Research Of Sentiment Classification Based On Attention Word Embeddings

With the rapid development of the Internet,it has provided more channels for people to access information,transfer information and exchange feelings.We use weibo to share what we see and hear every day.We use the Internet to shop and then write reviews.We book the hotel online before we travel and share the experience after we checked in.These data include peop le's basic necessities of life.They are extremely valuable for consumers,business organizations,and even government departments.How to effectively mine emotional information from vast amounts of data has become an urgent problem to be solved.Machine learning techniques provide many ways for sentiment classification problems.In particular the bloom of deep learning in recent years has brought new blood for solving sentiment classification.However,there are still many deficiencies to be solved.When using deep learning to deal with the problem of natural language processing,it must turn the text into a form that can be processed by a computer at first.At present,the most common way is word embeddings.Although word embeddings have achieved excellent performance on many tasks.Most methods for word embeddings training infer word vectors based on contextual information of words.In the Chinese domain,the meaning of word is also contained in the characters that make up it.In this paper,a novel method based on attention is proposed.Information of character embeddings is added into word embeddings.At the same time,the importance of different words is considered in the process of joining.Finally,in the tasks of similarity calculation,logical inference,text classification and so on,it is verified that the word vec tors obtained by this method have more excellent representation ability.After years of effort,many different models of sentiment classification have proposed,such as support vector machines,convolutional neural networks,recurrent neural networks and so on.These models extract knowledge from the data by different ways.We can integrate these models to get better results.Traditional ensemble learning commonly uses voting,averaging,or learning methods during combination step.These methods increase calculated quantity during testing.This paper proposes a teacher-student network.First train multiple individual classifiers.And then initialize a new neural network as a target classifier.When training target classifier,we not only consider the true labels,but also care about the output of previous classifiers.In this way,compress multiple classifiers into a classifier.While keeping the original performance,the method costs less calculated quantity.