a variety of methods have been de- veloped for improving the performance of sen- timent analysis models. The distributed repre- sentation for words has been proposed in 2003 (Bengio et al., 2003). This model trained by t- wo kinds of three-layer neural networks gener- ates vectors to represent words. Glove, which is the improvement of the model mentioned above, has been proposed in 2014 (Pennington et al., 2014). The improved representation of words gives contribution to the research of NLP tasks including sentiment analysis, and are used as the input of sentiment analysis models. There are several kinds of deep learning method- s to extract text features. Sequential models such as Recurrent Neural Network (Schmidhuber, 1990), Bidirectional Recurrent Neural Networks (Member et al., 1997), Long Short-Term Memo- ry (Hochreiter and Schmidhuber, 2012) and Gat- ed Recurrent Unit (Cho et al., 2014) mainly ex- tract serialization information of text. Multi-layer sequential models have also been proposed for sentiment analysis (Wang et al., 2016)(Tang et al., 2015)(He et al., 2016)(Yang et al., 2016). Tree- structured models extract structural information. The first tree-structured model named Recur- sive Neural Network has been proposed in 2012 (Socher et al., 2012b), followed by more model- s such as Matrix-Vector Recursive Neural Net- work (Socher et al., 2012a) and Recursive Neural Tensor Network (Socher et al., 2013). In 2015, Tree-structured Long Short-Term Memory Neural Network (Tree-LSTM), which combines the LST- M cell and tree-structured models, has been pro- posed …

Content truncated to 500 words