Research On End-to-end Models For Aspect-based Sentiment Analysis

E-commerce,as a new shopping and marketing channel,has led to an upsurge of review sites for a variety of services and products.In this context,Aspect Based Sentiment Analysis(ABSA)-that is to say,mining and summarizing opinions from text about specific entities and their aspects-can help consumers decide what to purchase and businesses to better monitor their reputation and understand the needs of the market.ABSA aims to identify all aspect-categories mentioned in the text and their corresponding sentiment polarities,namely Aspect-category Detection(ACD)and Sentiment Classification(SC).Most existing methods model these two tasks separately.Although we have achieved excellent experimental results on the two subtasks,it extracted incomplete user opinions.Also,the accuracy of the upstream aspect-category detection task will affect the result of the downstream sentiment classification task which reduces the practical value of these methods.In addition,the data labeling in ABSA is very expensive and the training data is limited which brings many challenges for ABSA.In address to this problems,this paper proposes a variety of end-to-end neural networks for aspect based sentiment analysis and leverages unlabeled data by modeling word representations.To summarize,our effort is in three folds:For sentence-level ABSA,this paper proposes three end-to-end neural networks for ABSA: end-to-end bi-directional long short-term memory neural network(Bi-LSTM),end-to-end attention-based neural networks and attention-based Bi-LSTM neural network.For review-level ABSA,this paper proposes a hierarchical neural network based on attention and Bi-LSTM.Experimental results show that for review-level ABSA,a hierarchical model is better than a single-level model.Experiments conducted on Sem Eval-2016 restaurant dataset show that our proposed model achieves state-of-the-art performance.Considering the lack of training data,this paper leverages unlabeled data by modeling word representations.The experimental results show that it is important to select a suitable word embedding method for ABSA.It can alleviate the problem of lack of training data and lead to performance gains.