Identifying Emotional Risk Signals in Adolescents Using Conversational Context and Lexical Weighting

Abstract

Adolescents often express emotional distress indirectly, and such cues are difficult to detect from surface text alone. This paper proposes a two-stage method for identifying emotional risk signals in Korean adolescent utterances by integrating KoBERT-based emotion recognition, conversational context, and PHQ-9-informed lexical weighting. In the first stage, utterances are encoded with KoBERT and classified into six emotion categories. In the second stage, conversational-context embeddings and PHQ-9 lexical embeddings condition a token-level attention module that generates a weighted sentence representation for risk-signal classification. The final routing strategy maps utterances into positive, general negative, and emotional risk signal classes. Experiments on the adolescent subset of the AI Hub Korean emotional dialogue corpus show that the proposed model achieves 79.4% Macro F1 and 75.0% F1 for the emotional risk signal class, outperforming the text-only baseline. The results indicate that combining contextual and clinically inspired lexical cues helps capture subtle risk-related expressions in adolescent conversations.