Long Context Question Answering via Supervised Contrastive Learning

TL;DR

This paper introduces a contrastive learning approach to improve long-context question answering by better identifying supporting evidence, leading to enhanced performance on benchmark datasets.

Contribution

It proposes a novel sequence-level contrastive supervision method for long-context QA models, improving evidence identification and overall accuracy.

Findings

Consistent performance improvements on HotpotQA and QAsper benchmarks.

Enhanced evidence sentence discrimination through contrastive loss.

Improved long-context QA accuracy across multiple transformer models.

Abstract

Long-context question answering (QA) tasks require reasoning over a long document or multiple documents. Addressing these tasks often benefits from identifying a set of evidence spans (e.g., sentences), which provide supporting evidence for answering the question. In this work, we propose a novel method for equipping long-context QA models with an additional sequence-level objective for better identification of the supporting evidence. We achieve this via an additional contrastive supervision signal in finetuning, where the model is encouraged to explicitly discriminate supporting evidence sentences from negative ones by maximizing question-evidence similarity. The proposed additional loss exhibits consistent improvements on three different strong long-context transformer models, across two challenging question answering benchmarks -- HotpotQA and QAsper.