Study And Implementation Of Chinese Short Text Emotion Classification Based On Multi-model Fusion Of Deep Learning

With the increasing popularity of the Internet,people are willing to express their views and communicate with each other on the Internet.They express their ideas in the form of texts containing emotions.Analyzing the emotions of these texts is very useful for online marketing,public opinion monitoring and so on.Comparing with English text emotion classification,the research of emotion classification of Chinese text has less concern,and its accuracy is generally lower than the accuracy of English text emotion classification.Therefore,it is of great research value for Chinese short text emotion classification.This paper mainly focuses on studying the influence of word representation,network structure,training strategy and loss function on the accuracy of text classification model based on deep learning.Then,we train serveral deep learning models,and create fusion model base on these models to achieve higher accuracy of Chinese short text emotion classification.Firstly,we analyze the influence of word representation on the accuracy of text classification model based on deep learning.We find that deep learning model will achieve higher accuracy if using pre-trained word representation for training.In view of the fact that maximum pooling method in textCNN network may lose important feature information,we propose K-Max-CNN network: the textCNN network with improved pooling method.We also propose four improved DCNN networks by referring to the characteristics of text CNN network.Experiment shows that the K-Max-CNN network and the four improved DCNN networks can achieve higher accuracy than original networks.Secondly,because deep learning training result is often unstable,we improve the mini-batch training method for deep learning model.Experiment shows that the currirulum learning method of "difficult before easy" can obtain more stable results,and can also improve the accuracy of the model.At the same time,In order to avoid bad influence of too many negative samples on model training process,we propose the improved focal loss to increase model accuracy,which makes sure that important samples can get more attention.Finally,we use the pre-trained word representation,the improved network structure and the improved focal loss,with the help of curriculum learning method of "hard first and then easy",to train several deep learning models,and create fusion models base on them for solving these tasks of subjective-vs-objective emotion classification,multi-emotion emotion classification and sentiment polarity classification.The results show that the accuracy of fusion model created by our proposed method is respectively 0.79%,2.85% and 2.05% higher than the accuracy of the one of highest accuracy in the comparing models in these three tasks.