Research On Short Text Classification Based On Deep Learning

With the continuous information construction in China and the development of Mobile Internet,information from the Internet grows explosively,the age of fragmented information which is based on short text information is coming.Therefore,the way to accurately extract required information becomes the focus of scholars and commercial company.This study presents the research in Chinese short text classification technologies,the training model of word embedding and character embedding is improved.In addition,the classic text classification model is enhanced effectively.Then,a classification model which is suitable for Chinese short text is designed and implemented.This paper introduces works in three parts:First,a novel training model of word embedding and character embedding is proposed based on Chinese characters and radicals information.In this paper,based on the research of the deficiency of word embedding in Chinese corpus,previous training model is improved by adding Chinese characters and radicals information,which provides extra cooccurrence information to word embedding.It is more suitable for Chinese text.Moreover,radical transformation mechanism which changes radical to the corresponding Chinese character makes word embedding effectively to identify words with connected semantic.Thus,words with similar semantic is closer in vector space so that the performance and explanatory is better.Based on this,character embedding,which reduces the effect to word embedding from word break error,provides richer semantic information for subsequent classification models.Second,a new feature extraction network is designed.The convolutional neural network and recurrent neural network are combined to design a new feature extraction network using Attention Model technology.The network employs k-max pooling and bidirectional cyclic neural network technology which has stronger feature extraction capability to identify and extract semantic feature from text data.Using Attention Model technology,Internet pays more attention to extract classification features and removes invalid features.Therefore,quality of feature vector is improved while classification effect of models is more obvious.Third,a dual-channel short text classification model is proposed combining word embedding and character embedding.This model adopts a dual-channel short text classification model,which combines text data of word embedding and character embedding,to extract text feature from two different texts.It greatly enriches text information of short text and enhanced the effect of classification.Besides,comparing with other similar models in experiments,the model is proved feasibly.