Persona-Driven Bilingual LLM Dialogues: Cultural Metaphors, Linguistic Divergence, and Emotional Disclosure

Large Language Models (LLMs) can take on personas that provide culturally informed advisor-client conversations that facilitate mental health therapy. This paper is a comparative study of bilingual, persona-based therapeutic conversations created by LLMs in the person of either a client or a counselor. We will compare an English-speaking big brother (Raoul) and a Chinese auntie persona ("Aunt Mei" or Mei Yi) in conversations concerning depression. We will use quantitative linguistic analysis, tracking sentiment trajectory, an exploratory self-disclosure scoring system, and a lexicon of culturally based metaphors to examine how these personas impact the conversation dynamic. The results show the two groups diverge in the language they use. The Chinese Auntie Mei persona uses multitudes of proverbs and social-relational types of communication, while the Western Raoul persona makes more use of first person plural ("we") inclusivity. These differences in language types impact the level of self-disclosure registered. Psychiatric clients talking in English exhibit much higher levels of use of both "I" pronouns and negative emotional words from the beginning. The Chinese clients, in contrast, reveal their feelings in a more indirect, gradual manner. The sentiment trajectory shows the sentiment of Chinese (ZH) clients rises steeply and quickly, while the sentiment trajectory of English (EN) clients proceeds more gradually. The main result of our study is that culturally grounded personas impact the therapeutic process in ways beyond mere linguistic flavor. They impact and frame the self-disclosure and emotional improvement trajectories themselves. We will discuss the implications for design of culturally informed bilingual therapeutic AI which respects local narrative forms while facilitating self-disclosure.