---
title: "Psycholinguistics & language acquisition — papers"
description: "Psycholinguistics & language acquisition: 62 papers of the Elamur Research library, the most cited first."
canonical: "https://research.elamur.ai/free/en/topics/psycholinguistics/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/topics/psycholinguistics/)

[home](https://md.research.elamur.ai/free/en/) / [topics](https://md.research.elamur.ai/free/en/topics/) / Psycholinguistics & language acquisition

# 🧒 Psycholinguistics & language acquisition

62 papers · page 1 of 2 · 1–25 · the most cited first · [all topics](https://md.research.elamur.ai/free/en/topics/)

How people read, hear and learn language, and models that behave like them.

💡 Explained: [Child-directed speech](https://md.research.elamur.ai/free/en/concepts/child-directed-speech/) · [Cognitive linguistics](https://md.research.elamur.ai/free/en/concepts/cognitive-linguistics/) · [Construction grammar](https://md.research.elamur.ai/free/en/concepts/construction-grammar/) · [Psycholinguistics](https://md.research.elamur.ai/free/en/concepts/psycholinguistics/)

A survey · about 17 minutes to read · every paper named here has its own page in this library

**In this article**

1. [What it is](#a-what-it-is)
2. [History and milestones](#a-history-and-milestones)
3. [Main problems and tasks](#a-main-problems-and-tasks)
4. [Methods and schools](#a-methods-and-schools)
5. [Datasets, resources and benchmarks](#a-datasets-resources-and-benchmarks)
6. [Landmark papers in this library](#a-landmark-papers-in-this-library)
7. [People](#a-people)
8. [Institutions and places](#a-institutions-and-places)
9. [Open questions](#a-open-questions)
10. [Neighbouring topics](#a-neighbouring-topics)
11. [Key works in this library](#a-key-works-in-this-library)

Psycholinguistics studies how people produce, understand and acquire language: how a reader builds the structure of a sentence word by word, why some sentences are hard, how memory limits shape comprehension, and how a child gets from hearing speech to speaking a grammar. Computational psycholinguistics adds explicit models to these questions. A model that assigns probabilities to words or structures makes predictions that can be checked against reading times, eye movements, brain responses and children's development.

The papers here sit at the meeting point of that tradition and natural language processing. In one direction, theories from psycholinguistics are tested with computational tools, above all with probabilities from language models. In the other, the methods of the psycholinguistics laboratory, such as controlled minimal pairs, garden-path sentences and developmental comparisons, are turned on neural networks to find out what they have learned. A third strand builds resources: annotated child-language corpora, eye-tracking and EEG datasets, and small benchmarks that put a learner on a child's diet of words.

Most of the work here is recent, but the most cited paper, from 2001, introduced the idea that still organises much of the field: surprisal.

## What it is

The topic covers two related programmes.

**Sentence processing** asks what makes a word or a structure easy or hard to process in real time. The main measures are reading times (from self-paced reading or eye-tracking), event-related brain potentials such as the N400 and P600, and acceptability judgments. The main theoretical families are:

- **Expectation-based accounts**, in which difficulty reflects how unpredictable a word is in its context. Surprisal theory, the negative log probability of a word given what came before, is the standard formulation.
- **Memory-based accounts**, in which difficulty reflects the cost of storing and retrieving earlier material, as in Dependency Locality Theory, cue-based retrieval and ACT-R models of parsing.
- **Information-theoretic accounts of communication**, such as uniform information density and the memory–surprisal trade-off, which ask how languages and speakers distribute information over time.

**Language acquisition** asks how children learn words and grammar from limited and noisy input. The central debate is between nativist positions, which argue that some grammatical knowledge must be innate because the input underdetermines it (the "poverty of the stimulus"), and empiricist or usage-based positions, which argue that general learning mechanisms can extract grammar from the statistics of speech. Computational models make this debate testable by asking what a learner with given biases can acquire from realistic data.

What distinguishes this topic from its neighbours is its object: human behaviour and human learning. [Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) is about building models; here, models are instruments or subjects of experiments. [Interpretability](https://md.research.elamur.ai/free/en/topics/interpretability/) looks inside networks for engineering reasons; here, the question is whether what is inside resembles human processing.

## History and milestones

The experimental study of sentence processing grew in the 1970s and 1980s around garden-path sentences ("The horse raced past the barn fell"), where readers commit to the wrong structure and must reanalyse. Serial parsing models explained such effects with special processing principles. Constraint-based views of the 1990s countered that frequency and context guide comprehension continuously.

Landmark papers on this topic: the 12 most cited, in time The 12 most cited papers on this topic span 2001–2023; the dot grows with the citations.

**The numbers**

| Paper | Citations |
| --- | --- |
| [2001 — A Probabilistic Earley Parser as a Psycholinguistic Model](https://md.research.elamur.ai/free/en/papers/hale-2001-probabilistic-earley-parser-as-psycholinguistic/) | 1,448 |
| [2012 — On Improving the Accuracy of Readability Classification using Insights from Second Language Acquisition](https://md.research.elamur.ai/free/en/papers/vajjala-2012-improving-accuracy-readability-classification-insights/) | 311 |
| [2018 — Predictive power of word surprisal for reading times is a linear function of language model quality](https://md.research.elamur.ai/free/en/papers/goodkind-2018-predictive-power-word-surprisal-reading/) | 294 |
| [2018 — RNNs as psycholinguistic subjects: Syntactic state and grammatical dependency](https://md.research.elamur.ai/free/en/papers/futrell-2018-rnns-psycholinguistic-subjects-syntactic-state/) | 75 |
| [2019 — Neural language models as psycholinguistic subjects: Representations of syntactic state](https://md.research.elamur.ai/free/en/papers/futrell-2019-neural-language-models-as-psycholinguistic/) | 291 |
| [2020 — What BERT Is Not: Lessons from a New Suite of Psycholinguistic Diagnostics for Language Models](https://md.research.elamur.ai/free/en/papers/ettinger-2020-what-bert-is-not-lessons/) | 843 |
| [2020 — SyntaxGym: An Online Platform for Targeted Evaluation of Language Models](https://md.research.elamur.ai/free/en/papers/gauthier-2020-syntaxgym-online-platform-targeted-evaluation/) | 145 |
| [2021 — BabyBERTa: Learning More Grammar With Small-Scale Child-Directed Language](https://md.research.elamur.ai/free/en/papers/huebner-2021-babyberta-learning-more-grammar-small/) | 91 |
| [2022 — What Artificial Neural Networks Can Tell Us About Human Language Acquisition](https://md.research.elamur.ai/free/en/papers/warstadt-2022-artificial-neural-networks-tell-us/) | 72 |
| [2023 — Why Does Surprisal From Larger Transformer-Based Language Models Provide a Poorer Fit to Human Reading Times?](https://md.research.elamur.ai/free/en/papers/oh-2023-why-does-surprisal-larger-transformer/) | 211 |
| [2023 — Testing the Predictions of Surprisal Theory in 11 Languages](https://md.research.elamur.ai/free/en/papers/wilcox-2023-testing-predictions-surprisal-theory-11/) | 158 |
| [2023 — Language acquisition: do children and language models follow similar learning stages?](https://md.research.elamur.ai/free/en/papers/evanson-2023-language-acquisition-children-language-models/) | 33 |

[A Probabilistic Earley Parser as a Psycholinguistic Model](https://md.research.elamur.ai/free/en/papers/hale-2001-probabilistic-earley-parser-as-psycholinguistic/) (2001) gave this view a precise form. Hale used a probabilistic context-free grammar and an incremental parser to compute, for every word, how much probability mass of possible analyses it eliminates, and proposed that this quantity predicts reading difficulty. Garden paths and the greater difficulty of object relative clauses fell out of corpus frequencies, with no special machinery. Roger Levy's later formulation (2008) connected surprisal to the change in the reader's beliefs about the sentence, and surprisal theory became the dominant expectation-based account.

Through the 2000s and early 2010s, the resources that made large-scale testing possible appeared: eye-tracking corpora of natural text such as the Dundee Corpus, and later the Natural Stories self-paced reading corpus, the Provo corpus and the multilingual MECO. On the acquisition side, the CHILDES database, assembled by Brian MacWhinney from the 1980s, provided transcripts of child–caregiver conversation in many languages. Computational acquisition models such as [A Probabilistic Model of Syntactic and Semantic Acquisition from Child-Directed Utterances and their Meanings](https://md.research.elamur.ai/free/en/papers/kwiatkowski-2012-probabilistic-model-syntactic-semantic-acquisition/) (2012) learned grammar and meaning jointly from such data.

Papers on this topic by year 62 dated papers from 2001 to 2026; the busiest year is 2026 with 23, and 90% of them appeared in 2018 or later.

**The numbers**

| Year | Papers |
| --- | --- |
| 2026 | 23 |
| 2025 | 10 |
| 2024 | 5 |
| 2023 | 8 |
| 2022 | 2 |
| 2021 | 2 |
| 2020 | 3 |
| 2019 | 1 |
| 2018 | 2 |
| 2017 | 1 |
| 2016 | 1 |
| 2012 | 2 |
| 2011 | 1 |
| 2001 | 1 |

From about 2016 the field turned to neural networks. Two programmes grew together. One measured surprisal with neural language models and asked how well it predicts reading times; [Predictive power of word surprisal for reading times is a linear function of language model quality](https://md.research.elamur.ai/free/en/papers/goodkind-2018-predictive-power-word-surprisal-reading/) (2018) found that better models gave better fits. The other treated networks as experimental subjects, as in [RNNs as psycholinguistic subjects: Syntactic state and grammatical dependency](https://md.research.elamur.ai/free/en/papers/futrell-2018-rnns-psycholinguistic-subjects-syntactic-state/) and [What BERT Is Not: Lessons from a New Suite of Psycholinguistic Diagnostics for Language Models](https://md.research.elamur.ai/free/en/papers/ettinger-2020-what-bert-is-not-lessons/).

The early 2020s brought two surprises. First, very large language models turned out to predict reading times worse than medium-sized ones, the "inverse scaling" result of [Why Does Surprisal From Larger Transformer-Based Language Models Provide a Poorer Fit to Human Reading Times?](https://md.research.elamur.ai/free/en/papers/oh-2023-why-does-surprisal-larger-transformer/). Second, interest in data efficiency led to the BabyLM Challenge (first held in 2023), which asks what a model can learn from roughly the number of words a child hears. Since then the papers here have multiplied, most of them on BabyLM-style learning, on the measurement of surprisal itself, and on attention as a model of memory.

## Main problems and tasks

**Predicting reading times.** Given a text and human reading measures, which predictors (surprisal, entropy, frequency, length, memory costs) explain the variation, and with what functional form? This is evaluated by the gain in regression fit over a baseline.

**Explaining specific effects.** Garden paths, subject versus object relative clauses, agreement attraction, filler–gap dependencies and island constraints are classic phenomena. A model must not only predict that a difficulty exists but match its size.

**Targeted evaluation of models.** Minimal pairs and controlled item sets test whether a model has learned a specific grammatical generalisation, independent of overall perplexity.

**Acquisition from realistic input.** What can a learner acquire from child-directed speech of realistic size, and which biases (architecture, memory limits, curriculum, grounding) are needed?

## Methods and schools

### Surprisal and its measurement

The working method of much of the field is to compute word-by-word surprisal from a language model and enter it into a mixed-effects regression of reading times along with frequency, length and spillover terms. The cross-linguistic test [Testing the Predictions of Surprisal Theory in 11 Languages](https://md.research.elamur.ai/free/en/papers/wilcox-2023-testing-predictions-surprisal-theory-11/) found that surprisal helps in every language studied, that contextual entropy often adds more, and that the relation looks linear.

Much recent work examines the instrument itself. Tokenisation is one issue: language models assign probabilities to subword tokens, while experiments care about words or character regions. [On the Proper Treatment of Tokenization in Psycholinguistics](https://md.research.elamur.ai/free/en/papers/giulianelli-2024-proper-treatment-tokenization-psycholinguistics/) proposes converting models to character level first, and [The Impact of Token Granularity on the Predictive Power of Language Model Surprisal](https://md.research.elamur.ai/free/en/papers/oh-2025-impact-token-granularity-predictive-power/) shows that the size of the vocabulary changes the fit substantially. A second issue is the source of probabilities: [Clozing the Gap: Exploring Why Language Model Surprisal Outperforms Cloze Surprisal](https://md.research.elamur.ai/free/en/papers/nair-2026-clozing-gap-exploring-why-language/) asks why model probabilities beat probabilities estimated from human completions. A third is what else a model knows: [Probing for Reading Times](https://md.research.elamur.ai/free/en/papers/tsipidi-2026-probing-reading-times/) reads reading times directly from hidden states and finds early layers useful for early eye-movement measures.

### The inverse scaling debate

The finding that larger models fit reading times worse split the field. One explanation is that large models predict rare content words and named entities too well, beyond what any human reader could. Another is that word frequency confounds surprisal: [The Frequency Confound in Language-Model Surprisal and Metaphor Novelty](https://md.research.elamur.ai/free/en/papers/momen-2026-frequency-confound-language-model-surprisal/) argues that small or early-training models look "optimal" because their surprisal is close to plain frequency. A third line attempts to rescue surprisal for hard cases. Neural surprisal underestimates garden-path slowdowns by a large factor, and [An Existence Proof for Neural Language Models That Can Explain Garden-Path Effects via Surprisal](https://md.research.elamur.ai/free/en/papers/yoshida-2026-existence-proof-neural-language-models/) shows that fine-tuning on reading times can close the gap, while warning that such flexibility makes the theory harder to falsify.

### Memory and attention

Memory-based theories have found a new formal vehicle in transformer attention, which retrieves earlier material by similarity of keys and queries much as cue-based retrieval theories propose. [Transformer Attention as a Unified Model of Encoding and Retrieval in Human Sentence Processing](https://md.research.elamur.ai/free/en/papers/parker-2026-transformer-attention-as-unified-model/) argues that attention captures interference at encoding as well as at retrieval. [Aggregated Transformer Attention Measures Predict Reading Times Beyond Surprisal](https://md.research.elamur.ai/free/en/papers/mielczarek-2026-aggregated-transformer-attention-measures-predict/) derives memory costs from attention and finds that only some ways of aggregating them help. [Probing the Attention Representation of Filler-Gap Dependency in Transformers](https://md.research.elamur.ai/free/en/papers/yao-2026-probing-attention-representation-filler-gap/) locates heads that track filler–gap dependencies in GPT-2.

### Models as psycholinguistic subjects

Following the classic experimental design, researchers write controlled items that differ in one property and compare the model's surprisal across conditions. [Neural language models as psycholinguistic subjects: Representations of syntactic state](https://md.research.elamur.ai/free/en/papers/futrell-2019-neural-language-models-as-psycholinguistic/) used this to show that recurrent models track open clauses and garden-path ambiguities, but with effects that shrink where human effects grow. [SyntaxGym: An Online Platform for Targeted Evaluation of Language Models](https://md.research.elamur.ai/free/en/papers/gauthier-2020-syntaxgym-online-platform-targeted-evaluation/) turned the method into shared infrastructure. Related work moves to meaning and pragmatics, for example [Expectations over Unspoken Alternatives Predict Pragmatic Inferences](https://md.research.elamur.ai/free/en/papers/hu-2023-expectations-over-unspoken-alternatives-predict/) on scalar implicature, and to neurons, as in [Unveiling Language Competence Neurons: A Psycholinguistic Approach to Model Interpretability](https://md.research.elamur.ai/free/en/papers/duan-2025-unveiling-language-competence-neurons-psycholinguistic/).

### Symbolic and Bayesian acquisition models

Before neural networks dominated, acquisition models were typically probabilistic grammars with explicit meanings. The Edinburgh line of work learns a combinatory categorial grammar lexicon from utterances paired with candidate logical forms, and reproduces both one-shot learning of new words and sudden shifts to the correct word order. [A Language-agnostic Model of Child Language Acquisition](https://md.research.elamur.ai/free/en/papers/langstream-2024-mahon-language-agnostic-model-child-language/) carries the approach to Hebrew. Unsupervised grammar induction, as in [The Importance of Category Labels in Grammar Induction with Child-directed Utterances](https://md.research.elamur.ai/free/en/papers/jin-2020-importance-category-labels-grammar-induction/), asks what structure can be recovered from caregiver speech with no annotation at all.

### Neural acquisition and the BabyLM programme

The neural programme trains small models from scratch on child-scale data and asks what they acquire. [What Artificial Neural Networks Can Tell Us About Human Language Acquisition](https://md.research.elamur.ai/free/en/papers/warstadt-2022-artificial-neural-networks-tell-us/) sets out the logic: a model stripped of a hypothesised advantage that still learns a phenomenon shows the phenomenon is learnable without that advantage, so positive results are more informative than negative ones. [BabyBERTa: Learning More Grammar With Small-Scale Child-Directed Language](https://md.research.elamur.ai/free/en/papers/huebner-2021-babyberta-learning-more-grammar-small/) showed that a small model trained on a few million words of child-directed speech can approach a much larger model on grammatical minimal pairs.

Whether child-directed speech itself is special is contested. Later studies such as [Is Child-Directed Speech Effective Training Data for Language Models?](https://md.research.elamur.ai/free/en/papers/feng-2024-is-child-directed-speech-effective/) and [Child-Directed Language Does Not Consistently Boost Syntax Learning in Language Models](https://md.research.elamur.ai/free/en/papers/padovani-2025-child-directed-language-does-not/) found little consistent advantage on comprehension benchmarks, while [Child-directed speech facilitates production, not comprehension, in BabyLMs](https://md.research.elamur.ai/free/en/papers/bunzeck-2026-child-directed-speech-facilitates-production/) argues that the benefit appears when models are asked to produce. Curricula modelled on development, by sentence complexity or by the order in which parts of speech emerge, have mostly failed to beat shuffled data, as in [BabyLM Challenge: Curriculum learning based on sentence complexity approximating language acquisition](https://md.research.elamur.ai/free/en/papers/oba-2023-babylm-challenge-curriculum-learning-based/). Biases modelled on memory have done better: [Developmentally-plausible Working Memory Shapes a Critical Period for Language Acquisition](https://md.research.elamur.ai/free/en/papers/mita-2025-developmentally-plausible-working-memory-shapes/) and [Human-like Fleeting Memory Improves Language Learning but Impairs Reading Time Prediction in Transformer Language Models](https://md.research.elamur.ai/free/en/papers/thamma-2026-human-like-fleeting-memory-improves/) both echo Elman's old "starting small" idea.

The poverty-of-the-stimulus debate is addressed head-on in [How poor is the stimulus? Evaluating hierarchical generalization in neural networks trained on child-directed speech](https://md.research.elamur.ai/free/en/papers/yedetore-2023-how-poor-is-stimulus-evaluating/), where networks trained on CHILDES chose a linear rather than a hierarchical rule for English questions.

### Grounded and social learning

Children learn words in scenes and in conversation. [Multimodal Semantic Learning from Child-Directed Input](https://md.research.elamur.ai/free/en/papers/lazaridou-2016-multimodal-semantic-learning-child-directed/) learned word meanings from real images of the objects present, with attention and a social cue. [Computational Language Acquisition with Theory of Mind](https://md.research.elamur.ai/free/en/papers/liu-2023-computational-language-acquisition-theory-mind/) gave a speaker agent an internal model of its listener in a reference game. [Pragmatic Modelling in Language Learning: Caregiver Question-Answer Feedback in Child-Directed Dialogue](https://md.research.elamur.ai/free/en/papers/bala-2026-pragmatic-modelling-language-learning-caregiver/) catalogues how caregivers respond to children's questions as development proceeds.

## Datasets, resources and benchmarks

| Resource | Kind |
| --- | --- |
| CHILDES | transcripts of child–adult conversation in many languages |
| Dundee Corpus | eye-tracking of English newspaper editorials |
| Natural Stories | self-paced reading of constructed narratives |
| Provo, MECO | eye-tracking with cloze norms; multilingual eye-tracking |
| BLiMP, Zorro | grammatical minimal pairs; Zorro uses child-accessible vocabulary |
| BabyLM | shared task with 10M and 100M word training budgets |

Several resources were built by papers here. [UD-English-CHILDES: A Collected Resource of Gold and Silver Universal Dependencies Trees for Child Language Interactions](https://md.research.elamur.ai/free/en/papers/ud-2025-yang-ud-english-childes-collected-resource/) is the first official Universal Dependencies treebank of child–adult speech. [Universal Dependencies and Semantics for English and Hebrew Child-directed Speech](https://md.research.elamur.ai/free/en/papers/ud-2022-szubert-universal-dependencies-semantics-english-hebrew/) and [Cross-linguistically Consistent Semantic and Syntactic Annotation of Child-directed Speech](https://md.research.elamur.ai/free/en/papers/langstream-2021-szubert-cross-linguistically-consistent-semantic-syntactic/) add parallel syntax and logical forms for English and Hebrew. [From CHAT to Coded CoNLL-U: A Reproducible Pipeline for the Syntactic Annotation and Querying of Child Language Data](https://md.research.elamur.ai/free/en/papers/bank-2026-stein-chat-coded-conll-u-reproducible/) parses and queries CHILDES transcripts, and [What Exactly do Children Receive in Language Acquisition? A Case Study on CHILDES with Automated Detection of Filler-Gap Dependencies](https://md.research.elamur.ai/free/en/papers/zhou-2026-what-exactly-do-children-receive/) counts filler–gap constructions across dozens of corpora. For other languages, [TafBERTa: Learning Grammatical Rules from Small-Scale Language Acquisition Data in Hebrew](https://md.research.elamur.ai/free/en/papers/babylm-2025-gelboim-tafberta-learning-grammatical-rules-small/) builds a Hebrew child-speech corpus with agreement minimal pairs, and [Recurrent Networks Are (Linguistically) Better? An (Ongoing) Experiment on Small-LM Training on Child-Directed Speech in Italian](https://md.research.elamur.ai/free/en/papers/fusco-2024-recurrent-networks-are-linguistically-better/) does the same for Italian. On the processing side, [A Corpus of Joint EEG and Self-Paced Reading of Natural Dutch Texts](https://md.research.elamur.ai/free/en/papers/ostergaard-2026-corpus-joint-eeg-self-paced/) records brain and behaviour simultaneously over whole texts.

## Landmark papers in this library

### A Probabilistic Earley Parser as a Psycholinguistic Model (2001)

[Hale's paper](https://md.research.elamur.ai/free/en/papers/hale-2001-probabilistic-earley-parser-as-psycholinguistic/) defined word-by-word processing difficulty as surprisal computed from prefix probabilities of a probabilistic grammar. With rule probabilities taken from treebank counts, the parser predicted the garden-path effect in reduced relatives and the subject–object relative asymmetry. It is the founding text of surprisal theory and by far the most cited paper here.

### What BERT Is Not (2020)

[What BERT Is Not: Lessons from a New Suite of Psycholinguistic Diagnostics for Language Models](https://md.research.elamur.ai/free/en/papers/ettinger-2020-what-bert-is-not-lessons/) adapted stimuli from N400 experiments to test BERT on commonsense inference, event roles and negation. BERT handled category membership well but largely ignored negation, which became a much-repeated example of what next-word prediction alone may not teach.

### On readability and second-language measures (2012)

[On Improving the Accuracy of Readability Classification using Insights from Second Language Acquisition](https://md.research.elamur.ai/free/en/papers/vajjala-2012-improving-accuracy-readability-classification-insights/) used lexical and syntactic complexity measures from learner-language research as features for grading texts, and built the WeeBit corpus. It remains a standard reference for readability assessment.

### Surprisal and language model quality (2018)

[Predictive power of word surprisal for reading times is a linear function of language model quality](https://md.research.elamur.ai/free/en/papers/goodkind-2018-predictive-power-word-surprisal-reading/) compared n-gram and LSTM models on the Dundee eye-tracking data and found that lower perplexity gave steadily better fits while the effect of surprisal stayed roughly linear. The result justified using ever better models, until the inverse scaling findings complicated it.

### Neural language models as psycholinguistic subjects (2018–2019)

The two papers by Futrell, Wilcox, Morita, Levy and colleagues ([2018](https://md.research.elamur.ai/free/en/papers/futrell-2018-rnns-psycholinguistic-subjects-syntactic-state/), [2019](https://md.research.elamur.ai/free/en/papers/futrell-2019-neural-language-models-as-psycholinguistic/)) established the experimental paradigm for testing networks with controlled items. Recurrent models tracked syntactic state and showed garden-path effects, but failed on reflexive binding and negative polarity licensing and showed garden-path effects that did not scale like human ones.

### Inverse scaling of surprisal (2023)

[Why Does Surprisal From Larger Transformer-Based Language Models Provide a Poorer Fit to Human Reading Times?](https://md.research.elamur.ai/free/en/papers/oh-2023-why-does-surprisal-larger-transformer/) tested many model sizes across three families and found that fit to self-paced reading and eye-tracking data falls as models grow. Residual analysis traced the problem to very low surprisal on nouns, adjectives and named entities. The paper reset the expectation that better language models are better cognitive models.

### Surprisal in eleven languages (2023)

[Testing the Predictions of Surprisal Theory in 11 Languages](https://md.research.elamur.ai/free/en/papers/wilcox-2023-testing-predictions-surprisal-theory-11/) used the MECO eye-tracking data across five language families and confirmed three predictions of the theory in all or most languages. It is the broadest cross-linguistic test of surprisal to date.

### BabyBERTa (2021)

[BabyBERTa](https://md.research.elamur.ai/free/en/papers/huebner-2021-babyberta-learning-more-grammar-small/) trained a small RoBERTa on about five million words of American child-directed speech and evaluated it on a new set of minimal pairs restricted to child vocabulary. Its near parity with RoBERTa-base, at a tiny fraction of the data, helped launch the BabyLM movement.

### How poor is the stimulus? (2023)

[This study](https://md.research.elamur.ai/free/en/papers/yedetore-2023-how-poor-is-stimulus-evaluating/) put the textbook argument for an innate hierarchical bias, English question formation, to networks trained on CHILDES. They modelled the corpus well yet generalised by linear order, a result widely cited in the nativism debate.

### Children and models in stages (2023)

[Language acquisition: do children and language models follow similar learning stages?](https://md.research.elamur.ai/free/en/papers/evanson-2023-language-acquisition-children-language-models/) tracked many GPT-2 runs through training and found a consistent order of acquisition across seeds, partly matching stages in Hebrew-speaking children, with learning of "late" skills starting early but proceeding slowly.

Citations of the papers on this topic, by when they were published Papers published in 2020–2024 gather 40% of the 4,235 citations of the papers on this topic.

**The numbers**

| Published | Papers and their citations |
| --- | --- |
| 2025–2029 | 33 papers, 84 citations |
| 2020–2024 | 20 papers, 1,691 citations |
| 2015–2019 | 5 papers, 671 citations |
| 2010–2014 | 3 papers, 341 citations |
| 2000–2004 | 1 papers, 1,448 citations |

## People

[Roger Levy](https://md.research.elamur.ai/free/en/authors/roger-levy/) of the Massachusetts Institute of Technology is the most prolific author here and a central figure of surprisal theory. His papers here include the psycholinguistic-subject studies, SyntaxGym, the eleven-language test and the work on scalar inference.

[Ethan Wilcox](https://md.research.elamur.ai/free/en/authors/ethan-gotlieb-wilcox/), now at Georgetown University, co-authored the subject-style studies and SyntaxGym and led the cross-linguistic test of surprisal theory.

[Richard Futrell](https://md.research.elamur.ai/free/en/authors/richard-futrell/) of the University of California, Irvine, led the two papers that treat recurrent networks as psycholinguistic subjects, with [Takashi Morita](https://md.research.elamur.ai/free/en/authors/takashi-morita/) and [Peng Qian](https://md.research.elamur.ai/free/en/authors/peng-qian/).

[John Hale](https://md.research.elamur.ai/free/en/authors/john-hale/), now at Johns Hopkins University, wrote the 2001 paper that introduced surprisal as a measure of processing difficulty.

[Allyson Ettinger](https://md.research.elamur.ai/free/en/authors/allyson-ettinger/), now at the University of Washington, brought N400-based diagnostics to BERT.

[Byung-Doh Oh](https://md.research.elamur.ai/free/en/authors/byung-doh-oh/), now at Nanyang Technological University, and [William Schuler](https://md.research.elamur.ai/free/en/authors/william-schuler/) of The Ohio State University documented inverse scaling and studied token granularity; Schuler also worked on grammar induction from child-directed speech with [Lifeng Jin](https://md.research.elamur.ai/free/en/authors/lifeng-jin/).

[Ryan Cotterell](https://md.research.elamur.ai/free/en/authors/ryan-cotterell/) of ETH Zurich, with [Mario Giulianelli](https://md.research.elamur.ai/free/en/authors/mario-giulianelli/), [Tiago Pimentel](https://md.research.elamur.ai/free/en/authors/tiago-pimentel/) and [Clara Meister](https://md.research.elamur.ai/free/en/authors/clara-meister/), brought formal rigour to the measurement of surprisal: tokenisation, entropy and probing of hidden states.

[Mark Steedman](https://md.research.elamur.ai/free/en/authors/mark-steedman/) and [Sharon Goldwater](https://md.research.elamur.ai/free/en/authors/sharon-goldwater/) of the University of Edinburgh lead the line of CCG-based acquisition models and, with [Omri Abend](https://md.research.elamur.ai/free/en/authors/omri-abend/) of the Hebrew University of Jerusalem, [Nathan Schneider](https://md.research.elamur.ai/free/en/authors/nathan-schneider/), [Ida Szubert](https://md.research.elamur.ai/free/en/authors/ida-szubert/) and [Louis Mahon](https://md.research.elamur.ai/free/en/authors/louis-mahon/), the annotation of English and Hebrew child-directed speech.

[Yohei Oseki](https://md.research.elamur.ai/free/en/authors/yohei-oseki/) of the University of Tokyo, with [Ryo Yoshida](https://md.research.elamur.ai/free/en/authors/ryo-yoshida/) and [Tatsuki Kuribayashi](https://md.research.elamur.ai/free/en/authors/tatsuki-kuribayashi/), works on cognitively motivated biases for small models and on garden-path effects.

[Jennifer Hu](https://md.research.elamur.ai/free/en/authors/jennifer-hu/) contributed to SyntaxGym and studied pragmatic inference with language-model expectations.

[R. Thomas McCoy](https://md.research.elamur.ai/free/en/authors/r-thomas-mccoy/) and [Robert Frank](https://md.research.elamur.ai/free/en/authors/robert-frank/) of Yale University, with [Tal Linzen](https://md.research.elamur.ai/free/en/authors/tal-linzen/), tested hierarchical generalisation from child-directed speech and analysed filler–gap input.

[Elior Sulem](https://md.research.elamur.ai/free/en/authors/elior-sulem/), now at Ben-Gurion University of the Negev, co-authored BabyBERTa and its Hebrew counterpart TafBERTa.

[Sina Zarrieß](https://md.research.elamur.ai/free/en/authors/sina-zarries/) of Bielefeld University, with [Omar Momen](https://md.research.elamur.ai/free/en/authors/omar-momen/) and [Bastian Bunzeck](https://md.research.elamur.ai/free/en/authors/bastian-bunzeck/), studied metaphor novelty, frequency confounds and the role of child-directed speech in production.

Authors with the most papers on this topic Roger Levy leads with 5 papers; the 12 authors shown wrote 39 of the papers on this topic between them, counting each paper once per author.

**The numbers**

| Author | Papers here · their citations |
| --- | --- |
| [Roger Levy](https://md.research.elamur.ai/free/en/authors/roger-levy/) | 5 papers · 696 citations |
| [Ethan Wilcox](https://md.research.elamur.ai/free/en/authors/ethan-gotlieb-wilcox/) | 4 papers · 669 citations |
| [Mark Steedman](https://md.research.elamur.ai/free/en/authors/mark-steedman/) | 4 papers · 33 citations |
| [Byung-Doh Oh](https://md.research.elamur.ai/free/en/authors/byung-doh-oh/) | 3 papers · 224 citations |
| [William Schuler](https://md.research.elamur.ai/free/en/authors/william-schuler/) | 3 papers · 224 citations |
| [Ryan Cotterell](https://md.research.elamur.ai/free/en/authors/ryan-cotterell/) | 3 papers · 182 citations |
| [Sharon Goldwater](https://md.research.elamur.ai/free/en/authors/sharon-goldwater/) | 3 papers · 30 citations |
| [Yohei Oseki](https://md.research.elamur.ai/free/en/authors/yohei-oseki/) | 3 papers · 19 citations |
| [Nathan Schneider](https://md.research.elamur.ai/free/en/authors/nathan-schneider/) | 3 papers · 7 citations |
| [Omri Abend](https://md.research.elamur.ai/free/en/authors/omri-abend/) | 3 papers · 6 citations |
| [Sina Zarrieß](https://md.research.elamur.ai/free/en/authors/sina-zarries/) | 3 papers · 1 citation |
| [Peng Qian](https://md.research.elamur.ai/free/en/authors/peng-qian/) | 2 papers · 436 citations |

## Institutions and places

The United States leads by authorship, with [Georgetown University](https://md.research.elamur.ai/free/en/universities/georgetown-university/), the [University of California San Diego](https://md.research.elamur.ai/free/en/universities/university-of-california-san-diego/), the [Massachusetts Institute of Technology](https://md.research.elamur.ai/free/en/universities/massachusetts-institute-of-technology/), [The Ohio State University](https://md.research.elamur.ai/free/en/universities/the-ohio-state-university/), [Purdue University West Lafayette](https://md.research.elamur.ai/free/en/universities/purdue-university-west-lafayette/) and the [University of Colorado Boulder](https://md.research.elamur.ai/free/en/universities/university-of-colorado-boulder/) among the current homes of authors here. Germany is second, with the [University of Tübingen](https://md.research.elamur.ai/free/en/universities/university-of-tubingen/) and [Bielefeld University](https://md.research.elamur.ai/free/en/universities/bielefeld-university/). The [University of Edinburgh](https://md.research.elamur.ai/free/en/universities/university-of-edinburgh/) anchors British work on acquisition, while [ETH Zurich](https://md.research.elamur.ai/free/en/universities/eth-zurich/), the [University of Copenhagen](https://md.research.elamur.ai/free/en/universities/university-of-copenhagen/) and the [University of Amsterdam](https://md.research.elamur.ai/free/en/universities/university-of-amsterdam/) contribute to the measurement of surprisal.

Japan has a distinct cluster: [The University of Tokyo](https://md.research.elamur.ai/free/en/universities/the-university-of-tokyo/), the [National Institute for Japanese Language and Linguistics](https://md.research.elamur.ai/free/en/labs/national-institute-for-japanese-language-and-linguistics/), the [National Institute of Informatics](https://md.research.elamur.ai/free/en/labs/national-institute-of-informatics/) and [Tohoku University](https://md.research.elamur.ai/free/en/universities/tohoku-university/). Israel is represented by the [Hebrew University of Jerusalem](https://md.research.elamur.ai/free/en/universities/hebrew-university-of-jerusalem/), which links the topic to Hebrew child-language data; Asia also by the [Hong Kong University of Science and Technology](https://md.research.elamur.ai/free/en/universities/hong-kong-university-of-science-and-technology/) and [Nanyang Technological University](https://md.research.elamur.ai/free/en/universities/nanyang-technological-university/).

Where the authors on this topic work now Georgetown University is the current home of authors of 7 papers on this topic; counted are the authors with a page here, by their latest institution.

**The numbers**

| Institution | Papers |
| --- | --- |
| [Georgetown University](https://md.research.elamur.ai/free/en/universities/georgetown-university/) | 7 papers |
| [University of California San Diego](https://md.research.elamur.ai/free/en/universities/university-of-california-san-diego/) | 6 papers |
| [Cape Eleuthera Institute](https://ror.org/00fs71k85) | 5 papers |
| [Massachusetts Institute of Technology](https://md.research.elamur.ai/free/en/universities/massachusetts-institute-of-technology/) | 5 papers |
| [ETH Zurich](https://md.research.elamur.ai/free/en/universities/eth-zurich/) | 4 papers |
| [Hong Kong University of Science and Technology](https://md.research.elamur.ai/free/en/universities/hong-kong-university-of-science-and-technology/) | 4 papers |
| [National Institute for Japanese Language and Linguistics](https://md.research.elamur.ai/free/en/labs/national-institute-for-japanese-language-and-linguistics/) | 4 papers |
| [The University of Tokyo](https://md.research.elamur.ai/free/en/universities/the-university-of-tokyo/) | 4 papers |
| [University of Edinburgh](https://md.research.elamur.ai/free/en/universities/university-of-edinburgh/) | 4 papers |
| [University of Tübingen](https://md.research.elamur.ai/free/en/universities/university-of-tubingen/) | 4 papers |

Countries of the authors on this topic United States comes first: authors working there now wrote 26 papers on this topic.

**The numbers**

| Country | Papers |
| --- | --- |
| United States | 26 papers |
| Germany | 12 papers |
| United Kingdom | 8 papers |
| Japan | 7 papers |
| Netherlands | 6 papers |
| Bahamas | 5 papers |
| China | 5 papers |
| Hong Kong | 5 papers |
| Israel | 5 papers |
| Switzerland | 4 papers |

Transactions of the ACL carries several central papers; the CMCL workshop specialises in sentence processing, BabyLM and CoNLL host acquisition modelling, and LREC the resource papers.

Where the papers on this topic were published Proceedings of the Fifteenth Language Resources and Evaluation Conference carries the most, 7 of 62.

**The numbers**

| Venue | Papers |
| --- | --- |
| Proceedings of the Fifteenth Language Resources and Evaluation Conference | 7 papers |
| Proceedings of the 15th Workshop on Cognitive Modeling and Computational Linguistics | 5 papers |
| Transactions of the Association for Computational Linguistics | 5 papers |
| Proceedings of the 30th Conference on Computational Natural Language Learning | 3 papers |
| Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) | 3 papers |
| Proceedings of the First BabyLM Workshop | 3 papers |
| Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing | 2 papers |
| Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) | 2 papers |
| Proceedings of the 12th International Conference on Computational Semantics (IWCS) –- Short papers | 1 paper |
| Proceedings of the 13th Conference of the European Chapter of the Association for Computational Linguistics | 1 paper |

## Open questions

**What is the right language model for a cognitive model?** Neither the biggest nor the smallest model fits human data best, and fit depends on tokenisation, training stage and frequency. Whether there is a principled way to choose the probability model, and what constraints keep surprisal theory falsifiable, are open.

**Is surprisal enough?** Garden-path effects are far larger in humans than standard model surprisal predicts, and relative clause asymmetries resist expectation-based explanation. Whether memory costs must be added, and whether attention offers the right formalism for them, is under active study.

**What is innate?** Small models trained on child-scale text often fail on hierarchical generalisations that children master. Whether this reflects missing biases, missing modalities (vision, prosody, interaction) or simply too little data is unsettled.

**Is child-directed speech special?** Results differ by benchmark, by language and by whether comprehension or production is measured. Frequency and register differences between corpora make fair comparisons difficult.

**Beyond English.** Most reading-time data and most acquisition models concern English. The eleven-language test and the Hebrew, Italian and Dutch resources here are a start; many language families remain untested.

**From prediction to understanding.** Models can match reading times without processing language as people do. Position papers such as [Toward Cognitive Alignment in Large Language Models: Integrating Linguistic Theory and Human Data](https://md.research.elamur.ai/free/en/papers/zaghouani-2026-cognitive-alignment-large-language-models/) call for agreement at the levels of behaviour, representation and development together.

## Neighbouring topics

The closest neighbour is [Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/). [Universal Dependencies](https://md.research.elamur.ai/free/en/topics/universal-dependencies/) shares the child-language treebanks. [Idioms & multiword expressions](https://md.research.elamur.ai/free/en/topics/multiword-expressions/) shares work on figurative language, including [Are idioms surprising?](https://md.research.elamur.ai/free/en/papers/philipp-2023-are-idioms-surprising/) and [A Survey of Idiom Datasets for Psycholinguistic and Computational Research](https://md.research.elamur.ai/free/en/papers/flor-2025-survey-idiom-datasets-psycholinguistic-computational/). [Interpretability](https://md.research.elamur.ai/free/en/topics/interpretability/) shares the analyses of neurons and attention heads. The Hebrew acquisition work connects to [Hebrew & Aramaic](https://md.research.elamur.ai/free/en/topics/hebrew-aramaic/) and [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/). [Tokenization & segmentation](https://md.research.elamur.ai/free/en/topics/tokenization/) shares the question of how subword units affect surprisal, and [Grammatical error correction & learner language](https://md.research.elamur.ai/free/en/topics/grammatical-error-correction/) shares error detection in child speech and learner-based complexity measures.

Single papers link the topic to [Implicature](https://md.research.elamur.ai/free/en/topics/implicature/) through scalar inference, to [Coreference resolution](https://md.research.elamur.ai/free/en/topics/coreference-resolution/) through gender-inclusive language in [Adapting Psycholinguistic Research for LLMs: Gender-inclusive Language in a Coreference Context](https://md.research.elamur.ai/free/en/papers/bartl-2025-adapting-psycholinguistic-research-llms-gender/), to [Historical linguistics & language change](https://md.research.elamur.ai/free/en/topics/language-change/) through [Modeling the Memory-Surprisal Trade-Off over Time: Communicative Efficiency Decreases with Lexico-Grammatical Change in Scientific English](https://md.research.elamur.ai/free/en/papers/steuer-2026-modeling-memory-surprisal-trade-off/), to [Sentiment & emotion](https://md.research.elamur.ai/free/en/topics/sentiment-emotion/) through personality prediction, and to [Dialogue systems](https://md.research.elamur.ai/free/en/topics/dialogue-systems/), [Syntax & grammar formalisms](https://md.research.elamur.ai/free/en/topics/syntax-grammar/), [Evaluation & benchmarks](https://md.research.elamur.ai/free/en/topics/evaluation-benchmarks/) and [Vision & language](https://md.research.elamur.ai/free/en/topics/multimodal/).

Neighbouring topics: papers shared Language models & pretraining shares the most: 20 of the 62 papers here (32%) stand on both.

**The numbers**

| Topic | Papers shared |
| --- | --- |
| [Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) | 20 shared · 32% |
| [Idioms & multiword expressions](https://md.research.elamur.ai/free/en/topics/multiword-expressions/) | 4 shared · 6% |
| [Universal Dependencies](https://md.research.elamur.ai/free/en/topics/universal-dependencies/) | 4 shared · 6% |
| [Hebrew & Aramaic](https://md.research.elamur.ai/free/en/topics/hebrew-aramaic/) | 3 shared · 5% |
| [Interpretability](https://md.research.elamur.ai/free/en/topics/interpretability/) | 3 shared · 5% |
| [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) | 3 shared · 5% |
| [Grammatical error correction & learner language](https://md.research.elamur.ai/free/en/topics/grammatical-error-correction/) | 2 shared · 3% |
| [Tokenization & segmentation](https://md.research.elamur.ai/free/en/topics/tokenization/) | 2 shared · 3% |
| [Coreference resolution](https://md.research.elamur.ai/free/en/topics/coreference-resolution/) | 1 shared · 2% |
| [Dialogue systems](https://md.research.elamur.ai/free/en/topics/dialogue-systems/) | 1 shared · 2% |

By shelf, the topic stands mostly on [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/), then on [Natural language processing](https://md.research.elamur.ai/free/en/shelves/nlp/), with a few papers on [AI & machine learning](https://md.research.elamur.ai/free/en/shelves/ai/).

The shelves this topic stands on Linguistics & languages: 40 of the 62 papers here.

**The numbers**

| Shelf | Papers |
| --- | --- |
| [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/) | 40 papers |
| [Natural language processing](https://md.research.elamur.ai/free/en/shelves/nlp/) | 19 papers |
| [AI & machine learning](https://md.research.elamur.ai/free/en/shelves/ai/) | 3 papers |
| [Other](https://md.research.elamur.ai/free/en/shelves/other/) | 1 paper |

## Key works in this library

- [A Probabilistic Earley Parser as a Psycholinguistic Model](https://md.research.elamur.ai/free/en/papers/hale-2001-probabilistic-earley-parser-as-psycholinguistic/) — the origin of surprisal theory.
- [What BERT Is Not: Lessons from a New Suite of Psycholinguistic Diagnostics for Language Models](https://md.research.elamur.ai/free/en/papers/ettinger-2020-what-bert-is-not-lessons/) — N400-style diagnostics for BERT.
- [On Improving the Accuracy of Readability Classification using Insights from Second Language Acquisition](https://md.research.elamur.ai/free/en/papers/vajjala-2012-improving-accuracy-readability-classification-insights/) — learner-language measures for readability.
- [Predictive power of word surprisal for reading times is a linear function of language model quality](https://md.research.elamur.ai/free/en/papers/goodkind-2018-predictive-power-word-surprisal-reading/) — better models, better fits (for a while).
- [Neural language models as psycholinguistic subjects: Representations of syntactic state](https://md.research.elamur.ai/free/en/papers/futrell-2019-neural-language-models-as-psycholinguistic/) — testing syntactic state in networks.
- [Why Does Surprisal From Larger Transformer-Based Language Models Provide a Poorer Fit to Human Reading Times?](https://md.research.elamur.ai/free/en/papers/oh-2023-why-does-surprisal-larger-transformer/) — inverse scaling of fit to reading.
- [Testing the Predictions of Surprisal Theory in 11 Languages](https://md.research.elamur.ai/free/en/papers/wilcox-2023-testing-predictions-surprisal-theory-11/) — the cross-linguistic test.
- [SyntaxGym: An Online Platform for Targeted Evaluation of Language Models](https://md.research.elamur.ai/free/en/papers/gauthier-2020-syntaxgym-online-platform-targeted-evaluation/) — infrastructure for targeted evaluation.
- [BabyBERTa: Learning More Grammar With Small-Scale Child-Directed Language](https://md.research.elamur.ai/free/en/papers/huebner-2021-babyberta-learning-more-grammar-small/) — grammar from a child's budget of words.
- [RNNs as psycholinguistic subjects: Syntactic state and grammatical dependency](https://md.research.elamur.ai/free/en/papers/futrell-2018-rnns-psycholinguistic-subjects-syntactic-state/) — LSTMs on garden paths, reflexives and NPIs.
- [What Artificial Neural Networks Can Tell Us About Human Language Acquisition](https://md.research.elamur.ai/free/en/papers/warstadt-2022-artificial-neural-networks-tell-us/) — the logic of model-based acquisition research.
- [Language acquisition: do children and language models follow similar learning stages?](https://md.research.elamur.ai/free/en/papers/evanson-2023-language-acquisition-children-language-models/) — stages of learning in GPT-2 and children.
- [Expectations over Unspoken Alternatives Predict Pragmatic Inferences](https://md.research.elamur.ai/free/en/papers/hu-2023-expectations-over-unspoken-alternatives-predict/) — scalar inference from model expectations.
- [How poor is the stimulus? Evaluating hierarchical generalization in neural networks trained on child-directed speech](https://md.research.elamur.ai/free/en/papers/yedetore-2023-how-poor-is-stimulus-evaluating/) — question formation without a hierarchical bias.
- [A Probabilistic Model of Syntactic and Semantic Acquisition from Child-Directed Utterances and their Meanings](https://md.research.elamur.ai/free/en/papers/kwiatkowski-2012-probabilistic-model-syntactic-semantic-acquisition/) — joint learning of grammar and meaning.
- [On the Proper Treatment of Tokenization in Psycholinguistics](https://md.research.elamur.ai/free/en/papers/giulianelli-2024-proper-treatment-tokenization-psycholinguistics/) — surprisal over character regions.
- [Is Child-Directed Speech Effective Training Data for Language Models?](https://md.research.elamur.ai/free/en/papers/feng-2024-is-child-directed-speech-effective/) — testing the value of child-directed input.
- [Developmentally-plausible Working Memory Shapes a Critical Period for Language Acquisition](https://md.research.elamur.ai/free/en/papers/mita-2025-developmentally-plausible-working-memory-shapes/) — growing memory as a learning bias.
- [UD-English-CHILDES: A Collected Resource of Gold and Silver Universal Dependencies Trees for Child Language Interactions](https://md.research.elamur.ai/free/en/papers/ud-2025-yang-ud-english-childes-collected-resource/) — a UD treebank of child–adult speech.
- [An Existence Proof for Neural Language Models That Can Explain Garden-Path Effects via Surprisal](https://md.research.elamur.ai/free/en/papers/yoshida-2026-existence-proof-neural-language-models/) — closing the garden-path gap, and its cost.
- [Transformer Attention as a Unified Model of Encoding and Retrieval in Human Sentence Processing](https://md.research.elamur.ai/free/en/papers/parker-2026-transformer-attention-as-unified-model/) — attention as cue-based memory.
- [Word predictability estimates from language models are not robust to tokenizer vocabulary](https://md.research.elamur.ai/free/en/papers/nguyen-2026-word-predictability-estimates-language-models/) — the fragility of surprisal estimates.

## All 62 papers on this topic

**1** [2](https://md.research.elamur.ai/free/en/topics/psycholinguistics/page/2/) [next →](https://md.research.elamur.ai/free/en/topics/psycholinguistics/page/2/)

1. [**A Probabilistic Earley Parser as a Psycholinguistic Model**John Hale*2001 · 1 448 citations*](https://md.research.elamur.ai/free/en/papers/hale-2001-probabilistic-earley-parser-as-psycholinguistic/)
2. [**What BERT Is Not: Lessons from a New Suite of Psycholinguistic Diagnostics for Language Models**Allyson Ettinger*2020 · 843 citations*](https://md.research.elamur.ai/free/en/papers/ettinger-2020-what-bert-is-not-lessons/)
3. [**On Improving the Accuracy of Readability Classification using Insights from Second Language Acquisition**Sowmya Vajjala, Detmar Meurers*2012 · 311 citations*](https://md.research.elamur.ai/free/en/papers/vajjala-2012-improving-accuracy-readability-classification-insights/)
4. [**Predictive power of word surprisal for reading times is a linear function of language model quality**Adam Goodkind, Klinton Bicknell*2018 · 294 citations*](https://md.research.elamur.ai/free/en/papers/goodkind-2018-predictive-power-word-surprisal-reading/)
5. [**Neural language models as psycholinguistic subjects: Representations of syntactic state**Richard Futrell, Ethan Wilcox, Takashi Morita, et al.*2019 · 291 citations*](https://md.research.elamur.ai/free/en/papers/futrell-2019-neural-language-models-as-psycholinguistic/)
6. [**Why Does Surprisal From Larger Transformer-Based Language Models Provide a Poorer Fit to Human Reading Times?**Byung-Doh Oh, William Schuler*2023 · 211 citations*](https://md.research.elamur.ai/free/en/papers/oh-2023-why-does-surprisal-larger-transformer/)
7. [**Testing the Predictions of Surprisal Theory in 11 Languages**Ethan G. Wilcox, Tiago Pimentel, Clara Meister, et al.*2023 · 158 citations*](https://md.research.elamur.ai/free/en/papers/wilcox-2023-testing-predictions-surprisal-theory-11/)
8. [**SyntaxGym: An Online Platform for Targeted Evaluation of Language Models**Jon Gauthier, Jennifer Hu, Ethan Wilcox, et al.*2020 · 145 citations*](https://md.research.elamur.ai/free/en/papers/gauthier-2020-syntaxgym-online-platform-targeted-evaluation/)
9. [**BabyBERTa: Learning More Grammar With Small-Scale Child-Directed Language**Philip A. Huebner, Elior Sulem, Fisher Cynthia, et al.*2021 · 91 citations*](https://md.research.elamur.ai/free/en/papers/huebner-2021-babyberta-learning-more-grammar-small/)
10. [**RNNs as psycholinguistic subjects: Syntactic state and grammatical dependency**Richard Futrell, Ethan Wilcox, Takashi Morita, et al.*2018 · 75 citations*](https://md.research.elamur.ai/free/en/papers/futrell-2018-rnns-psycholinguistic-subjects-syntactic-state/)
11. [**What Artificial Neural Networks Can Tell Us About Human Language Acquisition**Alex Warstadt, Samuel R. Bowman*2022 · 72 citations*](https://md.research.elamur.ai/free/en/papers/warstadt-2022-artificial-neural-networks-tell-us/)
12. [**Language acquisition: do children and language models follow similar learning stages?**Linnea Evanson, Yair Lakretz, Jean-Rémi King*2023 · 33 citations*](https://md.research.elamur.ai/free/en/papers/evanson-2023-language-acquisition-children-language-models/)
13. [**Expectations over Unspoken Alternatives Predict Pragmatic Inferences**Jennifer Hu, Roger Levy, Judith Degen, et al.*2023 · 27 citations*](https://md.research.elamur.ai/free/en/papers/hu-2023-expectations-over-unspoken-alternatives-predict/)
14. [**How poor is the stimulus? Evaluating hierarchical generalization in neural networks trained on child-directed speech**Aditya Yedetore, Tal Linzen, Robert Frank, et al.*2023 · 27 citations*](https://md.research.elamur.ai/free/en/papers/yedetore-2023-how-poor-is-stimulus-evaluating/)
15. [**A Probabilistic Model of Syntactic and Semantic Acquisition from Child-Directed Utterances and their Meanings**Tom Kwiatkowski, Sharon Goldwater, Luke Zettlemoyer, et al.*2012 · 27 citations*](https://md.research.elamur.ai/free/en/papers/kwiatkowski-2012-probabilistic-model-syntactic-semantic-acquisition/)
16. [**On the Proper Treatment of Tokenization in Psycholinguistics**Mario Giulianelli, Luca Malagutti, Juan Luis Gastaldi, et al.*2024 · 23 citations*](https://md.research.elamur.ai/free/en/papers/giulianelli-2024-proper-treatment-tokenization-psycholinguistics/)
17. [**Computational Language Acquisition with Theory of Mind**Andy Liu, Hao Zhu, Emmy Liu, et al.*2023 · 21 citations*](https://md.research.elamur.ai/free/en/papers/liu-2023-computational-language-acquisition-theory-mind/)
18. [**SemEval-2026 Task 10: PsyCoMark – Psycholinguistic Conspiracy Marker Extraction and Detection**Mattia Samory, Felix Soldner, Veronika Batzdorfer*2026 · 20 citations*](https://md.research.elamur.ai/free/en/papers/samory-2026-semeval-2026-task-10-psycomark/)
19. [**Is Child-Directed Speech Effective Training Data for Language Models?**Steven Y. Feng, Noah D. Goodman, Michael C. Frank*2024 · 16 citations*](https://md.research.elamur.ai/free/en/papers/feng-2024-is-child-directed-speech-effective/)
20. [**The Impact of Token Granularity on the Predictive Power of Language Model Surprisal**Byung-Doh Oh, William Schuler*2025 · 12 citations*](https://md.research.elamur.ai/free/en/papers/oh-2025-impact-token-granularity-predictive-power/)
21. [**Child-Directed Language Does Not Consistently Boost Syntax Learning in Language Models**Francesca Padovani, Jaap Jumelet, Yevgen Matusevych, et al.*2025 · 9 citations*](https://md.research.elamur.ai/free/en/papers/padovani-2025-child-directed-language-does-not/)
22. [**Developmentally-plausible Working Memory Shapes a Critical Period for Language Acquisition**Masato Mita, Ryo Yoshida, Yohei Oseki*2025 · 9 citations*](https://md.research.elamur.ai/free/en/papers/mita-2025-developmentally-plausible-working-memory-shapes/)
23. [**BabyLM Challenge: Curriculum learning based on sentence complexity approximating language acquisition**Miyu Oba, Akari Haga, Akiyo Fukatsu, et al.*2023 · 9 citations*](https://md.research.elamur.ai/free/en/papers/oba-2023-babylm-challenge-curriculum-learning-based/)
24. [**Multimodal Semantic Learning from Child-Directed Input**Angeliki Lazaridou, Grzegorz Chrupała, Raquel Fernández, et al.*2016 · 9 citations*](https://md.research.elamur.ai/free/en/papers/lazaridou-2016-multimodal-semantic-learning-child-directed/)
25. [**Unveiling Language Competence Neurons: A Psycholinguistic Approach to Model Interpretability**Xufeng Duan, Xinyu Zhou, Bei Xiao, et al.*2025 · 8 citations*](https://md.research.elamur.ai/free/en/papers/duan-2025-unveiling-language-competence-neurons-psycholinguistic/)

**1** [2](https://md.research.elamur.ai/free/en/topics/psycholinguistics/page/2/) [next →](https://md.research.elamur.ai/free/en/topics/psycholinguistics/page/2/)

- [← previous topic Part-of-speech tagging](https://md.research.elamur.ai/free/en/topics/part-of-speech-tagging/)
- [next topic → Question answering](https://md.research.elamur.ai/free/en/topics/question-answering/)
- [↑ back to topics](https://md.research.elamur.ai/free/en/topics/)