# Learning to Filter Context for Retrieval-Augmented Generation

**Authors:** Zhiruo Wang, Jun Araki, Zhengbao Jiang, Md Rizwan Parvez, Graham Neubig  
**arXiv:** [2311.08377](https://arxiv.org/abs/2311.08377)  
**Date:** 2023-11-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2311.08377v1/

## Key points

- FILCO is a method designed to enhance context for generation models by filtering out irrelevant passages during test time
- The effectiveness of FILCO is demonstrated across knowledge-intensive tasks like extractive question answering, complex multi-hop QA, fact verification, and dialog generation
- Promising results are shown in tasks like NQ, TQA, HotpotQA, ELI5, FEVER, and WoW using metrics like Exact Match and Unigram F1
- Implementing FILCO requires training models for context filtering and output generation with varying computational resources depending on model architecture and size
- FILCO outperforms existing methods like RAG, FID, EVI in scenarios where top-5 retrieved passages are filtered by full passages or sentences

## Summary

, , , , 

In the study "Learning to Filter Context for Retrieval-Augmented Generation," the authors introduce FILCO, a method designed to enhance the quality of context provided to generation models by identifying useful information through lexical and information-theoretic approaches. By training context filtering models, FILCO aims to filter out irrelevant passages during test time, ultimately improving overall output quality. The effectiveness of FILCO is demonstrated across various knowledge-intensive tasks such as extractive question answering, complex multi-hop and long-form QA, fact verification, and dialog generation. The authors emphasize that their method shows promising results across different tasks like NQ, TQA, HotpotQA, ELI5, FEVER, and WoW using automatic metrics like Exact Match and Unigram F1. However, they encourage further evaluation through neural- or human-based assessments due to potential inaccuracies in automated measures. Furthermore, implementing FILCO requires training models for both context filtering and output generation which may vary in computational resources depending on the chosen model architecture and size. Additionally, the study compares FILCO's performance with existing methods such as RAG, FID, EVI., revealing superior results in various scenarios when providing top-5 retrieved passages filtered by either full passages or sentences. The authors also discuss related work in augmented generation techniques where additional contexts have proven effective but stress the importance of optimizing granularity and strategy for retrieval to enhance generation accuracy. Overall,<kgd> Learning to Filter Context for Retrieval-Augmented Generation</kgd> presents a novel approach in improving context quality for generation models through efficient filtering mechanisms. The method showcases promising outcomes across diverse knowledge-intensive tasks and encourages further exploration and validation before generalizing conclusions to specialized domain datasets.