Speech-based Psychological Crisis Assessment using LLMs
A new research paper published on arXiv introduces a Large Language Model (LLM)-based framework designed to automate crisis level classification for psychological support hotlines. Addressing the limitations of human operators, such as inconsistent judgments and staffing constraints, this system aims to enhance the quality and efficiency of mental health emergency services. The proposed method features a novel paralinguistic injection technique that integrates non-verbal emotional cues from speech into text transcripts, allowing the LLM to capture critical acoustic nuances. Additionally, the researchers implemented a reasoning-enhanced training strategy where the model generates diagnostic reasoning chains as an auxiliary task, serving as a regularizer to boost classification accuracy. Combined with data augmentation, the final system achieved a macro F1-score of 0.802 and an accuracy of 0.805 in a three-class classification task under 5-fold cross-validation. This technological advancement represents a significant step toward scalable, automated mental health support tools, leveraging AI to assist in critical decision-making processes during psychological crises.
Wire timeline
Speech-based Psychological Crisis Assessment using LLMs
A new research paper published on arXiv introduces a Large Language Model (LLM)-based framework designed to automate crisis level classification for psychological support hotlines. Addressing the limitations of human operators, such as inconsistent judgments and staffing constraints, this system aims to enhance the quality and efficiency of mental health emergency services. The proposed method features a novel paralinguistic injection technique that integrates non-verbal emotional cues from speech into text transcripts, allowing the LLM to capture critical acoustic nuances. Additionally, the researchers implemented a reasoning-enhanced training strategy where the model generates diagnostic reasoning chains as an auxiliary task, serving as a regularizer to boost classification accuracy. Combined with data augmentation, the final system achieved a macro F1-score of 0.802 and an accuracy of 0.805 in a three-class classification task under 5-fold cross-validation. This technological advancement represents a significant step toward scalable, automated mental health support tools, leveraging AI to assist in critical decision-making processes during psychological crises.
cs.AI updates on arXiv.org