| With the rapid development of internet,more and more users express a large number of comments on the internet platform.The purpose of the aspect-level sentiment classification is to predict different sentiments of a text in different aspects.Recent aspect-level sentiment classifications are mainly based on the supervised learning,relying on a large number of labeled samples.So how to perform semi-supervised aspect-level sentiment classification with unlabeled samples is very important.Our paper takes the review text as the research object and conducts the semi-supervised aspect-level sentiment classification research based on the variational autoencoder at the first time.The main work includes the following two parts:(1)For the problem of that existing aspect-level sentiment classification models usually use a single vector to represent each word but a single word embedding cannot distinguish the different aspects and sentiments which a word expresses.Our paper proposes a recurrent neural network model(referred to as ASWAR)based on aspect-sentiment word and attention.The model firstly regards a topic as an aspect,and utilizes a joint sentiment topic model(JST)to capture the aspect and sentiment assignment of a word.Then,we input the word-aspect assignment on the topical word embedding model(TWE)and train a specific-aspect word embedding considering the aspect and the contextual of the word simultaneously.And we decide a one-hot sentiment vector to each word according to the word-sentiment assignment.Finally,on the basis of the original attention-based LSTM model(ATAE-LSTM),we introduce specific-aspect word embedding and sentiment vector and allow the model to recognize the different aspects and different sentiments of a word.And we use the attention mechanism on the recurrent neural networks with LSTM unit and GRU unit to get the important part of the text in response to a given aspect and model the interdependence between aspects and sentiments of the word and the given aspect.And then we improve the accuracy of aspect level sentiment classification to a certain degree.(2)Most existing semi-supervised learning methods are based on a generative model.But when the model assumptions and data distribution are inconsistent,the accuracy of the model is low.For this issue,our paper proposes a semi-supervised aspect-level sentiment classification model based on variational autoencoder(referred to as AL-SSVAE).Based on the variational autoencoder,the model adds an aspect-level sentiment classifier and takes the aspect information into account in the encoder and decoder.AL-SSVAE model firstly uses the ASWAR model with LSTM unit in(1)(ASWA-LSTM)as the classifier and also encode the text.Then we use the specific-aspect word embedding to represent each word and connect the sentiment vector of the word in the decoder,and we also add the label and aspect embedding to reconstruct the input at each time step.Therefore,the model has the ability to recognize the corresponding aspects and sentiments of words during training.And the model can capture rich global semantic information and sentiment characteristics accurately and then achieve the semi-supervised aspect-level sentiment classification of the text. |
