How Often Are Errors in Natural Language Reasoning Due to Paraphrastic Variability?

Neha Srikanth, Marine Carpuat, Rachel Rudinger

Abstract

Large language models (LLMs) have demonstrated remarkable capabilities in natural language reasoning, yet their performance often remains sensitive to the specific phrasing of a prompt. This paper investigates paraphrastic variability—the phenomenon where a model's output changes when presented with different but semantically equivalent versions of the same problem. We introduce Paraphrastic Consistency ($P_C$), a metric designed to quantify this brittleness, and PARANLU, a comprehensive dataset of over 15,000 human and model-generated paraphrases. Our analysis reveals that while extensive pretraining enhances consistency, task-specific fine-tuning can paradoxically increase a model's sensitivity to minor linguistic variations. These findings suggest that current benchmarks may conflate logical reasoning ability with linguistic decoding stability.

Introduction

In the evaluation of Natural Language Processing (NLP) systems, researchers often equate a correct prediction with a fundamental understanding of the underlying task. However, most benchmarks rely on a single linguistic expression for any given reasoning problem. This approach fails to distinguish between a model's logical failure and its inability to map a specific phrasing to a known concept. If a model succeeds on one phrasing but fails on a paraphrase, its "reasoning" is likely an artifact of specific lexical patterns rather than robust logic.

This study addresses this gap by isolating the impact of paraphrastic variability. By asking the same question in multiple ways, we can determine how much of a model's error rate is attributable to the difficulty of the logic versus the choice of words.

Methodology

Paraphrastic Consistency ($P_C$)

The authors propose $P_C$ as a core metric for reliability. It is defined as the probability that a model’s predictions for two different paraphrases of the same problem match—regardless of whether those predictions are correct. To further refine this, the researchers utilize Variance Decomposition to calculate the "Variance Attributable to Paraphrasing" (VAP). This mathematical framework allows for the separation of errors caused by logical complexity from those caused by surface-level linguistic changes.

Experimental Design

To ensure the evaluation was not biased toward simple examples, the researchers employed AFLITE, an adversarial filtering algorithm. This allowed for the selection of a diverse range of reasoning problems across varying difficulty levels. The study focused on label-preserving paraphrasing, where the surface form of a sentence is modified while strictly maintaining the logical relationship (e.g., ensuring that a statement intended to strengthen a hypothesis still does so after being reworded).

The PARANLU Dataset

The study introduces PARANLU, a dataset specifically designed to test robustness against paraphrastic variability. It contains:

  • 7,782 human-written paraphrases: These were rigorously validated for quality and label preservation.
  • 7,295 model-elicited paraphrases: Generated using GPT-3 and QCPG to compare human-level linguistic diversity against machine-generated variations.

PARANLU was built upon established reasoning benchmarks, including α-NLI (Abductive Natural Language Inference) and δ-NLI (Defeasible NLI), which includes specialized subsets like δ-SNLI, δ-ATOMIC, and δ-SOCIAL.

Key Insights and Results

The Fine-Tuning Paradox

A significant finding of this research is that accuracy and consistency do not always move in tandem. While fine-tuning a model on a specific task generally improves its accuracy, it often decreases its paraphrastic consistency. As models learn the complex decision boundaries of a specific dataset, they appear to become more "jittery," developing a heightened sensitivity to specific lexical artifacts that do not change the underlying logic.

The Role of Pretraining

In contrast to fine-tuning, the volume of pretraining data is a strong predictor of consistency. The researchers observed that $P_C$ grows rapidly during the early stages of pretraining (from 1M to 100M tokens). While returns diminish after 1B tokens, the trend suggests that the ability to recognize the invariance of meaning across different phrasings is a fundamental linguistic skill acquired during general-purpose training.

Human vs. Machine Stress-Testing

Models demonstrated significantly higher consistency when tested against machine-generated paraphrases compared to human-written ones. Human paraphrases tend to be more syntactically and lexically diverse, suggesting that current automated methods for evaluating model robustness may be overly optimistic.

Performance Benchmarks

High-capacity models like RoBERTa and DeBERTa outperformed simpler architectures (e.g., BiLSTM) in both accuracy and consistency. However, even top-tier models like RoBERTa-large showed a $P_C$ of approximately 0.9. This indicates a 10% probability that the model will change its answer simply because a prompt was reworded, highlighting a significant reliability gap for real-world applications.

Conclusion and Future Work

The study concludes that paraphrastic variability is a major source of error in modern NLU models. Until models can demonstrate high paraphrastic consistency, measurements of their "reasoning" abilities will remain confounded by linguistic instabilities.

Future research is required to extend these findings beyond binary classification into generative and open-ended reasoning tasks. Additionally, while this study focused on paraphrasing the "target" sentences, investigating the impact of paraphrasing the "context" or background information may reveal even greater levels of model brittleness. Achieving true linguistic invariance remains a critical hurdle for the development of trustworthy artificial intelligence.

Summary updated

BibTeX
@article{srikanth2024how,
  author = {Neha Srikanth and Marine Carpuat and Rachel Rudinger},
  title = {{How Often Are Errors in Natural Language Reasoning Due to Paraphrastic Variability?}},
  journal = {Trans. Assoc. Comput. Linguistics},
  year = {2024},
  doi = {10.1162/tacl\_a\_00692},
}
Figure