Abstract
In the last half-decade, the field of natural language processing (NLP) has undergone two major transitions: the switch to neural networks as the primary modeling paradigm and the homogenization of the training regime (pre-train, then fine-tune). Amidst this process, language models have emerged as NLP's workhorse, displaying increasingly fluent generation capabilities and proving to be an indispensable means of knowledge transfer downstream. Due to the otherwise opaque, black-box nature of such models, researchers have employed aspects of linguistic theory in order to characterize their behavior. Questions central to syntax—the study of the hierarchical structure of language—have factored heavily into such work, shedding invaluable insights about models' inherent biases and their ability to make human-like generalizations. In this paper, we attempt to take stock of this growing body of literature. In doing so, we observe a lack of clarity across numerous dimensions, which influences the hypotheses that researchers form, as well as the conclusions they draw from their findings. To remedy this, we urge researchers make careful considerations when investigating coding properties, selecting representations, and evaluating via downstream tasks. Furthermore, we outline the implications of the different types of research questions exhibited in studies on syntax, as well as the inherent pitfalls of aggregate metrics. Ultimately, we hope that our discussion adds nuance to the prospect of studying language models and paves the way for a less monolithic perspective on syntax in this context.
Introduction
Since the "deep learning revolution" of the mid-2010s, Natural Language Processing (NLP) has largely abandoned hand-coded linguistic features in favor of massive, opaque neural architectures like BERT and GPT. While these models demonstrate remarkable fluency, they remain "black boxes." A central question in modern NLP research is whether these models truly "know" syntax—the abstract, hierarchical rules governing language—or if they are simply sophisticated statistical pattern-matchers.
Kulmizev and Nivre (2021) characterize the current state of this research as a paradox: "Schrödinger’s Tree." In this framework, syntax appears simultaneously present and absent depending on the experimental lens applied. This paper synthesizes the current literature to resolve these contradictions and provides a framework for more rigorous future inquiry.
The Three Paradigms of Syntactic Evaluation
Research into neural syntax generally falls into three methodological categories:
- Targeted Syntactic Evaluation (TSE): This approach uses "minimal pairs" to test model behavior. For example, a model is presented with two sentences: "The keys to the cabinet are on the table" and "The keys to the cabinet is on the table." If the model assigns a higher probability to the grammatically correct version, it is seen as possessing syntactic knowledge.
- Probing: Researchers attempt to "extract" syntactic structures directly from a model's internal mathematical representations (hidden states). By training a secondary "probe" to reconstruct a dependency tree from these embeddings, researchers test if the model encodes hierarchical information.
- NLU Evaluation: This assesses how much a model relies on syntax to solve downstream tasks, such as the GLUE benchmark. This often involves "corrupting" the input—such as shuffling word order—to see if the model's performance collapses.
Coding Properties vs. Syntactic Structure
A primary source of confusion in the literature is the conflation of coding properties with syntactic structure. Coding properties are the surface-level cues that signal syntax, such as word order in English or morphological case in Finnish. Syntactic structure, conversely, refers to the abstract, hierarchical relationship between words.
High performance on a subject-verb agreement task (TSE) does not necessarily prove a model has mastered abstract grammar; it may simply indicate sensitivity to specific coding properties. This distinction is crucial for cross-linguistic studies, as different languages prioritize different properties to convey the same structural information.
The NLU Paradox and the "Schrödinger" Effect
The most striking contradiction in recent research is the "NLU Paradox." While structural probes (e.g., Hewitt & Manning, 2019) successfully recover complex dependency trees from BERT's embeddings—suggesting the information is present—other studies (e.g., Sinha et al., 2021) show that models can solve many NLU tasks even when the input is a "word salad" of shuffled tokens.
This suggests that while models can represent syntax (Capacity), they do not always need it to solve tasks (Necessity). Many current benchmarks can be bypassed using lexical heuristics or keyword spotting, leading to the "Schrödinger" effect: syntax is "alive" when we look for it with probes, but "dead" when we test if the model actually requires it for downstream success.
Moving Beyond Aggregate Metrics
The paper identifies several pitfalls in current evaluation methods, most notably the reliance on aggregate metrics. A single accuracy score on a benchmark like BLiMP can mask significant failures. A model might excel at simple, local agreement but fail entirely at complex, nested dependencies.
Furthermore, the choice of linguistic formalism matters. When probing for trees, the results vary depending on whether the researcher uses Universal Dependencies (UD) or Surface-Syntactic Universal Dependencies (SUD). The model's "internal tree" may look different depending on the theoretical lens used to interpret it.
Conclusion and Future Directions
To move the field forward, researchers must distinguish between three distinct questions: Does the model learn syntax (Observation), can it learn syntax (Capacity), and does it need syntax (Necessity)?
Future work should prioritize:
- Typological Diversity: Moving beyond English to see how models handle diverse coding properties like inflectional morphology.
- Causal Explanations: Moving from mere observation of syntactic information to proving that this information actually causes the model's behavior.
- Granular Analysis: Moving away from monolithic aggregate scores toward a more nuanced understanding of specific linguistic phenomena.
By addressing these conceptual gaps, the NLP community can move toward a less monolithic and more scientifically rigorous understanding of how neural models process the complexities of human language.
Summary updated