Source-linked AI summary
Syntactic Structure from Deep Learning
Tal Linzen, Marco Baroni
TL;DR
The paper asks whether deep networks trained on raw language data acquire human-like grammatical knowledge, a question relevant to debates about innate linguistic structure. It surveys syntactic evaluations and analyses of internal representations, finding substantial syntactic ability alongside systematic deviations from idealized human competence. The authors argue that combining linguistic analysis with deep-learning research can inform both language-processing systems and theories of language.
Problem
The paper addresses whether DNNs trained on raw data acquire human-like grammatical knowledge, bearing on debates about the innate structure needed for language acquisition.
Method
The paper surveys linguistic and psycholinguistic evaluations of DNN syntax, plus analyses of the internal processes underlying their grammatical behavior.
Results
DNNs achieve high accuracy on challenging syntactic tasks, but their behavior generally deviates from idealized syntactic competence and relies on heuristics.
Takeaways & Limitations
DNN analyses can help vet human-language-processing stimuli and may offer insights into language and what is unique about human linguistic abilities.
Takeaways & Limitations
DNN training data differ from child-directed speech in both size and nature, and typical word-prediction training lacks perceptual and social grounding.
Abstract
from arXiv · showhide
Modern deep neural networks achieve impressive performance in engineering applications that require extensive linguistic skills, such as machine translation. This success has sparked interest in probing whether these models are inducing human-like grammatical knowledge from the raw data they are exposed to, and, consequently, whether they can shed new light on long-standing debates concerning the innate structure necessary for language acquisition. In this article, we survey representative studies of the syntactic abilities of deep networks, and discuss the broader implications that this work has for theoretical linguistics.
1. INTRODUCTION
Deep learning’s success in language applications motivates renewed examination of whether neural networks acquire syntactic knowledge without linguistically symbolic representations. The survey evaluates this question with theoretically informative linguistic and psycholinguistic paradigms, finding substantial but incomplete syntactic learning.
- Motivation: Deep neural networks succeed in NLP applications despite using learned vector representations rather than linguistically familiar symbolic structures.They encode words and sentences as sequences of real numbers instead of parse trees or logical formulas.
- Contribution: The survey reassesses arguments for rich innate structure by directly testing DNN syntactic knowledge on theoretically significant constructions.The authors distinguish these cases from common constructions that mainly support practical application performance.
- Contribution: Across subject-verb agreement, filler-gap dependencies, and other phenomena, contemporary DNNs learn substantial syntax but fall short of human competence.
2. DEEP LEARNING FOR LANGUAGE PROCESSING
DNNs process language by transforming vector representations through learned numerical computations, with training objectives ranging from classification and next-word prediction to sequence generation. RNNs maintain sequential hidden states, while gating and attention provide mechanisms for handling longer dependencies.
- Neural-network foundations: Artificial neural networks compute functions over sequences of real numbers using weighted nonlinear units whose weights are learned from examples.Training repeatedly compares predicted outputs with expected outputs and adjusts the weights to reduce their discrepancy.
- Representing sequences: Word embeddings encode words as compact, dense vectors, while recurrent neural networks process word sequences through a left-to-right hidden state.The hidden state summarizes the prefix seen so far but creates a bottleneck because earlier states are not directly available.
- Representing sequences: Gated RNNs such as LSTMs and GRUs control hidden-state updates, enabling them in principle to track dependencies spanning many words.LSTMs are the predominant RNN architecture in the studies reviewed.
- Attention and transformers: Attention lets networks use previous hidden states when computing the next state, and transformers rely on attention without recurrence.The relative linguistic abilities of transformers and RNNs under equal training data remain unclear.
- Training and use settings: In classifier, language-model, and seq2seq settings, DNNs respectively predict labels, assign probabilities to possible next words, or generate output sequences.Language models require only an unannotated text corpus, whereas classifiers use annotated examples and seq2seq systems map input sequences to output sequences.
3. LONG-DISTANCE AGREEMENT
Long-distance agreement tests whether DNNs represent hierarchical dependencies rather than merely following nearby words. The reviewed studies show strong agreement performance across architectures and training regimes, alongside failures on rare or multiply embedded structures and reliance on heuristics.
- Agreement evaluation: Number prediction evaluates whether a DNN can select singular or plural verb forms from a sentence prefix according to the subject’s syntactic structure.Intervening nouns called attractors test whether the model resists a nearby but irrelevant number cue.
- Agreement results: GRUs and convolutional networks also performed successfully, indicating that the original agreement result does not depend crucially on LSTM-specific features.The convolutional model succeeded without the recurrent mechanism central to LSTMs.
- Unsupervised learning: LSTMs trained only for next-word prediction learned agreement across four languages and across dependencies beyond subject-verb agreement.These evaluations included English, Hebrew, Italian, and Russian, plus adjective-noun agreement where available.
- Unsupervised learning: Performance degraded only mildly on semantically implausible “colorless green ideas” prefixes, suggesting agreement computation without strong lexical or semantic cues.The study therefore attributes substantial agreement learning to word prediction alone, without syntactic supervision or semantic grounding.
- Limitations: LSTM language models use heuristics, perform poorly on some infrequent nested dependencies, and therefore do not achieve perfect syntactic competence.Reported heuristics include expecting short embedded clauses and over-weighting the first noun’s number.
4. OTHER SYNTACTIC PHENOMENA
Research beyond agreement finds that neural networks can model several syntactic dependencies, but sensitivity is uneven across constructions and evaluation settings. Results are strongest for some filler-gap and syntactic-state patterns, while island constraints and hierarchical generalization remain inconsistent.
- DNNs were evaluated on syntactic phenomena including filler-gap dependencies, island constraints, syntactic state, garden-path sentences, and auxiliary fronting.
- Filler-gap dependencies: Filler-gap acceptability was only marginally affected by licensor-gap distance, and networks learned that one wh-phrase cannot license multiple gaps.
- Island constraints: Networks showed reduced gap expectations in some island contexts, but neither tested LSTM captured the Subject Island Constraint asymmetry.
- Island constraints: The survey reports mixed island findings across studies and suggests that construction choice and evaluation measures may help explain the disagreement.
- Syntactic state: LSTMs tracked syntactic state, showing high surprise when subordinate clauses lacked matrix clauses and at the disambiguating verb in NP/Z garden paths.
- Auxiliary fronting: Standard sequential RNNs showed mixed auxiliary-fronting results, whereas tree-based RNNs robustly preferred hierarchical generalizations only with explicit correct trees for both input and output.
5. WHAT DO THE NETWORK’S INTERNAL REPRESENTATIONS ENCODE?
The survey examines what neural networks encode internally using black-box probing, diagnostic classifiers, structural probes, and unit ablations. These methods reveal syntax-related representations and mechanisms, while showing that decodability alone does not establish behavioral use.
- Black-box output analysis is complemented by internal-state analysis, which is attractive because artificial-network activations are directly accessible and more granular than comparable human neuroscientific measurements.
- Diagnostic classifiers: Diagnostic classifiers test whether linguistic distinctions can be recovered from an existing network’s vector representations, indicating potential access to that information.
- Probing results: Probing recovered subject plurality and parse-tree depth, while structural probes found parse-distance metrics for LSTM and transformer representations but not simpler baselines.
- Interpretive caveat: Successful decoding does not show that the network uses the decoded information, because irrelevant and behaviorally relevant auxiliaries were both recoverable with high accuracy.
- Behavioral linkage: For subject-verb agreement, lower probe accuracy on erroneous predictions and successful state interventions linked decoded plurality to the network’s behavior.
- Causal analysis: Unit ablation was used to test whether individual neurons causally affect long-distance agreement performance.
- Agreement mechanisms: Lakretz and colleagues found a syntax-based agreement circuit, but embedded agreement remained difficult when intervening attractors had to be handled by linear-distance units.
- Implications: Detailed agreement mechanisms may help formulate predictions about human processing of syntactic dependencies.
6. DISCUSSION
The survey argues that DNNs can acquire substantial syntactic knowledge, but their success does not by itself establish human-like grammar or learning from data alone. Interpreting their behavior requires attention to architectural biases, training data, and the distinction between competence and performance.
- DNNs are not tabulae rasae: their initial weights and architectures impose biases, but these differ from traditional linguistic assumptions.Therefore, syntactic success cannot establish that statistical learning from data alone is sufficient, and the crucial architectural factors remain unknown.
- The survey distinguishes supervised from unsupervised training and corpus-extracted from synthetic data because each setup changes interpretation and introduces different confounds.Corpus data may contain spurious correlations, whereas synthetic test sentences may differ from training data in irrelevant lexical or distributional properties.
- DNN training data differ from children’s experience in both size and nature, including the absence of perception and social interaction in standard NLP settings.The supported cognitive question is therefore how much syntax can be learned from huge amounts of written linguistic data alone.
- DNNs achieve high accuracy on challenging syntactic tasks, yet nearly all reviewed studies find behavior diverging from idealized grammatical competence.In agreement, their behavior appears to rely on complex heuristics rather than a fully fledged context-free grammar.
- DNNs can help vet human-processing stimuli and generate hypotheses about human syntactic performance from shared error patterns.A DNN succeeding on a task despite falling short of human competence may indicate that the task does not probe full grammatical ability.
- Under shallow views of syntax, differences between DNN and human syntactic knowledge may be quantitative rather than qualitative.This perspective treats humans as possessing a larger and more sophisticated set of heuristics, linking DNN behavior to construction grammar and related formalisms.
- The authors call for greater collaboration between linguistics and deep learning to improve DNNs and clarify what they learn about grammar.They also caution that negative results from attempts to inject linguistic constraints may reflect current implementation difficulties rather than ineffective linguistic priors.