Detecting Moments of Change and Suicidal Risks in Longitudinal User Texts Using Multi-task Learning

Abstract

This work describes the classification system proposed for the Computational Linguistics and Clinical Psychology (CLPsych) Shared Task 2022. We propose the use of multitask learning approach with a bidirectional long-short term memory (Bi-LSTM) model for predicting changes in user's mood (Task A) and their suicidal risk level (Task B). The two classification tasks have been solved independently or in an augmented way previously, where the output of one task is leveraged for learning another task, however this work proposes an 'all-inone' framework that jointly learns the related mental health tasks. Our experimental results (ranked top for task A) suggest that the proposed multi-task framework outperforms the alternative single-task frameworks submitted to the challenge and evaluated via the timeline based and coverage based performance metrics shared by the organisers. We also assess the potential of using various types of feature embedding schemes that could prove useful in initialising the Bi-LSTM model for better multitask learning in the mental health domain. * Equal contributions. 1 https://www.who.int/news/item/02- 03-202203- -covid-19pandemic-triggers-25-increase-in-prevalence-of-anxiety-anddepression-worldwide (Accessed on 25.5.2022

Towards Suicide Prevention from Bipolar Disorder with Temporal Symptom-Aware Multitask Learning

“…Although other baselines perform better than STATENet by exploiting sequential data, our model surpasses them by learning the temporal dynamics of BD symptoms. BD symptom task: The results show that our proposed model improves BD symptom prediction performance compared to UoS [3]. While our model directly utilizes contextualized post embeddings to predict BD symptoms in each post, UoS considers post sequences.…”

“…Future Suicidality Assessment Using Social Media Data. While most of the work has focused on identifying the current suicidality revealed in a given post from social media [3,43,44,64,65 3 BIPOLAR DISORDER DATA…”

“…A sequence of posts can show the progressive mood states, which is important information for assessing future suicidality [32]. To generate the semantic representation of each post, we employ the pre-trained Sentence-BERT (SBERT) [61], which showed promising results in detecting moments of change in the mood [3] and representing historical tweets [65]. SBERT is a modification of the pre-trained BERT network that uses siamese and triplet network structures to derive semantically meaningful sentence embeddings by computing the mean of output vectors for all tokens to derive a fixed-size sentence embedding.…”

“…• UoS [3]: UoS is the best performing model at the CLPsych 2022 shared task with Zirikly et al [79] to capture moments of change in a suicidal individual's mood. The obtained embeddings from the pretrained Sentence BERT are fed into a biLSTM layer and a multi-head attention layer.…”

“…Notably, 𝑠 3 has identical posts with 𝑠 1 (i.e., 𝑝 3 1 −𝑝 3 4 ), but the model differently gives attention weights to them. By focusing on a manic symptom in 𝑝 3 4 , the model forecasts lower suicidality than 𝑠 1 , which implies user A's mental status is shown to be improved. Moreover, we find that our model tends to focus on recent posts, giving more attention to the manic symptoms of 𝑝 1 4 , which is a new observation compared to a previous work that claimed that manic episodes less contribute to future suicidality [50]; a future validation is required.…”