Language

Language processing covers how people produce speech and how they understand what they hear and read. Its central dispute, across production and comprehension alike, is between staged models, in which each level of processing finishes its work before passing a single result on, and interactive or cascaded models, in which levels work in parallel and influence each other (Goldrick, n.d.; Wagers & Dillon, 2025). In speaking, a staged model of word production competes with interactive models; a review concludes that the truth lies between them (Levelt et al., 1999; Goldrick, n.d.). In recognising spoken words, researchers disagree on whether knowledge of words feeds back to the perception of speech sounds (Norris et al., 2018). In understanding sentences, single-analysis “garden-path” models compete with constraint-based models, and the role of prediction is contested: a measure of word predictability called surprisal explains some reading difficulty but not all (Wagers & Dillon, 2025; van Schijndel & Linzen, 2021), and a well-known prediction effect failed to replicate (Nieuwland et al., 2018).

History

The production debate grew from two empirical traditions that disagreed. Speech errors, such as saying “rat” for “cat”, favoured interaction between levels, while reaction-time data from Levelt, Roelofs and Meyer were fitted well by staged, discrete models (Goldrick, n.d.). Levelt, Roelofs and Meyer’s 1999 theory divides word production into stages, is implemented as a computer model called WEAVER++, and was tested mainly with the time people take to name pictures (Levelt et al., 1999). Dell’s 1986 model, the main interactive alternative, allows activation to flow back from sounds to words (Nozari & Pinet, 2020).

In sentence comprehension, the garden-path model holds that readers commit to a single analysis of a sentence at each point, and must reanalyse when it proves wrong; constraint-based models describe comprehension as the simultaneous interaction of several levels of representation (Wagers & Dillon, 2025). Traxler describes three later developments that move away from purely bottom-up parsing: anticipation of upcoming words, Bayesian or “noisy-channel” inference, in which readers weigh what they perceive against what is likely to have been meant, and “good-enough” parsing, in which readers settle for a shallow interpretation (Traxler, 2014).

Core concepts

Speaking a word

To say a word, a speaker must select it from memory and retrieve its sounds. All major models agree that words other than the target are partly activated along the way; they differ on feedback. Only models that allow feedback, such as Dell’s, activate other words through shared sounds (Nozari & Pinet, 2020). Despite large behavioural, brain-imaging and electrophysiological literatures, it remains unclear how the production system handles these co-activated items, and a review identifies three common problems in how the evidence is interpreted (Nozari & Pinet, 2020).

Discrete two-step models and fully cascading models differ in two testable predictions: whether distractors from the same category cause interference, and whether the sound forms of more than one word become available at once (Bölte et al., 2015).

Recognising spoken words

Spoken-word recognition is split on feedback. Magnuson and colleagues argue that activation feeding back from the word level to the level of speech sounds helps recognition. Norris, McQueen and Cutler argue that such feedback is unnecessary and contradicted by behavioural evidence (Norris et al., 2018), and that the opposing argument is tied to one class of model, misses the point about parsimony and ignores crucial data (Norris et al., 2018).

Understanding sentences

Surprisal is a measure of how unexpected a word is in its context, computed from the probability a language model assigns to it. Levy argued that surprisal theory, first proposed by Hale in 2001, is equivalent to reallocating resources across parallel probabilistic analyses of a sentence, which ties it to constraint-based parallel models (Levy, 2008). Readers also use prior beliefs in interpretation: comprehenders misinterpret anomalous sentences in the direction of what is plausible (Traxler, 2014).

Regular and irregular past tense

How people form the past tense of English verbs became a test case between rules and associations. Pinker and Ullman’s words-and-rules account holds that irregular forms (“went”) are stored in declarative memory, while regular forms (“walked”) are computed by a rule using the procedural system (Pinker & Ullman, 2002). Single-route accounts, associated with connectionism, hold that both kinds come from one associative mechanism (Yoon, 2023).

Questions, controversies, and new developments

Staged or cascading production?

This question is contested. Goldrick’s review concludes that both the strictly discrete and the fully interactive positions are wrong: activation does cascade, and some feedback flows from sounds to words, but both are limited, and feedback from words to meaning is absent or insignificant (Goldrick, n.d.).

The mixed-error effect is the most cited evidence for interaction: errors that are related to the target both in meaning and in sound, such as “rat” for “cat”, occur more often than the sum of purely semantic and purely phonological errors would predict. The effect is itself inconsistent, with null results in spontaneous, experimentally induced and aphasic error corpora (Goldrick, n.d.). Goldrick argues that it needs only a minimal step beyond a discrete system, cascading activation, not full feedback, and that very strong cascading predicts patterns the data do not show (Goldrick, n.d.). Roelofs, defending the staged model, is reported to answer that mixed errors arise from mistakenly selecting two words, not from feedback (Bölte et al., 2015).

Experiments with pictures give inconsistent results. A facilitation effect of phonologically related distractor pictures reported by Morsella and Miozzo was not replicated by Jescheniak and colleagues, and pictures from the same category produced facilitation, interference or no effect in different studies (Bölte et al., 2015). Semantic interference in picture-word tasks can be read as competition when a word is selected, as staged models say, or as an effect in an output buffer after selection, as cascading models allow, so the effect does not settle the debate (Bölte et al., 2015). Bölte and colleagues found that distractor pictures activate their own word forms, which they say is hard to reconcile with discrete serial models (Bölte et al., 2015).

Can surprisal explain garden paths?

A garden-path sentence leads the reader into a wrong analysis and causes a slowdown when the error is detected. Surprisal from a fully parallel parser predicts that garden-path effects exist, but drastically underpredicts their size and their relative severity across sentence types; van Schijndel and Linzen conclude that prediction alone may not replace mechanisms of reanalysis (van Schijndel & Linzen, 2021). A 2026 preprint reports that language models fine-tuned for the purpose do capture human garden-path slowdowns through surprisal, offered as an existence proof against that conclusion (Yoshida et al., 2026). With 409 surprisal estimates and 368 readers, Timkey and colleagues found that surprisal explains early effects of syntactic disambiguation but not the added cost of rereading, so it captures routine structure-building but not the detection and correction of errors (Timkey et al., 2026).

What shape is the predictability effect?

Surprisal theory predicts that reading time rises with the logarithm of a word’s improbability. Across six reading datasets, Shain and colleagues found a logarithmic effect, which favours probabilistic-inference accounts over accounts in which readers pre-activate likely words (Shain et al., 2024). Brothers and Kuperberg found a linear relation, contradicting surprisal theory and favouring proportional pre-activation (Brothers & Kuperberg, 2021). The function linking predictability to reading time is contested.

Do readers predict specific words?

DeLong and colleagues (2005) reported that the brain’s N400 response, an electrical signal sensitive to how expected a word is, differed already at the article before a predictable noun (“a” versus “an”), suggesting that readers pre-activate the sound of the upcoming noun. A nine-lab preregistered replication with 334 participants reproduced the effect at the noun but not at the article (Nieuwland et al., 2018). Evidence for routine pre-activation of word sounds, one pillar of predictive accounts, thus failed to replicate.

Rules or analogy in the past tense?

Evidence on the past tense points both ways, and the question is contested.

• In an ERP study, violations of regular past tenses and of phrase structure produced left anterior negativities, a brain response, while irregular violations did not; the authors read this as distinct processing of regulars and irregulars (Newman et al., 2007).
• An N400 priming study also found a regular-irregular dissociation, but argued that it fits single-system models better than dual-system ones (Justus et al., 2008).
• In stroke patients, deficits with the regular past tense were associated with damage to regions linked to articulatory sequencing and phonological short-term memory, not with grammatical difficulty (Meteyard et al., 2013).
• A behavioural study of children’s past tenses of novel verbs concluded that regular forms are explained by analogy over stored examples, not a formal rule (Blything et al., 2018).

Whether modern neural networks resolve the debate is an open question (Yoon, 2023).

Replication record

Broader connections

The past-tense debate is a central case in the dispute between rules and associations described in Connectionism. Language and emotion are argued to be interactive or even constitutive of each other rather than separate, with the evidence favouring the constitutive view (Satpute & Lindquist, 2021). The storage of irregular forms in declarative memory ties the words-and-rules account to the memory systems described in Memory.

Further reading

• Goldrick (n.d.): a review of the staged-versus-interactive debate in production.
• Nozari and Pinet (2020): how production models handle co-activated words.
• Traxler (2014): prediction, noisy-channel and good-enough accounts of comprehension.
• Nieuwland et al. (2018): the large replication of the article prediction effect.