Studies On Entity Level Sentiment Analysis Based On Dilated CNN

Effective analysis of implicit sentiment information in text is one of the focuses in the field of natural language processing.Specifically,entity level sentiment analysis aims at reasoning the sentiment polarities conveyed by the entities in the text,taking into account the opinions on product and service in more detail,which is of great value.Therefore,to overcome the deficiencies of the existing models in the task of entity level sentiment analysis,this thesis explores the entity level sentiment analysis form the specific entity and unspecific entity aspects,as follows:(1)Specific entity: Considering the inefficiency of the recurrent neural network and the limited feature extraction scope of the convolutional neural network(CNN),a sparse attention based separable dilated convolutional neural network(SA-SDCNN)is studied.Firstly,multi-channel word embedding,consisting of semantic word embedding and sentiment word embedding,is applied to encode the input text and the specific entity.The separable dilated convolution module is constructed,in which the separable dilated convolutional neural network with diverse dilation rates is adopted to expand the scope of the semantic interaction,so as to reduce model parameters while obtaining multi-scale semantic dependencies.Additionally,according to the position information of the specific entity and the semantic features of the input text,a sparse attention mechanism is designed for the extraction of the sentiment features of the specific entity.Finally,the comparative experiments on several real datasets show that the proposed model with model parameters of 17.9 K could obtain classification accuracies of 73.57% and 81.36% on Laptop and Restaurant datasets,respectively,and also indicate the effectiveness of the model.(2)Unspecific entity: Considering that the existing methods model the task of entity level sentiment analysis as two sub-tasks of entity extraction and sentiment polarity discrimination,ignoring the interaction between the two sub-tasks,the task of entity level sentiment analysis is transformed into a sequence labeling problem by adopting a selfattention based hierarchical dilated convolutional neural network(SA-HDCNN),which could realize the recognition of entities and corresponding sentiment polarities collaboratively.SA-HDCNN is mainly composed of an encoding module,a feature extraction module and a decoding module.Specifically,the encoding module maps the text as a word embedding matrix containing both semantic and sentiment information.In the feature extraction module,the hierarchical dilated convolutional neural network and self-attention mechanism are responsible for the extraction of the multi-level semantic features and the capture of the inter-word dependencies in the same semantic space,respectively.The decoding module annotates the words in the text according to the text features and the constraint rules of labels.The experimental results show that the proposed model can achieve F1 scores of 63.33% and 72.85% on Review and Twitter datasets respectively without relying on any domain special features and prior information,which reveals the feasibility of the proposed model.