A two-stage multi-task learning model for predicting and interpreting received answers in online mental health communities

Online mental health communities (OMHCs) have emerged as a vital platform for individuals with mental health issues to seek supports through questioning. However, there is a scarcity of research focusing on improving the quality of help-seekers’ questions. Furthermore, previous methods have overlooked the complexity of answerers’ decision-making process when answering to questions. To this end, we propose a Two-stage Multi-task Learning Approach (TsMLA) that predicts and interprets received answers for the given question. Our model considers interactions between different views of information by introducing early and late fusion mechanisms to capture memory activation between titles and descriptions. Then, we incorporate the number of answers over a specific period as an auxiliary task to enhance predictive performance. More importantly, motivated by answerers’ decision-making behavior, we design a two-stage approach to consider interactions between different pieces of features. Our experiments demonstrate that TsMLA significantly outperforms several benchmark methods and obtains better performance. Case studies demonstrate the interpretability of our model. Our results have critical implications for health information system to design computational artifacts. Furthermore, our proposed model contributes to management of OMHCs and help-seekers’ experience.