PaperlayerView exact revision on arXiv

Source-linked AI summary

Do Language Models Exhibit Human-like Structural Priming Effects?

Jaap Jumelet, Willem Zuidema, Arabella Sinclair

arXiv:2406.04847v2cs.CL

TL;DR

The paper asks whether linguistic factors that shape human structural priming also influence language-model predictions. It measures priming at sentence and token levels and finds human-like frequency predictors alongside model asymmetries and lexical effects.

  • Problem

    The paper asks which linguistic factors shape language-model predictions and whether they reflect human and corpus findings on structural priming.

    Source: §1 · p. 1, §7 · p. 8

  • Method

    The authors measure sentence- and token-level priming and examine lexical overlap, surprisal, and verb preference across language models.

    Source: §1 · p. 1, §3.1 · p. 3, §6.1 · p. 7

  • Results

    Across models, surprisal and verb preference show human-like inverse-frequency effects predictive of priming; asymmetry runs opposite to humans, and function-word overlap predicts priming.

    Source: §7 · pp. 8–9

  • Takeaways & Limitations

    Model priming reflects an abstract inverse-frequency pattern associated with human implicit learning, making it more intricate than simple repetition.

    Source: §7 · p. 9

    Generality remains untested across other English constructions, other languages, broader vocabulary, and more complex sentences; the model sample is too small for meaningful inference.

    Source: Limitations · pp. 9–10

Abstract

from arXiv · show

We explore which linguistic factors -- at the sentence and token level -- play an important role in influencing language model predictions, and investigate whether these are reflective of results found in humans and human corpora (Gries and Kootstra, 2017). We make use of the structural priming paradigm, where recent exposure to a structure facilitates processing of the same structure. We don't only investigate whether, but also where priming effects occur, and what factors predict them. We show that these effects can be explained via the inverse frequency effect, known in human priming, where rarer elements within a prime increase priming effects, as well as lexical dependence between prime and target. Our results provide an important piece in the puzzle of understanding how properties within their context affect structural prediction in language models.

1 Introduction

Structural priming is increased likelihood of repeating a structure after exposure to a congruent sentence. The paper examines how human-relevant frequency and lexical factors influence language-model priming.

1 Introduction

  • Structural priming makes speakers more likely to repeat a structure after recently encountering a sentence with that structure.The paper contrasts a congruent prime with an incongruent prime.Source: §1 · p. 1
  • Language models can show structural priming, treating the prime as an in-context demonstration that influences a target sentence.Prior studies report priming in large language models (Prasad et al., 2019; Sinclair et al., 2022; Michaelov et al., 2023).Source: §1 · p. 1
  • Human priming is stronger with shared words and unusual primes, motivating tests of surprisal and structural preference as predictors.These factors are drawn from corpus-linguistic work.Source: §1 · p. 1
  • The paper studies lexical overlap and priming asymmetry, locating effects at the token level and testing inverse-frequency predictors in models.It reports that surprisal and verb preference are predictive of priming.Source: §1 · p. 1

2 Structural Priming

The paper frames model priming against human production and corpus research, emphasizing inverse frequency, lexical dependence, and prior evidence of asymmetrical priming in language models.

2 Structural Priming

  • Structural priming is studied across human production, comprehension, and corpus analyses.The paper provides background on human priming and priming in language models.Source: §2 · p. 2

2.1 Properties of Priming

  • Because language models are trained on human-produced data, the paper focuses on production- and corpus-based priming patterns.The authors note that model behavior need not align directly with comprehension findings.Source: §2.1 · p. 2
  • Human priming follows inverse frequency: surprising primes and a verb’s dispreferred structure tend to produce stronger priming.The paper investigates whether these effects are found in modern language models.Source: §2.1 · p. 2
  • Shared words and semantic overlap can boost structural priming, while human dative studies distinguish content-word repetition from function-word repetition.A common account links overlapping words to activation of syntactic frames associated with verbs.Source: §2.1 · p. 2

2.2 Structural Priming in Language Models

  • Prior language-model studies report cumulative, recency-sensitive priming boosted by lexical-semantic overlap, with asymmetries across alternative structures.The paper tests whether dative asymmetry extends across state-of-the-art models and relates to inverse-frequency effects (Sinclair et al., 2022).Source: §2.2 · p. 3

3 Measures, Data & Models

The paper measures sentence-level and token-level priming using dative prime-target pairs, then compares models across controlled lexical-overlap and semantic-similarity conditions.

3.1 Sentence-level Priming Effect

  • Sentence-level priming effect compares a target’s log probability after a congruent prime with its log probability after an incongruent prime.The target sentence probability is computed by summing log probabilities across its tokens.Source: §3.1 · p. 3

3.2 Token-level Priming Effect

  • The token-level metric assigns priming effects to individual target tokens, whose scores sum to the sentence-level priming effect.This decomposition identifies which target tokens contribute to the overall effect.Source: §3.2 · p. 3

3.3 The Prime-LM Corpus

  • The Prime-LM dative corpus provides 15,000 prime-target pairs per subset for isolating lexical overlap and semantic similarity.Conditions include a no-overlap baseline, semantic-similarity subsets, and shared nouns, verbs, determiners, or prepositions.Source: §3.3 · pp. 3–4

3.4 Models

  • The study considers autoregressive LLMs, including GPT2-large, Llama-2-7b, Falcon-7b, and Mistral-7b, and tests aligned versions for models marked with an asterisk.Source: §3.4 · p. 4

4 Exp1: Measuring Structural Priming

Across the tested models, sentence-level priming is asymmetric, while greater lexical overlap makes effects stronger and more balanced.

4 Exp1: Measuring Structural Priming

  • The experiment extends a sentence-level approach to a wider range of large, contemporary language models.A subsequent section examines priming at a finer-grained level.Source: §4 · p. 4
  • ρ = -0.72 to -0.77 for the correlation between PO and DO priming effects across the models, with only 26 to 38% of sentences positive in both directions.The analysis uses the Core condition of Prime-LM; the relationship is shown in Figure 1A.Source: §4 · p. 4
  • Newer, larger models show stronger skew and asymmetry than GPT2-large, unlike human findings that are generally positive for both structures.The direction of model asymmetry differs from the human pattern.Source: §4 · p. 4
  • Increasing lexical overlap moves all models toward stronger, more balanced priming, especially when verbs or function words overlap.The reported changes are shown in Figure 1C–E.Source: §4 · p. 4

5 Exp2: Locating Structural Priming

Token-level priming scores locate where sentence-level effects arise: shared pre-divergence positions contribute inverse effects, while lexical overlap boosts predictions at specific target tokens.

Structural Divergence

  • Before PO and DO targets diverge at N2, their token-level scores are inversely proportional and reflect structural bias rather than prime–target congruence.These pre-divergence scores partly explain the strong negative sentence-level correlation.Source: §5 · p. 5
  • ρ: -0.76 to -0.52 for Llama-2-chat; measuring from target divergence reduces the negative PO–DO correlation, and priming becomes more balanced.Figure 1 also shows greater balance with increased lexical overlap.Source: Figure 1 · p. 5, §5 · p. 6

Lexical Dependence

  • The most consistent balanced effects occur from the start of the second noun phrase, suggesting models narrow structural predictions later in the sentence.In Core, skewed priming at N2 suggests semantic role contributes less to model structural prediction than expected.Source: §5 · p. 6
  • With noun overlap, w-PE rises at N2 and N3 but not other tokens; verb overlap boosts verb/preposition and N2/N3 scores.The authors interpret the latter pattern as reflecting expected semantic roles at N2.Source: §5 · p. 6
  • Unlike reported human production and comprehension findings, prepositional overlap strongly boosts priming in the models.Source: §5 · p. 6

6 Exp3: Explaining Structural Priming

The section applies production- and corpus-informed linear mixed-effects models to sentence- and token-level factors that may predict structural priming. Lexical overlap, prime surprisal, and verb structural preference predict priming, with some parallels and differences relative to human findings.

6.1 Priming Factors

  • Predictors cover lexical dependence and inverse frequency, using token overlap and similarity, sentence similarity, sentence surprisal, and verb structural preference.Verb preference is based on average probability differences between alternating structures across Core-condition prime sentences; most verbs favor PO in the models.Source: §6.1 · pp. 6–7
  • Models vary in verb preference and correlate weakly with human preference orders; most verbs favor PO, unlike some reported American English production and corpus preferences.The authors note that PO preference may be confounded by transitive verb phrases followed by a prepositional modifier.Source: §6.1 · p. 7, Figure 4 · p. 7

6.2 Modeling Priming Effects

  • The analysis fits two LMMs to 30,000 items, balances Core and altered conditions, and includes by-model random intercepts; each model explains part of priming variation.The authors leave broader exploratory analysis for future work and focus on confirmatory hypothesis testing.Source: §6.2 · p. 7
  • Lexical overlap is the strongest predictor; prime surprisal and opposite-structure verb preference also predict priming, while only DO target surprisal does.The LMMs predict sδ-PE(PO) and sδ-PE(DO); their R2 values are 0.257 for PO and 0.227 for DO.Source: §6.2 · pp. 7–8, Figure 5 · p. 8
  • Noun similarity predicts priming weakly, sentence similarity does not, and prepositional overlap is a strong predictor unlike human dative findings.Shared determiners are also consistently important, despite receiving less attention in human literature.Source: §6.2 · p. 8
  • Prime surprisal predicts priming in both PO and DO directions, while target surprisal predicts only DO priming, consistent with an inverse frequency effect.The effect indicates that less frequent or plausible primes are associated with increased priming.Source: §6.2 · p. 8

7 Discussion & Conclusion

The authors find that language models show asymmetric structural priming and that human-linked frequency and lexical factors predict model priming, though function-word overlap is especially predictive. They argue that inverse-frequency patterns suggest priming is more than simple repetition, while noting open questions about scope and mechanism.

7 Discussion & Conclusion

  • Models show asymmetric priming, more pronounced in newer, larger models, with a direction opposite to human dative priming; balanced effects appear later within targets than expected for humans.Token-level priming effects help locate where the asymmetry arises.Source: §7 · pp. 8–9
  • Human-linked lexico-semantic and frequency predictors also predict model priming, but function-word overlap is especially predictive compared with human dative findings.The authors report inverse-frequency effects of prime surprisal and verb preference across models, consistent with human findings.Source: §7 · p. 9
  • The authors interpret inverse-frequency effects as consistent with error-based implicit learning, while noting controlled data distributions are needed to test for indirect explanations.They argue that the patterns make model priming more intricate than simple repetition.Source: §7 · p. 9
  • The analysis measures comprehension by testing predictions on fixed targets; generation-based priming remains a proposed direction for further study.The authors note that comprehension and production in LLMs share an output distribution.Source: §7 · p. 9
  • The authors frame their results as evidence about where human linguistic patterns are replicated in models and suggest studying hierarchical representations and in-context learning mechanisms.They identify these investigations as future work.Source: §7 · p. 9

Limitations

The authors limit their conclusions to the analyses conducted and identify several open questions about generalizability and model comparisons.

Limitations

  • Whether the findings generalize beyond the English constructions studied remains open.The authors also question whether dataset constraints affect the outcomes.Source: Limitations · p. 9
  • Generality across wider vocabulary and more complex sentences remains untested because of constraints in the dataset used.The authors pose these as open questions rather than established failures of generalization.Source: Limitations · p. 9
  • Model-size effects need more systematic testing, and the base and alignment-tuned model sample is too small for meaningful inference.The authors describe the observed size effects as interesting but not conclusive.Source: Limitations · p. 10

B Sentence-level asymmetry effects

Increased semantic similarity primarily boosts priming for the dominant DO structure, while combined noun and verb similarity has a larger effect than the separate conditions would suggest.

B Sentence-level asymmetry effects

  • Increased semantic similarity boosts the dominant DO structure’s PE score while leaving PO PE unaffected.The combined increase in noun and verb similarity has a greater impact than the individual conditions combined.Source: B Sentence-level asymmetry effects · p. 13
  • The effect of increasing noun and verb similarity together is not linearly additive; its impact exceeds the individual conditions combined.Source: B Sentence-level asymmetry effects · p. 13

C Token-level PE for Increased Semantic Similarity

D Preference Order Correlation

Language models show substantial variation in structural preference order, with no significant correlation to the reported human order. The section also identifies the LMM results reported for the two constructions.

D Preference Order Correlation

  • No language model significantly correlates with human structural preference order, and preference orders vary within and across models.Only Falcon-7b and its instruction-tuned variant retain high preference overlap; other aligned models diverge strongly from their base models.Source: D Preference Order Correlation · pp. 13, 15, Figure 9 · p. 15

E Linear Mixed-effect Model Summary

F Model Size and Alignment

In the tested models, larger models show more skewed core priming and greater susceptibility to lexical boosting than GPT2, but the sample is too small for broad generalizations. Base and alignment-tuned models can differ substantially in priming for the same prime-target pair.

F Model Size and Alignment

  • The sample is too small for broad generalizations, but tested larger models show more skewed core priming and greater lexical-boosting susceptibility than smaller GPT2.The passages report no strong pattern distinguishing alignment tuning from base models, while noting substantial priming differences for some same-base model pairs.Source: F Model Size and Alignment · p. 15
  • Base and alignment-tuned models can show substantial priming-effect differences for the same prime-target pair, even when derived from the same base.The results report no strong overall pattern distinguishing alignment tuning from base models.Source: F Model Size and Alignment · p. 15
  • Tables 1 and 2 present raw LMM results predicting sδ-PE for PO and DO, respectively.Source: Table 1 · p. 16, Table 2 · p. 16
Loading 2406.04847v2…