Finding Generalizable Evidence by Learning to Convince Q&A Models

Ethan Perez, Siddharth Karamcheti, Rob Fergus, Jason Weston, Douwe Kiela, Kyunghyun Cho

Abstract

We propose a system that finds the strongest supporting evidence for a given answer to a question, using passage-based question-answering (QA) as a testbed. We train evidence agents to select the passage sentences that most convince a pretrained QA model of a given answer, if the QA model received those sentences instead of the full passage. Rather than finding evidence that convinces one model alone, we find that agents select evidence that generalizes; agent-chosen evidence increases the plausibility of the supported answer, as judged by other QA models and humans. Given its general nature, this approach improves QA in a robust manner: using agent-selected evidence (i) humans can correctly answer questions with only ~20% of the full passage and (ii) QA models can generalize to longer passages and harder questions.

Introduction

Answering complex questions often requires navigating vast amounts of text to identify core arguments or facts. This distillation process is time-consuming for humans and computationally challenging for machines. This paper addresses the problem of automatically discovering the fundamental evidence required to answer a question.

The authors frame this as a "persuasion game": if an AI agent is tasked with selecting specific sentences to maximize a QA model's confidence in a particular answer, does that text represent genuine evidence or merely adversarial noise? By evaluating this extracted text against humans and diverse models, the study investigates whether neural networks can learn to identify universally valid evidence.

The Persuasion Framework

The system splits the QA task into two distinct roles: Judges and Evidence Agents.

  • Judge Models: These are pre-trained QA models (e.g., TF-IDF, fastText, or BERT) that act as the target to be convinced. They process input text and output a probability distribution over possible answers.
  • Evidence Agents: For each multiple-choice option, an agent is assigned to "defend" it. The agent's goal is to select a subset of sentences from the original passage that maximizes the judge's confidence in its assigned answer.
The persuasion framework: Evidence Agents select sentences to convince a Judge Model of a specific answer.

The authors evaluate two types of agents:

  1. Search Agents: These perform an exhaustive search at test time, testing every sentence against the judge to find the one that maximizes the target answer's probability.
  2. Learned Agents: These are BERT models trained to predict a sentence's influence on the judge, acting as a function approximator to avoid exhaustive querying.

Experimental Setup

The study utilizes multiple-choice QA datasets, which provide explicitly defined competing answers for agents to defend:

  • RACE: Contains 100,000 formal reading comprehension questions from Chinese middle and high school English exams.
  • DREAM: Consists of 10,000 informal, dialogue-based reading comprehension questions.

Generalizability of Evidence

A central finding is that evidence optimized for one specific neural network is not merely exploiting model-specific quirks. Instead, it captures fundamental, model-agnostic properties of the text.

Generalization heatmap: Evidence optimized for one judge model (rows) often successfully convinces other judge models (columns).

Key insights include:

  • Cross-Model Generalization: As shown in the heatmap above, evidence that convinces a BERT-based judge often also convinces simpler models like TF-IDF, and vice versa. High diagonal values indicate model-specific optimization, but significant off-diagonal values demonstrate generalization.
  • Human Alignment: Learned agents are particularly effective at finding human-interpretable evidence. While search agents might pick "adversarial" sentences that exploit statistical biases (e.g., the presence of specific keywords), learned agents smooth out these quirks, leading to evidence that humans find more persuasive.
  • Regularization Effect: Restricting a QA model to only agent-selected sentences acts as a powerful regularizer. By removing extraneous or distracting text, the model can better generalize to harder or longer questions outside its training distribution.

Performance and Efficiency

The framework demonstrates significant improvements in both human efficiency and model robustness:

  • Human Persuasion: When shown only a single agent-selected sentence, humans choose the supported answer significantly more often than random chance (up to 42% on RACE and 50% on DREAM).
  • Human Efficiency: Humans using combined agent-selected evidence can maintain ~90% accuracy while reading only ~20% of the full passage.
  • Robustness to Length and Difficulty: A BERT judge trained on short passages (≤12 sentences) improved its accuracy on long passages (≥27 sentences) from 44.1% to 50.2% when using agent-selected evidence. Similarly, generalization from middle-school to high-school questions improved from 50.7% to 53.0%.

Discussion and Future Directions

While promising, the approach has limitations. Search-based agents can be susceptible to adversarial examples, selecting sentences that exploit linguistic biases rather than providing semantic support. Furthermore, the current reliance on multiple-choice formats makes direct application to generative QA more complex.

Future work could expand this "debate" format to assist humans in fact-verification, open-domain QA, and the summarization of massive documents where reading the full text is impractical.

Summary updated

Abstract

We propose a system that finds the strongest supporting evidence for a given answer to a question, using passage-based question-answering (QA) as a testbed. We train evidence agents to select the passage sentences that most convince a pretrained QA model of a given answer, if the QA model received those sentences instead of the full passage. Rather than finding evidence that convinces one model alone, we find that agents select evidence that generalizes; agent-chosen evidence increases the plausibility of the supported answer, as judged by other QA models and humans. Given its general nature, this approach improves QA in a robust manner: using agent-selected evidence (i) humans can correctly answer questions with only ~20% of the full passage and (ii) QA models can generalize to longer passages and harder questions.
BibTeX
@inproceedings{perez2019finding,
  author = {Ethan Perez and Siddharth Karamcheti and Rob Fergus and Jason Weston and Douwe Kiela and Kyunghyun Cho},
  title = {{Finding Generalizable Evidence by Learning to Convince Q\&A Models}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2019},
  url = {https://openreview.net/forum?id=Lp3Y-Haix_B},
}
Figure