PaperlayerView exact revision on arXiv

Source-linked AI summary

From Exposure to Expectation: Frequency, Surprisal, and Language Across Development in Spanish

Francisco Portillo López

arXiv:2608.22452v1cs.CL

TL;DR

It remains unclear whether surprisal's established association with adult reading also explains why children acquire some words earlier than others. The paper compares surprisal and frequency in Spanish corpus-based analyses of acquisition timing for 225 nouns and adult fixation durations. Surprisal’s predictive association differed reliably across the two domains (z = 3.63, p < .001), and it added no acquisition-timing information beyond cumulative lexical exposure and word length.

  • Problem

    It remains unclear whether surprisal's established association with adult reading also explains why children acquire some words earlier than others.

    Source: §1 · pp. 2, 5, 6

  • Method

    The paper compares surprisal and frequency in Spanish corpus-based analyses of acquisition timing for 225 nouns and adult fixation durations.

    Source: §2.1.1 · p. 6, §4.5 · p. 25

  • Results

    Surprisal’s predictive association differed reliably across the two domains (z = 3.63, p < .001), and it added no acquisition-timing information beyond cumulative lexical exposure and word length.

    Source: §4 · p. 20, §4.5 · p. 25

  • Takeaways & Limitations

    The findings suggest that surprisal’s predictive association differs across the two domains.

    Source: §3.3 · p. 17, §4.5 · p. 25

    The apparent developmental contrast may reflect richer context available for surprisal estimation in Study 2 rather than a genuine developmental difference.

    Source: §4.3 · p. 24

Abstract

from arXiv · show

Surprisal, the negative log-probability a language model assigns to a word given its preceding context, reliably predicts adult reading times. Does it contribute as much to explaining when children acquire individual words? Frequency reflects a learner's cumulative exposure to a word, whereas surprisal reflects how predictable a single occurrence is given its context. We investigate this question across two corpus-based studies of Spanish. In Study 1, we modeled age of acquisition (AoA) for 225 Spanish nouns using lexical frequency and contextual diversity from child-directed speech, plus surprisal from three language models differing in architecture and training language (BETO, BERTIN, mGPT). Frequency strongly predicted AoA (r=-.597, p<.001); surprisal added little beyond frequency and word length, including in a naturalistic-context analysis. In Study 2, we modeled adult fixation durations in the Chilean Spanish subsample of the Multilingual Eye-movement Corpus (MECO Wave 2), using mGPT surprisal alongside two independent frequency measures. Surprisal robustly predicted longer fixation durations after controlling for frequency and word length, consistent across both frequency sources. A matched word-type-level comparison showed the surprisal-behavior association was stronger in reading than in acquisition (z=3.63, p<.001). The findings suggest cumulative lexical exposure and contextual predictability play different roles across the language trajectory: frequency is particularly informative about when early lexical representations are acquired, whereas surprisal captures moment-to-moment processing difficulty in an already-established linguistic system. We discuss this pattern in relation to usage-based and entrenchment-based accounts of lexical development and to the evaluation of language models as models of human language behavior.

1 Introduction

The introduction frames two Spanish studies asking whether LLM-derived surprisal and cumulative word frequency explain word acquisition and adult reading in the same way. It distinguishes frequency from contextual predictability and considers whether surprisal’s explanatory value varies across developmental and behavioral levels.

1 Introduction

  • Frequency represents cumulative experience with a word, whereas surprisal measures the predictability of one occurrence in context; the measures can therefore make different predictions about acquisition and processing.Usage-based accounts link repeated encounters to strengthening or entrenchment of lexical representations.Source: §1 · p. 3
  • The paper compares surprisal’s explanatory value for lexical acquisition and adult online processing in Spanish, while treating its acquisition analysis as descriptive rather than a test adjudicating learning theories.The authors frame the comparison as a step toward establishing the relationship between surprisal and age of acquisition after accounting for frequency.Source: §1 · pp. 4–6
  • Study 1 tests whether surprisal from BETO, BERTIN, and mGPT predicts Spanish children’s noun age of acquisition beyond frequency and contextual diversity.It also examines mGPT surprisal in naturalistic child-directed contexts to check whether the association depends on standardized carrier frames.Source: §1 · p. 5
  • Study 2 tests whether mGPT surprisal predicts adult fixation duration in Chilean Spanish MECO Wave 2 beyond frequency and word length.Two independent frequency norms serve as robustness checks for the surprisal effect.Source: §1 · p. 5
  • The authors compare left-to-right autoregressive and bidirectional masked models, and calculate word-level surprisal by summing constituent subword-token surprisals.They flag subword tokenization as a non-trivial modeling choice.Source: §1 · pp. 5–6
  • Surprisal is expected to add limited information about AoA variation once frequency is controlled.Source: §1 · p. 6
  • mGPT surprisal is expected to predict adult reading difficulty beyond frequency and word length.Source: §1 · p. 6
  • A stronger surprisal association with adult reading than acquisition would support evaluating model-derived statistics at developmental and behavioral levels.Source: §1 · p. 6

2 Study 1: Surprisal and Age of Acquisition in Spanish · 2.1 Method

Study 1 examines age of acquisition for Spanish nouns using child-directed speech measures and language-model surprisal. Frequency is strongly associated with earlier acquisition, while contextual diversity contributes less independently once frequency is considered.

2.1.1 Materials

  • The study examined AoA for 225 Spanish nouns using Spanish CDI norms from Wordbank; corpus-derived frequency and contextual-diversity analyses included 224 nouns.Source: §2.1.1 · pp. 6–7

2.1.2 Predictor measures

  • Lexical frequency counted each noun’s occurrences in child-directed speech, while contextual diversity counted the distinct utterances or contexts in which it occurred.Because these measures were strongly related, their relationships with AoA were evaluated separately and their incremental contributions examined in regression models.Source: §2.1.2 · p. 7
  • Contextual diversity showed a similar relationship with age of acquisition, but its independent contribution was greatly reduced after accounting for frequency.Source: §2.1.2 · p. 7
  • Surprisal was the negative log probability of a target word given preceding context, estimated with BETO, BERTIN (De la Rosa et al., 2022), and mGPT (Shliazhko et al., 2022).The models differed in architecture and training data; BETO and BERTIN were Spanish-trained models, whereas mGPT was trained on 60 languages.Source: §2.1.2 · pp. 7–8
  • Primary surprisal estimates came from standardized carrier frames; an additional mGPT analysis averaged target-word surprisal across up to ten natural child-directed-speech contexts per noun.Source: §2.1.2 · p. 8

2.1.3 Statistical analysis

  • AoA was analyzed using Spearman rank correlations and regression models testing whether mGPT surprisal added predictive value beyond lexical frequency and word length; standardized-context and natural-context analyses used 225 and 224 nouns, respectively.Source: §2.1.3 · p. 8

2.1.2 Predictor measures

  • Spearman r = −.597, p < .001: child-directed-speech frequency was negatively associated with age of acquisition for 225 Spanish nouns.The figure labels selected early- and late-acquired nouns for illustration.Source: §2.1.2 · p. 7, Figure 1 · p. 9

2.2 Results

Across 224 nouns, child-directed-speech frequency was associated with earlier age of acquisition (AoA), while LLM surprisal added little independent explanatory value. A small raw surprisal association in naturalistic contexts did not translate into reliable incremental prediction beyond frequency and word length.

2.2.1 Frequency and contextual diversity

  • Frequency in child-directed speech correlated negatively with AoA (ρ = −.597, p < .001), indicating that more frequent words tended to be acquired earlier.Contextual diversity also related strongly to AoA, but that relationship was substantially reduced when frequency was taken into account.Source: §2.2.1 · p. 9

2.2.2 LLM surprisal

  • BETO surprisal was unrelated to AoA (ρ = −.036, p = .590), while BERTIN’s association was small and non-significant (ρ = .102, p = .128), and primary mGPT surprisal was essentially unrelated (ρ = −.048, p = .469).Adding mGPT surprisal to a baseline model containing frequency and word length produced almost no additional explanatory value (ΔR2 = .0006); its coefficient was not significant (β = −.024).Source: §2.2.2 · p. 10
  • Natural-context mGPT surprisal had a small positive raw association with AoA (ρ = .153, p = .022, N = 224), but added only ΔR2 = .0008 after controls.The baseline model explained R2 = .3677 and the full model R2 = .3685; the surprisal coefficient was not significant (β = −.0209, SE = .0393, t = −.532, p = .595).Source: §2.2.2 · p. 11
  • Among 154 nouns with at least five valid contexts, the natural-context association was not significant (ρ = .132, p = .104), and adjusted ΔR2 = .0076 remained small.The mGPT coefficient remained non-significant (β = −.0737, SE = .0569, p = .197).Source: §2.2.2 · p. 11

2.3 Interim discussion

In Study 1, child-directed speech frequency was more informative about Spanish noun acquisition timing than surprisal: its association with AoA persisted where surprisal added little beyond frequency and word length. Naturalistic contexts showed a small bivariate surprisal association that disappeared after controls.

2.3 Interim discussion

  • Frequency in child-directed speech predicted earlier AoA, whereas standardized-context mGPT surprisal was unrelated to AoA and added virtually nothing beyond frequency and word length.Source: §2.3 · pp. 11–12
  • Contextual diversity was associated with AoA, but its explanatory contribution diminished substantially after controlling for frequency.Source: §2.3 · p. 12
  • In naturalistic child-directed speech contexts, mGPT surprisal had a small bivariate AoA association that disappeared after controlling for frequency and word length and was not robust in the five-context subset.Source: §2.3 · p. 12
  • The findings are compatible with usage-based and entrenchment accounts in which repeated encounters strengthen lexical representations.They do not imply that children ignore contextual expectations; rather, moment-to-moment predictability provided limited information about acquisition timing once cumulative exposure was considered.Source: §2.3 · p. 12
  • Study 2 tests whether mGPT surprisal predicts adult moment-to-moment reading difficulty, probing whether its limited acquisition contribution reflects a broader failure to capture psychologically relevant Spanish structure.Source: §2.3 · p. 12

3 Study 2: Surprisal and Adult Reading Times in Spanish

This section examines whether mGPT surprisal explains adult Chilean Spanish fixation durations beyond frequency and word length, and situates that reading-time effect alongside its association with AoA. It considers contextual surprisal as more directly expressed in online processing than in AoA’s cumulative distributional measure, without implying that surprisal is unrelated to learning.

3.1.1 Participants and materials

  • The Chilean Spanish MECO Wave 2 sample comprised 45 participants reading 12 passages; after exclusions, 46,442 word-level observations remained.The outcome was total fixation duration, aggregated across fixations on each word during first-pass reading and rereading, then log-transformed.Source: §3.1.1 · p. 13

3.1.2 Predictor measures

  • mGPT surprisal was computed for each word from the preceding passage context, summing surprisal across constituent subword tokens when needed.This measured contextual predictability as words unfolded in the continuous texts participants read.Source: §3.1.2 · pp. 13–14
  • EsPal and SUBTLEX-ESP frequency measures correlated at r = .95, so the authors fit parallel models rather than including both norms together.Convergence across the two frequency sources served as a robustness test of the surprisal relationship.Source: §3.1.2 · p. 14

3.1.3 Statistical analysis

  • Linear mixed-effects models predicted log fixation duration from surprisal, frequency, and word length, with crossed random intercepts for participant and word type.Models A and B used EsPal and SUBTLEX-ESP frequency, respectively.Source: §3.1.3 · p. 14

3.2 Results

  • mGPT surprisal predicted longer fixation durations in both models: Model A (EsPal) β = 0.0114 (SE = .00067, t = 17.19, p < .001) and Model B (SUBTLEX-ESP) β = 0.0114 (SE = .00066, t = 17.33, p < .001).The effect remained after controlling for frequency, word length, and crossed random effects for participant and word type.Source: §3.2 · p. 15

3.2.1 Relative magnitude of predictors

  • In the MECO Wave 2 Chilean Spanish sample, standardized effects predicting log total fixation duration ranked word length strongest (β = 0.292, 95% CI [0.267, 0.316]), surprisal second (β = 0.112, 95% CI [0.099, 0.125]), and lexical frequency negative (β = −0.051, 95% CI [−0.082, −0.020]).All three effects were statistically significant.Source: §3.2.1 · p. 16, Figure 3 · p. 17

3.3 Interim discussion

  • The dissociation does not make surprisal unrelated to learning or frequency and predictability wholly independent; contextual surprisal appears more directly expressed in online processing than in AoA’s cumulative distributional measure.Source: §3.3 · p. 17
  • mGPT surprisal predicted longer adult fixation durations after controls for frequency and word length, while its naturalistic-context AoA association disappeared after those controls.Source: §3.3 · pp. 17–18, Figure 4 · p. 19

4 General Discussion

The discussion considers how frequency and surprisal relate to word acquisition and adult reading behavior, including the stronger surprisal association reported for reading. It addresses limits on developmental interpretation and proposes matched-context and more developmentally representative studies to test the contrast.

4 General Discussion

  • In Study 1, frequency predicted age of acquisition while surprisal added little beyond frequency and word length; in Study 2, higher surprisal predicted longer adult fixations after controls.In the standardized adult-reading model, surprisal was the second-largest predictor among the three lexical variables, after word length.Source: §4 · pp. 19–20
  • Naturalistic CDS-context mGPT surprisal had a small bivariate AoA association (r = .153, p = .022), which disappeared after controlling for frequency and word length (β = −.021, p = .595, ∆R2 = .0008).The association was not robust among nouns with at least five available contexts (r = .132).Source: §4 · pp. 19–20

4.1 Entrenchment, expectation, and the acquisition–processing distinction

  • The discussion distinguishes cumulative word exposure, which was strongly associated with AoA, from contextual predictability, which better captured processing of already-established lexical knowledge.The authors present entrenchment as a theoretical interpretation, not a demonstrated developmental mechanism, because neither study experimentally manipulated frequency or predictability.Source: §4.1 · pp. 20–21
  • The authors argue that model evaluation should be task- and population-dependent, pairing adult reading-time benchmarks with developmental measures such as vocabulary composition, AoA, and learning trajectories.mGPT surprisal predicted adult reading behavior but provided negligible incremental information about AoA after cumulative frequency and word length were considered.Source: §4.1 · p. 22

4.2 A formal dissociation between the two domains

  • To formally compare the domains, the authors matched the statistical quantity across datasets by aggregating MECO data to the word-type level and using Spearman correlations.This comparison was intended to go beyond contrasting one significant result with one nonsignificant result.Source: §4.2 · p. 23
  • Surprisal correlated more strongly with adult fixation duration than with acquisition (r = .221 vs. r = −.048), a difference confirmed by Fisher r-to-z (z = 3.63, p < .001).The comparison used unadjusted word-type-level correlations, not the covariate-adjusted regression estimates from the two studies.Source: §4.2 · p. 23
  • With naturalistic child-directed speech contexts, the acquisition correlation rose to r = .153 but disappeared after controlling for frequency and word length.The authors conclude that acquisition-association magnitude partly depends on surprisal estimation, while its independent explanatory value remains limited.Source: §4.2 · p. 23

4.3 Limitations

  • The cross-study contrast may partly reflect unequal context length: Study 2 used full preceding passages, whereas Study 1 used standardized frames or up to ten child-directed speech contexts.The authors also note that the acquisition and reading studies differed in Spanish variety and register, which may contribute to the contrast.Source: §4.3 · p. 24
  • Only mGPT was tested in both studies, and the correlational designs cannot establish that frequency causes earlier acquisition or surprisal causes longer reading times.BETO and BERTIN converged with the weak acquisition effect but were not evaluated on the MECO reading materials; the entrenchment-versus-expectation account is therefore an interpretation, not a causal conclusion.Source: §4.3 · p. 24

4.4 Future directions

  • Future work should equalize surprisal context windows across developmental and adult datasets, using the same model and estimator to test whether the contrast persists.A surviving contrast would make a developmental interpretation more compelling.Source: §4.4 · pp. 24–25
  • Researchers should test models trained on realistic child-directed input and combine developmental and processing measures within the same language variety and, where possible, participants or longitudinal populations.These approaches could distinguish effects of adult-trained models from a more fundamental distinction between frequency-based learning and expectation, and test whether information predicting acquisition later predicts skilled online processing.Source: §4.4 · p. 25

4.5 Conclusion

  • Across two corpus-based Spanish studies, frequency robustly predicted noun acquisition, whereas surprisal predicted adult reading times independently of frequency, with the cross-study association difference statistically reliable.The design cannot determine whether that contrast is developmental or reflects methodological asymmetries, especially unequal surprisal context length.Source: §4.5 · p. 25
  • The authors interpret the pattern as compatible with entrenchment-based early learning and expectation-based prediction becoming behaviorally dominant once the lexicon is established.They caution against treating cognitive plausibility as a general-purpose property of language models and advocate evaluation across multiple developmental populations.Source: §4.5 · p. 26

Ethics Statement

The analyses used publicly available, de-identified secondary data, involved no new human-participant data collection, and relied on original data collection approved by the respective corpus teams.

Ethics Statement

  • The study used publicly available, de-identified secondary data from Wordbank CDI norms, CHILDES transcripts, and the MECO Wave 2 eye-tracking corpus.Source: Ethics Statement · p. 26
  • Because the author collected no new data from human participants, institutional review was not required for these analyses.Source: Ethics Statement · p. 26
  • The respective corpus teams conducted and ethically approved all original data collection.Source: Ethics Statement · p. 26

Funding

The research received no specific grant from any funding agency.

Funding

  • The research received no specific grant from any funding agency.Source: Funding · p. 26

Open Science Statement

Both studies were exploratory, corpus-based analyses of secondary data and were not preregistered. Their statistical comparisons are hypothesis-generating as well as hypothesis-testing, and the results are provisional pending independent replication.

Data Availability

The study materials and resources are publicly available: CDI norms via Wordbank, CHILDES corpora via TalkBank, and MECO Wave 2 data via OSF; analysis code for both studies is also available.

Data Availability

  • CDI norms are publicly available via Wordbank.https://wordbank.stanford.eduSource: Data Availability · p. 27
  • CHILDES corpora are publicly available via TalkBank.https://childes.talkbank.orgSource: Data Availability · p. 27
  • MECO Wave 2 data (release 2.0) are publicly available via OSF, and analysis code for both studies is available.The OSF link is https://osf.io/3527a/ (Siegelman et al., 2025).Source: Data Availability · p. 27

Related papers

Loading 2608.22452v1…