关键词: 深度学习, 情感分析, 递归自编码, 递归神经网络

Abstract: During Tibetan sentiment analysis in past, the algorithm always ignores some important information like sentences structure and words order etc, which lead low accuracy of sentiment analysis. To deeply get more sentiment details, this paper proposes a novel approach of Tibetan sentiment analysis based on deep learning. Firstly, one word in Tibetan is represented by a word vector while one sentence is represented by a matrix which is composed by its word vectors; Secondly, the matrix is turned into a vector which contains most important details such as sentence meaning and words order etc, through an unsupervised recursive auto encoder algorithm; Finally, the classifier in output layer is trained by supervised method which uses the word vectors and its sentiment tags. In the experiment part, this paper discusses the selection of word vector dimensions and reconstruction error weights, studies corpus amount how to affect algorithm accuracy, and analyzes the relation between corpus amount and training time. The experimental results demonstrate that the proposed method can improve accuracy up 8.6% compared with semantic space model which is almost the best in traditional machine learning algorithm.

Key words: deep learning, sentiment analysis, recursive auto encoder, recursive neural networks