Depression Tendency Processing Based on LSTM Technique for Text Emotion Recognition

Abstract

Depression is a common mental disorder that can lead to suicide in specific severe cases. A large number of suicides occur each year due to a lack of timely observational attention and treatment worldwide. This paper indicates that social media posts should be monitored for implicit depression and categorized immediately to reduce the risk of suicide and improve public mental health. In the data preprocessing, the NLTK library is used to slice the labeled social media text collected from Kaggle. Words not relevant to sentiment analysis are deleted. Words from sentiment lexicon can represent the positive and negative and filter the noise. The original data was randomly divided into training set(70%) and prediction set(30%). In the machine learning stage(for classification results and evaluation), a bidirectional LSTM mode( used to solve the problem of gradient explosion or disappearance from the preamble text.)l is used. The accuracy, precision, recall and F1-score are 72.85%, 73.88%, 72.37%, 73.12%. They can be components of confusion matrix to evaluate the model. It is difficult for CNN algorithm to capture good n-gram feature results. The quality of the model can be improved by using higher quality and larger numbers of datasets and noise handling, perfecting a better sentiment lexicon and adjusting model parameters, to prevent overfitting. Finally, the most suicidal users(top 3.6% high score) in the text to prevent the tragedy of suicide can be found and more attention to public mental health and safety should be paid.