Research On Health Issue Analysis Methods Based On Social Media Data

In the past decades,with the rapid development of various social media,more and more users post information related to health issues through social media.Research based on artificial intelligence has achieved great success,but there are also many challenges.On the one hand,through the social media data released by the public,we can analyze the public attitudes towards some public health issues.However,this health information is usually unlabeled,which limits the application of deep learning methods.Meantime,labeling large-scale vaccination-related dataset is expensive.On the other hand,through users' information containing personal health issues,some studies focus on the detection of potential depression,anorexia and other mental disorders.However,the existing mental health detection models do not consider the impact of the data imbalance on the model and neglect to incorporate a large amount of prior knowledge in the knowledge graph to further improve the model performance.In order to solve the problem of lacking labeled dataset,this paper proposes three methods based on transfer learning to analyze the attitudes towards public health issues.One is to transfer static word vectors and dynamic ELMo vectors respectively,and then process through the bidirectional gated recurrent neural network and an attention mechanism.The others are to fine-tune the pre-trained language models GPT and BERT with limited labeled data.Finally,the experimental results of the three methods on the Twitter HPV dataset show that the three proposed methods based on transfer learning perform well in HPV vaccination attitude analysis task,and the fine-tuned BERT model achieves the best performance.In order to solve the problem that the existing mental health detection models neglect to consider the impact of the data imbalance and incorporate the domain knowledge,this paper proposes a knowledge-enhanced ensemble learning model for mental health detection.First,a new sentence containing prior knowledge is obtained by injecting matched triple information into the original sentence.Then,a knowledgeenhanced base classifier is obtained through the BERT embedding layer,bidirectional gated recurrent neural network layer,and attention layer.Finally,the Ada Boost ensemble learning algorithm is used to integrate the base classifier to reduce the impact of data imbalance on the model and obtain the final mental health detection model.The best performance of the model in the RSDD depression detection task and the CLEF e Risk 2018 anorexia detection task shows that it has certain practicality in the field of psychiatric detection.