Research On Aspect-level Sentiment Analysis Method Of Product Reviews Based On Attention Mechanis

With the development of the Internet at home and abroad,platforms such as online shopping,online ordering and online taxi hailing have been widely favored.It is customary for people to elaborate their experience and feelings about the goods after spending via e-commerce websites,which results in numerous review texts.When we want to quickly understand the product information,we can infer the quality of the product through each review,which not only helps users quickly locate the evaluation of a specific aspect of the product,but also helps merchants efficiently optimize a certain aspect of the product.Therefore,aspect-level sentiment analysis has also received high attention in the industry.Traditional sentiment analysis methods not only require a lot of manual marking,but also feature extraction,and the final result is not highly accurate.Deep learning methods have gradually become the first choice of scholars.Aspect-level sentiment analysis is a fine-grained classification task within the sentiment analysis task.This task can extract the aspect word and the emotional polarity corresponding to the aspect word from the given sentence.Although most of the current studies have achieved good results,most researchers ignore the role of sentence position information,so this paper proposes aspect level sentiment analysis based on joint aspect and position-level attention mechanism network,which combines position information to strengthen feature information,thereby improving the final performance.However,this approach employs a long and short-term memory network,though this method can have a long dependence on the sequence information,the problem that the long and short-term memory network is not parallelized makes it computationally intensive.With this issue in mind,we have proposed an aspect-level emotional analysis that is based on a network of location-based bidirectional attention mechanisms.The main work is as follows:(1)Some models ignore the importance of location information to aspect terminology.To solve this problem,we add a location information feature to the model.At the same time,existing research shows that the interaction between aspect words and context is very important for sentiment analysis at aspect level.Using this idea,this paper proposes a joint aspect and positional level attention mechanism.At the same time,the model adopts the joint method to establish the aspect feature and position feature model.On the one hand,when entering word vector information,the interaction between aspect words and context can be clearly captured.On the other hand,this method can improve the importance of position information in sentences,thereby improving the information retrieval ability of the model.In addition,the model utilizes a hierarchical attention mechanism to extract feature information and obtain different emotional polarities,which is similar to filtering useless information again.Experiments with training datasets show that in the aspect level sentiment classification task,our model has improved performance after comparing multiple benchmark models on restaurant datasets and laptop datasets.(2)In previous work,many researchers have extracted features based on variations of recurrent neural networks.Recurrent neural networks not only have shortcomings in parallel processing,but also have the problem of gradient vanishing when the sequence length exceeds a certain limit.In addition,most studies have overlooked the important role of location information in aspect words.Therefore,we propose an attention network based on positional syntax dependency matrix to overcome the shortcomings of recurrent neural networks.Firstly,the multi-attention mechanism and convolution network are adopted to extract the features in the network.Second,the model combines positional correlation matrices and context embedding,thereby enhancing the position information of words in sentences.Finally,the network uses the two-way attention mechanism to locate and identify aspect words more accurately,so as to better judge the emotional direction corresponding to aspect words.In conclusion,the performance of our proposed model has been enhanced to some degree after testing the datasets and comparing it with other models.