Multilingual and Cross-cultural interactions for context-aware, and bias-controlled dialogue systems for safety-critical applications

Challenge

We are confronted today with a strong change in paradigms relevant to the emergence of LLM (Large Language Model) technologies that are taking up the scenes and presenting themselves as a big part of the future of the digital space. It is evident that, on a regulatory level, the EU is launching strategic and legislative initiatives that aim to regulate and emphasise the role of these solutions. However, the technical powerhouse of the technology is still the American ecosystem. With ELOQUENCE, we try to better understand unstructured dialogues and translate them into explainable, safe, knowledge-grounded, trustworthy and bias-controlled language models in accordance with the strategies and future legislations of the EU. ELOQUENCE will thus work upon the solutions that paved the way for a robust ecosystem in the domain of conversational agents such as the chatGPT, LaMDA2, or LLaMA3. The project will thus exploit pre-trained LLMs and tailor them to specific domains, with considerations given to green computing and eventual regulations on AI. This will result in a more trust-worthy LLM with mitigated impact from the usual memory distortion.

Objectives

ELOQUENCE is set to achieve its goal by putting in place use cases to validate the technologies it will produce. These will be segmented into four settings, the first being a language model learning through decentralised training in smart homes. This setting will pay close attention to the fundamental rights of privacy and data protection. The project will also delve into context-aware language model detecting biases. The project will also aim at retrieval-augmented LLMs as virtual agents, capable of understanding the user’s goals, making calls and responding. And finally the project will aim to develop support call centres through AI-based supervision of multimodal dialogues. By integrating conversational AI agents in contact centres or smart home assistants we can improve outcomes of AI technologies to better identify and resolve the reason of a call, or any request of action.

Our Role

Privanova will lead Ethics Compliance, including the establishment of an independent Ethics Advisory Board for the project. Drawing on its expertise, Privanova will also spearhead Data Protection Compliance, ensuring adherence to relevant regulations, and coordinate the consortium’s efforts to develop and maintain a comprehensive, up-to-date Data Management Plan. Additionally, Privanova will coordinate the creation of the ELOQUENCE Community of Experts, which will include a diverse representation of all relevant stakeholder groups. All project partners will support this initiative by providing contacts. Privanova will also lead Standardisation Recommendations by collaborating with standards-developing organisations (SDOs) to align the project’s technical, ethical, privacy, and interoperability work with relevant standards.