A Survey of Mental Health Intent Recognition Approaches

Abstract

As mental health conditions are increasing around the world, digital interventions are desperately required in a scale-form. The main focus of these solutions is intent recognition, the capacity to understand user intentions, whether it is a particular support request or indications of suicidal thoughts. The paper will overview the technical history of the area, beginning with the early machine learning methods and concluding with the current state-of-the-art transformer models, discussing the important datasets, multimodal methods and data ethical issues. One of the major issues of our analysis is the so-called deployment gap. We point to a significant mismatch: where NLP models are commonly found to be over 90 percent accurate in carefully controlled settings, they are often found to fail to do so in the real clinical scenarios. Exploring this discrepancy, we claim that the practical utility of academic metrics ought to be prioritized to responsible AI. It is a singularly thorough work that provides a six-decade historical point of view (between ELIZA and LLMs), a comparative analysis of methodologies and actual performance data, and a realistic assessment of the constraints that prevent its use in clinical settings.