Aspect-level Sentiment Analysis Based On The Deep Neural Network

With the advent of the era of mobile internet,the quantity of online customer reviews grow rapidly.In order to explore the opinion of customers,sentiment analysis algorithm won the favor of the researchers.Aspect-level sentiment analysis is a fine-grained task of sentiment analysis,is to determine sentiment polarity based on specific aspect.This article mainly about the algorithm improvement based on neural networks.Besides,we built a Chinese dataset for aspect-level sentiment analysis,hoping to promote the research in the field of Chinese.The main contribution and research are listed below:1.The improvement of word embedding.Today,word2vec is widely used to get word embedding.This article extended one of the frameworks of word2vec,which is named CBOW based on Negative Sampling,by adding a sentiment analysis module.By this way,we can get word embedding which can express semantics and sentiment simultaneously.2.The improvement of aspect-level sentiment analysis algorithm.First of all,this article proposed Twofold-Scanning mechanism,enhance the comprehension ability of the neural network model for customer reviews with long length.Secondly,this article presented an adaptive attention mechanism to make the model can "recognition" different aspects of customer reviews more accurately,by adding the attention layer.In addition,we implemented the "Wide and Deep" model to learn logistic regression model and deep neural networks jointly to combine their respective superiorities:memorization ability and generalization ability.3.Creation of Chinese dataset.This article set up a Chinese corpora tagging platform.We created the first Chinese dataset in the field of aspect-level sentiment analysis.And,we have done some experiments on the dataset.However,the scale of the Chinese dataset is relatively small,so we should do more work to make it better.