Deep cascaded multitask framework for detection of temporal orientation, sentiment and emotion from suicide notes

Area of Science:

• Computational psychiatry
• Natural Language Processing (NLP)
• Machine Learning

Background:

• Rising global suicide rates necessitate advanced computational tools for early identification of at-risk individuals.
• Accurate analysis of emotional states in suicide notes is crucial for timely psychiatric intervention.
• Existing methods often struggle with the complexity of emotional nuances in suicide-related text.

Purpose of the Study:

• To investigate the role of temporal orientation and sentiment classification in enhancing emotion recognition for suicide risk assessment.
• To develop and evaluate a novel multitask learning framework for joint analysis of these tasks.
• To improve the accuracy of identifying individuals requiring urgent psychiatric treatment.

Main Methods:

• Development of a multi-layer cascaded shared-private attentive network for multitask learning.
• Joint training of primary task (emotion recognition) with auxiliary tasks (temporal orientation and sentiment classification).
• Experimentation on a diversified suicide notes corpus (CEASE-v2.0) for robust evaluation.

Main Results:

• The proposed multitask framework achieved a 3.78% improvement in the Emotion task compared to state-of-the-art systems.
• Achieved a cross-validation Mean Recall (MR) of 60.90% for emotion recognition.
• Empirical and qualitative analyses confirmed a strong correlation between learning temporality, sentiment, and emotion recognition.

Conclusions:

• Jointly learning temporal orientation and sentiment classification significantly enhances emotion recognition in suicide notes.
• The novel multitask attentive network provides a more effective computational approach for suicide risk assessment.
• This research offers a promising direction for developing more accurate and timely computational tools in mental healthcare.