Research On Adverse Drug Reaction Discovery Based On Social Media

Adverse drug reactions(ADRs)refer to the harmful effects produced after the patient takes the drug under the guidance of the physician.The ADRs are usually not related to the purpose of the medication.Identifying ADRs in a timely manner can prevent patients from ADRs and also assist doctors in the rational use of drugs.With the development of the Internet,some patients are willing to share their feelings on social media after medication.Therefore,a large number of patient medication data has been generated.It is more real-time than other resources and provides real data for ADR research.Automatically obtaining ADR information from social media can effectively promote the development in the biomedical domain.However,the small scale of the existing annotated datasets and the problem of social media data itself have limited the development of the field.To solve the above problem,this dissertation applies text mining technology to focus on three tasks:ADR detection,ADR named entity recognition,and ADR relation extraction.First,the ADR detection task can determine the texts containing adverse reactions from massive texts to narrow the scope of research.Then,the ADR named entity recognition task identifies the drugs and adverse reaction entities in the text.Finally,the ADR relation extraction task is used to extract the type of relation between the drug and adverse reaction entity,so as to discover the potential ADR.If a patient has ADRs,it's usually accompanied by sentimental fluctuations.The social media data released by this patient usually contains rich sentimental information.Existing methods integrate sentiment features into ADR detection tasks just by considering sentence-level sentiment polarity scores,ignoring the fine-grained word-level sentiment features.To solve this problem,an adversarial neural network with sentiment-aware attention(ANNSA)model is proposed to enhance the sentimental element in social media.With the sentiment-aware attention,the ANNSA method can obtain the fine-grained sentiment features and realize the effective integration of sentiment features in the ADR detection task.Moreover,it can capture the inner sentiment information of the text and enhance the sentiment expressed in the text.To address the issue of small social media datasets,this dissertation applies the adversarial perturbation mechanism to the ANNSA.Experimental results show that the ANNSA method can effectively integrate sentiment features,and improve the performance of ADR detection tasks based on social media data.For the ADR named entity recognition task,the previous methods usually ignore the characteristics of social media data:serious colloquial phrase and spelling errors.To address the problem of social media,this dissertation proposes an adversarial transfer network with bilinear attention(ATN-BA)model.The ATN-BA is exploited to transfer the normative text expression and professional medical knowledge from PubMed dataset to social media dataset.In ATN-BA,a shared feature extractor equipped with a data discriminator is introduced to alternatively encode social media and PubMed datasets.With adversarial training,the shared feature extractor can obtain the shared features of different data sources.In addition,the bilinear attention mechanism can select the task-specific features from the private features and shared features,identify the adverse drug reaction entities more effectively.Experimental results show that the ATN-BA method can learn the shared features between different sources and improve the performance of ADR identification from social media.For the complex semantic and implicit syntactic structure in social media,this dissertation proposes a gated iterative capsule network(GICN)model for ADR relation extraction task.Firstly,the GICN method uses convolutional neural networks to focus on the local features of the text.Then,the gated capsule network structure is used to capture the deep semantic information.The gated iteration unit in the gated capsule network can fully consider the contextual information of sentences and reserves important information for ADR detection.The parameters in capsule units are updated by the dynamic routing mechanism,which determines the selection process of the upper capsule to the lower one,and reduces the influence of the noise on the final classification performance.Experimental results show that the GICN can solve the problem of informal expression,and obtain a significant improvement on the social media ADR relation extraction task.