RAG-LER: Retrieval Augmented Generation with LM Enhanced Reranker
Abstract: Large language models (LLMs) have demonstrated impressive capabilities across diverse NLP tasks, yet they still struggle with hallucination due to limited parametric knowledge. Retrieval Augmented Generation (RAG) addresses this issue by integrating non-parametric data stores. However, straightforward integration of information retrieval or end-to-end training of these components often leads to suboptimal results or computational inefficiency.
In this work, we introduce RAG-LER, a framework that enhances an LM's context understanding and improves the quality and accuracy of provided passages through an LM-supervised re-ranker. RAG-LER fine-tunes a pre-trained LM to follow instructions and discriminately use provided information. It then leverages this fine-tuned LM to generate ranking scores, which serve as supervised labels for training the re-ranker.
By harnessing LLMs' strong capabilities, our approach eliminates the need for manual human labeling in re-ranker training while achieving improved performance. Experiments demonstrate that RAG-LER outperforms existing retrieval-augmented LMs on open-domain QA and fact-checking tasks, while exhibiting consistently improved performance when applied to different LMs, highlighting its versatility and effectiveness.
Paper Type: Long
Research Area: Language Modeling
Research Area Keywords: Language Modeling, Information Retrieval and Text Mining, Question Answering
Languages Studied: English
Submission Number: 37
Loading