<![CDATA[This study develops and evaluates a retrieval-based therapist-facing session copilot for live counseling support using the public bilingual Psy-Insight dataset. Rather than providing autonomous psychotherapy or relying on a generative large language model (LLM), the system assists human therapists by ranking historical responses, retrieving interpretable rationales, and providing conservative contextual support. The reproducible pipeline combines TF-IDF representations, class-balanced LinearSVC routers, nearest-neighbor rationale retrieval, and label-aware response ranking without LLM fine-tuning. Experiments use all 520 English and 431 Chinese sessions (6,208 and 5,776 turns, respectively) with session-level train/dev/test splits. Psychotherapy routing achieves strong macro-F1 scores of 0.897 in English and 0.757 in Chinese, whereas strategy routing remains weak (0.253 and 0.268). Label-aware rationale retrieval improves ROUGE-L from 0.145 to 0.152 in English and from 0.142 to 0.151 in Chinese. The best response-ranking approach presents retrieved reasoning in parallel rather than through reasoning-fused reranking, increasing MRR from 0.498 to 0.541 in English and from 0.519 to 0.523 in Chinese while maintaining low latency (6.98–14.64 ms/query). These results demonstrate computational feasibility but do not establish therapeutic safety or clinical effectiveness.]]>
Copyrights © 2025
