D2.3 Methodology for jointly training FIR and response generator

This report summarizes research within the ELOQUENCE project on factual information retrieval (FIR) with conversational/speech queries and the design of a modern, LLM-based dialogue manager. A key contribution is a speech-enabled methodology for the joint training of information retrieval and response generation components, which utilizes a connector-based architecture that was developed within the WP2 and presented in previous deliverables. The resulting system outperforms text-only baselines, although it introduces challenges in maintaining retrieval performance. To address data scarcity and mitigate system bias, the report introduces novel methods for creating controlled training data, including a framework for generating psychologically coherent user personas based on Italian demographic data and the SDialog Python library for creating reproducible, orchestrated synthetic dialogues. Finally, the report outlines the architecture for an LLM-based Dialogue Manager designed to orchestrate the conversational system and discusses its future integration into ELOQUENCE pilots.