Preserving Knowledge Invariance: Rethinking Robustness Evaluation of Open Information Extraction

Ji Qi, Chuchun Zhang, Xiaozhi Wang, Kaisheng Zeng, Jifan Yu, Jinxin Liu, Jiuding Sun, Yuxiang Chen, Lei How, Juanzi Li, Bin Xu

Abstract

Open Information Extraction (OpenIE) models often struggle with "syntactic drift"—the variation in sentence structure for the same semantic content. We introduce ROBUST, a benchmark designed to evaluate knowledge invariance through a clique-based structure. Our evaluation reveals significant performance degradation across state-of-the-art models and LLMs, highlighting a critical need for structurally robust extraction methods.

Introduction

Open Information Extraction (OpenIE) is the task of pulling structured knowledge—usually in the form of $(a_1, p, a_2, ...)$ tuples—out of unstructured text. While significant strides have been made in extraction accuracy, most benchmarks evaluate models on a per-sentence basis. This ignores a critical real-world challenge: knowledge invariance. In practice, the same fact can be expressed in numerous ways. A model might successfully extract a fact from a simple active-voice sentence but fail if that same fact is presented in a complex passive construction or a relative clause.

The Challenge of Knowledge Invariance

The primary reason models fail is a sensitivity to sentence structure, or "syntactic drift." When the syntactic distribution shifts away from the training data, performance plummets. As shown in Figure 1, a model like OpenIE6 might perform well on a standard test sentence but fail on its paraphrases, even though the underlying knowledge remains identical.

Extraction results of OpenIE6 for three semantically equivalent sentences.

The ROBUST Benchmark

To address this, we developed the ROBUST benchmark using a novel "knowledge-invariant clique" structure. ROBUST is built upon the foundation of the CaRB dataset and consists of 4,971 sentences organized into 1,272 cliques.

Methodology

  1. Clique Generation: Starting with seed sentences, we used a syntactically controllable paraphraser to generate multiple versions of each sentence, filtered for high syntactic diversity and semantic consistency.
  2. Human-in-the-loop Annotation: Human experts corrected the generated sentences and annotated ground-truth knowledge tuples for every sentence in the clique.
  3. Robustness Metric: Instead of averaging scores, we propose a metric that evaluates the worst performance within a clique. A model is only considered truly robust if it extracts the same knowledge accurately across all variations.
An example of a robustness clique showing syntax, knowledge, and extractions.

Experimental Results

We tested six representative OpenIE models and ChatGPT (GPT-3.5). The results reveal a universal "frustrating degradation" in performance:

  • Universal Degradation: Every model showed a dramatic performance drop on ROBUST compared to CaRB, with an average F1 score decrease of 18 percentage points.
  • LLM Fragility: Even ChatGPT (GPT-3.5) shows a significant robustness gap, with its score dropping from 0.516 on CaRB to 0.275 on ROBUST in a 10-shot setting.
  • SpanOIE's Stability: SpanOIE demonstrated the least degradation (13.35 F1 drop), likely because its strategy of enumerating geometric spans is more resistant to syntactic shifts than labeling-based architectures.

Conclusion

Traditional metrics hide fragility by averaging results. By measuring clique-wise worst-case performance, ROBUST exposes how easily models are tripped up by simple rephrasing. Our study highlights a need for new model architectures that explicitly prioritize knowledge invariance over surface-level pattern matching.

Summary updated

Abstract

The robustness to distribution changes ensures that NLP models can be successfully applied in the realistic world, especially for information extraction tasks. However, most prior evaluation benchmarks have been devoted to validating pairwise matching correctness, ignoring the crucial measurement of robustness. In this paper, we present the first benchmark that simulates the evaluation of open information extraction models in the real world, where the syntactic and expressive distributions under the same knowledge meaning may drift variously. We design and annotate a large-scale testbed in which each example is a knowledge-invariant clique that consists of sentences with structured knowledge of the same meaning but with different syntactic and expressive forms. By further elaborating the robustness metric, a model is judged to be robust if its performance is consistently accurate on the overall cliques. We perform experiments on typical models published in the last decade as well as a popular large language model, the results show that the existing successful models exhibit a frustrating degradation, with a maximum drop of 23.43 F1 score. Our resources and code are available at https://github.com/qijimrc/ROBUST.
BibTeX
@misc{qi2023preserving,
  author = {Ji Qi and Chuchun Zhang and Xiaozhi Wang and Kaisheng Zeng and Jifan Yu and Jinxin Liu and Jiuding Sun and Yuxiang Chen and Lei How and Juanzi Li and Bin Xu},
  title = {{Preserving Knowledge Invariance: Rethinking Robustness Evaluation of Open Information Extraction}},
  year = {2023},
  eprint = {2305.13981},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2305.13981},
}
Figure