RESEARCH DOSSIER
Computational Models of Language Processing
Abstract. Computational Models of Language Processing critically synthesizes formal accounts of how humans perceive, predict, interpret, learn, and produce language over time. It traces the transition from cohort and interactive-activation models of spoken words to distributed learning, probabilistic parsing, working-memory retrieval, surprisal, rational communication, and contemporary neural language models. The methodological focus is the relation between precisely specified computations and human evidence: response times, eye movements, developmental learning, event-related potentials, intracranial activity, and cross-linguistic distributions. The analysis distinguishes algorithmic success, predictive fit, cognitive process claims, and neural implementation, and evaluates the effects of model training, linguistic representation, context, task design, and measurement. Shared next-word prediction and contextual sensitivity provide useful bridges between artificial and human language systems; differences in grounding, experience, adaptation, and causal mechanism remain essential targets of study. The result is a critical narrative research synthesis built from identifiable original studies and scholarly reviews; it is not an original behavioral experiment, a PRISMA systematic review, or a meta-analysis.
Keywords: computational psycholinguistics; language processing; spoken word recognition; TRACE; probabilistic parsing; surprisal; sentence comprehension; working memory; statistical learning; Bayesian pragmatics; rational speech acts; language models; reading times; N400; neural encoding; cognitive model validation; Artificial Era
Scientific Overview
Computational models of language processing turn claims about comprehension and production into explicit operations over time, information, memory, and linguistic alternatives. A model specifies what representations are available, how evidence changes them, and which observable consequences follow when a listener or reader encounters a particular signal. The central psychological question concerns the organization of human language behavior under uncertainty: how does partial acoustic or visual input produce graded expectations, lexical choices, syntactic commitments, conceptual interpretations, and appropriate communicative responses? This monograph compares mechanistic proposals rather than treating accurate text prediction as a complete explanation. The distinction matters because successful output may be generated through different internal processes, and a psychological model must earn its explanatory interpretation through independent human observations Levy (2008) McClelland & Elman (1986).
Three classes of evidence define the research field. Controlled experiments isolate contrasts among sounds, word candidates, structural ambiguities, or referential alternatives; naturalistic corpora preserve the statistical environments in which people ordinarily read and listen; physiological recordings help relate model variables to time-dependent brain activity. These measurements answer partially different questions. A fixation indicates an allocation of visual attention within a task, a reading-time effect indexes a composite of perceptual and linguistic work, and an electrophysiological response reveals a temporally structured neural correlate. Their convergence can constrain explanatory theories more strongly than any single score. Model-based inference requires an explicit mapping between predicted quantities and observed behavior, together with tests of competing explanations Tanenhaus et al. (1995) Demberg & Keller (2008) Frank et al. (2015).
The synthesis proceeds from the definition and historical development of computational psycholinguistics through lexical competition, incremental parsing, statistical learning, pragmatic reasoning, and the evaluation of neural systems. The principal criteria are conceptual specificity, reproducibility, psychometric validity, predictive performance on held-out data, and transparent limits on causal claims. Prediction is central because human processing continuously depends on context, yet prediction is only one component of a larger system involving action, reference, learning, and social coordination. The contemporary Artificial Era increases the importance of separating shared computational principles from assumptions about equivalence of internal organization. Empirical evidence about people and artificial models will be considered at the same level of scrutiny Schrimpf et al. (2021) Goldstein et al. (2022).
Levels of Computational Explanation
A computational description begins with a functional problem, such as identifying a spoken word from noisy evidence or selecting a referent given an utterance. An algorithmic description then specifies representations, state transitions, and decision procedures. Implementation concerns the physical or neural substrate supporting these operations. Distinguishing these questions prevents a frequent inference error: a statistical model can reproduce a behavioral curve without showing that humans compute the same intermediate states. Conversely, a process model may capture the structure of observed decisions even when its biological implementation remains hypothetical. Explicit conditional probabilities, activation updates, memory retrieval functions, and choice policies provide the necessary bridges between conceptual explanations and measurements Jurafsky (1996) Lewis & Vasishth (2005).
A psycholinguistic theory gains precision when it predicts contrasts rather than simply restating successful behavior. For instance, a graded lexical-competition model should predict how temporarily plausible competitors differ from unrelated words over successive segments of speech. A parser should specify the effects of syntactic ambiguity and context before the ambiguity is resolved. A probabilistic pragmatic model should predict what changes when a speaker has access to additional possible utterances. Such theories can fail informatively. A fit limited to one sentence type or one communicative game does not demonstrate general adequacy, and a high score without a psychologically meaningful dependent variable does little to establish human relevance McClelland & Elman (1986) Frank & Goodman (2012).
Model parameters also require substantive interpretation. A temperature parameter, decay constant, attention distribution, or retrieval threshold can describe variation in behavior, but a fitted number does not automatically identify a discrete cognitive mechanism. Parameter correlations and flexible functional forms can permit different models to reproduce the same response patterns. Strong comparison therefore combines held-out prediction, selective manipulations, parameter recovery where possible, and cross-task tests. Interpretability requires showing that an inferred variable remains stable under relevant changes of material and experimental context. These are requirements of scientific inference and apply equally to hand-crafted symbolic systems, connectionist models, Bayesian accounts, and modern neural language models Hu et al. (2020) Shain et al. (2024).
Historical Development of Formal Language Models
Formal accounts of language processing emerged through several intersecting traditions: psycholinguistic experiments, mathematical theories of information, linguistic grammars, cognitive architectures, and computational simulations. Early spoken-word accounts organized recognition around the unfolding input and a changing set of compatible lexical candidates. Marslen-Wilson's account separated access, selection, and integration, making the temporal course of lexical recognition a theoretical object rather than treating recognition as a completed decision. This move motivated experiments that could expose competitors during partial input. Later connectionist work made the representational state itself dynamic, allowing the same network to embody both constraints and temporal competition Marslen-Wilson (1987) McClelland & Elman (1986).
The late twentieth century also saw a decisive extension from localized recognition rules to learned distributed representations. Seidenberg and McClelland modeled printed-word recognition and naming with connections between orthographic and phonological units, and Elman explored how recurrent networks form context-dependent internal states while processing sequences. These studies established that complex patterns of linguistic behavior could emerge from learning in networks with explicitly specified training experience. Their historical significance lies in opening a method for testing assumptions about representations and acquisition through simulation. The adequacy of each particular model remains an empirical question: learning a useful pattern does not prove that a human learner employs identical units or learning rules Seidenberg & McClelland (1989) Elman (1990).
Probability-based models subsequently united expectations about word identity, lexical category, and syntactic structure. Jurafsky formulated probabilistic access and disambiguation across linguistic levels, while Hale showed how a probabilistic parser generates word-by-word predictions of processing load. Levy clarified the information-theoretic relationship between context-dependent expectation and processing difficulty. The historical development is therefore cumulative across competing model families: network dynamics, explicit grammatical alternatives, and probabilistic inference each contribute testable explanations for different dimensions of language. Contemporary neural language modeling extends this tradition with new training scale and representational flexibility, while renewing longstanding problems of cognitive plausibility, explanation, and generalization Jurafsky (1996) Hale (2001) Levy (2008).
Spoken Word Recognition and Lexical Competition
Spoken words arrive as continuous acoustic signals whose informative cues are distributed through time. A listener must integrate emerging sound patterns with knowledge of possible word forms, resolve competitors sharing initial segments, and determine when enough evidence is available to identify a candidate. The cohort approach made these operations explicit by distinguishing the initial activation of word forms from later discrimination and contextual integration. Its central contribution is temporal: lexical access can begin before a word ends, and alternative candidates need not disappear at the same moment. Measuring this evolving candidate set requires more than final identification accuracy, because a correct response can conceal a prolonged period of uncertainty Marslen-Wilson (1987).
Computational accounts differ on where contextual constraints enter the recognition system. In a strongly bottom-up architecture, acoustic evidence governs lexical selection, with semantic context exerting effects at later integration. Interactive systems allow higher-level expectations to change activation at earlier representational levels. Both frameworks can reproduce context-sensitive behavior under suitable assumptions; discriminating them requires temporal data and carefully controlled evidence. Misleading predictions about early activation, competitor suppression, or restoration of uncertain segments should be tested with manipulations that separate perception from response bias. This distinction is central when lexical models are compared with general-purpose text models trained without the actual physical variability of speech McClelland & Elman (1986) Marslen-Wilson (1987).
The visual-world paradigm provides a particularly direct measurement of partial lexical and referential competition. When people hear instructions while viewing possible objects, their eye movements reveal the unfolding influence of linguistic evidence and environmental affordances. Tanenhaus and colleagues showed that visual context can influence interpretation during the earliest stages of spoken instruction comprehension. Yet an eye fixation follows oculomotor constraints, task goals, object salience, and linguistic processing jointly. A model must therefore predict the temporal distribution of fixations under an explicit observation model, and experiments should manipulate competitor sets independently of changes in acoustic or syntactic content Tanenhaus et al. (1995).
Interactive Activation and the TRACE Architecture
TRACE offers a historically important computational formulation of speech perception as continuously updated activation across levels representing acoustic features, phonemes, and words. Units interact through excitatory and inhibitory connections, and the pattern of activation unfolds across time-sliced input. This architecture captures graded support, lexical rivalry, and context-dependent processing within a common dynamical system. Its theoretical distinctiveness lies in the claim that perception is an ongoing constraint-satisfaction process, with previous evidence retained in an evolving processing trace. A simulation can expose intermediate phoneme and lexical states that would remain invisible in an account built only around final correct classifications McClelland & Elman (1986).
The model's explanatory value is clearest when one asks why a familiar word can become active despite ambiguous sound and how stronger rivals suppress weaker alternatives. Lateral inhibition and bidirectional excitation provide specific mechanisms for these patterns, making predictions about how stimulus changes affect relative activation. At the same time, interactive feedback is a substantive hypothesis, not a mandatory ingredient of any successful model. Empirical consequences depend on processing time, strength of competing candidates, and the response measurement used. It is possible for a system with primarily feedforward operations to produce apparently interactive behavior through later decisions, so evaluating feedback requires carefully timed dissociations McClelland & Elman (1986) Marslen-Wilson (1987).
TRACE is also a methodological lesson about scale and explicitness. A relatively compact network permits investigators to examine activation trajectories, change parameters, and observe identifiable failures. Larger models can encode richer language statistics but are often harder to interrogate mechanistically. The relevant comparison concerns the explanations each can support: a specialized perceptual architecture represents acoustic-phonetic competition in detail, whereas an autoregressive text model estimates distributions over token sequences. A finding that both predict lexical context effects would establish a shared statistical regularity, while independent evidence would still be required to identify equivalent perceptual operations. This distinction organizes the larger research agenda Goldstein et al. (2022) Schrimpf et al. (2021).
Distributed Representations and Sequential Learning
Distributed models represent linguistic information as patterns of activity rather than fixed entries that must be retrieved in isolation. In Seidenberg and McClelland's account of word recognition and naming, orthographic and phonological units interact through learned connections, yielding graded sensitivity to properties of letter strings and pronunciations. Such models invite an explanation of regularities, exceptions, and generalization through the statistical structure of training exposure. They also require a precise statement of what counts as a learned representation and which aspects of performance depend on architectural choices. Successful simulation of familiar-word naming does not, by itself, settle the learning mechanisms through which children acquire literacy Seidenberg & McClelland (1989).
Recurrent networks expanded this program by integrating temporal context into internal representations. Elman's work explored how a network with feedback from previous hidden states could extract useful sequential categories without a pre-specified inventory of grammatical symbols at the hidden layer. The resulting states combine recent input with learned expectations about the continuation of a sequence. This offers an illuminating alternative to accounts that first assign every symbol a discrete category and then compute grammatical combinations. Nevertheless, the degree to which such representations remain stable across unfamiliar constructions, speakers, and languages depends on training distributions and testing procedures Elman (1990).
The central question for distributional explanations is what experience permits a learner to infer, and under what generalization conditions. Exposure frequency, lexical diversity, sampling structure, and the relation between surface forms and communicative functions all affect learning. A model trained on a narrow corpus can produce human-like errors within that corpus while failing when systematic contrasts are rearranged. Strong evaluation therefore includes new lexical items, reordered contexts, altered statistical contingencies, and explicit comparisons with symbolic and probabilistic competitors. The research aim is to specify the inductive resources that make linguistic learning possible, rather than inferring psychological mechanism from optimization success alone Seidenberg & McClelland (1989) Elman (1990) Hu et al. (2020).
Probabilistic Lexical and Syntactic Inference
Probabilistic psycholinguistics treats comprehension as inference over multiple candidate interpretations conditioned on the evidence encountered so far. A lexical item can be compatible with several categories or senses, and a sentence fragment may license competing syntactic structures. Jurafsky's model formalized access and disambiguation across lexical and syntactic levels, offering a common probabilistic language for phenomena previously explained by relatively separate models. The critical commitment is that preference changes with evidence and learned distributional regularities. A realistic account must specify which alternatives remain active, how their probabilities are updated, and what processing costs follow from these updates Jurafsky (1996).
Computational parsers operationalize grammatical proposals by assigning probabilities to alternative structural analyses. Their predictions can be compared with the disruption people experience in temporarily ambiguous sentences. A probabilistic Earley parser is especially useful because it constructs a distribution over analyses incrementally, rather than assigning a single tree after the final word. Hale connected the changing probability mass of predictions with cognitive load, and Levy described the resulting expectation-based account in more general information-theoretic terms. These models gain psychological relevance when their predictions are evaluated on behavioral materials designed to separate structure, plausibility, frequency, and memory constraints Hale (2001) Levy (2008).
The probability distribution is itself a model-dependent object. A grammar trained on edited newspaper text, a neural network trained on mixed web documents, and a human listener with specialized background knowledge may assign very different probabilities to the same word. Thus, surprisal estimates carry assumptions about the training corpus and available context. Comparing models on human language processing requires controlling lexical frequency, sentence position, contextual constraint, and material-specific artifacts. Differences in language-model perplexity can change estimated predictor reliability without demonstrating that human processing uses a representation specific to one parser. The underlying theory is strengthened by robust contrasts across estimators and datasets Demberg & Keller (2008) Shain et al. (2024).
Incremental Parsing and Word Surprisal
Surprisal quantifies how unexpected a linguistic event is under an explicit probability distribution: for a word w in context c, surprisal is −log P(w | c). This measure provides a principled link between uncertainty and information gained when an incoming word reduces the set of possible continuations. In incremental comprehension, a low-probability continuation often induces greater processing difficulty. Hale's probabilistic parsing proposal and Levy's expectation-based analysis formalized this relation for syntactic contexts, connecting garden-path effects to changes in probability across competing structures. The same mathematical quantity can be estimated by models whose internal representations differ substantially Hale (2001) Levy (2008).
Evidence for surprisal includes both laboratory materials and naturalistic reading data. Smith and Levy examined the relation between contextual probability and word reading time and found support for a logarithmic relation. Shain and colleagues later assessed a much larger body of reading-time observations and evaluated the shape of these effects. These results provide strong motivation for treating word predictability as a continuous variable rather than dividing every event into predictable and unpredictable categories. Yet the precise slope and temporal signature of the relation can depend on measurement modality, spillover, word frequency, and the chosen model for probabilities Smith & Levy (2013) Shain et al. (2024).
The existence of a surprisal effect does not establish that prediction is the only difficulty-generating operation. Integration demands, retrieval interference, semantic plausibility, and reanalysis can covary with unexpectedly encountered words. A useful empirical design creates contrasts in which the candidate mechanisms diverge. One may compare structurally similar continuations with different conditional probabilities, or manipulate the number and similarity of intervening dependencies while controlling lexical predictability. Regression-based comparisons are informative when stimulus-level dependencies and uncertainty in model predictions are handled carefully. The scientific target is a falsifiable process account, not merely a significant coefficient for a probability-derived regressor Demberg & Keller (2008) Lewis & Vasishth (2005) Frank & Bod (2011).
Working Memory Retrieval and Structural Interference
Human sentence comprehension depends on maintaining and retrieving information that may be separated by intervening words. An activation-based model developed by Lewis and Vasishth treats syntactic processing as skilled retrieval from memory under limited cognitive resources. Linguistic representations gain and lose accessibility, and retrieval cues can match multiple stored items. Similarity-based interference explains why some dependencies become difficult even when sentence length alone is insufficient to predict a slowdown. The proposal is a mechanistic account because it defines specific memory operations and derives expectations about the latency and accuracy of completing structural relations Lewis & Vasishth (2005).
Memory-oriented accounts emphasize a different explanatory variable from simple surprisal. Processing difficulty may arise when the reader retrieves an earlier noun to integrate it with a verb, particularly if several nouns share diagnostic features. Even a highly expected continuation can require effortful retrieval, while an unexpected word may be resolved quickly if relevant context is accessible. This does not place memory and prediction in separate psychological systems by necessity. It defines experimentally distinguishable sources of processing cost that can interact within one comprehension episode. Controlled similarity manipulations and sentence types with varying structural dependencies are needed to adjudicate the contribution of retrieval Lewis & Vasishth (2005) Levy (2008).
Quantitative model comparison must also consider what is measured: reading-time means, response distributions, eye-movement regressions, or error types. The activation-based approach has generated explicit simulation results for structural interference and center embedding, but extending these fits across languages and tasks requires careful testing. Parameter tuning that improves a particular dependency pattern can conceal poor performance on another. A robust research program estimates models on one set of contrasts and evaluates them on independently selected materials. The ultimate goal is to explain the combined effects of expectation, retrieval demands, and the dynamically evolving representation of a sentence Lewis & Vasishth (2005) Demberg & Keller (2008) Shain et al. (2024).
Information-Theoretic Approaches to Processing Cost
Information theory supplies general quantities that can connect linguistic expectations with cognitive activity. Surprisal measures the informational consequence of the particular item observed, whereas entropy summarizes uncertainty across all possible continuations before the next event arrives. These quantities are mathematically related but psychologically distinct: a highly uncertain context does not guarantee that the next word will be individually surprising, and a clear prediction can be violated by an unexpected continuation. Researchers must distinguish prospective uncertainty, event-specific prediction error, and the computational work associated with revising a representation. Levy's account made the importance of this distinction central to psycholinguistic theory Levy (2008).
Naturalistic corpora permit tests of these proposals across thousands of words, but such tests require careful identification of confounds. Demberg and Keller compared information-based and locality-oriented processing theories against eye-tracking observations, illustrating both the opportunity and the limitations of observational model evaluation. Natural discourse provides varied syntactic structures and lexical contexts, yet the same word also differs in length, position, frequency, discourse role, and visual layout. A credible computational analysis needs a statistical observation model that accounts for these factors and evaluates uncertainty in fitted effects. Out-of-sample prediction is more informative than a post hoc explanation of a few salient sentences Demberg & Keller (2008).
Information-based efficiency can also concern the structure of produced messages. Jaeger examined why speakers optionally realize syntactic material, linking such variation to expected information density. Piantadosi and colleagues showed cross-language associations between word length and average information content, broadening the discussion from momentary comprehension to lexical organization. These findings illuminate constraints on communication, but corpus regularities alone do not reveal whether particular speakers consciously optimize the same numerical objective. Production choices are shaped by communicative goals, learned conventions, articulatory constraints, and interaction with listeners. Model building must connect abstract efficiency principles to plausible decisions and measurable communicative outcomes Jaeger (2010) Piantadosi et al. (2011).
Statistical Learning of Linguistic Regularities
Statistical learning models ask how listeners discover useful units and dependencies without being told in advance where each word begins. In a continuous speech stream, the transitional probabilities among syllables can help distinguish cohesive sequences from boundaries between words. Saffran, Aslin, and Newport demonstrated that eight-month-old infants could use sequential statistical structure to discriminate patterns after brief exposure to artificial-language materials. This result established an experimentally tractable learning capacity that computational models could formalize. It also provides a specific behavioral achievement rather than a general proof that every component of language acquisition follows from the same mechanism Saffran et al. (1996).
Formal learning theories must define both the evidence available and the representations learners can construct from it. A system may track adjacent transition probabilities, inferred latent categories, broader sequences, or relations between sounds and referents. Training distributions can be arranged so that some candidate rules agree on observed examples but diverge on new combinations. Testing these combinations reveals whether the model generalized a structural principle or retained surface correlations. Developmental studies introduce further constraints because age, task demands, attention, and exposure conditions influence performance. The strongest explanation connects an explicit learning rule to identifiable individual observations across more than one laboratory task Saffran et al. (1996) Elman (1990).
Language learning occurs within a communicative and cognitive environment. Infants hear highly structured speech from particular speakers, observe referents, and participate in interactions that provide pragmatic information beyond raw sound sequences. Computational models that omit these factors may still isolate a genuine learning capacity, but the scope of their conclusions is correspondingly limited. A full account should explain how statistical sensitivities are deployed during recognition, category formation, and later generalization. It should also distinguish mechanisms that exploit local frequency from mechanisms that infer relations across intervening materials. The historical influence of artificial-language studies lies in making such distinctions experimentally controllable Saffran et al. (1996) Christiansen & Chater (2016).
Temporal Learning and the Processing Bottleneck
The stream of spoken language unfolds rapidly, and recent input becomes less accessible as new information arrives. Christiansen and Chater developed the Now-or-Never bottleneck as a theoretical constraint: successful processing requires the rapid recoding of fleeting input into organized chunks. The proposal links immediate comprehension with language learning by arguing that learners develop capacities for processing patterns while immersed in the flow of communication. Computationally, chunking provides a way to compress information into usable units whose internal details need not remain explicitly present at every later stage. The theory raises measurable questions about memory duration, chunk boundaries, and individual differences in processing Christiansen & Chater (2016).
The bottleneck account generates distinctive hypotheses about how representation changes over time. When several units arrive in rapid succession, a comprehender may maintain compressed structures rather than an exact trace of every phonetic segment. Processing costs should therefore depend on the availability of relevant chunks and on the ability to update them as new input arrives. Recurrent neural networks provide one family of mechanisms for maintaining context-sensitive state, while explicit parsers and working-memory systems offer alternative implementations. Comparing these approaches requires experiments that manipulate temporal rate, intervening material, and opportunities for rehearsal or reanalysis Elman (1990) Lewis & Vasishth (2005) Christiansen & Chater (2016).
The conceptual value of a bottleneck depends on specifying the magnitude and behavioral consequences of representational loss. Human participants can sometimes recover earlier linguistic content, particularly when it is meaningful or associated with an external cue. A strong theoretical account must explain such recovery within a model of temporally constrained resources rather than assuming universal immediate disappearance. Language acquisition and online processing also operate across different timescales, which raises questions about how short episodes shape durable knowledge. Carefully designed experiments and models can relate recoding, statistical learning, retrieval, and consolidated expectations without treating a vivid metaphor as a completed mechanism Christiansen & Chater (2016) Saffran et al. (1996).
Bayesian Pragmatics and Rational Speech Acts
Understanding a sentence requires inferences about why a speaker chose a particular utterance in a particular situation. A listener may infer a referent from the literal possibilities licensed by the words and from assumptions about the speaker's communicative goals. Frank and Goodman tested a Bayesian approach to simple language games in which listener judgments are shaped by the informative choices a cooperative speaker is expected to make. This formalizes a social aspect of comprehension: the same expression can suggest different interpretations when available alternatives and speaker knowledge change. The model's predictions can be evaluated against systematic distributions of human responses Frank & Goodman (2012).
Rational Speech Acts accounts represent nested reasoning between speakers and listeners. A literal listener assigns candidate referents based on linguistic meaning; a pragmatic speaker tends to choose utterances that guide a listener toward intended outcomes; a pragmatic listener reasons backward from the observed utterance to the speaker's likely intention. The model may incorporate costs for producing messages, uncertainty about referents, and graded preferences. These ingredients help explain implicature and context dependence, as discussed in the framework's broader synthesis. Yet its scientific validity depends on whether participants actually use such recursive inferences under specified task conditions, not on the mathematical elegance of the recursive formulation Goodman & Frank (2016).
An important methodological distinction separates predicted equilibrium behavior from online psychological process. A normative Bayesian account can fit final judgments while remaining silent about the time taken to compute them, the memory requirements, or the order of intermediate decisions. Experiments using time pressure, alternative-utterance availability, communicative feedback, and population differences can probe these missing dimensions. The model is most compelling when it predicts new choices after changes in context that were not used to fit its parameters. Integrating probabilistic pragmatics with incremental lexical processing remains an important task because human interpretation unfolds while the utterance is still being produced Frank & Goodman (2012) Goodman & Frank (2016) Tanenhaus et al. (1995).
Speech Production and Communicative Efficiency
Computational language processing includes the production of utterances, where a speaker chooses among grammatically possible forms under constraints of time, meaning, and anticipated listener needs. Information-theoretic accounts propose that speakers may distribute linguistic information in ways that improve communicative efficiency. Jaeger's analysis of spontaneous speech linked the optional realization of syntactic material to the information carried by the emerging utterance. This research demonstrates that production choices can be analyzed probabilistically rather than assigned to isolated stylistic preferences. Its interpretation requires considering alternative explanations based on availability, ambiguity, and processing dependencies Jaeger (2010).
Efficiency also appears in the statistical organization of lexical forms. Piantadosi, Tily, and Gibson compared properties of words across multiple languages and found that average information content offers an informative predictor of word length. The result suggests a connection between stable linguistic conventions and the distribution of communicative demands. But the historical formation of a lexicon involves transmission across speakers and generations, and a cross-language association does not isolate an individual decision rule. A mechanistic model should predict how forms arise, change, or become favored under particular communicative pressures, and should be evaluated against plausible alternative linguistic histories Piantadosi et al. (2011) Futrell et al. (2015).
Production models can also inform comprehension when both partners adapt to one another. A speaker's choice changes the distribution of expressions a listener is likely to hear, and the listener's understanding can influence subsequent phrasing and repair. Rational-pragmatic models formalize these dependencies through speaker and listener expectations, while information-density accounts describe allocation of predictable and less predictable material. The synthesis offers a research program on interactive optimization rather than a claim that all communicative behavior maximizes one universal objective. Experiments must test communication quality, comprehension, adaptation, and the costs of producing particular forms in addition to conventional measures of grammaticality Goodman & Frank (2016) Jaeger (2010).
Situated Comprehension and Referential Context
Language processing takes place in situations with visible objects, ongoing actions, interpersonal goals, and prior discourse. Tanenhaus and colleagues used spoken instructions together with eye tracking to show that visual context could influence interpretation at early moments of sentence comprehension. This evidence constrains purely sequential accounts in which situational information becomes relevant only after syntactic analysis has reached a final decision. Computational models of situated language must specify how linguistic evidence, candidate referents, and task context jointly determine attention and interpretation. Such models can generate predictions for fixation patterns rather than relying solely on eventual compliance with an instruction Tanenhaus et al. (1995).
A central difficulty is distinguishing an effect of context on an early linguistic representation from an effect on subsequent choice or attention. Eye movements are continuous and informative, but their latency and interpretation depend on oculomotor planning, object salience, and the task's response structure. Computational observation models can relate hidden interpretive states to measurable gaze distributions and permit comparisons between alternative accounts. Referential scenes with strategically selected competitors are particularly useful because they can change the usefulness of a phrase without changing its surface form. Naturalistic stimuli improve relevance while introducing new correlated variables that require modeling Tanenhaus et al. (1995) Marslen-Wilson (1987).
Situated language illustrates why a complete theory must connect multiple timescales. Acoustic evidence supports word candidates, syntactic expectations constrain grammatical continuation, and knowledge of objects guides the selection of a plausible interpretation. Pragmatic expectations further depend on what the speaker is assumed to know. A model representing only one of these variables may predict a narrow laboratory outcome without capturing the full human task. The objective is a layered explanatory system in which the contributions of signal processing, memory, probabilistic inference, and communicative knowledge can be disentangled experimentally while remaining coordinated during ordinary comprehension McClelland & Elman (1986) Levy (2008) Frank & Goodman (2012).
Behavioral Measurement and Model-Based Observation
A computational model becomes a psychological hypothesis when its internal variables are linked to measurements. Common observations include response latency, lexical decision accuracy, word-by-word self-paced reading, eye fixation duration, anticipatory gaze, and graded judgments of meaning. Each measure has its own causal pathway from latent comprehension state to recorded response. Reaction time includes visual encoding, motor preparation, and decision components; reading time can reflect anticipatory attention and later integration; acceptability judgments include metalinguistic strategies. A defensible model therefore specifies which component of human behavior its predictions describe and why changes in that component should be observable under the task Smith & Levy (2013) Tanenhaus et al. (1995).
Experimental control can expose differences among mechanisms that naturally covary. To test lexical competition, investigators can manipulate competitor similarity while holding the target's frequency or acoustic quality approximately constant. To test syntactic predictions, they can vary structural probabilities independently of dependency length when materials permit. To evaluate pragmatics, they can manipulate what possible utterances a speaker could have selected. These interventions constrain the interpretation of effects more directly than correlations in unmanipulated corpora. However, strict control can remove properties central to natural interaction, so ecological validity is an additional dimension of evidence rather than an optional embellishment McClelland & Elman (1986) Lewis & Vasishth (2005) Frank & Goodman (2012).
Individual variation requires hierarchical statistical analysis. People differ in reading skill, exposure, working-memory access, and familiarity with specific constructions, while words and sentences differ in length, lexical frequency, and discourse context. When a model is evaluated against repeated observations, dependencies among participants and items must be represented in the analysis. Even a reliable group-average effect can conceal qualitatively different strategies. Researchers should report held-out predictive performance, uncertainty around effect estimates, and residual patterns, and should analyze whether adding model features yields a meaningful improvement over conventional lexical and linguistic baselines Demberg & Keller (2008) Shain et al. (2024).
Naturalistic Reading Time and Surprisal Benchmarks
Naturalistic reading corpora give computational theories a test under diverse linguistic contexts rather than a handful of carefully designed sentences. Demberg and Keller used eye-tracking observations to compare accounts of syntactic processing cost, demonstrating how broad-coverage predictors can be related to human reading. The advantage is variation in lexical context, sentence structure, and discourse, including combinations that might be rare in a laboratory experiment. The corresponding challenge is that natural language variables are highly correlated. A credible statistical benchmark must compare appropriate baselines, model data dependencies, and establish whether improvements generalize to unseen readers and texts Demberg & Keller (2008).
Word predictability requires particular care because a measured effect depends on how probabilities are estimated. Smith and Levy argued from reading-time data that a logarithmic transformation of conditional probability offered a good account of processing cost. Shain and colleagues revisited this issue using large-scale evidence, directly examining functional form and statistical generalizability. This line of research shifts evaluation toward quantitative questions: does an information-theoretic model predict a continuous response, and does it do so after lexical and contextual factors are taken into account? The goal is to evaluate models against response distributions, not merely show that an unexpected word is sometimes difficult Smith & Levy (2013) Shain et al. (2024).
Modern language models can supply probability estimates with impressive coverage, but a lower perplexity is not automatically a better account of reading time. Training data may overlap with test corpora, tokenization may divide a human-recognized word into several units, and context length may exceed what participants could remember. The evaluation protocol should aggregate token probabilities appropriately, identify data contamination where possible, and compare multiple architectures under the same statistical observation model. If a language model predicts human latency, the inference is that its estimated quantity shares information with that latency. Stronger process claims require discriminating tests beyond predictive correlation Hu et al. (2020) Schrimpf et al. (2021).
Electrophysiological Evidence and Semantic Integration
Event-related brain potentials provide time-sensitive evidence about language processing. Kutas and Hillyard's foundational study reported a negative-going response associated with semantic incongruity, helping establish the N400 as a crucial object of investigation in sentence comprehension. The finding made it possible to assess linguistic expectations through physiology rather than only behavior after a sentence was understood. Yet an event-related waveform reflects the aggregate response to a stimulus under a task, and its psychological interpretation requires experiments that distinguish lexical properties, semantic fit, and attention. A component label is an empirical pattern rather than a direct readout of one uniquely identified algorithm Kutas & Hillyard (1980).
Computational accounts estimate word surprisal from sentence context and test whether this quantity predicts neural response amplitudes. Frank and colleagues examined naturally occurring sentences and reported a relation between surprisal estimates and the N400, while comparing different types of language models. Such analyses add evidence that probabilistic context sensitivity reaches beyond overt reading-time behavior. Nevertheless, the mapping is complex: the N400 is affected by the state of semantic knowledge, lexical association, and context-dependent accessibility. Correlation with surprisal demonstrates a useful quantitative constraint without determining the precise sequence of neural computations that produces the waveform Frank et al. (2015) Kutas & Hillyard (1980).
Electrophysiological validation benefits from converging analyses of timing, topography, stimulus manipulations, and individual-level response variation. One model can predict the same mean amplitude as another while differing substantially in whether it expects anticipation before word onset, rapid updating afterward, or later reinterpretation. Time-resolved analyses can reveal such disagreements when measurement uncertainty and temporal filtering are acknowledged. Researchers must guard against constructing physiological predictors from text that contains information unavailable to the participant at the corresponding moment. The relevant standard is alignment between the model's causal information state and the listener's actual experience Goldstein et al. (2022) Frank et al. (2015).
Neural Encoding Models of Language Comprehension
Neural encoding studies test whether features derived from computational systems predict patterns of brain activity during language comprehension. Schrimpf and colleagues compared neural network language models with human neural recordings and behavioral measures, reporting that models trained for predictive language tasks could capture important variance in human responses. This approach provides a quantitative bridge between an artificial representation and neural measurements. Its interpretation depends on the encoding procedure, the quality and composition of human recordings, the architecture of the model, and the distinction between statistically shared information and equivalent causal mechanisms Schrimpf et al. (2021).
Goldstein and colleagues examined naturalistic listening with electrocorticography and investigated anticipatory and post-onset relations between a language model and neural activity. Their evidence identifies computational parallels involving next-word expectations, responses to unexpected input, and context-sensitive representations. The study used a specific listening task and patient recordings, so the generality of its conclusions must be judged against population and measurement limits. Neural similarities are valuable observations that can motivate stronger experimental tests. They do not imply that artificial model units and biological neurons have a one-to-one relationship or that training histories are equivalent Goldstein et al. (2022).
Encoding accuracy is sensitive to shared statistical structure among candidate models. Different neural networks may learn correlated word representations, and many linguistic variables are themselves associated with frequency, length, position, and semantic similarity. An encoding model can predict brain signals through any feature carrying relevant covariance. Strong neuroscientific inference therefore calls for controlled baselines, partitioned training and test stimuli, temporally appropriate regressors, and independent manipulations. Where possible, complementary representational analyses and lesion-style model interventions can test whether the purportedly explanatory feature contributes uniquely. The psychological conclusion should match the evidence: a relation between measurable representations and responses rather than an unsupported identity of processes Schrimpf et al. (2021) Goldstein et al. (2022).
Modern Neural Language Models as Cognitive Candidates
Neural language models assign probabilities to sequences, learn context-sensitive representations, and can generalize beyond individual training examples. These capabilities make them useful as computational estimators of predictability and as candidate theories of representational organization. However, their scientific role must be specified. A model may be a measurement instrument for surprisal, a source of feature representations for neural encoding, a proposed algorithmic analogue of comprehension, or an artificial communicative system studied in its own right. Evidence supporting one role does not automatically validate the others. Clear research questions prevent a single benchmark score from carrying incompatible explanatory claims Hu et al. (2020) Schrimpf et al. (2021).
Systematic linguistic testing can reveal important differences between general language-model quality and structural competence. Hu and colleagues compared model architectures and training conditions across a battery of syntactic generalization tests, showing that predictive performance and grammatical generalization cannot simply be assumed to move together. This is consequential for cognitive modeling because people generalize structural relations to novel expressions, and a model that fits common text may rely on superficial correlates of those structures. A useful evaluation suite includes controlled minimal contrasts, variable-depth dependencies, unfamiliar lexical combinations, and tests developed independently of model training Hu et al. (2020).
Modern models also differ from human learners in exposure, embodiment, objectives, and the way their outputs are evaluated. A system optimized on large text corpora may offer strong predictions of lexical probabilities without matching how children learn from speech and interaction. A well-fitting neural encoding model can reveal shared information dimensions while leaving the origin and functional role of those dimensions unsettled. Research should therefore compare models under explicit task alignment and report the limits of access to training data, tokenization, and computational resources. Contemporary systems expand the theoretical search space, but scientific explanation still depends on discriminative evidence from human behavior Saffran et al. (1996) Goldstein et al. (2022) Hu et al. (2020).
Structural Generalization and Competing Explanations
One of the most consequential disputes concerns whether the linguistic predictions observed in human comprehension depend on explicit hierarchical representations. Frank and Bod compared probabilistic models with different structural commitments and reported that hierarchical variants did not explain unique reading-time variation beyond sequential alternatives in their analysis. This finding challenges simple assumptions that a model's grammatical hierarchy must be directly recoverable from every measure of processing time. It does not show that hierarchy is absent from human grammar or irrelevant to all processing tasks. The inference is narrower: under the tested materials, estimators, and statistical comparison, hierarchical predictions offered limited additional explanatory value Frank & Bod (2011).
Other experimental and computational results emphasize the cognitive importance of structural relations. Working-memory retrieval models derive difficulty from how grammatical dependencies are represented and accessed; systematic tests of neural networks show that models vary in generalization across syntactic constructions despite similar predictive objectives. These findings motivate distinguishing three questions: whether a model represents hierarchical structure, whether that representation improves word prediction, and whether the same representation causally contributes to human processing. The questions may receive different answers for different constructions and tasks. A rigorous comparison builds materials on which sequential and hierarchical accounts make genuinely divergent predictions Lewis & Vasishth (2005) Hu et al. (2020) Frank & Bod (2011).
The appropriate evidential approach is a program of adversarial model comparison. Researchers can preregister the target contrast, assess competing models before analyzing human outcomes, and use held-out behavioral observations to evaluate their distinct predictions. Residual analysis matters as much as overall goodness of fit: systematic failure for long-distance agreement, temporarily misleading contexts, or rare constructions can identify limitations masked by a favorable aggregate score. Computational models should also be tested after changes to training distribution and linguistic register. A narrow success is informative when carefully bounded; a broad explanation requires predictive durability across the contexts in which its mechanism is claimed to operate Hu et al. (2020) Shain et al. (2024).
Causal Testing and Identifiability of Mechanisms
Computational models often occupy an ambiguous evidential position because many parameter settings and architectures can reproduce the same observations. A difference in aggregate fit may reflect representational power, training data, a convenient measurement transformation, or an actual process advantage. Identifiability is therefore a substantive psychological concern. If two architectures assign similar probabilities to the same words, reading-time correlations alone cannot determine which architecture resembles human cognition. Stronger inference requires tests designed to expose different internal computations through observable consequences. Manipulating competitor identity, structural interference, and communicative alternatives can supply the necessary discriminating evidence Levy (2008) Lewis & Vasishth (2005) Frank & Goodman (2012).
Computational ablation is one useful method when its causal interpretation is kept precise. Removing feedback, restricting memory capacity, changing retrieval cues, or excluding syntactic information reveals how a model's predictions depend on particular design elements. But an ablation demonstrates a mechanism's importance within the tested architecture, rather than its necessity in human cognition. Corresponding human experiments must manipulate independent variables thought to engage the proposed mechanism and measure outcomes under comparable information states. A strong causal argument thus combines internal simulation interventions with external behavioral interventions and identifies where the analogy might break down McClelland & Elman (1986) Lewis & Vasishth (2005) Hu et al. (2020).
Reproducibility further depends on stable definitions and complete methodological reporting. Investigators need access to stimulus lists, model version, preprocessing, token aggregation, fitted parameters, cross-validation design, and statistical analysis code when sharing is permissible. Neural recordings introduce additional issues of sensor selection, trial exclusion, temporal smoothing, and choice of response windows. Researchers should compare against low-complexity baselines and examine calibration as well as ranking or correlation. Robustness checks against alternate corpora, participants, and linguistic domains reveal whether a model captures an enduring psychological regularity or an idiosyncrasy of a dataset Schrimpf et al. (2021) Goldstein et al. (2022) Shain et al. (2024).
Cross-Linguistic Constraints and Population Generalization
A model of human language processing should ultimately explain more than written English. Languages vary in word order, morphological complexity, lexical segmentation, orthographic systems, and the informational role of context. Futrell and colleagues found evidence for shorter-than-chance syntactic dependency lengths across a broad set of languages, offering a quantitative constraint relevant to processing-efficiency theories. Such results can reveal regularities not apparent from one language's laboratory materials. At the same time, cross-linguistic corpus analysis requires attention to annotation conventions, sampling, linguistic relatedness, and whether the chosen baseline captures realistic alternative structures Futrell et al. (2015).
Different processing tasks impose different accessibility constraints across populations. Word-by-word reading times cannot be generalized directly to spoken dialogue, developmental acquisition, or signed languages, where perceptual and temporal structures differ. Research must consider literacy, age, learning history, sensory modality, and cultural expectations. A neural language model trained mostly on a narrow collection of languages may assign poorly calibrated probabilities to another language while still generating superficially grammatical sentences. Human validity requires matched exposure and task conditions where feasible, and transparent assessment of the populations represented by the evidence Saffran et al. (1996) Christiansen & Chater (2016) Hu et al. (2020).
Cross-linguistic comparison also offers unusually powerful opportunities to distinguish universal processing constraints from historically contingent conventions. Dependency locality, information density, statistical learning, and contextual inference can each be formulated as general principles, but their observable manifestations need not be identical. A model should predict what changes when morphology makes grammatical dependencies explicit, when flexible word order shifts anticipation, or when communicative alternatives carry different cultural expectations. Cross-population studies should therefore test theoretically specified interactions, document departures from predictions, and resist treating a shared average tendency as evidence of one invariant internal algorithm Futrell et al. (2015) Jaeger (2010) Goodman & Frank (2016).
Human and Artificial Language Processing in the Artificial Era
The Artificial Era introduces artificial language systems as important scientific comparators for human cognition. Their ability to estimate continuations, maintain context-dependent representations, and produce pragmatically plausible responses creates opportunities to test theories that previously required simplified simulations. Goldstein and colleagues reported neural correlates of processing principles also present in a deep language model, and Schrimpf and colleagues identified useful correspondence between model features and language-related human measurements. These are significant cross-system observations. They support the investigation of shared computational regularities while leaving open detailed questions about learning histories, perceptual grounding, and the causal origin of behavior Goldstein et al. (2022) Schrimpf et al. (2021).
Comparative interpretation requires distinct levels of attribution. A mathematical resemblance concerns information-processing structure; behavioral correspondence concerns responses in particular tasks; process correspondence requires evidence about the internal transformation of representations; neural correspondence concerns relations to biological measurements. Human comprehension also develops through bodily perception, action, social participation, and linguistic communities. Artificial models can exhibit advanced linguistic generalization under training regimes that differ dramatically from human development. The scientific aim is to identify which abstractions transfer across systems and which claims depend on the specific organization of each system. A common vocabulary can support comparison when the operational definitions remain explicit Saffran et al. (1996) Hu et al. (2020) Goldstein et al. (2022).
Several research opportunities follow. Models can be tested on real-time interactive tasks, adaptation to unfamiliar speakers, referential repair, incremental belief revision, and robust transfer between modalities. Comparisons should measure how predictions change after new evidence and whether a system reproduces distinctive patterns of human errors, latency, and sensitivity to context. The interface between computational psycholinguistics and the psychology of artificial systems should thus be built through shared experimental paradigms and carefully distinguished constructs. Reliable artificial language performance is itself an important research object, but its interpretation as a model of human cognition requires additional empirical validation Frank & Goodman (2012) Tanenhaus et al. (1995) Schrimpf et al. (2021).
Conclusion
Computational models of language processing offer an increasingly precise science of how linguistic evidence is transformed into recognition, expectation, interpretation, learning, and communicative choice. The field's foundations include cohort and interactive-activation models of word perception, distributed sequence learning, probabilistic lexical and syntactic inference, activation-based memory retrieval, and Bayesian accounts of pragmatic reasoning. Each tradition identifies a different set of variables that can be measured or manipulated. Their scientific value lies in linking explicit computational assumptions to discriminating behavioral predictions, rather than in providing interchangeable metaphors for understanding. Historical developments become cumulative when models clarify and test the boundaries of preceding explanations Marslen-Wilson (1987) McClelland & Elman (1986) Jurafsky (1996) Goodman & Frank (2016).
The strongest evidence for contemporary quantitative theories comes from convergence among controlled experiments, naturalistic reading data, developmental learning studies, electrophysiology, and neural recordings. Surprisal provides a useful link between contextual probability and measured language processing, while memory retrieval, structural generalization, and situated reference establish additional demands on any comprehensive account. Neural language models supply powerful predictors and representations, but model performance and cognitive equivalence remain distinct empirical propositions. Scientific progress depends on held-out evaluation, carefully defined observation models, causal contrasts, transparent training provenance, and investigation of failure cases Smith & Levy (2013) Shain et al. (2024) Lewis & Vasishth (2005) Goldstein et al. (2022).
The unresolved research agenda concerns integration across representational levels and timescales: how learning shapes anticipation, how memory interacts with prediction, how social inference changes interpretation before an utterance ends, and how neural activity supports these processes. It also concerns comparisons across languages, populations, modalities, and artificial systems. The decisive methodological principle is that each proposed mechanism must be tied to an observable consequence capable of distinguishing it from a plausible rival. On this foundation, computational psycholinguistics can continue developing explanatory models of human language while providing a disciplined comparative framework for the emerging language capacities of the Artificial Era Christiansen & Chater (2016) Futrell et al. (2015) Schrimpf et al. (2021).
References
- Christiansen, M. H., & Chater, N. (2016). The Now-or-Never bottleneck: A fundamental constraint on language. Behavioral and Brain Sciences, 39, e62. https://doi.org/10.1017/S0140525X1500031X
- Demberg, V., & Keller, F. (2008). Data from eye-tracking corpora as evidence for theories of syntactic processing complexity. Cognition, 109(2), 193–210. https://doi.org/10.1016/j.cognition.2008.07.008
- Elman, J. L. (1990). Finding structure in time. Cognitive Science, 14(2), 179–211. https://doi.org/10.1207/s15516709cog1402_1
- Frank, S. L., & Bod, R. (2011). Insensitivity of the human sentence-processing system to hierarchical structure. Psychological Science, 22(6), 829–834. https://doi.org/10.1177/0956797611409589
- Frank, M. C., & Goodman, N. D. (2012). Predicting pragmatic reasoning in language games. Science, 336(6084), 998. https://doi.org/10.1126/science.1218633
- Frank, S. L., Otten, L. J., Galli, G., & Vigliocco, G. (2015). The ERP response to the amount of information conveyed by words in sentences. Brain and Language, 140, 1–11. https://doi.org/10.1016/j.bandl.2014.10.006
- Futrell, R., Mahowald, K., & Gibson, E. (2015). Large-scale evidence of dependency length minimization in 37 languages. Proceedings of the National Academy of Sciences, 112(33), 10336–10341. https://doi.org/10.1073/pnas.1502134112
- Goldstein, A., Zada, Z., Buchnik, E., Schain, M., Price, A., Aubrey, B., Nastase, S. A., Feder, A., Emanuel, D., Cohen, A., Jansen, A., Gazula, H., Choe, G., Rao, A., Kim, C., Casto, C., Fanda, L., Doyle, W., Friedman, D., ... Hasson, U. (2022). Shared computational principles for language processing in humans and deep language models. Nature Neuroscience, 25, 369–380. https://doi.org/10.1038/s41593-022-01026-4
- Goodman, N. D., & Frank, M. C. (2016). Pragmatic language interpretation as probabilistic inference. Trends in Cognitive Sciences, 20(11), 818–829. https://doi.org/10.1016/j.tics.2016.08.005
- Hale, J. (2001). A probabilistic Earley parser as a psycholinguistic model. In Proceedings of the Second Meeting of the North American Chapter of the Association for Computational Linguistics (pp. 159–166). Association for Computational Linguistics. https://doi.org/10.3115/1073336.1073357
- Hu, J., Gauthier, J., Qian, P., Wilcox, E., & Levy, R. P. (2020). A systematic assessment of syntactic generalization in neural language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (pp. 1725–1744). Association for Computational Linguistics. https://doi.org/10.18653/v1/2020.acl-main.158
- Jaeger, T. F. (2010). Redundancy and reduction: Speakers manage syntactic information density. Cognitive Psychology, 61(1), 23–62. https://doi.org/10.1016/j.cogpsych.2010.02.002
- Jurafsky, D. (1996). A probabilistic model of lexical and syntactic access and disambiguation. Cognitive Science, 20(2), 137–194. https://doi.org/10.1016/S0364-0213(99)80005-6
- Kutas, M., & Hillyard, S. A. (1980). Reading senseless sentences: Brain potentials reflect semantic incongruity. Science, 207(4427), 203–205. https://doi.org/10.1126/science.7350657
- Levy, R. (2008). Expectation-based syntactic comprehension. Cognition, 106(3), 1126–1177. https://doi.org/10.1016/j.cognition.2007.05.006
- Lewis, R. L., & Vasishth, S. (2005). An activation-based model of sentence processing as skilled memory retrieval. Cognitive Science, 29(3), 375–419. https://doi.org/10.1207/s15516709cog0000_25
- Marslen-Wilson, W. D. (1987). Functional parallelism in spoken word-recognition. Cognition, 25(1–2), 71–102. https://doi.org/10.1016/0010-0277(87)90005-9
- McClelland, J. L., & Elman, J. L. (1986). The TRACE model of speech perception. Cognitive Psychology, 18(1), 1–86. https://doi.org/10.1016/0010-0285(86)90015-0
- Piantadosi, S. T., Tily, H., & Gibson, E. (2011). Word lengths are optimized for efficient communication. Proceedings of the National Academy of Sciences, 108(9), 3526–3529. https://doi.org/10.1073/pnas.1012551108
- Saffran, J. R., Aslin, R. N., & Newport, E. L. (1996). Statistical learning by 8-month-old infants. Science, 274(5294), 1926–1928. https://doi.org/10.1126/science.274.5294.1926
- Schrimpf, M., Blank, I. A., Tuckute, G., Kauf, C., Hosseini, E. A., Kanwisher, N., Tenenbaum, J. B., & Fedorenko, E. (2021). The neural architecture of language: Integrative modeling converges on predictive processing. Proceedings of the National Academy of Sciences, 118(45), e2105646118. https://doi.org/10.1073/pnas.2105646118
- Seidenberg, M. S., & McClelland, J. L. (1989). A distributed, developmental model of word recognition and naming. Psychological Review, 96(4), 523–568. https://doi.org/10.1037/0033-295X.96.4.523
- Shain, C., Meister, C., Pimentel, T., Cotterell, R., & Levy, R. (2024). Large-scale evidence for logarithmic effects of word predictability on reading time. Proceedings of the National Academy of Sciences, 121(10), e2307876121. https://doi.org/10.1073/pnas.2307876121
- Smith, N. J., & Levy, R. (2013). The effect of word predictability on reading time is logarithmic. Cognition, 128(3), 302–319. https://doi.org/10.1016/j.cognition.2013.02.013
- Tanenhaus, M. K., Spivey-Knowlton, M. J., Eberhard, K. M., & Sedivy, J. C. (1995). Integration of visual and linguistic information in spoken language comprehension. Science, 268(5217), 1632–1634. https://doi.org/10.1126/science.7777863