Timely identification of suicide risk is an important and challenging application area for modern machine learning and data science. Written suicide notes provide a rare, high–signal source of information about the cognitive and emotional states that precede lethal self-harm, but they also raise strict requirements on model reliability, transparency, and responsible use. Existing work on the CEASE family of corpora has progressed from traditional classifiers to deep recurrent multitask models and, more recently, VAD-assisted transformer architectures, yet current systems typically optimise only for classification accuracy, treat discrete and dimensional affective constructs in a fragmented way, and pay limited attention to probability calibration or human-centred explainability. In this paper we introduce a Calibrated VAD-Aware Multitask Transformer (C-VAMT) for sentence-level analysis of suicide notes, situated at the intersection of natural language processing, deep learning, and responsible AI. Our model jointly predicts (i) multi-label fine-grained emotions, (ii) three-way sentiment polarity, and (iii) continuous Valence–Arousal–Dominance (VAD) scores within a single transformer-based architecture. We enrich the encoder with lexicon-derived VAD features and propose a label-graph attention mechanism that explicitly models dependencies among emotion and sentiment labels. A hybrid supervision scheme combines large-scale lexicon-based VAD estimates with a smaller set of human-validated VAD and sentiment annotations. To support risk-sensitive decision-making, we incorporate temperature scaling and Monte Carlo dropout for calibrated uncertainty estimates, and we derive token-level rationales via integrated gradients, which are evaluated by clinicians. Experimental results on CEASE-v2.0 show that C-VAMT improves macro-F1 and mean recall for multi-label emotion recognition, particularly on rare emotion categories, while also reducing expected calibration error and producing clinically plausible explanations. The proposed framework illustrates how advanced deep learning, multitask learning, and explainable AI techniques can be integrated into a coherent, human-centred pipeline for a societally critical application of computer and data sciences.
Timely identification of suicide risk is an important and challenging application area for modern machine learning and data science. Written suicide notes provide a rare, high–signal source of information about the cognitive and emotional states that precede lethal self-harm, but they also raise strict requirements on model reliability, transparency, and responsible use. Existing work on the CEASE family of corpora has progressed from traditional classifiers to deep recurrent multitask models and, more recently, VAD-assisted transformer architectures, yet current systems typically optimise only for classification accuracy, treat discrete and dimensional affective constructs in a fragmented way, and pay limited attention to probability calibration or human-centred explainability. In this paper we introduce a Calibrated VAD-Aware Multitask Transformer (C-VAMT) for sentence-level analysis of suicide notes, situated at the intersection of natural language processing, deep learning, and responsible AI. Our model jointly predicts (i) multi-label fine-grained emotions, (ii) three-way sentiment polarity, and (iii) continuous Valence–Arousal–Dominance (VAD) scores within a single transformer-based architecture. We enrich the encoder with lexicon-derived VAD features and propose a label-graph attention mechanism that explicitly models dependencies among emotion and sentiment labels. A hybrid supervision scheme combines large-scale lexicon-based VAD estimates with a smaller set of human-validated VAD and sentiment annotations. To support risk-sensitive decision-making, we incorporate temperature scaling and Monte Carlo dropout for calibrated uncertainty estimates, and we derive token-level rationales via integrated gradients, which are evaluated by clinicians. Experimental results on CEASE-v2.0 show that C-VAMT improves macro-F1 and mean recall for multi-label emotion recognition, particularly on rare emotion categories, while also reducing expected calibration error and producing clinically plausible explanations. The proposed framework illustrates how advanced deep learning, multitask learning, and explainable AI techniques can be integrated into a coherent, human-centred pipeline for a societally critical application of computer and data sciences.
