Research Of Document-level Sentiment Classification Based On LSTM

With the rapid development of the Internet Web 2.0,it generates a lot of valuable sentiment information about users,events and products,so sentiment analysis is particularly important.For paragraph-level or sentence-level sentiment analysis,there are many good solutions,but for document-level sentiment,one of the remaining challenges is to model long texts under a recurrent architecture.Although Long Short-Term Memory neural network has achieved a great success in many sequence problems,it is not powerful enough to handle the overflow and to capture key sentiment messages from relatively far time-steps.In order to model long texts,we propose a hierarchical neural networks structure based on LSTM and an improved method of it.This article has the three following contributions to document-level sentiment analysis.First,we propose a new neural network model with two hidden LSTM layers.The first layer learns sentence vectors to represent semantics of sentences and in the second layer,the relations of sentences are encoded in document representations.Secondly,in order to improve the model,we propose an idea based on sentiment dictionary to filter out some objective sentences and reduce the noise of training model.Finally,we use Tree-LSTM to model words to sentences to consider the spatial structure between words in one sentence.The experimental results show that our models outperform the state-of-the-art models on three publicly available document-level review datasets and tree-LSTM can get a better sentence representation than standard LSTM.