---
title: "Psycholinguistics — definition and key papers"
description: "Psycholinguistics is the study of the mental processes by which people produce, comprehend and acquire language."
canonical: "https://research.elamur.ai/free/en/concepts/psycholinguistics/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/concepts/psycholinguistics/)

[home](https://md.research.elamur.ai/free/en/) / [concepts](https://md.research.elamur.ai/free/en/concepts/) / Psycholinguistics

# 🧩 Psycholinguistics

Concept · 18 key papers · [all concepts](https://md.research.elamur.ai/free/en/concepts/)

Psycholinguistics is the study of the mental processes by which people produce, comprehend and acquire language. It combines experimental methods, such as reading times, eye tracking, priming and brain recordings, with formal models that predict where and why language processing becomes difficult.

Also: psychology of language, language psychology, experimental psycholinguistics, computational psycholinguistics (for its modelling branch)

## What it is

Psycholinguistics sits between linguistics, cognitive psychology and neuroscience. It asks three questions: how a listener or reader builds an interpretation word by word, how a speaker turns an idea into a sentence, and how these abilities develop. It does not describe what a grammar permits. Instead, it studies the real-time machinery that uses linguistic knowledge under limits of memory, attention and time.

Its evidence is mostly behavioural or physiological, gathered under controlled conditions. The main methods are these:

- Self-paced reading. Participants reveal a text one word at a time, and the time spent on each word is recorded.
- Eye tracking. This records fixations and regressions (backward eye movements to earlier text) during natural reading. Common measures include: first-fixation duration; gaze duration, the sum of first-pass fixations on a word; go-past time, which runs from first entering a word until leaving it rightward and includes regressions; total reading time. First-fixation and gaze duration are called early-pass measures.
- Priming. Prior exposure to a word or structure speeds or biases later processing. This reveals which representations items share.
- Event-related potentials (ERPs). These are scalp-recorded brain responses time-locked to a word. The best known is the N400, a negativity peaking about 400 ms after word onset whose size is linked to how expected a word is in its context.
- Cloze norms. People complete sentence fragments, and their answers estimate how predictable each word is.

Researchers build controlled stimulus sets that differ in a single property, such as an ambiguous versus an unambiguous verb form. A region of interest is the stretch of text over which reading measures are collected. Classic test items include garden-path sentences, which lure readers into a wrong analysis. In *The horse raced past the barn fell*, readers first take *raced* as the main verb. It actually begins a reduced relative clause, that is, a relative clause with *who/that was* left out. Relative-clause contrasts are another standard case. Object relatives such as *The reporter who the senator attacked…* are harder than subject relatives such as *The reporter who attacked the senator…*.

## The key idea

A central organising claim of modern psycholinguistics is that comprehension is incremental and predictive. Readers update expectations at every word, and the effort a word costs reflects how unexpected it is. Surprisal theory turns this into a measurable quantity:

```
surprisal(w_t) = −log₂ P(w_t | w_1 … w_{t−1}) = log₂ ( α_{t−1} / α_t )
RT(w_t) ≈ β₀ + β₁ · surprisal(w_t) + β₂ · controls
```

The terms are as follows:

- w_t is the word at position t, and w_1 … w_{t−1} is the context before it.
- P is the probability that the comprehender, or a model standing in for the comprehender, assigns to the next word.
- α_t is the prefix probability of the sentence up to and including word t. The ratio of consecutive prefix probabilities therefore measures how much probability mass the new word eliminates.
- With a base-2 logarithm, surprisal is measured in bits.
- RT is a reading measure for the word, such as gaze duration.
- β₀ is an intercept, and β₁ is the per-bit cost of surprisal.
- β₂ weights control predictors such as word length and frequency.

Hale defined difficulty as this log-ratio of prefix probabilities. The regression link to reading measures was developed in later work. In his account, a word's cost is set by the probability mass of the analyses it rules out. Prefix probabilities come from a probabilistic Earley parser, a chart parser that works through the words strictly left to right. The parser runs over probabilistic context-free grammars whose rule probabilities were estimated from counts in a treebank, a corpus of syntactically parsed sentences. Some grammars were small and hand-written, and one was read off the whole corpus ([Hale 2001](https://md.research.elamur.ai/free/en/papers/hale-2001-probabilistic-earley-parser-as-psycholinguistic/)). The parser keeps all analyses in parallel. It can therefore explain garden-path difficulty without a separate reanalysis mechanism ([Hale 2001](https://md.research.elamur.ai/free/en/papers/hale-2001-probabilistic-earley-parser-as-psycholinguistic/)). A related quantity, contextual entropy, is the expected surprisal over possible next words. It is often added as a second predictor ([Wilcox et al. 2023](https://md.research.elamur.ai/free/en/papers/wilcox-2023-testing-predictions-surprisal-theory-11/)).

## History and founding papers

Psycholinguistics emerged in the early 1950s. A 1953 seminar named and organised it within behaviourist and information-theoretic psychology (Osgood & Sebeok 1954). In the 1960s, Chomsky's (1957, 1965) generative grammar reshaped the field, and many experiments tested whether grammatical transformations had psychological costs. Lexical priming paradigms trace back to Meyer and Schvaneveldt (1971). Garden-path research grew out of Bever (1970) and the structural parsing principles of Frazier (1978). Kutas and Hillyard (1980) first reported the N400 component.

The probabilistic turn came with Hale. Working from a small set of constructions, he argued that frequencies of phrase-level structures can do the work of dedicated processing principles. His model predicted the reduced-relative garden path and the greater difficulty of object relatives compared with subject relatives ([Hale 2001](https://md.research.elamur.ai/free/en/papers/hale-2001-probabilistic-earley-parser-as-psycholinguistic/)). One test sentence was *The banker told about the buy-back resigned*. For it, a grammar read off the whole corpus, with 532 NP rules and 120 S rules, still put surprisal at *resigned* close to ten times the value at the nouns. With a smaller hand-written grammar modelled on generalized phrase structure grammar, mean surprisal was about 5.0 for object relatives and about 2.1 for subject relatives ([Hale 2001](https://md.research.elamur.ai/free/en/papers/hale-2001-probabilistic-earley-parser-as-psycholinguistic/)). Levy (2008) later generalised the idea into a broad theory of expectation-based processing.

The next phase replaced grammars with statistical language models. Goodkind and Bicknell compared seven models on the Dundee eye-tracking corpus. These were n-gram models, which predict a word from the few words before it, an LSTM (a recurrent neural network) and two mixtures of the two. Perplexity measures how poorly a model predicts text, so lower is better. As model quality improved, the gain in fit to gaze durations grew linearly with the models' per-word log-likelihood. The surprisal effect itself kept a roughly constant size and an approximately linear shape ([Goodkind & Bicknell 2018](https://md.research.elamur.ai/free/en/papers/goodkind-2018-predictive-power-word-surprisal-reading/)).

Around the same time, a "models as subjects" programme began. Futrell and colleagues ran controlled psycholinguistic designs on LSTMs, treating per-word surprisal as a proxy for reading time. The networks showed human-like garden-path effects. However, they failed to learn two kinds of structural condition. One governs reflexives: *himself* needs a suitable antecedent. The other governs negative polarity items, words such as *ever* that need a licensor such as negation ([Futrell et al. 2018](https://md.research.elamur.ai/free/en/papers/futrell-2018-rnns-psycholinguistic-subjects-syntactic-state/)). Ettinger took a similar approach with stimuli drawn from N400 experiments. BERT retrieved hypernyms (category words such as *bird* for *robin*) well but largely ignored negation ([Ettinger 2020](https://md.research.elamur.ai/free/en/papers/ettinger-2020-what-bert-is-not-lessons/)).

## How it is used today

Much current work uses neural language models both as measuring instruments and as candidate theories of processing.

One line of work tests the scope of surprisal. The MECO eye-tracking data cover eleven languages from five families. In every one of them, surprisal raised the held-out log-likelihood of regression models of reading times (delta log-likelihood). Contextual entropy added predictive power in most languages, and a linear link fitted as well as more flexible functions ([Wilcox et al. 2023](https://md.research.elamur.ai/free/en/papers/wilcox-2023-testing-predictions-surprisal-theory-11/)). In English reading data, model-derived surprisal also predicts eye-tracking measures better than surprisal estimated from human cloze responses. The proposed reasons are the coarse resolution of cloze data and the finer handling of meaning and rare words by models ([Nair & Oh 2026](https://md.research.elamur.ai/free/en/papers/nair-2026-clozing-gap-exploring-why-language/)).

For large Transformers, the positive relation between model quality (lower perplexity) and fit to reading times reverses. The evidence comes from 17 Transformer models in the GPT-2, GPT-Neo and OPT families. Larger, lower-perplexity variants fitted self-paced reading and eye-tracking data worse, in a monotonic log-linear pattern. The errors concentrated on nouns, adjectives and named entities, which large models find too predictable ([Oh & Schuler 2023](https://md.research.elamur.ai/free/en/papers/oh-2023-why-does-surprisal-larger-transformer/)). Adding human-like memory decay to Transformers improved learning on developmental-scale data but worsened reading-time fit ([Thamma & Heilbron 2026](https://md.research.elamur.ai/free/en/papers/thamma-2026-human-like-fleeting-memory-improves/)).

Garden paths remain a key stress test. The LSTM models tested show behavioural evidence of tracking incremental syntactic state. So does an RNNG, a neural model that builds explicit syntactic trees. Unlike humans, however, their garden-path effects shrink as the ambiguous region grows longer ([Futrell et al. 2019](https://md.research.elamur.ai/free/en/papers/futrell-2019-neural-language-models-as-psycholinguistic/)). Huang et al. (2024) found that slowdowns predicted from model surprisal were only about 1/10 to 1/30 of observed human self-paced reading slowdowns on such sentences. One study showed that fine-tuning models on human reading times can close this gap on held-out items ([Yoshida et al. 2026](https://md.research.elamur.ai/free/en/papers/yoshida-2026-existence-proof-neural-language-models/)).

Methodological refinement is a third line. Subword tokenization splits words into smaller units. Researchers have examined how it interacts with regions of interest ([Giulianelli et al. 2024](https://md.research.elamur.ai/free/en/papers/giulianelli-2024-proper-treatment-tokenization-psycholinguistics/)). Others vary token granularity. In Mamba-2 models whose tokenizers were trained on English Wiki-40B, vocabularies near 8,000 entries gave the most predictive surprisal for the naturalistic reading corpora tested ([Oh & Schuler 2025](https://md.research.elamur.ai/free/en/papers/oh-2025-impact-token-granularity-predictive-power/)). A further approach probes hidden states directly. Early layers predict early-pass eye measures better than surprisal does ([Tsipidi et al. 2026](https://md.research.elamur.ai/free/en/papers/tsipidi-2026-probing-reading-times/)).

Psycholinguistic paradigms are also exported to new uses:

- auditing gender bias in coreference ([Bartl et al. 2025](https://md.research.elamur.ai/free/en/papers/bartl-2025-adapting-psycholinguistic-research-llms-gender/))
- locating neurons behind implicit causality ([Duan et al. 2025](https://md.research.elamur.ai/free/en/papers/duan-2025-unveiling-language-competence-neurons-psycholinguistic/))
- scoring metaphor novelty ([Momen et al. 2026](https://md.research.elamur.ai/free/en/papers/momen-2026-surprisal-metaphor-novelty-judgments-moderate/))
- building idiom norms ([Flor et al. 2025](https://md.research.elamur.ai/free/en/papers/flor-2025-survey-idiom-datasets-psycholinguistic-computational/))
- labelling psycholinguistically motivated markers of conspiracy talk ([Samory et al. 2026](https://md.research.elamur.ai/free/en/papers/samory-2026-semeval-2026-task-10-psycomark/))

| Finding | Detail | Source |
| --- | --- | --- |
| Better n-gram/LSTM models fit gaze durations better, linearly | seven models; Dundee gaze durations | ([Goodkind & Bicknell 2018](https://md.research.elamur.ai/free/en/papers/goodkind-2018-predictive-power-word-surprisal-reading/)) |
| Very large Transformers fit reading times worse | 17 models; self-paced reading and go-past durations | ([Oh & Schuler 2023](https://md.research.elamur.ai/free/en/papers/oh-2023-why-does-surprisal-larger-transformer/)) |
| Surprisal predicts reading in 11 languages | MECO eye tracking; five language families | ([Wilcox et al. 2023](https://md.research.elamur.ai/free/en/papers/wilcox-2023-testing-predictions-surprisal-theory-11/)) |
| Model garden-path effects are too small, unless the model is tuned | predicted slowdowns about 1/10–1/30 of human ones; no vocabulary size removed the shortfall | ([Yoshida et al. 2026](https://md.research.elamur.ai/free/en/papers/yoshida-2026-existence-proof-neural-language-models/); [Oh & Schuler 2025](https://md.research.elamur.ai/free/en/papers/oh-2025-impact-token-granularity-predictive-power/)) |

## Open questions

Is surprisal enough? Memory-based accounts, rooted in Gibson's (1998) locality theory, still compete as explanations of relative-clause asymmetries. Models tuned to reproduce garden-path slowdowns did not capture the asymmetry between subject and object relatives ([Yoshida et al. 2026](https://md.research.elamur.ai/free/en/papers/yoshida-2026-existence-proof-neural-language-models/)). That same flexibility raises a second worry. If a model's probabilities can be adjusted freely to match human data, surprisal theory risks becoming unfalsifiable. Proposed remedies include training on human-scale corpora or requiring that the parse distributions a model posits be psychologically real ([Yoshida et al. 2026](https://md.research.elamur.ai/free/en/papers/yoshida-2026-existence-proof-neural-language-models/)).

Which model counts as human-like? The answer may depend on the training data, the architecture and the tokenizer ([Oh & Schuler 2023](https://md.research.elamur.ai/free/en/papers/oh-2023-why-does-surprisal-larger-transformer/); [Oh & Schuler 2025](https://md.research.elamur.ai/free/en/papers/oh-2025-impact-token-granularity-predictive-power/)). No single feature derived from a model wins across languages and reading measures ([Tsipidi et al. 2026](https://md.research.elamur.ai/free/en/papers/tsipidi-2026-probing-reading-times/)).

Cross-linguistic coverage is still narrow. The largest test so far was dominated by Indo-European, high-resource languages ([Wilcox et al. 2023](https://md.research.elamur.ai/free/en/papers/wilcox-2023-testing-predictions-surprisal-theory-11/)). Cloze-to-reading comparisons exist only for English ([Nair & Oh 2026](https://md.research.elamur.ai/free/en/papers/nair-2026-clozing-gap-exploring-why-language/)).

Meaning-dependent phenomena also remain open. Negation is one example. BERT's neglect of negation resembles the N400, which in the original human experiments was also largely insensitive to negation. It departs, however, from human cloze completions and truth judgments ([Ettinger 2020](https://md.research.elamur.ai/free/en/papers/ettinger-2020-what-bert-is-not-lessons/)). The relation between word-level predictability, N400 amplitude and priming effects is likewise unresolved.

## Key papers

The works this page cites, the most cited first; citations as counted by OpenAlex plus the papers of this library that cite the work, read 2026-10-23.

- [A Probabilistic Earley Parser as a Psycholinguistic Model](https://md.research.elamur.ai/free/en/papers/hale-2001-probabilistic-earley-parser-as-psycholinguistic/) — John Hale · 2001 · 1,448 citations
- [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/) — Allyson Ettinger · 2020 · 843 citations
- [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/) — Adam Goodkind, Klinton Bicknell · 2018 · 294 citations
- [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/) — Richard Futrell, Ethan Wilcox, Takashi Morita et al. · 2019 · 291 citations
- [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/) — Byung-Doh Oh, William Schuler · 2023 · 211 citations
- [Testing the Predictions of Surprisal Theory in 11 Languages](https://md.research.elamur.ai/free/en/papers/wilcox-2023-testing-predictions-surprisal-theory-11/) — Ethan G. Wilcox, Tiago Pimentel, Clara Meister et al. · 2023 · 158 citations
- [RNNs as psycholinguistic subjects: Syntactic state and grammatical dependency](https://md.research.elamur.ai/free/en/papers/futrell-2018-rnns-psycholinguistic-subjects-syntactic-state/) — Richard Futrell, Ethan Wilcox, Takashi Morita et al. · 2018 · 75 citations
- [On the Proper Treatment of Tokenization in Psycholinguistics](https://md.research.elamur.ai/free/en/papers/giulianelli-2024-proper-treatment-tokenization-psycholinguistics/) — Mario Giulianelli, Luca Malagutti, Juan Luis Gastaldi et al. · 2024 · 23 citations
- [SemEval-2026 Task 10: PsyCoMark – Psycholinguistic Conspiracy Marker Extraction and Detection](https://md.research.elamur.ai/free/en/papers/samory-2026-semeval-2026-task-10-psycomark/) — Mattia Samory, Felix Soldner, Veronika Batzdorfer · 2026 · 20 citations
- [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/) — Byung-Doh Oh, William Schuler · 2025 · 12 citations
- [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/) — Xufeng Duan, Xinyu Zhou, Bei Xiao et al. · 2025 · 8 citations
- [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/) — Michael Flor, Xinyi Liu, Anna Feldman · 2025 · 5 citations
- [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/) — Marion Bartl, Thomas Brendan Murphy, Susan Leavy · 2025 · 3 citations
- [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/) — Ryo Yoshida, Shinnosuke Isono, Taiga Someya et al. · 2026 · 1 citation
- [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/) — Sathvik Nair, Byung-Doh Oh · 2026 · 1 citation
- [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/) — Abishek Thamma, Micha Heilbron · 2026 · 1 citation
- [Probing for Reading Times](https://md.research.elamur.ai/free/en/papers/tsipidi-2026-probing-reading-times/) — Eleftheria Tsipidi, Samuel Kiegeland, Francesco Ignazio Re et al. · 2026 · 1 citation
- [Surprisal and Metaphor Novelty Judgments: Moderate Correlations and Divergent Scaling Effects Revealed by Corpus-Based and Synthetic Datasets](https://md.research.elamur.ai/free/en/papers/momen-2026-surprisal-metaphor-novelty-judgments-moderate/) — Omar Momen, Emilie Sitter, Berenike Herrmann et al. · 2026 · 1 citation

## Related concepts

- [Child-directed speech](https://md.research.elamur.ai/free/en/concepts/child-directed-speech/) — Child-directed speech is the speech register adults and older children use when talking to young children. It typically has higher and more…
- [Cognitive linguistics](https://md.research.elamur.ai/free/en/concepts/cognitive-linguistics/) — Cognitive linguistics is an approach to language that treats grammar and meaning as products of general human cognition, including…
- [Language model](https://md.research.elamur.ai/free/en/concepts/language-model/) — A language model is a probability distribution over sequences of tokens (words, subwords or characters). It is usually factored into…
- [Implicature](https://md.research.elamur.ai/free/en/concepts/implicature/) — Implicature is the part of what a speaker communicates that goes beyond the literal content of the words. Hearers infer it by assuming the…

## More papers

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

Shelves: [🗣️ Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/)

- [← previous concept Part-of-speech tagging](https://md.research.elamur.ai/free/en/concepts/part-of-speech-tagging/)
- [next concept → Scaling laws](https://md.research.elamur.ai/free/en/concepts/scaling-laws/)
- [↑ back to concepts](https://md.research.elamur.ai/free/en/concepts/)