Abstract:

Text sentiment analysis is an important task in the field of network public opinion analysis, product evaluation and data mining. With the growth of data volume, the traditional methods such as manual engineering and NLP tools cannot handle the task due to their low accuracy and high costs. Therefore, we propose a deep learning method named convolution neural network (CNN) to deal with it. The traditional CNN does not consider the structural information of sentences and suffers from overfitting. Aiming at the two problems, we first design a piecewise convolution neural network (PCNN) to combine the structural features, in which the feature vector of a sentence is divided into several segments and does the maxpooling for each of them. Then we introduce the Dropout algorithm to prevent the model from overfitting and extend its generalization abilities. We use two datasets in our experiments: Chinese hotel reviews and the Stanford Sentiment TreeBank. Experimental results on the two datasets show that both the PCNN and the Dropout can enhance the performance. The proposed model can achieve 91% accuracy on the Chinese dataset and 45.9% on the English dataset, which are higher than all of the baseline systems.

Key words: sentiment analysis, deep learning, piecewise convolution neural network, piecewisepooling, Dropout algorithm