The aim of the project ReSpiRa is to exploit structured knowledge bases to perimeter the behaviour of the LMM and strongly verticalise their generative capacity on the content of interest. With this in mind, therefore, techniques for the automatic extraction of knowledge and the construction of Knowledge Graphs (with a focus on relation extraction) will be studied. Relationship extraction (Cabot and Navigli, 2021) is the task of extracting triplets of relationships between entities from raw text, without a given entity range, usually also called end-to-end relationship extraction. In parallel, methodologies for exploiting this knowledge in LLM will be investigated. Two research directions are envisaged here: (i) instruction tuning of LLMs to improve their reliability and specialisation and (ii) designing tasks in such a way that unreliable and/or incorrect LLM outputs can be detected. Both approaches will exploit structured knowledge external to the model.
Project Duration: 18 months
Period: Mar 2024 – Sep 2025
Type of Funding: INNOVATION GRANT – Determina CTS del 07/03/2024
