Research On Text Sentiment Classification Based On Deep Neural Network

Text sentiment analysis,as an important branch of Natural Language Processing,is devoted to extracting users' opinions from a large number of unstructured texts and classifying them according to their sentiment polarity.Among different methods of sentiment classification,the methods based on rules and sentiment lexicon are a bit tedious and the traditional machine learning methods need the features of manual selection.Both of them need to be improved.Recently,the application of deep neural networks and word vectors has made the result of sentiment classification further improved,but there are still some defects.The training of model also requires large amounts of labeled data.To solve the existing problems,we take on-line users' reviews as material and do the following two parts of work:(1)We propose a model which combines Long Short-Term Memory Network with sentiment lexicon and attention mechanism(called ALE-LSTM and WALE-LSTM).Although word vector can capture the semantic and grammatical information of a word,it is hard to accurately represent individual words by only using word vectors in sentiment classification task.Since word vectors are trained according to the co-occurrence of words in the external corpora,if two words have similar contexts and different sentiment polarity,they may have similar word vectors.Inaccurate sentiment representation of a single word can affect the prediction of overall sentiment polarity of the whole text.Our model first uses sentiment lexicon to train a word sentiment classifier and obtains the word's sentiment vector by using the classifier.Then,the word's sentiment vector and the original word vector are concatenated as the final input of LSTM.Furthermore,in order to improve the problem that the LSTM is biased to save recently input information and can't save the input information of long history intervals,a general sentiment classification model with attention mechanism is proposed to selectively save important sentiment information for classification.Experiments show that ALE-LSTM and WALE-LSTM model can obtain higher accuracy of sentiment classification.(2)We propose a Fuzziness Based Domain-Adversarial Neural Network with Auto-Encoder(called Fuzzy-DAAE).Most sentiment classification methods based on deep neural networks and word vectors require a large amount of labeled training data.However,in some emerging areas,the construction of labeled data is time-consuming and labor intensive.Domain Adaption algorithm aims at using the labeled data of related domain(source domain)to improve the performance of target domain(only has a small amount of labeled data or even no).However,the existing Domain Adaption algorithm pay more attention to the common features of different domains and ignore characteristics of samples themselves.Our proposed model not only uses a gradient reverse layer to achieve the adversarial training which makes the domain classifier unable to identify the differences between domains to obtain the domain-invariable features,but also uses an auto-encoder to reconstruct original inputs to maintain the characteristics of samples.Besides,in order to introduce some sentiment supervised information of target domain samples,we also add some unlabeled target domain samples and their predicted labels to the original training data according to their fuzziness and then retrain the whole model.Experiments show that Fuzzy-DAAE model is effective.It is worth noting that Fuzzy-DAAE model can be used for any other Domain Adaptation tasks,not limited to cross-domain sentiment classification.