Evaluating Object Detection in Chatbot Interactions for Mental Health Assessment

Weilin Sun^*

• School of Electronic Information Engineering, Taiyuan University of Science and Technology,, Taiyuan, China

The increasing integration of conversational agents into behavioral and mental health domains necessitates advanced evaluation methodologies tailored to the complexity of these interactions. Conventional metrics, including BLEU and ROUGE, are insufficient for assessing the nuanced and context-dependent nature of mental health dialogues, as they fail to account for persona alignment, therapeutic coherence, and contextual relevance. To overcome these limitations, this study introduces an innovative evaluation framework that combines formal task modeling, interlocutor-aware embedding, and knowledge-grounded reasoning. At the core of this framework lies the Interlocutor-Aware Latent Evaluation Network (IALENet), a dual-encoder architecture designed to capture dialogue quality within a shared latent semantic space that reflects interaction dynamics. IALENet integrates speaker-role embeddings and temporal attention mechanisms to extract evaluative features aligned with conversational fluency and therapeutic consistency. To enhance this, the Evaluation-Augmented Dialogue Alignment (EADA) strategy incorporates external knowledge bases and discourse priors, ensuring evaluations are calibrated to psychological expectations. By modeling response coherence, pragmatic relevance, and semantic fluency through structured augmentation and residual alignment, this approach produces interpretable and context-sensitive metrics that closely align with expert human assessments. Experimental results demonstrate that the proposed system achieves superior robustness and granularity across diverse chatbot-generated dialogue datasets in mental health settings,providing a scalable and theoretically grounded dialogue evaluation framework specifically designed for mental health chatbot interactions.