Reseaech And Implementation Of Sentiment Feature Extraction Method Based On Actor-critic Mechanism

The rapid development of the Internet and the popularity of computers,mobile phones and other devices in the era of Web2.0 have promoted the proliferation of user-generated content on the Internet.These data containing a huge amount of information,are of great significance and research value to provide personalized services for users.Sentiment analysis is an important means to excavate the content of the text,which is mainly to identify the subjective feelings that the document wants to express.Fine-grained Sentiment Analysis,which is the Aspect-level Sentiment Analysis,extracting the sentiment polarity of a given aspect from the text,has become a focus in the industry in recent years.Firstly,this thesis analyzes the current research status of aspect-level sentiment analysis task at home and abroad,and summarizes two major problems of existing methods based deep learning:most neural networks combining attention mechanism will inevitably introduce noise.Furthermore,the inability to process sentiment words that are sensitive to a given aspect,whose sentiment can vary from aspect to aspect.For dealing with existing two problems,this thesis designs an innovative and practical algorithms of Sentiment Feature extraction named AC-FDN(Feature Distillation Network Based Actor-Critic Mechanism)to reduce the noise and extract the emotional features related to a given aspect.The algorithm first designs a feature extraction network FDN(Feature Distillation Network),which introduces the context nonlinear projection layer(CNPL)on the context embedding encoded by BILSTM while further emphasizing the information of the given aspect,which can effectively alleviate the problems caused by the sensitivity of emotion words.Moreover,FDN designs a new Double-Gate Layer(DGL)to realize the interaction between the given aspect and the context features,filter the noise brought by words unrelated to the given aspect,and amplify the emotional features related to the given aspect.But FDN cannot directly delete irrelevant words from the text level,only filter the irrelevant emotions from the feature level.On the basis of FDN,AC-FDN algorithm further combines the Actor-Critic mechanism,which first deletes the words unrelated to the text structure and semantic emotional representation at the text level and reduces the text noise once.Then,AC-FDN uses FDN to perform a second feature-level noise reduction on the denoised text,that is,to filter out the noise caused by emotional words irrelevant to a given aspect.In this thesis,many experiments are carried out on three datasets of different sizes.The experimental results confirm that the model proposed in this thesis can achieve a significant improvement effect compared with the benchmark model,and reach the best.Two different methods are also used to conduct visualization experiments to further verify the effectiveness of the model substructure.