Context Filtering with Reward Modeling in Question Answering

TL;DR

This paper presents a reward modeling-based context filtering method for question answering that improves token efficiency and performance by removing irrelevant information from contexts, especially benefiting low-resource scenarios.

Contribution

It introduces a novel context filtering approach using reward modeling to enhance QA performance without relying on costly human evaluation.

Findings

6.8-fold increase in EM Per Token (EPT) metric

Significant performance improvement over baseline

Enhanced token efficiency in low-resource settings

Abstract

Question Answering (QA) in NLP is the task of finding answers to a query within a relevant context retrieved by a retrieval system. Yet, the mix of relevant and irrelevant information in these contexts can hinder performance enhancements in QA tasks. To address this, we introduce a context filtering approach that removes non-essential details, summarizing crucial content through Reward Modeling. This method emphasizes keeping vital data while omitting the extraneous during summarization model training. We offer a framework for developing efficient QA models by discerning useful information from dataset pairs, bypassing the need for costly human evaluation. Furthermore, we show that our approach can significantly outperform the baseline, as evidenced by a 6.8-fold increase in the EM Per Token (EPT) metric, which we propose as a measure of token efficiency, indicating a notable token-efficiency boost for low-resource settings.