پيش‌بيني سبك همدلي جهت شخصي‌سازي پرامپت در چت‌بات فارسي با استفاده از تيپ شخصيتي و ابعاد فرهنگي كاربر جهت بهبود علائم PMS

Predicting empathic style for pro‎mp‎t personalization in a Persian chatbot using users’ personality types an‎d cultural dimensions to improve PMS symptoms

مهندسي فناوري اطلاعات

Premenstrual syndrome (PMS) is a collection of physical, mood-related, an‎d behavioral symptoms that occur during the days preceding menstruation an‎d can affect women’s psychological well-being, daily activities, an‎d quality of life. Estimates indicate that the global average preva‎lence of premenstrual syndrome is approximately 48%, while its reported preva‎lence in Iran is 98%. Given differences in individuals’ communication needs, the uniform responses of large language model (LLM)-based chatbots do not necessarily align with their individual characteristics an‎d communication styles. Therefore, this study proposes a method for predicting empathy style to personalize pro‎mp‎ts in a Persian-language chatbot, using users’ personality types an‎d cultural dimensions, with the aim of improving premenstrual syndrome symptoms. In this method, users’ personality, cultural, an‎d premenstrual syndrome-related symptom features were extracted from questionnaire data. To automatically extract symptom-related features from text, a symptom extraction model was developed with a mean F1-score of 70%. Furthermore, after synthetic data generation an‎d optimization, the empathy-style prediction model, trained on 726 samples using questionnaire-derived features, achieved an accuracy of 85%. The predicted empathy style was then used to personalize responses, an‎d empathy-style-specific pro‎mp‎t templates were designed an‎d optimized using behavioral guidelines an‎d few-shot examples. To eva‎luate the proposed method, the performance of the personalized chatbot was compared with that of a baseline version in terms of symptom changes an‎d response quality. Response quality was eva‎luated by a psychology specialist an‎d a large language model. The results showed that, for the mood instability component, the mean improvement in the personalized group was 0.488, compared with 0.008 in the baseline group. This difference was statistically significant an‎d was associated with an effect size of 0.539. In the subgroup that interacted with the personalized model over multiple days, a mean improvement of 0.526 in physical symptoms was observed, with an effect size of 0.804 an‎d a statistically significant difference. Moreover, the personalized version demonstrated better performance in terms of perceived empathy an‎d perceived effectiveness. Specifically, the mean perceived empathy score increased from 2.121 in the baseline group to 3.650 in the personalized group, while the mean perceived effectiveness score increased from 2.394 to 3.675. These differences were associated with effect sizes of 0.805 an‎d 0.748, respectively. The response-quality eva‎luation also demonstrated the superiority of the personalized version. The mean overall score in the specialist eva‎luation increased from 3.26 in the baseline group to 4.18 in the personalized group, while the mean score in the LLM-based eva‎luation increased from 3.22 to 4.74. Overall, the findings indicate that integrating premenstrual syndrome symptom-related feature extraction, empathy-style prediction, an‎d personalized response generation by a large language model can provide a suitable approach for developing empathetic Persian-language chatbots that are aligned with the communication needs of users experiencing premenstrual syndrome.

6

167957