Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RE-RAG: Improving Open-Domain QA Performance and Interpretability with Relevance Estimator in Retrieval-Augmented Generation

Kiseung Kim, Jay-Yoon Lee

TL;DR

Open-domain QA with retrieval-augmented generation suffers when irrelevant contexts distract the model. RE-RAG adds a relevance estimator that provides both relative context relevance and a confidence measure, trained with weak supervision from QA data. The RE is used to rerank contexts and to weight answer marginalization, and jointly trained with the generator using three losses; decoding strategies exploit confidence to mark unanswerable cases or to leverage LLM param knowledge. Experiments on Natural Questions and TriviaQA show that RE-RAG improves over RAG and FiD-based baselines, generalizes to unseen data, and enables efficient decoding strategies. The work advances interpretability and reliability of retrieval-augmented QA, with practical implications for deploying RAG with LLMs.

Abstract

The Retrieval Augmented Generation (RAG) framework utilizes a combination of parametric knowledge and external knowledge to demonstrate state-of-the-art performance on open-domain question answering tasks. However, the RAG framework suffers from performance degradation when the query is accompanied by irrelevant contexts. In this work, we propose the RE-RAG framework, which introduces a relevance estimator (RE) that not only provides relative relevance between contexts as previous rerankers did, but also provides confidence, which can be used to classify whether given context is useful for answering the given question. We propose a weakly supervised method for training the RE simply utilizing question-answer data without any labels for correct contexts. We show that RE trained with a small generator (sLM) can not only improve the sLM fine-tuned together with RE but also improve previously unreferenced large language models (LLMs). Furthermore, we investigate new decoding strategies that utilize the proposed confidence measured by RE such as choosing to let the user know that it is "unanswerable" to answer the question given the retrieved contexts or choosing to rely on LLM's parametric knowledge rather than unrelated contexts.

RE-RAG: Improving Open-Domain QA Performance and Interpretability with Relevance Estimator in Retrieval-Augmented Generation

TL;DR

Open-domain QA with retrieval-augmented generation suffers when irrelevant contexts distract the model. RE-RAG adds a relevance estimator that provides both relative context relevance and a confidence measure, trained with weak supervision from QA data. The RE is used to rerank contexts and to weight answer marginalization, and jointly trained with the generator using three losses; decoding strategies exploit confidence to mark unanswerable cases or to leverage LLM param knowledge. Experiments on Natural Questions and TriviaQA show that RE-RAG improves over RAG and FiD-based baselines, generalizes to unseen data, and enables efficient decoding strategies. The work advances interpretability and reliability of retrieval-augmented QA, with practical implications for deploying RAG with LLMs.

Abstract

The Retrieval Augmented Generation (RAG) framework utilizes a combination of parametric knowledge and external knowledge to demonstrate state-of-the-art performance on open-domain question answering tasks. However, the RAG framework suffers from performance degradation when the query is accompanied by irrelevant contexts. In this work, we propose the RE-RAG framework, which introduces a relevance estimator (RE) that not only provides relative relevance between contexts as previous rerankers did, but also provides confidence, which can be used to classify whether given context is useful for answering the given question. We propose a weakly supervised method for training the RE simply utilizing question-answer data without any labels for correct contexts. We show that RE trained with a small generator (sLM) can not only improve the sLM fine-tuned together with RE but also improve previously unreferenced large language models (LLMs). Furthermore, we investigate new decoding strategies that utilize the proposed confidence measured by RE such as choosing to let the user know that it is "unanswerable" to answer the question given the retrieved contexts or choosing to rely on LLM's parametric knowledge rather than unrelated contexts.
Paper Structure (29 sections, 8 equations, 2 figures, 14 tables)

This paper contains 29 sections, 8 equations, 2 figures, 14 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Traditional RAG overview
  4. RE-RAG framework
  5. Relevance Estimator
  6. Reranking of contexts by relevance
  7. Answer marginalization with context RE
  8. Joint training of RE-RAG
  9. Experimental Setup
  10. Dataset
  11. Evaluation Metric
  12. Baseline
  13. Model
  14. Experiment Results
  15. Main Results
  16. ...and 14 more sections

Figures (2)

  • Figure 1: Overview of our proposed RE-RAG framework. The black lines represent the flow of information and the red lines represent the flow of gradients.
  • Figure 2: The relationship between confidence score and accuracy by model size. RAG means that the model utilizes contextual knowledge and Parametric means that the model utilizes only parametric knowledge without external knowledge.