Abstract
This study investigates whether Large Language Models (LLMs) possess a genuine semantic understanding of noun compounds (NCs) or if they primarily rely on "parroting" definitions from their training data. By distinguishing between Noun Compound Interpretation (NCI) for common terms and Noun Compound Conceptualization (NCC) for novel or rare terms, the authors evaluate GPT-3 and T5 models. While GPT-3 achieves near-human performance on common compounds, its accuracy drops on novel concepts where data overlap is minimal. N-gram analysis confirms that LLMs rely heavily on memorization for common compounds, suggesting that true compositional reasoning remains a significant challenge.
Introduction
Noun compounds, such as apartment dog or olive oil, are a staple of the English language. Understanding them requires uncovering the implicit relationship between the constituent nouns. For instance, a "chocolate bunny" is chocolate shaped like a bunny. Traditionally, this task is known as Noun Compound Interpretation (NCI).
However, the ubiquity of LLMs introduces a "data contamination" risk. Because these models are trained on massive web corpora, they likely encounter definitions for common compounds thousands of times. To test for true compositional logic versus simple retrieval, the authors propose the Noun Compound Conceptualization (NCC) task, which focuses on rare or "ad-hoc" compounds that a human can easily imagine but that are absent from the web, such as a chocolate crocodile.
Methodology
The researchers compared two distinct approaches:
- Supervised Learning: A T5-large model fine-tuned on a cleaned version of the SemEval-2013 NCI dataset.
- Few-shot Prompting: GPT-3 (text-davinci-002) provided with only 10 examples of paraphrases.
To ensure rigorous evaluation, the authors moved beyond automatic metrics like ROUGE or METEOR, which often prioritize stylistic mimicry over semantic accuracy. Instead, they utilized human evaluation via Amazon Mechanical Turk. Furthermore, they conducted an n-gram overlap analysis against the C4 web corpus to quantify how much of the models' outputs were copied directly from existing training data.
The "Parroting" Hypothesis
The study's most striking finding involves the correlation between model correctness and data overlap. For common compounds (NCI), GPT-3's correct outputs showed a massive overlap with the C4 corpus—up to 94% for 3-grams. This suggests the model is often retrieving a memorized definition rather than reasoning from first principles.
When faced with novel compounds (NCC), this overlap drops significantly. While GPT-3 still performs well, the decrease in accuracy indicates that its reasoning capabilities are less robust than its retrieval capabilities. In some NCC failures, the model completely ignored one of the nouns (e.g., defining a literal crocodile instead of a "chocolate crocodile"), failing to compose the two concepts.
Key Results and Metrics
- Performance Gap: GPT-3 achieved 95.64% human acceptability on common compounds (NCI) but dropped to 83.81% on novel compounds (NCC).
- Metric Deception: Automatic metrics like ROUGE-L favored the T5 model because it mimicked the training data's syntax. However, human evaluators found T5's actual interpretations to be significantly inferior to GPT-3's, highlighting the limitations of standard NLP benchmarks for semantic tasks.
- Human Baseline: Interestingly, GPT-3's NCC score (83.81%) outperformed the human crowdworker baseline (73.33%), though the authors suggest this reflects the difficulty of crowdsourcing creative tasks rather than LLMs surpassing human imagination.
Conclusion
While LLMs appear to "solve" noun compound interpretation, much of this success is attributable to high-fidelity memorization of web data. The drop in performance on the NCC task reveals that while models can compose new concepts, they are significantly more reliable when they can "parrot" existing knowledge. Future work in compositional AI must focus on these novel, low-resource scenarios to move beyond sophisticated pattern matching toward true conceptual understanding.
Summary updated