Empathetic Text Style Transfer Layer for Conversational Agents a modular architecture and new resource for the italian language

The integration of Emotional Intelligence (EI) into Artificial Intelligence (AI) systems represents a central challenge in the advancement of Human-Computer Interaction (HCI), particularly in domains such as healthcare where communication requires both factual accuracy and appropriate emotional attenuation. Despite advances in natural language generation, current conversational agents rarely exhibit genuine empathetic behavior, often producing neutral or affectively inadequate responses. This limitation is even more evident for the Italian language, where the development of empathy‑aware systems is constrained by the scarcity of high‑quality annotated resources and the absence of parallel corpora for controlled stylistic generation. Moreover, incorporating empathetic capabilities directly into existing domain‑specific conversational agents entails significant risks for system reliability and imposes high retraining and maintenance costs. To address these constraints, this thesis introduces a modular framework based on an Empathetic Text Style Transfer. This architectural paradigm decouples the generation of empathetic expression from the underlying dialogic and domain‑expert components, enabling the enhancement of pre‑existing conversational agent infrastructures without modifying their internal logic or compromising factual robustness. A key contribution of this work is the construction and rigorous human evaluation of IDRE (Italian Dialogue for Empathetic Responses), a novel dataset consisting of curated triplets—user query, neutral response, and empathetic reformulation specifically designed to support supervised training for empathetic text style transfer while guaranteeing semantic fidelity. This research bridges a critical gap in Italian NLP resources by introducing IDRE, the first corpus for empathetic text style transfer in this language. Extensive experimental analysis comparing ten distinct model architectures reveals that fine-tuning on the IDRE dataset significantly outperforms few-shot learning strategies, enabling even compact models (approximately 1 billion parameters) to achieve high levels of lexical diversity and emotional resonance. The results confirm that this modular capability is not only robust within the medical domain but generalizes effectively to financial, legal, and social contexts. Ultimately, this work provides a scalable, cost-effective, and computationally efficient framework for deploying human-centric AI, offering a sustainable alternative to resource-intensive commercial models.