Contextual Ranking Models for Retrieval Augmented Natural Language Inference

• Grant number: 25/01118-2
• Support Opportunities: Scholarships in Brazil - Post-Doctoral
• Start date: April 01, 2025
• End date: March 31, 2027
• Field of knowledge: Physical Sciences and Mathematics - Computer Science
• Principal Investigator: Daniel Carlos Guimarães Pedronette
• Grantee: Francielle Alves Vargas
• Host Institution: Instituto de Geociências e Ciências Exatas (IGCE). Universidade Estadual Paulista (UNESP). Campus de Rio Claro. Rio Claro , SP, Brazil
• Associated research grant: 24/04890-5 - Robust Augmented Retrieval for Natural Language Inference over Transformer-based Models, AP.R

Abstract

Considering the huge challenges posed by the complexity of language aspects to the development of AI-based algorithms capable of comprehending and grasping a language, the emergence of LLMs represents a significant advancement. However, despite the significant advancements and the large spectrum of applications, LLMs face challenges, with important associated limitations. Especially in handling domain-specific or highly specialized queries, which can lead to the generation of incorrect information, or hallucinations. In this scenario, Retrieval Augmented Generation (RAG) has been established as a relevant alternative , providing LLMs with a factual context associated with specific tasks. In general, the entire RAG system is composed of two main components : the retriever and the generator.While the evolution of RAG is promising, it also faces challenges. In this direction, pre and post-processing approaches capable of improving retrieval effectiveness are of fundamental importance and consist in the main focus of this project. The main objective consists in to investigate the follow post-retrieval steps: (i) the use of re-ranking approaches for textual data on the post-processing step of evidence retrieval; (ii) the use of rank aggregation approaches in order to fuse the distinct retrieval results computed by query augmentation techniques.