Source-linked AI summary
Neural Language Models as Psycholinguistic Subjects: Representations of Syntactic State
Richard Futrell, Ethan Wilcox, Takashi Morita, Peng Qian, Miguel Ballesteros, Roger Levy
TL;DR
The paper asks whether neural language models incrementally represent syntactic state, a question relevant to understanding their learned linguistic representations. It applies controlled psycholinguistic tests to artificial complex sentences across four models and finds that large-data LSTMs represent syntactic state over long spans comparably to the RNNG, unlike the small-data LSTM, which does so weakly or not at all.
Problem
The paper investigates whether neural language models learn incremental syntactic-state representations and robust syntactic generalizations beyond training-set peculiarities.
Method
The study uses targeted psycholinguistic evaluations of artificial complex sentences, comparing two large-data LSTMs, an RNNG, and a matched small-data LSTM.
Results
Large-data LSTMs represent syntactic state over large text spans comparably to the RNNG, whereas the small-data LSTM does so weakly or not at all.
Takeaways & Limitations
The findings suggest that coarse syntactic structure requires syntactic supervision or large data, while exploiting fine-grained lexical cues requires large data.
Takeaways & Limitations
The models’ garden-path pattern is opposite to human reading-time behavior, which shows larger effects when ambiguity resolution is delayed.
Abstract
from arXiv · showhide
We deploy the methods of controlled psycholinguistic experimentation to shed light on the extent to which the behavior of neural network language models reflects incremental representations of syntactic state. To do so, we examine model behavior on artificial sentences containing a variety of syntactically complex structures. We test four models: two publicly available LSTM sequence models of English (Jozefowicz et al., 2016; Gulordava et al., 2018) trained on large datasets; an RNNG (Dyer et al., 2016) trained on a small, parsed dataset; and an LSTM trained on the same small corpus as the RNNG. We find evidence that the LSTMs trained on large datasets represent syntactic state over large spans of text in a way that is comparable to the RNNG, while the LSTM trained on the small dataset does not or does so only weakly.
1 Introduction
The paper asks whether neural language models learn incremental syntactic-state representations like those captured by stack-based symbolic grammars. It uses targeted psycholinguistic tests to examine syntactically complex sentences and compares generic LSTMs with an explicitly structured RNNG.
- Motivation: The study addresses whether neural language models learn robust syntactic generalizations rather than only training-set-specific patterns.This question matters for understanding the content of neural representations and the interpretability, accountability, and controllability of NLP systems.
- Research focus: Unlike prior work on word-to-word dependencies, this paper tests incremental syntactic state maintained across sentence processing.The target is knowledge analogous to a stack-based incremental parse state, such as remaining inside a subordinate clause.
- Approach: The experiments use artificially constructed sentences designed to expose behavior dependent on syntactic state representations.They examine subordinate clauses and garden-path effects involving main-verb/reduced-relative and NP/Z ambiguities.
- Models: The study compares generic LSTMs without explicit syntactic structure against an RNNG that calculates context-free syntactic representations while assigning word probabilities.This comparison tests whether explicit syntactic representation affects sensitivity to syntactic state.
2 General methods
The paper probes neural language models using word surprisal as a psycholinguistic analogue of processing difficulty. It evaluates four English models differing in architecture and training data, using targeted sentences and regression analyses.
- Surprisal: Surprisal measures the log inverse probability a model assigns to each word, in bits.The current word is evaluated given the model’s hidden state before consuming it, and surprisal contributes directly to language-modeling loss.
- Psycholinguistic measure: The study treats language-model surprisal as an analogue of human word-by-word reading time.This links model expectations about upcoming words to psycholinguistic measures of online processing difficulty.
- Experimental analysis: Experiments use targeted sentences whose word-by-word surprisal profiles reveal whether models represent particular syntactic states.The profiles are analyzed with regression analysis and, generally, linear mixed-effects models.
- Models tested: Four English models include two large-data LSTMs, an RNNG, and a matched small-data LSTM.The models are trained with a language-modeling objective and differ in architecture, training corpus, and data scale.
3 Subordinate clauses
The subordinate-clause experiments test whether models detect and maintain the requirement for a following matrix clause. All models show some licensing, but only the large GRNN and the RNNG reliably show the corresponding no-matrix penalty and clearer long-span behavior.
- 3 Subordinate clauses: Subordinators cue a subordinate clause whose syntactic requirements must be maintained while processing intervening words.The construction therefore provides a test of stack-like incremental syntactic state.
- Experimental measures: The matrix licensing effect compares [SUB, MATRIX] with [NOSUB, MATRIX], while the no-matrix penalty compares [SUB, NOMATRIX] with [NOSUB, NOMATRIX].A licensing effect is negative; a no-matrix penalty is positive.
- Results: All models showed a facilitative matrix licensing effect, but only GRNN and the RNNG showed a significant no-matrix penalty.The licensing effect was significant for all models (p < .001), whereas the other models did not significantly penalize an ungrammatical continuation.
- Conclusion: Full representation of subordinate-clause requirements appears to require either large amounts of data or explicit syntactic structure.All models partially represented the licensing relationship, but only GRNN and the RNNG fully captured the relevant behavior.
- Maintenance and degradation: All models used the subordinator as a cue and maintained subordinate-clause knowledge through intervening material up to the comma.The long-span experiment added increasingly complex PP, SRC, and ORC modifiers to test degradation.
- Maintenance and degradation: GRNN’s matrix licensing interaction weakened as intervening material became longer and more syntactically complex, while other models behaved less interpretably.The interaction was significant for subject postmodifiers (p < 0.001) and object postmodifiers (p < 0.01).
4 Garden path effects
Garden-path experiments test whether models maintain incremental syntactic state and use lexical and structural cues to update it. All models show basic sensitivity to local ambiguities, but large LSTMs and the RNNG more consistently capture fine-grained syntactic effects, while state degradation patterns differ from humans.
- Garden-path effects arise when a disambiguating word forces revision of a locally likely but incorrect syntactic parse.They provide evidence that a model represented a syntactic state that made the disambiguating word surprising.
- 4.1.1 NP/Z Garden Path Effect: All models show NP/Z garden-path effects, but only the large LSTMs appear sensitive to embedded-verb transitivity.The comma manipulation contrasts clause-boundary cues, while transitivity tests whether models maintain verb argument-structure information across six intervening words.
- 4.1.2 Maintenance and degradation of state: All models show a basic garden effect in the SHORT condition, whereas TinyLSTM lacks a significant effect in the LONG condition.The length-by-comma interaction is significant for GRNN and TinyLSTM, but not for JRNN or the RNNG.
- 4.1.2 Maintenance and degradation of state: The length-and-comma interaction patterns suggest that GRNN and TinyLSTM show the clearest evidence of syntactic-state degradation.This pattern resembles degradation observed in subordinate-clause state information, but differs from human digging-in effects, where longer ambiguities increase garden-path effects.
- 4.2 Main Verb/Reduced Relative Ambiguity: The three large LSTMs and the RNNG replicate the human-like garden-path effect from verb-form ambiguity, while TinyLSTM shows a much smaller effect.Even with an unambiguous participial verbform, all models except TinyLSTM show a significant garden-path effect, suggesting that networks treat morphology as a noisy relative-clause cue.
5 General Discussion and Conclusion
The models all show basic incremental syntactic-state representations, but their ability to capture richer state effects depends on model training conditions and available data. The authors also present this psycholinguistic methodology as a broader tool for probing black-box NLP representations.
- All models showed clear evidence of basic incremental syntactic-state representations.
- Only the large LSTMs sometimes inferred clause boundaries from verb argument structure and morphological verb form.
- Only GRNN and the RNNG fully captured the proper behavior of subordinate clauses.
- Coarse-grained syntactic structure appears to require syntactic supervision or large data, whereas fine-grained lexical cues require large data.
- The psycholinguistic methodology offers a lens on black-box internal representations and a basis for systematic tests of NLP linguistic competence.