TL;DR
Current Open Information Extraction (OpenIE) models perform exceptionally well on standard benchmarks but prove highly fragile when the same sentence is rephrased. To address this, the authors introduce ROBUST, a new evaluation benchmark organized into "knowledge-invariant cliques"—groups of sentences that convey identical facts using different syntactic structures. Testing reveals a frustrating performance drop across all major OpenIE models and large language models (LLMs), with $F_1$ scores plummeting by up to 23 points.
What Problem Does This Paper Address?
Open Information Extraction (OpenIE) is a foundational NLP task that extracts structured facts from unstructured text in the form of $n$-ary tuples, such as (subject, predicate, object). While modern OpenIE models score highly on standard benchmarks like CaRB, these evaluations suffer from a major blind spot: they test models on a sentence-by-sentence basis using text that closely mirrors the training distribution.
In the real world, the same core piece of information can be expressed in countless ways. If a model successfully extracts a fact from one sentence but fails when the grammar is shuffled, it isn't truly robust.
As shown above in Figure 1, the state-of-the-art model OpenIE6 handles the standard sentence ($s_1$) perfectly but misses critical arguments when the sentence structure drifts ($s_2$ and $s_3$), despite the underlying meaning remaining completely unchanged.
Methods
The authors introduce ROBUST (Robust OpenIE Benchmark with Ubiquitous Syntactic Transformations). Instead of evaluating independent sentences, ROBUST evaluates models using knowledge-invariant cliques. A clique is a cluster of sentences that look different syntactically but share the exact same semantic tuples.
To build this benchmark, the authors designed a multi-stage pipeline:
- Syntactic Generation: They used AESOP, a syntactically controllable paraphrasing model, to generate 10 structurally diverse variations for every source sentence in the CaRB dataset.
- Expressive Filtering: They applied a heuristic search strategy based on BLEU scores to eliminate redundant or overly similar sentences, keeping only the most diverse variations.
- Human Annotation: Human experts corrected grammatical errors in the machine-generated paraphrases and manually annotated the ground-truth $n$-tuples to guarantee semantic equivalence.
The Robustness Metric
Traditional metrics average performance across all sentences. ROBUST introduces a more stringent metric: a model's robustness score for a given clique is determined by its worst performance on any single sentence within that clique. If a model fails on even one syntactic variation, its robustness score drops significantly.
Datasets
The ROBUST benchmark is built directly on top of CaRB (which originates from OIE2016).
- ROBUST Statistics: 1,272 cliques, containing a total of 4,971 sentences and 16,191 high-quality human-annotated knowledge extractions.
- On average, each clique contains 3.877 sentences.
Key Insights and Ideas
- The Illusion of SOTA: Current OpenIE models are heavily overfitted to specific syntactic patterns. When those patterns shift, performance degrades drastically.
- Architecture Dictates Robustness: Not all architectures fail equally. SpanOIE, a model that enumerates geometric spans rather than relying on sequence labeling, showed the smallest performance drop. This suggests that span-enumeration strategies are inherently more resilient to syntactic drift.
- LLMs Face the Same Hurdles: Even ChatGPT (tested via
gpt-3.5-turbo) experiences a massive performance drop on ROBUST. However, the authors discovered an actionable fix: providing ChatGPT with syntactically diverse demonstrations in a few-shot prompt significantly boosts its robustness.
Key Results
The authors evaluated six traditional OpenIE models (spanning rule-based, neural-based, and pre-trained language model-based architectures) alongside ChatGPT.
- Universal Degradation: Across all models, evaluating on ROBUST caused an average drop of 20% in precision, 15% in recall, and 18% in $F_1$ score.
- OpenIE6: The top-performing supervised model fell from an $F_1$ of 55.22 on CaRB to 35.78 on ROBUST (a 19.44 point drop).
- ChatGPT (10-shot): Achieved a strong $F_1$ of 51.6 on CaRB, but plummeted to 27.5 on ROBUST.
Limitations and Future Work
While this paper exposes a critical vulnerability in OpenIE systems, the authors note a few limitations:
- Zero-Shot Evaluation: Due to a lack of open-source code for certain contemporary zero-shot models, they could not fully explore how all zero-shot architectures handle this benchmark.
- Scope: The dataset focuses entirely on OpenIE. However, the concept of "knowledge invariance" under syntactic drift is a universal challenge in NLP. Future work should extend this clique-based robustness evaluation to other tasks like relation extraction, question answering, and text summarization.
Summary updated