This paper explores the performance enhancement of psychotherapy chatbots using large-scale language models (LLMs). To overcome the limitations of existing rule-based chatbots, we present two studies that apply expert-written conversational scripts to LLMs. The first study compared rule-based, pure LLMs, and LLMs incorporating expert scripts through fine-tuning and prompting. We found that the scripted LLMs demonstrated superior performance in terms of empathy, conversational relevance, and adherence to therapeutic principles. The second study proposed a more flexible script-strategy-aligned generation (SSAG) method, reducing reliance on expert scripts while maintaining therapeutic effectiveness. SSAG can contribute to reducing expert effort and enhancing the scalability of psychotherapy chatbots.
