Información del proyecto

ID: PID2023-150584OB-C21

Coordinador institucional

Financiación

Más información en

Análisis de autorías institucional

Pardo MuÑoz, Jose ManuelMiembro del equipo de trabajoMartin Fernandez, IvanMiembro del equipo de trabajoEsteban Romero, SergioMiembro del equipo de trabajoGil Martin, ManuelParticipanteFerreiros López, JavierParticipanteMontero Martínez, Juan ManuelInvestigador principalFernández Martínez, FernandoInvestigador principal

Compartir

Armonizando Flexibilidad y Conformidad en Sistemas de Inteligencia Artificial Conversacional

Investigadores/as: Montero Martínez, Juan Manuel (Investigador principal (IP)); Ferreiros López, Javier (Participante); GIL MARTIN, MANUEL (Participante); ESTEBAN ROMERO, SERGIO (Miembro del equipo de trabajo); MARTIN FERNANDEZ, IVAN (Miembro del equipo de trabajo); PARDO MUÑOZ, JOSE MANUEL (Miembro del equipo de trabajo); Lebai NA, Syaheerah (Miembro del equipo de trabajo); Fernández Martínez, Fernando (Investigador principal (IP))

Afiliaciones

Resumen

The proposed project, TrustBoost, addresses the evolving landscape of Conversational AI, focusing on the integration of Language Model based Systems (LLMs) with rule-based constraints to enhance trustworthiness, compliance, and adaptability. The current prominence of Conversational AI, spanning dialog systems, chatbots, and virtual assistants, has been significantly shaped by the development of LLMs, allowing for more natural language interactions. While LLMs offer unprecedented adaptability and versatility, their lack of strict adherence to business rules poses trust challenges. Intent based systems, on the other hand, excel in compliance but lack the flexibility and naturalness of LLMs. Trust in Conversational AI is vital for user engagement, and the lack of adherence to business rules can lead to diminished trust and potential consequences. TrustBoost aims to strike a balance by endowing LLM-powered systems with the ability to comply with business/domain rules while adapting to user needs. The project recognizes the challenges of incorporating rule-based constraints into LLM training, emphasising the need for fine-tuning processes and interpretability. To address these challenges, TrustBoost also focuses on the affective branch, aiming to dynamically tailor responses based on user emotional cues, enhancing personalization and empathy. Recognizing emotional cues is a complex challenge, requiring advancements in NLP techniques and the integration of multimodal data. TrustBoost emphasises contextualising emotional cues within conversations and users' overall situations to provide context-aware interactions. The project acknowledges the limitations of current LLMs in understanding emotional context and proposes advancements in emotional state recognition algorithms. In summary, TrustBoost seeks to advance Conversational AI by addressing the dual challenges of compliance and adaptability. Through the integration of rule-based constraints, prompt-based learning, and emotional intelligence, the project aims to create more trustworthy, compliant, and emotionally aware Conversational AI systems, ultimately enhancing user experience and engagement. (Objectives)

Palabras clave

Instituciones participantes

Indicios de calidad

Ítems relacionados