JMIR Publications

ABSTRACT

Background:

Suicide has become the fifth leading cause of death worldwide. With the development of the Internet, social media has become an imperative source for studying psychological illnesses such as depression and suicide.

Objective:

Many methods have been proposed for suicide risk assessment. However, most of the existing methods cannot grasp the key information of the text. To solve this problem, we propose an efficient method to extract the core information from social media posts for suicide risk assessment.

Methods:

This paper proposes a multi-feature fusion recurrent attention model for suicide risk assessment. We use the bidirectional long short-term memory network to create the text representation with context information from social media posts. And we introduce the self-attention mechanism to extract the core information. Then we fused linguistic features to improve our model.

Results:

We evaluated our model on the dataset delivered by CLPsych 2019 share task. The experimental results show that our model improves the risk-F1, urgent-F1, and existence-F1 by 3.3%, 0.9%, and 3.7%, respectively.

Conclusions:

We found that BiLSTM performs well in the long text representation, and the attention mechanism can identify the key information in the text. The external features can complete the semantic information lost by the neural network during feature extraction and further improve the performance of the model. The experimental results show that our model performs better than the state-of-the-art method. Our work has a good theoretical and practical value of suicidal risk assessment.