Isolating Abstract Syntax from Lexical and Semantic Shortcuts

20 concepts15 papers1 proposal

When a large language model correctly predicts the next word in a complex sentence, does it do so because it understands the underlying hierarchical rules of grammar, or because it is exploiting statistical shortcuts?

In natural language, form (syntax) and meaning (semantics) are deeply intertwined. As Bailly et al. point out, word co-occurrences in real-world datasets are driven simultaneously by grammatical rules and real-world context. If a model correctly handles a long-range subject-verb agreement—such as matching "the dogs in the park" with the plural verb "bark"—it is difficult to tell whether it is navigating a hierarchical syntactic tree or simply relying on the strong semantic correlation between "dogs" and "bark."

To truly understand whether neural networks possess abstract linguistic competence, researchers have developed techniques to disentangle syntax from its lexical and semantic wrappers.

The Illusion of Syntactic Competence

Standard evaluations often give a false impression of a model's grammatical mastery. When models are tested on natural, meaningful sentences, they can use semantic common sense as a crutch to resolve structural dependencies. Li et al. and Lasri et al. demonstrate that while transformers capture a non-trivial amount of grammatical structure, they remain highly sensitive to local lexical patterns and memorized word associations, frequently stumbling when distracting words break those familiar patterns.

This entanglement also compromises the tools we use to inspect these models. "Syntactic probes"—diagnostic classifiers trained to extract parse trees from a model's internal layers—often achieve high scores not because they have found pure structural rules, but because they are exploiting semantic shortcuts.

Maudslay et al. exposed this reliance by testing probes on "Jabberwocky" sentences: sentences that are grammatically flawless but semantically meaningless. When the semantic shortcuts are stripped away, the performance of standard probes plummets. This indicates that the perceived syntactic understanding was largely inflated by familiar word relationships. Similarly, Alishahi et al. found that in spoken language models, syntactic representations remain heavily entangled with specific word identities, meaning high performance on structural tasks may be driven by word-level patterns rather than generalized hierarchy.

Standard Sentence:   "The chef cooked the delicious meal." 
                      ▲      ▲          ▲         ▲
                      └──────┴──────────┴─────────┴─ Strong semantic & syntactic links

Jabberwocky/Nonce:   "The colorless green ideas sleep furiously."
                      ▲         ▲     ▲     ▲     ▲
                      └─────────┴─────┴─────┴─────┴─ Pure syntactic links (No semantic cues)

Stripping the Crutch: Nonce Sentences and Minimal Pairs

To isolate pure syntax, researchers must systematically remove semantic and lexical cues so that a model must rely on structural rules alone.

One classic approach relies on "nonce" sentences, famously epitomized by Noam Chomsky's phrase, "The colorless green ideas sleep furiously." Gulordava et al. pioneered this method in neural networks, demonstrating that recurrent networks can track complex hierarchical agreements even when sentences lack any semantic meaning. Extending this to transformers, Arps et al. used nonce datasets to prove that models like BERT and RoBERTa do encode abstract constituency trees that can be linearly extracted without any semantic assistance.

Another approach controls for the underlying "message" mathematically. Hu et al. formalize string probability as a combination of a sentence's semantic message and its grammaticality. Because raw probabilities are noisy, they argue that evaluating grammatical competence requires minimal pairs—sentences that differ by only a single structural feature while keeping the semantic message constant.

Alternatively, we can strip away lexical noise directly from the input. Daultani et al. suggest replacing specific entities with coarse categories (e.g., swapping "Google" for "ORGANIZATION"). By removing rare or highly contextual words, they allow automated systems to judge sentence acceptability based purely on connective grammar.

Looking Under the Hood for Shared Machinery

If language models are truly learning abstract syntax, their internal representations should treat structurally identical sentences similarly, regardless of the specific vocabulary used. Even when models exhibit human-like behavior, such as structural priming (the tendency to repeat a grammatical structure), it can be difficult to determine whether they are responding to abstract rules or local lexical cues like shared function words Jumelet et al..

This conflation also extends to cognitive modeling. Standard language models fail to distinguish between the unpredictability of a specific word and the unpredictability of its grammatical role, underestimating the difficulty humans face when resolving purely syntactic ambiguities Arehalli et al.. Furthermore, when unsupervised models attempt to induce dependency trees from statistical relationships, their representations often struggle to separate pure grammatical principles from semantic noise Vigly et al..

To see if models overcome this, Boguraev et al. applied causal interpretability methods to investigate whether models use the same internal "machinery" across different filler-gap constructions. They discovered that large language models actually converge on abstract, shared internal representations for these complex syntactic dependencies, proving that their abilities are not just a collection of surface-level lexical tricks.

However, this abstraction is not always perfect. Veenboer et al. applied collostructional analysis to BERT and found that while the model captures high-level conceptual content, it often fails to distinguish between distinct grammatical forms that express the exact same meaning. This suggests that while models can form abstract representations, their internal spaces still tend to prioritize semantic content over pure structural features.

Engineering Explicit Disentanglement

Instead of just probing for syntax after training, can we force models to separate form from meaning by design?

Input Sentence ──► [ VGVAE Encoder ] ──┬──► Latent Variable z_semantic ──► Paraphrase Reconstruction
                                       └──► Latent Variable z_syntactic ──► Word-Order Prediction

Chen et al. addressed this by developing the vMF-Gaussian Variational Autoencoder (VGVAE). The architecture explicitly splits a sentence's representation into two distinct latent variables: one for semantics and one for syntax. By training the model on multi-task objectives like paraphrase reconstruction (which preserves meaning but changes structure) and word-order prediction (which focuses on structure), the authors successfully forced the network to isolate formal grammatical rules from the specific vocabulary used to express them, achieving cleaner similarity calculations for both facets.

Key Foundational Work

Colorless Green Recurrent Networks Dream Hierarchically (2018) — contributed 1 concept elements to this direction.
A Multi-Task Approach for Disentangling Syntax and Semantics in Sentence Representations (2019) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_lexical_syntactic_entanglement_6eaf2dfa,
  title = {lexical-syntactic entanglement},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/lexical-syntactic-entanglement/rdr_6eaf2dfacfcd4232813381944de90d20},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_lexical_syntactic_entanglement_6eaf2dfa_v1,
  title = {lexical-syntactic entanglement},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/lexical-syntactic-entanglement/rdr_6eaf2dfacfcd4232813381944de90d20/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake