Research And Implementation Of Sentiment Analysis Based On Deep Learning

Sentiment Analysis,also known as Opinion Mining,has become a hot topic in Natural Language Processing Fields.It aims to mine subjective emotional information from large scale of corpus,and shows great application value in commodity evaluation and public opinion evaluation etc.The traditional methodology of Sentiment Analysis is based on sentiment lexicon,which predict the sentiment of texts by rules.The quality of sentiment lexicon method is mostly determined by the quality of sentiment lexicon,which is hard to build and usually can't work well.Recent years,deep learning technologies is widely used and got breakthrough success in NLP fields.Rich hidden features of the texts can be mined by training neural networks with large corpuses,which lead to advanced results in sentiment analysis tasks.With the aging tendency of population and the shortage of labor,social robots,especially service robots have caught extensive interests for replacing some work of human.How to improve the interactivity of the social robots is urgently needed to be researched.This thesis improves the interactivity of the social robot “Caibao”,developed by Center for Robotics of UESTC,by sentiment analysis based on Deep Learning technologies.This thesis applied deep learning technologies to construct sentiment models firstly,and then implements the sentiment analysis system of “Caibao”.With the sentiment analysis system based on deep learning,“Caibao” provides much better interaction services by reactive to the sentiment states of users.The detail works of this thesis are as follows:(1)This thesis proposes a sentiment analysis model based on bidirectional GRU networks and Self-Attention mechanism.For the application scenario of social robots and the lack of local corpus,the microblog sentiment dataset is chosen as the task data set.This thesis proposes a sentiment based on basic LSTM network,which lacks of backward information.Then a bidirectional LSTM cell is used to extend the network layer of the model,which can get both forward and backward information of the texts.Attention mechanism is used to predict the importance for the sentiment expression of the whole sentence.However,bi-LSTM and Attention mechanism makes the model too complex to train.Bidirectional GRU network is used to reduce the complexity of the model and SelfAttention mechanism is used to break the external dependence of the sentiment weight vector.The bi-GRU and Self-Attention model shows more accuracy in the experiment on task dataset.(2)This thesis design and realize the sentiment analyze system of “Caibao” based on the sentiment analyze model proposed above.This thesis analyzes the background and requirements of “Caibao” and its sentiment analysis system.The “Caibao” robot system,human-computer interaction system and sentiment analysis system are designed based on the system requirements.Then the emotional human-robot interaction method with both animated expression and robot action was designed.The other part of sentiment analyze system like database and APIs were also realized.At last,the sentiment analyze system was deployed and several sample tests was completed.