Controlled Text Generation for Emotional Support an‎d Enhancing Positivity in Persian

10/22/2026 12:00:00 AM

The present study aims to design an‎d implement an intelligent, controllable framework for generating empathetic an‎d supportive texts in Persian by leveraging the capabilities of large language models (LLMs) in the domain of mental health. The main challenge addressed in this research lies in the limitations of current models, including their superficial understan‎ding of users’ psychological states an‎d the scarcity of data resources an‎d specialized eva‎luation frameworks for the Persian language. To this end, three dedicated datasets were designed an‎d collected, namely PsychoLexeva‎l, PsychoLex-Query, an‎d PsychoLex-Dialog, which were used to eva‎luate the psychological knowledge of LLMs, assess their performance in single-turn interactions, an‎d examine system effectiveness in multi-turn conversations with long-term memory, respectively. The proposed framework, referred to as PsychoLexTherapy, simulates the mental processes of psychotherapists across three therapeutic approaches: Cognitive Behavioral Therapy (CBT), Reality Therapy (RT), an‎d Person-Centered Therapy (PCT). This framework incorporates two key modules: a therapy-approach selec‎tor for determining the most appropriate response strategy, an‎d a long-term memory manager for maintaining coherence an‎d continuity across interactions. eva‎luation results indicate that PsychoLexTherapy, particularly in its long-term memory–enhanced version, consistently outperforms multiple baselines (including simple pro‎mp‎ting, empathy chains, an‎d empathetic agents) across all psychological an‎d linguistic criteria such as empathy, coherence, cultural alignment, an‎d personalization. These findings mark a significant step toward developing intelligent systems that are empathetic, culturally adaptive, an‎d tailored to the emotional support an‎d mental health needs of Persian-speaking users.

توليد كنترل شده متن , پشتيباني عاطفي , مدل‌هاي زباني بزرگ , روان‌درماني

psychotherapy , Large Language Models , Emotional Support , controlled text generation

Mohammad Amin Abbasi

Hassan Naderi