Research Of Short Text Classification Based On Convolutional Neural Networks

In recent years,with the popularity and rapid development of the Internet,the need for text classification has become more and more urgent.Text classification is the process of automatically classifying texts of unknown categories based on the contents and the semantics.The goal of this paper is to solve the short text classification problem and improve it based on CNN model.The main contents of this paper are as follows:In current Natural Language Processing research,people can combine different neural network structure and classification algorithm when using Convolution Neural Network(CNN)to conduct text classification tasks.Thus,this paper proposed a hybrid CNN-ELM model for short text classification.Firstly,this model used word vectors to represent sentence as the input data.Secondly,it extracted features through CNN and completed features optimization with Highway network.Finally,it used error minimization extreme learning machine(EM-ELM)as a classifier to complete text classification task.According to the experimental results in various English datasets,CNN-ELM model is more suitable for short text classification tasks than traditional machine learning models and deep learning models.At present,most CNN models used in text classification are shallow models and have limited ability to express short texts especially on the Internet.This paper proposes a character-level deep CNN model(Char Deep CNN)for short text classification.This model improves the disadvantages of shallow neural networks in Deep Learning.Firstly,it uses character-level features as input,and then uses a multi-layer CNN structure to complete feature extraction.Experiments show that Char Deep CNN model performs better than classic deep learning models on datasets of different scales.Finally,this paper proposes a Hybrid Deep CNN model which combines characterlevel features,deep CNN structures,ELM classification algorithm and Highway networks for large datasets.Based on Char Deep CNN model,this model uses Highway networks for feature optimization,and uses EM-ELM classifier to complete the classification tasks.Experiments on large datasets show that Hybrid Deep CNN model can further improve the performance of Char Deep CNN model.