With the development of the Internet,people participate more into the it,leading to a progressive number of User Generated Content(UGC).The sentiment analysis with UGC presents fascinating commercial value and academic value.As a significant branch of Machine Learning,Reinforcement Learning(RL)plays a better role in imitating humans'learning process.However,there exists a huge gap between RL and text classification.Consequently,this paper manage to mix RL and text sentiment analysis.This paper analyzes multiple UGC dataset and propose a complete scheme for pretreatment.To extract text feature better,word embedding and LDA are used as word representation in experiments.To make a better use of human knowledge,we build sentiment dictionaries,which are used at the training step.This paper proposes a new Reinforcement Learning(RL)model,Word-level Sentiment LSTM(WS-LSTM).We regard sentiment analysis as getting emotional fluctuation during reading a text.WS-LSTM contrains an action-choose layer with an action-critic layer to mix RL and text sentiment analysis,and a state layer with full-connection layer helping it unite RL and supervise learning.The parameters are trained based on the classification results.The results show that WS-LSTM plays a better role in each dataset.As to the sentiment action choosing,this paper improves the traditional?-greedy algorithm.We add sentiment element to make it more useful.This algorithm helps model choose sentiment action more accurately and get improvement on the result.As to the evaluation criterion for word-level sentiment analysis,this paper builds a new evaluation scheme based on sentiment intensity and successes in using the experiment result to prove its validity.With the evaluation scheme,we are able to evaluate the word-level sentiment result. |