Abstract
Traditional linguistic theory suggests that human language processing relies on hierarchical tree structures rather than simple linear sequences. While Large Language Models (LLMs) achieve remarkable performance, it remains debated whether they utilize similar latent hierarchical representations. This paper presents a novel "one-shot word deletion task" to probe these hidden structures. Our findings reveal that both humans and LLMs spontaneously prioritize grammatical constituents over linear positions, suggesting an emergent structural bias in pre-trained models that mirrors human linguistic intuition.
Introduction
In linguistics, sentences are viewed as hierarchies of nested units called constituents (e.g., a noun phrase within a verb phrase). While LLMs like ChatGPT are trained on massive linear text corpora, previous research using "probing" techniques has shown they encode structural information. However, whether they actively use this information to solve tasks was previously unclear. We address this by testing for latent structure without technical jargon, ensuring results are not biased by explicit grammatical instructions.
Methods: The One-Shot Word Deletion Task
To test for hierarchical processing, we introduced a one-shot word deletion task. Participants (humans or LLMs) are shown a single demonstration: a sentence where a specific constituent is deleted (e.g., “She had an idea” becomes “She had”). They are then asked to apply the same underlying rule to a new test sentence, such as “We put some orders together.”
If the participant deletes "some orders" (a constituent), it indicates hierarchical parsing. Deleting "order together" (a non-constituent) suggests a reliance on linear sequence. We further analyzed these patterns using the CKY algorithm to reconstruct full "deletion-based" trees and compared performance against a linear LSTM baseline.
Key Results
Our experiments across English and Chinese datasets (including the Penn Treebank and Chinese Treebank) yielded several critical insights:
- Spontaneous Constituent Preference: Both humans and LLMs deleted constituents at rates significantly higher than chance ($p < 0.001$). The reconstructed trees closely matched gold-standard linguistic trees, outperforming random baselines.
- Structural Bias in LLMs: Unlike the linear LSTM, which required approximately 50 demonstrations to learn the constituent-based rule, LLMs and humans were "one-shot" learners. This suggests that structural hierarchy is an emergent property already present in pre-trained LLMs.
- Cross-Linguistic Adaptability: The study identified language-specific strategies. In English, participants favored top-down parsing (parent-category rules), while in Chinese, they shifted toward bottom-up parsing (node-category rules). Notably, bilingual speakers could switch strategies depending on the language of the test sentence.
Discussion and Conclusion
The results demonstrate that syntax often trumps semantics in these latent representations; both humans and LLMs prioritized structural rules even when the resulting sentence was nonsensical. While LLMs showed a slight deficit in semantic plausibility compared to humans, their ability to navigate complex hierarchical structures is robust. This word deletion task offers a promising tool for evaluating tacit linguistic knowledge in diverse populations, including children and clinical patients, and provides a framework for understanding the internal mechanics of state-of-the-art AI.
Summary updated