Loriaux, Quentin (2024) Questions / Answers : Language models helped by information retrieval ; detection of the need to retrieve PRE - Research Project, ENSTA.
![]()
| PDF 1072Kb |
Abstract
Nowadays, AI models such as ChatGPT are part of our everyday life and we sometimes rely too much on the knowledge they impart us. The field of Questions/Answers is the place where researchers try to enhance the quality and make more reliable what these models generate. It is at the University of Montreal that I worked on the question of how to enhance the Retrieval Augmented Generation (RAG) in order to make models answers to questions more accurate, thanks to automated retrieval of external knowledge. Firstly, I analyzed a recent framework called "DRAGIN" which determines dynamically the need to use RAG thanks to the attention mechanism. Secondly, I modified his framework to create a new way of detecting this need, based on the substitution of sentences. This new framework I named DRAGINUS will still need to be tested and enhanced in order to be popularized among RAG techniques.
Item Type: | Thesis (PRE - Research Project) |
---|---|
Uncontrolled Keywords: | Natural Language Processing (NLP), Large Language Model (LLM), Information Retrieval (IR), Retrieval Augmented Generation (RAG) , fewshot, hallucination, attention |
Subjects: | Information and Communication Sciences and Technologies |
ID Code: | 10029 |
Deposited By: | Quentin LORIAUX |
Deposited On: | 28 août 2024 19:06 |
Dernière modification: | 28 août 2024 19:06 |
Repository Staff Only: item control page