PaperlayerView exact revision on arXiv

Source-linked AI summary

Modeling the Developmental Shift in Telicity Acquisition

Ellie Xia, Parisa Kordjamshidi, Alan Hezao Ke

arXiv:2609.17996v1cs.CL

TL;DR

Telicity acquisition requires learners to map surface and semantic cues to event structure, but the computational trajectory of that mapping is not well understood. The paper uses GPT-2 surprisal over paired temporal-adverbial diagnostics to label telicity and trains classifiers on child and adult child-directed speech. The child model relies on a deterministic post-verbal determiner cue, whereas the adult model relies more on lexical semantics, a pattern the authors interpret as consistent with Syntactic Bootstrapping.

  • Problem

    The computational trajectory by which L1 learners map surface and semantic cues to telicity remains poorly understood.

    Source: §1 · p. 2

  • Method

    The paper labels telicity using GPT-2 surprisal differences for paired temporal-adverbial diagnostics, then compares child and adult feature-based logistic-regression models.

    Source: §1 · p. 2, §3.2 · p. 4

  • Results

    The child model reaches 100% across all metrics through a deterministic post-verbal determiner cue, while the adult model reports accuracy 0.93 and macro F1-score 0.84.

    Source: §4 · p. 6

  • Takeaways & Limitations

    The authors interpret the shift from children’s surface-level determiner cue to adults’ lexical-semantic emphasis as supporting Syntactic Bootstrapping.

    Source: §7 · p. 9

    GPT-2 is trained on adult web text rather than child-language input, and manual CHILDES transcripts may omit multimodal cues such as gestures.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Acquiring telicity, which is the distinction between bounded (e.g., ate an apple) and unbounded (e.g., ate apples) events, requires first language (L1) learners to map surface-level and semantic cues to abstract event structures, but the computational trajectory of this mapping is not well understood. We introduce a Difference in Surprisal method that uses GPT2 token surprisal over paired temporal adverbial diagnostics (in an hour versus for an hour) to automatically label telicity across English CHILDES corpora, validated against expert linguist judgments. Using these labels, we train diagnostic logistic regression classifiers on 12 syntactic and lexical semantic features to compare how child speech and child-directed speech encode telicity. The two models diverge: the child model reaches near perfect accuracy through a single deterministic cue, the presence of a post-verbal determiner, while the adult model relies more heavily on verb class and other lexical semantic features, with the determiner cue neutralized. This trajectory supports Syntactic Bootstrapping: learners first exploit high-frequency structural cues as a scaffold to bootstrap, before developing fully compositional, verb-based event structures.

1 Introduction

Telicity distinguishes bounded events with endpoints from unbounded events, and learners must connect this distinction to morphosyntactic and semantic cues. The study uses computational modeling of naturalistic child and child-directed speech to examine how those cues develop.

1 Introduction

  • Telicity distinguishes events with an inherent culmination from events characterized by ongoing duration without an endpoint.Temporal adverbials diagnose the distinction: “in an hour” suits telic predicates, while “for an hour” selects atelic predicates.Source: §1 · p. 1
  • Acquiring telicity requires learners to track the interface between syntactic cues, such as determiners, and semantic cues, such as boundedness and verb class.The introduction also identifies resultative particles as surface cues learners must map to aspectual semantics.Source: §1 · pp. 1–2
  • The study targets the gap between formal theories of event boundedness and the distribution of telicity cues in naturalistic child and child-directed speech.It tests hypotheses concerning object boundedness, verb properties, particles, and combinations of cues using CHILDES and language-model analysis.Source: §1 · p. 2
  • The proposed pipeline uses LLM-derived surprisal labels and logistic-regression probes to compare the cues used in child and adult speech.The authors frame the expected developmental pattern as a shift from syntactic bootstrapping toward semantic composition.Source: §1 · p. 2

2 Related Work

Prior work describes telicity acquisition as uneven across cue types and documents computational models’ limits in representing compositional aspect. The paper situates its approach at the intersection of these developmental and computational accounts.

2 Related Work

  • Telicity spans verb class, morphology, and syntax, and prior research has examined this interface through developmental psycholinguistics and computational modeling.The cited research includes Bloom et al. (1980), Shirai and Andersen (1995), van Hout (1998), Zhao et al. (2021), and Kim et al. (2024).Source: §2 · p. 2

2.1 Theoretical Acquisition of Telicity

  • Prior work reports earlier acquisition of transparent mappings than complex compositional ones, including earlier interpretation of verb particles than object quantization.Children’s use of tense-aspect morphology is initially restricted to semantic classes; reported object-quantization judgments often do not become adult-like until after age five.Source: §2.1 · p. 3

2.2 Computational Modeling of Telicity

  • Earlier aspect classifiers use hand-crafted indicators or sequential labeling, while distributional models capture coarse telic environments but miss hierarchical composition and contextual boundaries.The surveyed work also describes structured models for coercion and pretrained transformers that rely disproportionately on surface cues relative to human judgments.Source: §2.2 · pp. 3–4

3 Methodology

The methodology pairs GPT-2 surprisal-based telicity labels with feature-based diagnostic classifiers. It compares models trained separately on child and adult child-directed speech to examine which cues predict telicity.

3 Methodology

  • The pipeline first generates binary telicity labels from surprisal differences, then trains diagnostic classifiers on those labels and compares child and adult feature weights.The authors use LLM annotations as targets for the feature-based classifiers.Source: §3 · p. 4

3.1 Data and Preprocessing

  • The corpus comes from English-NA and English-UK CHILDES, partitioned into child utterances and adult speech directed to children.Preprocessing removes non-linguistic annotations, fragments under three tokens, utterances without verbs, and utterances exceeding 30 tokens.Source: §3.1 · p. 4

3.2 Telicity Labels via LLM-Based Surprisal

  • The Difference in Surprisal method labels predicates by comparing GPT-2 surprisal for temporal diagnostics, with positive ∆S indicating greater surprisal for “for an hour” and thus a more telic event.The score is discretized into binary telic or atelic labels; averaging token surprisal over phrase length is intended to avoid length-dependent bias.Source: §3.2 · pp. 4–5

3.3 Features for Telicity Classification

  • The 12 features represent syntactic configuration, lexical verb class, and morphosyntactic cues, including determiners, particles, goal phrases, and aspect marking.Verb classes distinguish unergative activities, unaccusative change-of-state verbs, and transitive verbs whose telicity varies with their object.Source: §3.3 · p. 5

3.4 Diagnostic Probing Models

  • The study compares child and parent speech models by using logistic-regression coefficients to quantify the importance of linguistic cues for predicting telicity labels.A high positive coefficient indicates that a feature is a strong predictor of telicity for that group.Source: §3.4 · pp. 5–6

4 Results

The child model’s perfect test performance is attributed to a single deterministic determiner cue, while the adult model performs robustly but less perfectly. The models therefore differ in how their features predict the LLM-derived labels.

4 Results

  • 100% across all metrics for the Child Model reflects a deterministic post-verbal determiner rule, not superior generalization; the Adult Model reaches accuracy 0.93 and macro F1-score 0.84.In the child subset, the presence of a determiner implies telicity.Source: §4 · p. 6

5 Expert Validation

The Difference in Surprisal labels were compared with expert judgments, with the default threshold showing a predictive signal and an optimized boundary improving alignment. Expert agreement was high, while some judgments reflected ambiguity in aspectual interpretation.

5 Expert Validation

  • Expert labels were derived from three linguists’ judgments of compatibility with “in an hour” versus “for an hour,” with utterances labeled telic at TI >= 0.5.The evaluation sampled 200 utterances, split evenly between child and child-directed speech.Source: §5 · p. 6

5.1 Inter-Rater Reliability and Human Consensus

  • At threshold 0, surprisal differences correlated with human judgments (Pearson r = 0.266) and yielded 70.5% classification accuracy, while underestimating telic events.The pre-threshold confusion matrix identifies 120 true negatives for atelic events.Source: §5 · p. 6, Table 2 · p. 6
  • At least two of three linguists agreed on 96.00% of utterances, but all three agreed on 43.50%, with disagreements often involving permitted dual readings or contextual licensing.Ambiguity flags applied to 69 of 200 items and were evenly split between child and parent speech.Source: §5.1 · p. 7
  • 87.50% lenient accuracy against human majority consensus and 88.50% average per-rater accuracy show alignment between the model and expert judgments.Table 3 reports these validation outcomes for 200 utterances.Source: §5.1 · p. 7, Table 3 · p. 7
  • The optimized boundary ϵ = −0.198 raised correlation to r = 0.2999; for 53 unambiguous items, correlation rose from r = 0.2666 to r = 0.4188.The latter increase was statistically significant (p = 0.0018).Source: §5.1 · pp. 7–8
  • The child model’s determiner-based rule was interpreted as evidence that post-verbal syntactic configuration can serve as a structural cue for bootstrapping telicity.The authors frame this result within the Syntactic Bootstrapping account.Source: §5.1 · p. 8

6 Discussion

The discussion describes a developmental shift: child telicity predictions rely on post-verbal determiners, while adult predictions rely more on verb semantics. It presents this contrast as computational evidence for syntactic bootstrapping.

6.1 “Determiner clue” in Child Speech

  • Child Model accuracy is 1.00, with n_determiner_after_verb coefficient 16.07 versus −0.02 in the Parent Model.The discussion interprets the adult model's near-zero determiner coefficient in light of more syntactically complex and longer parent utterances, where determiners are more variably distributed.Source: §6.1 · p. 8
  • The discussion presents the child model's structural cue reliance as computational evidence for the Syntactic Bootstrapping Hypothesis.The hypothesis proposes that children learn verb meanings by observing the grammatical structures in which verbs occur.Source: §6.1 · p. 8

6.2 Structural cues: Distance Matters

  • verb_det_dist has β = 3.78, and the discussion reports “perfect separation” caused by determiners.It describes determiners as a frequent, closed-class cue in the immediate post-verbal domain, contrasted with open-class verbs' greater variety and sparser individual token frequencies.Source: §6.2 · p. 8

6.3 Developmental Shift: From Syntax to Semantics

  • The Adult model's Verb_Class coefficient is β = 3.752, while its n_determiner_after_verb coefficient drops near zero relative to the Child Model.The discussion frames this as a transition from syntactic cues in child production to semantic cues in adult production.Source: §6.3 · p. 8
  • The Child Model relies on explicit determiners for telicity, a high-precision but brittle rule that can fail with mass nouns such as “popcorn.”The discussion links the child pattern to H1 and contrasts it with the Adult Model's use of inherent verb aspect, such as recognizing arrive as telic regardless of other cues, linked to H4.Source: §6.3 · pp. 8–9

7 Conclusion

The conclusion presents a computational link between formal event semantics and first-language acquisition theory. It reports a shift from children's determiner-based cue to adults' reliance on lexical semantics, and identifies cross-language testing as future work.

7 Conclusion

  • The study links formal event semantics and first-language acquisition theory through a computational approach.The conclusion describes the work as a computational bridge between these areas.Source: §7 · p. 9
  • The conclusion characterizes children as bootstrapping event boundedness through a high-precision, surface-level “Determiner Clue,” while adults prioritize inherent lexical semantics.Source: §7 · p. 9
  • Future work will test whether this developmental pattern holds across languages with richer or lesser morphology, including Spanish, German, Mandarin, and Thai.The stated aim is to assess whether the Determiner Clue is universal or specific to English grammar.Source: §7 · p. 9

Limitations

The study's labeling model and transcript data constrain what its findings represent. GPT-2 reflects adult-web-text norms, and manual transcripts may omit multimodal cues such as gestures.

Limitations

  • GPT-2 was trained on adult web text, so its surprisal measures alignment with adult norms rather than child language input or production.Source: Limitations · p. 9
  • Manual CHILDES transcripts may overlook multimodal cues, including gestures.Source: Limitations · p. 9
  • The study's labeling-model alignment is with adult-web-text norms, not the language input or production typical in child acquisition.Source: Limitations · p. 9

Ethical considerations

The human evaluation involved three expert linguists at Michigan State University. Participation was voluntary, and participants received a task and objective description and gave written informed consent.

Ethical considerations

Chenchen Song. 2018. Severing telicity from result.

The listed works span telicity acquisition, aspectual bootstrapping, computational event classification, and cross-linguistic studies of event structure.

A Preprocessing and Annotation Statistics

The section reports corpus subset sizes and telicity-label distributions, then describes the expert-annotated sample’s label distribution by rater and against model predictions.

Loading 2609.17996v1…