Depression Detection Based On Time-Aware Social Media Text

With the development of the Internet in recent years,depression detection for users has attracted more and more attention in the field of social media text.The content of social media texts is complex and various,and the data on depression is extremely unbalanced.In the sequence model,the text depression detection based on the Recurrent Neural Networks(RNN)still has the following problems and challenges:the standard RNN defaults the time interval to be equal between the adjacent units,and the validity of time in the text depression detection has not been verified;in the field of clinical text processing,the existing model,Time-Aware LSTM(TLSTM),does not consider the memory weight and time globally.The main work and innovation of this thesis are as follows:(1)To address the problems of large noisy text and unbalanced data,this thesis fully preprocesses the data on three data sets(RSDD^[1],Shen et al.^[2],Wolohan et al.^[3]data sets).Experiments include the removal of web links,removal of numbers,re-moval of stop words,removal of text emojis,the replacement of slang and acronyms,tokenization,lemmatization.In addition,this thesis carries out data screening and un-dersampling to eliminate the performance impact of data imbalance in the experiments.(2)Aiming at the status that the time interval has not been considered in the field of text depression detection,this thesis applies TLSTM to model the users,and ex-tracts the semantic relationships between users'adjacent texts through the time aware mechanism to enhance LSTM's ability to process sequence data in the field of de-pression detection on social media text.Experiments have indicated that in the field of depression detection on social media texts,the sequence model LSTM is effective,and the time aware mechanism(uneven time interval)can improve the performance of text depression detection effectively.(3)Based on the above work,this thesis proposes a self attention-based time aware LSTM model(SATA-LSTM).Furthermore,the thesis proposes a text depression de-tection model based on SATA-LSTM.SATA-LSTM solves the limitations of TLSTM processing time interval and LSTM updating unit memory weights.It utilizes self at-tention mechanism and global time aware mechanism to extract the weights of LSTM memory units.Experimental results and analysis have proved that the performance of SATA-LSTM is better than other baseline models on the five basic evaluation indica-tors(Accuracy,Precision,Recall,F1,AUC);in addition,this thesis also proves the effectiveness of the self attention and time aware mechanisms in SATA-LSTM.In order to verify the effectiveness of time-aware and SATA-LSTM,experiments are carried out on multiple social media text data sets.Experiments'results show that SATA-LSTM can improve the performance on depression detection and effectively identify users at risk of depression.This thesis provides a new alternative scheme and idea for text depression detection,which has certain reference value for the application of auxiliary methods about clinical depression diagnosis.