Research On BiGRU Based A Spect-level Sentiment Classification Method For Drug Reviews

With the rise of online medical field,people began to comment on the use of drugs,efficacy through online platform.It is helpful to understand the effectiveness and side effects of drugs,and provide reference for treatment and nursing.However,due to the indirect difficulties in drug reviews,such as strong subjectivity,obscure sentiment tendency and certain medical vocabulary,most of the existing researches focus on coarse-grained sentiment classification at sentence or document level,which can not more accurately and comprehensively understand the attitude of patients.The research of fine-grained(aspect-level)sentiment classification is more limited,which may be due to the lack of relevant datasets.When the existing aspect level sentiment classification methods apply to drug reviews,there are problems such as low accuracy due to neglecting the meaning of aspect words and insufficient training due to sparse datasets;in addition,the differences of users,the characteristics of drugs and so on will affect the sentiment classification.Only using the attention mechanism cannot dynamically obtain the rich feature expression of the text,which limits the accuracy of classification.Only using the attention mechanism ca not dynamically obtain the rich feature expression of the text,which limits the accuracy of the classification.To solve the above problems,this paper studies the aspect-level sentiment classification task based on the drug review text.The details are as follows:(1)In order to explore the method of sentiment classification based on drug reviews,this paper proposes a dataset of aspect-level drug reviews sentiment classification by manual annotation,called SentiDrugs,and verifies the validity of dataset by Kappa value.The data set SentiDrugs contains 4028 aspect-level drug reviews,each of which involves one or more aspects.Sentiment polarity divides into three categories:negative,neutral and positive.The data set provides an important research foundation for the research of relevant classification methods.(2)Aiming at the problems that the existing models ignore the positive effect of the semantics of aspect words on sentiment classification results and the limited scale of aspect-level drug review datasets,this paper proposes a sentiment classification model PM-DBiGRU based on double BiGRU and knowledge transfer.The BiGRU network is established for aspect words and drug reviews respectively to and then the sentence representation of specific aspect words is further obtained by using the attention mechanism,so that the model can judge the sentiment polarity according to the semantics of aspect words.Then it uses the attention mechanism to express the reviews about the specific aspect words.At the same time,this paper proposes two methods of knowledge transfer to learn semantic knowledge from large-scale short-text level drug reviews with semantic similarity,in order to improve the classification effect as much as possible under the condition of limited aspect level dataset.One is to transfer the weight learned from the pre-training model based on short-text level to aspect-level model hierarchically,the other is to use multi-task learning in order to share deep semantic knowledge by training aspect-level and short-text level tasks together.The experimental results demonstrate that the proposed PM-DBiGRU performs better than the contrast model in the dataset SentiDrugs.(3)PM-DBiGRU model can classify sentiment polarity according to the semantics of aspect words,and use knowledge transfer method to obtain domain knowledge,which achieves better classification effect.In addition,the unstructured expression of drug reviews makes the sentence structure of some drug reviews complex and obscure,the fixed coding problem of attention mechanism makes the model unable to obtain more effective information dynamically to a certain extent.To solve the above problems,this paper futher proposes a BiGRU sentiment classification model DRBiGRU-UD based on dynamic routing and external information(user information and drug information).In addition to considering aspect words,the model also extracts the implicit sentiment features of user and drug information in sentences to deepen the understanding of the sentence.In addition,based on attention mechanism,dynamic routing algorithm is used to dynamically capture rich text spatial features such as semantic information,word location information and grammatical structure to better classify.Finally,the weighted loss function is utilized for representing the importance of aspect words,users and drug information to results.The experimental results show that the DRBiGRU-UD model is effective in dataset SentiDrugs.