[Paper Review] Can NLI Models Verify QA Systems' Predictions?

Jifan Chen, Eunsol Choi+1 more|arXiv (Cornell University)|Apr 18, 2021

Topic Modeling59 references4 citations

TL;DR

This paper proposes a novel framework that leverages natural language inference (NLI) models to verify the correctness of question answering (QA) system predictions by converting QA instances into premise-hypothesis pairs through question conversion and decontextualization. The approach improves QA model calibration and identifies cases where answers are correct for the wrong reasons, with gains of up to 5.5 F1 points in selective prediction settings.

ABSTRACT

To build robust question answering systems, we need the ability to verify whether answers to questions are truly correct, not just "good enough" in the context of imperfect QA datasets. We explore the use of natural language inference (NLI) as a way to achieve this goal, as NLI inherently requires the premise (document context) to contain all necessary information to support the hypothesis (proposed answer to the question). We leverage large pre-trained models and recent prior datasets to construct powerful question converter and decontextualization modules, which can reformulate QA instances as premise-hypothesis pairs with very high reliability. Then, by combining standard NLI datasets with NLI examples automatically derived from QA training data, we can train NLI models to judge the correctness of QA models' proposed answers. We show that our NLI approach can generally improve the confidence estimation of a QA model across different domains, evaluated in a selective QA setting. Careful manual analysis over the predictions of our NLI model shows that it can further identify cases where the QA model produces the right answer for the wrong reason, or where the answer cannot be verified as addressing all aspects of the question.

Motivation & Objective

• To address the lack of robustness in QA systems that often produce plausible but incorrect answers due to distribution shifts or spurious patterns.
• To improve confidence calibration in QA models by using NLI as a verification mechanism to distinguish between correct and incorrect reasoning.
• To investigate whether NLI-based verification can detect cases where QA models arrive at the right answer using insufficient or irrelevant context.
• To evaluate the compatibility of NLI models trained on mixed datasets (including automatically generated QA-derived NLI examples) with out-of-domain QA generalization.
• To enable explicit error analysis by isolating information mismatches between question and supporting context using NLI's entailment criterion.

Proposed method

• Convert (question, context, answer) triplets into (premise, hypothesis) NLI pairs using a two-stage pipeline: question conversion and decontextualization.
• Employ a strong pre-trained T5 model fine-tuned for question-to-declarative-sentence conversion to improve hypothesis formulation quality.
• Apply a decontextualization model to rewrite the answer sentence from the context so it stands alone without document support, preserving semantic integrity.
• Construct a large-scale NLI dataset by combining standard NLI datasets (e.g., SNLI, MNLI) with automatically generated NLI pairs derived from QA training data.
• Train an NLI model on the combined dataset to predict entailment, contradiction, or neutral relationships between premise and hypothesis.
• Use the trained NLI model as a verifier to re-rank or recalibrate predictions from a base QA model, especially in low-confidence or out-of-domain settings.

Experimental results

Research questions

• RQ1 Can NLI models effectively verify QA system predictions by assessing whether the answer is logically entailed by the supporting context?
• RQ2 To what extent does incorporating QA-derived NLI examples improve the generalization and calibration of NLI-based verification across out-of-domain QA datasets?
• RQ3 How well do the entailment judgments of NLI models align with the actual information requirements of questions, especially when context is incomplete?
• RQ4 What are the failure modes of the NLI verification pipeline, and how do they differ from errors in the original QA model?
• RQ5 Can the NLI verification framework detect cases where a QA model gives the correct answer based on incorrect or partial reasoning?

Key findings

• The NLI-based verification framework improves the F1 score of a base QA model from 81.6 to 87.1 in selective prediction settings on the top 20% of most confident predictions.
• The decontextualization and question conversion modules are critical to pipeline performance, with errors in these stages accounting for only a small fraction of overall NLI model errors.
• Approximately 10% to 15% of verification failures stem from insufficient context, indicating that the model often misjudges answers when key information is missing.
• Manual analysis reveals that the NLI model successfully identifies cases where the QA model produces the correct answer for the wrong reason, particularly when the premise lacks full information to support the hypothesis.
• The combined use of standard NLI datasets and automatically generated QA-derived NLI examples results in a more robust verifier that generalizes better across domains than relying solely on in-domain QA data.
• The approach enables explicit error localization, allowing system designers to distinguish between correct answers derived from valid reasoning versus those based on partial or incorrect evidence.