Source-linked AI summary
Disentangling Statistical Preemption from Entrenchment in Language Models' Avoidance of Overgeneralization
Yixuan Wang, Freda Shi, Kanishka Misra
TL;DR
The paper asks how learners avoid overgeneralizations such as transitive uses of laugh without explicit negative evidence. Using controlled rearing experiments with CHILDES-trained language models, it separates preemptive from non-preemptive input and finds no verb-specific preemption, weak abstract effects, and evidence that competing structures function as indirect positive evidence. These results motivate models with finer-grained indirect-negative sensitivities and human tests of abstract preemption.
Problem
Learners constrain overgeneralizations despite little explicit negative evidence, but it remains unclear whether this reflects preemption by near-synonymous constructions or broader entrenchment.
Method
The study uses controlled rearing experiments on language models trained on CHILDES, selectively removing verb-specific or abstract preemptive versus non-preemptive usages.
Results
Models avoid overgeneralizations but show no clear verb-specific preemption; abstract manipulations yield weak effects, while training dynamics support indirect positive rather than negative evidence.
Takeaways & Limitations
The findings motivate sensitivity to fine-grained indirect negative evidence in neural learners and new human experiments testing abstract preemption.
Takeaways & Limitations
The analysis assumes that competing constructions are preemptive and treats near-synonymous constructions as the relevant preemptive cases, leaving competition detection and broader preemption for future work.
Abstract
from arXiv · showhide
How do learners avoid overgeneralizations such as Tom laughed me without explicit negative evidence? Constructionists have posited two proposals that describe indirect negative evidence against overgeneralizations: preemption (which privileges exposure to near-synonymous construction---e.g., she made him laugh) vs. entrenchment (all exposures to a verb's grammatical usages, including cases like He laughed). We disentangle these hypotheses by running controlled rearing experiments on LMs trained on child-caregiver conversations, where we systematically remove preemptive vs. non-preemptive evidence. We find that while LMs avoid overgeneralizations, they do not show preemption at a verb-specific level, instead showing weak but non-zero evidence of abstract preemption. Combined with results from analyzing the LMs' training dynamics, we find that LMs treat competing structures as indirect positive---as opposed to negative---evidence in the verb-specific condition. Insofar as preemption is the more plausible route to avoiding overgeneralizations in humans, our results point the need for there to be sensitivities to indirect negative evidence in neural network learners, and suggest new human experiments to test abstract preemption.
1 Introduction
Learners eventually constrain productive generalizations despite little explicit negative evidence, motivating competing accounts based on preemption and entrenchment. This paper uses controlled rearing of language models to causally distinguish these accounts and finds no verb-specific preemption, but weak abstract effects.
- Children overgeneralize argument-structure patterns before eventually restricting them, despite caregivers rarely systematically correcting such errors.
- Preemption treats a near-synonymous construction as evidence against a competing usage, whereas entrenchment counts all grammatical usages of the verb.
- Because preemptive evidence is a subset of entrenching evidence, prior human studies struggle to disentangle the proposals and establish their causal status.
- The study trains language models on child-caregiver conversations while selectively removing preemptive or equally frequent non-preemptive evidence.
- Verb-specific manipulations show no preemption, whereas abstract manipulations produce relatively larger but generally weak effects.
- Training dynamics suggest models treat competing structures as indirect positive evidence, and the authors propose human experiments to test abstract preemption.
2 Data, Model, and Methods
The experiments use lemmatized CHILDES conversations and autoregressive language models, manipulating periphrastic causative versus other usages while evaluating unseen transitive overgeneralizations with negative log-likelihood.
- The training data are English CHILDES conversations between children and caregivers, with morphological and syntactic annotations used to identify relevant constructions.
- Lemmatization maps inflected forms such as laughed and laughing to the same verb token, allowing them to contribute collective evidence.
- The models are 115M-parameter GPT-2 architectures trained for three epochs with five random seeds.
- Disentangling Preemption from Entrenchment: REMOVE-P deletes a verb’s periphrastic causative usages, while REMOVE-NON-P deletes the same number of other usages, keeping raw frequency reductions equal.
- Evaluation: The study focuses on six generally intransitive verbs and evaluates unseen transitive and periphrastic causative sentence pairs constructed from CHILDES nouns.
- Measure: Overgeneralization is measured as the average NLL of unseen transitive sentences, with higher NLL indicating stronger suppression of the unconventional form.
3 Preconditions to Not Overgeneralizing
Before testing indirect evidence, the authors verify that CHILDES-trained models learn the expected distinction: transitive uses of target intransitive verbs are less likely than periphrastic causatives.
- The models must first show the appropriate generalization behavior before their responses to indirect evidence can be interpreted.
- The precondition compares NLL for transitive evaluation sentences with NLL for their periphrastic causative counterparts.
- The evaluation contrasts six target intransitive verbs with alternating and transitive-biased verbs, whose expected preferences differ.
- Across training, intransitive verbs show generally positive ΔNLLs, transitive verbs generally negative values, and alternating verbs intermediate values.
4 Verb-specific and Abstract Preemption
The controlled-rearing experiments compare preemptive and non-preemptive evidence in verb-specific and abstract conditions. Verb-specific results oppose preemption, while abstract effects are relatively stronger but generally weak, and alternating verbs retain direct transitive evidence.
- Design: The experiments compare models trained with preemptive evidence kept versus removed when evaluating unconventional transitive usages.The analysis uses NLLs and contrasts KEEP-P with REMOVE-P across target verbs and manipulation conditions.
- Verb-specific condition: In the verb-specific condition, target intransitive verbs show largely negative rather than positive preemption effects.Transitive usages generally receive lower NLL when periphrastic causative evidence is removed, contrary to the preemption prediction.
- Abstract condition: The abstract condition shows relatively stronger positive preemption effects than the verb-specific condition.The effect reflects accumulated periphrastic causative evidence across verbs rather than evidence associated with one verb alone.
- Alternating verbs: Alternating verbs break and stop show more seemingly positive verb-specific preemption, but direct transitive usages remain available after removal.Their unaffected transitive cases can support transitive preferences in REMOVE-P, unlike the generally intransitive target verbs.
- Interpretation: The verb-specific findings suggest that periphrastic causatives function as indirect positive evidence linking transitive and causative structures.Removing them removes evidence that the two structures can be related, rather than producing the blocking pattern predicted by preemption.
- Interpretation: Encoding structural similarity precedes encoding competition, so recognizing related constructions is a prerequisite for treating them as competitors.This distinction motivates examining whether the models first learn a relation between the structures.
5 Prospective Moments of Indirect Negative (or Positive?) Evidence
The training-dynamics analysis identifies batches containing periphrastic causatives and measures model behavior immediately before and after those batches. Changes in losses for causative and transitive constructions are generally highly correlated, supporting an indirect-positive-evidence account.
- Analysis setup: Prospective Moments of Indirect Negative Evidence are training batches containing constructions hypothesized to discourage overgeneralization.The model is evaluated before and after encountering each relevant batch to estimate its effect on overgeneralization behavior.
- Analysis setup: For each relevant batch, the analysis measures pre-to-post NLL changes for periphrastic causative and transitive evaluation sentences.The study also computes Pearson correlation and pairwise sign agreement between the two NLL changes.
- Results: Generally high correlations and agreement across verbs show that both constructions change in the same direction after periphrastic causative exposure.This pattern corroborates the hypothesis that periphrastic causatives provide indirect positive rather than negative evidence.
6 Implications and Speculations
The models show no verb-specific preemption but do show weaker abstract preemptive signals across verbs. Their training dynamics instead suggest that competing constructions can function as indirect positive evidence, while language-only input may be insufficient for communicative sensitivity.
- 6.1 Semantic cues against overgeneralization: Preemption may require semantic or environmental cues that were unavailable in the language-only CHILDES training input.The authors suggest testing whether environmental annotations or explicit semantic information can facilitate this sensitivity.
- 6.1 Semantic cues against overgeneralization: The training analysis finds positive correlations between changes in transitive and periphrastic causative NLLs after periphrastic causative exposures.This pattern supports treating competing structures as indirect positive rather than negative evidence.
- 6.2 Morphological blocking: Morphological blocking may offer another route to preemption, but the relevant forms may not be represented in the models’ BPE or lemmatized vocabularies.The discussion contrasts argument-structure preemption with examples such as childs and goed.
- 6.3 Preemption from abstract cues: At the verb-specific level, the models do not show preemption, but exposures from other verbs provide evidence for more abstract preemption.This suggests that indirect evidence may accumulate across verbs rather than operate only through a target verb’s own competing constructions.
- 6.3 Preemption from abstract cues: Human corpus estimates of preemption might be strengthened by incorporating indirect evidence from other verbs rather than considering only verb-level usage.The authors present this as a question for future analysis, not as a reason to reject prior human findings.
7 Conclusion
The paper demonstrates how controlled rearing can separate closely related explanations of statistical learning that human experiments cannot causally disentangle. It uses transformer language models to study how input shapes generalization behavior.
- 7 Conclusion: Controlled rearing allows the study to disentangle subtly different proposals for indirect negative evidence in statistical learners.The authors emphasize its value because this causal separation is difficult with human learners.
- 7 Conclusion: The work contributes to understanding how transformer language models’ input shapes their generalization behavior.The conclusion places the study within broader efforts to use statistical learners for in-principle insights about statistical learning.
Limitations
The limitations concern the narrow empirical scope, model generalization, and assumptions used to identify preemptive signals from corpus data.
- Single Phenomenon: The study examines one phenomenon, so models may show different generalization behavior on other constructions.The authors note that controlled rearing can be applied to additional constructions in future work.
- Generalization of our conclusion across models: The model analyses cover only one instance of the transformer language models trainable on CHILDES.The absence of individual-verb preemption therefore does not establish that preemption cannot arise in transformers or neural networks generally.
- Generalization of our conclusion across models: The findings may be stronger with larger datasets or models, and one connectionist model showing both effects relies on handcoded semantic features rather than naturalistic learning.These scope conditions limit direct generalization from the present experiments.
- Operationalizing preemptive signals: The study assumes that all instances of the competing construction are contexts in which either construction could have been used.Future work should distinguish cases where constructions necessarily compete from cases where they do not.
A Details for Training
The study trains language models on CHILDES-derived data using parsed constructions, synthetic evaluation sentences, and controlled diagnostics of training behavior.
- Training setup: Models are trained with GPT-2 and PyTorch Lightning on CHILDES data, with each run requiring about 1.5 hours on one NVIDIA L40s GPU.The training setup uses 1024-token sequences and sentence-boundary separators.
- Construction identification: Periphrastic causatives with let and make are identified when a non-to-infinitive complement verb has an object or its own subject.The procedure also requires let to have a subject within its predicate to exclude hortatives such as “let’s go.”
- Construction identification: For get, cause, and force, the same criteria apply except that the complement verb must be a to-infinitive.Have is handled similarly to let and make, with an external LLM used to filter false positives.
- Evaluation synthesis: Synthetic periphrastic-causative sentences are formed by prompting Qwen 3.5 27B to choose natural objects from candidate lists for sampled subjects and verbs.The system prompt targets simple scenes suitable for young children’s literature.
- Evaluation: The evaluation compares transitive and periphrastic causative sentence pairs across intransitive, alternating, and transitive verb classes.The study also tracks final-checkpoint ΔNLLs across five seeds and 100 sentence pairs per verb.
E Additional Results for Abstract Effect from Intransitive Verbs
The abstract manipulation removes periphrastic causative usages for all identified intransitive verbs while preserving usages of alternating verbs.
- Abstract manipulation: Removing periphrastic causatives for 174 intransitive verbs eliminates 4,675 usages, or 62.7% of all such usages in CHILDES.Periphrastic causatives of alternating verbs remain intact in this intermediate manipulation.
F Additional Results for Training Trajectories
Training-trajectory analyses show convergence without clear overfitting, while overgeneralization tendencies can fluctuate during training.
- Training trajectories: Figure 9 tracks NLLs of overgeneralized transitive forms across controlled-rearing conditions during training.The figure uses solid lines for REMOVE-P models and dashed lines for KEEP-P models.
- Training trajectories: The models largely converge within 3 epochs without significant overfitting, but overgeneralization tendencies remain volatile during training.The smoothed trajectories are not suitable for comparing the final NLLs used in the main experiments.
- Condition contrasts: Figure 8 compares KEEP-P and REMOVE-P estimated marginal means across verb-specific, abstract-intransitive-only, and abstract conditions, with positive values indicating preemption.The figure includes 95% confidence intervals.
G Additional Results Without Lemmatization
Additional analyses examine BPE tokenization and reassess a prior preemption measure, finding that its apparent effect can reflect changes in the conventional sentence’s likelihood.
- BPE robustness: The study repeats its main analyses with an 8,192-token BPE tokenizer because subword tokenization can burden learning connections among word forms.The BPE analyses are conducted as a robustness check against the primary tokenization setting.
- Comparison with prior work: Guo et al. measure overgeneralization using the surprisal difference between unconventional transitive and conventional periphrastic-causative sentences, and report preemption effects in LMs.Their method combines correlations with preemptive usages and fine-tuning on conventional versus unconventional examples.
- Methodological reassessment: The paper argues that Guo et al.’s fine-tuning cannot establish how preferences were learned because the models may already have formed them during pretraining.The authors also identify overlap between the manipulated conventional form and the measure used to assess overgeneralization.
- Methodological reassessment: The authors argue that disentangling preemption from entrenchment requires direct training-data manipulation and a measure focused exclusively on the unconventional form.Their controlled-rearing methods are designed around these requirements.
- Alternative measure: Under the alternative NLL-difference measure, apparent preemption for go, laugh, and sneeze is driven by lower periphrastic-causative NLL in KEEP-P models.The transitive NLL itself is lower in KEEP-P than REMOVE-P, contrary to an interpretation based solely on stronger avoidance of the unconventional form.