Suicidal ideation recognition based on sentence completion test via coding- and topic-enhanced model

Suicidal ideation refers to the thoughts related to suicide, including but not limited to specific plans for death and desires for suicide. Its recognition is of great significance in preventing individuals from suicide. In the context of large-scale screening on suicide ideation, self-report scale is the most used approach, but the subjects are easy to conceal real information. Though some automated methods based on social media platforms are put forward, they are difficult to cover all the populations that need to be tested. To this challenge, in this paper, a new perspective on suicidal ideation recognition via sentence completion test (SCT) is provided. SCT contains some sentence fragments and requires subjects to complete them, having implicitness and being suitable for large-scale screening, but its use depends on automated scoring method. Therefore, based on a self-developed SCT, a dataset is collected, containing 1,399 individuals’ responses on both the SCT and one classical self-report scale about suicidal ideation. To support the prediction with reasonable evidences for such a psychometric task, considering that individual suicidal ideation may be reflected by the different response patterns of each item or the general topic contained in all items, a coding- and topic-enhanced model for suicidal ideation recognition is proposed. The strategies of contrastive learning and focal loss are leveraged to establish different representations of different response patterns in the SCT and solve the class-imbalanced problem. To verify the effectiveness, extensive experiments are conducted, demonstrating that the proposed method achieves a feasible and practical performance.