Research On End-to-end Aspect Term Sentiment Analysis Based On Multi-task Learning

With the popularity of the Internet,more and more people choose online consumption,entertainment,social and learning,while leaving a large number of comments on various Internet platforms.These massive Internet comment text data contains the user's views and attitudes on different things,which can provide important information for different people to make decisions.Aspect term sentiment analysis is a text mining technology,which can extract fine-grained sentiment information from comment text.There are three subtasks in aspect term sentiment analysis: aspect term extraction(AE),opinion term extraction(OE)and aspect-level sentiment classification(ASC).Most of existing studies only focused on one of AE or ASC only.Although aspect term sentiment analysis can be done in a pipeline manner in practical applications,such approach does not fully exploit joint information from these subtasks,and there is error propagation.For this problem,the paper mainly studies the end-to-end aspect term sentiment analysis,solves three sub-tasks of AE,OE and ASC simultaneously by using multi-task learning model.In view of the feature extraction and subtask interaction in the task,we propose two multi-task learning models:(1)Attention-based multi-task learning network for aspect term sentiment analysis(AMN-ATSA).In the model,multi-head attention network is used as the feature extractor,which has stronger feature extraction ability and parallel computing ability than RNN.At the same time,for the different connections between the sub-tasks of AE,OE and ASC in the aspect-based sentiment analysis,a suitable interaction based on the attention mechanism is designed.Ways to enhance the ability to exploit joint information between subtasks.Experiments on semeval-2014 and semeval-2015 show that,the AMN-ATSA model outperforms other benchmark models,and the training speed is also much faster than other benchmark models which based on RNN.(2)Syntax-enhanced multi-task learning network for aspect-based sentiment analysis(SMN-ATSA).Aiming at the problem that it is difficult to accurately match aspect term and corresponding opinion term in the case of multi-aspect sentence,we designed an syntax-enhanced interactive attention mechanism.At the same time,an enhanced message passing mechanism is proposed,which updates the shared features by passing the task related features with higher information dimension,so that the subtask module can obtain more abundant information in the iterative process.Experimental results show that,compared with various end-to-end benchmark models,the SMN-ATSA model has the best end-to-end indicators on all data sets.Also,the effectiveness of the two mechanisms in the model was proved by ablation experiments.