A new short text sentimental classification method based on multi-mixed convolutional neural network

Almost all e-commerce platforms provide online product comments service. The comments are some words of mouth about the products. Customers usually reference the comments to make an informal decision when buying similar products. In addition, companies would like to know the feedbacks about the products through the comments. However, the typical big data characteristic and noisy short text data characteristic make the online comment data analysis becoming a challenge work. In this work, we propose a Multi-Mixed Convolutional Neural Network (MMCNN) model to analyze the sentiment of online product comments. We mix the convolution and pooling features in mixed layer to enhance effectiveness of the online comments sentimental analysis. The skip-gram model is used to train the word vector. Because the length of each comment is not fixed, two new empirical matrix filling methods are designed which are cyclic matrix filling and random matrix filling. We apply our approach for two datasets which are online comments about infant power crawled from www.jd.com and online reviews about hotel crawled from www.elong.com. Experiment results demonstrate the effectiveness of our approach in comparison with Support Vector Machine, Maximum Entropy, Naive Bayesian and classic CNN.