Year: 2024
Type: dissertation
Abstract: Large Language Models (LLMs) can generate textual information that can be factually incorrect. This occurs when the model fails to accurately represent or reason about the real world. Retrieval-Augmented Generation (RAG) allows an LLM to refer to external data during the generation process to enhance and improve the factual correctness and contextual understanding of textual information. As of right now existing generation models are not up to par with academic standards and excessively hallucinate information that they don't have in their training. The challenge is to minimize the hallucination problem of LLM and to generate accurate answers as it is very important in the medical field. This research proposes a RAG framework that uses Large Language Models for accurately generate answers for medical datasets. The framework combines Retrieval-Augmented Generation techniques, LLMs, and knowledge base stored in a vector database. The LLM models are implemented using Llama 3, Gemma2, Mistral, GPT2. The results of these models, integrated with RAG, are evaluated based on the following metrics: 'BLEU', 'ROUGE-1', 'BERT P', 'BERT R', 'BERT F1', 'Perplexity' score. This research will benefit medical researchers and doctors in getting semantically correct medical information without hallucinations.
Author: Hussain, Madni Ali
Institution: National College of Ireland
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7005668061