Sentence processing is the moment-by-moment construction of grammatical structure and meaning as spoken, written, or signed language unfolds. A comprehender identifies words, establishes their relationships, assigns participants to events, revises provisional interpretations, and integrates the result with context. These operations begin before a sentence is complete. They are usually fast and efficient, yet their limitations become visible in ambiguity, long-distance dependencies, unexpected continuations, and misleading interpretations. Psycholinguistics investigates the timing and computational organization of these processes with behavioral, eye-movement, neural, and computational evidence. Levy (2008); Reichle (2021).
A simple contrast illustrates why grammar is central. In “The scientist admired the musician,” word order and grammatical relations establish who admires whom. In “The musician the scientist admired smiled,” the same two nouns must temporarily be maintained while the reader determines which noun is the object of “admired” and which is the subject of “smiled.” Comprehension involves structure-building under time pressure. Understanding the meanings of separate words is necessary, but determining how the words combine is an additional cognitive achievement.
What Sentence Processing Means in Psycholinguistics
Grammar, parsing, and interpretation
Syntax describes the structural relationships that permit words and phrases to form sentences. Parsing refers to the processes through which a language user constructs or accesses those relationships in real time. The resulting representation can encode phrase boundaries, grammatical functions, dependencies between distant elements, and the semantic roles of participants. Semantic composition combines lexical meanings according to a structural analysis and other interpretive constraints. These processes interact with reference, discourse history, world knowledge, and communicative goals.
A representation of “The boy chased the dog” must distinguish the boy as the agent of the chase and the dog as its target. Changing the word order in English usually changes the roles. In another language, case endings or agreement marking can carry more of that information. A parser consequently uses language-specific evidence; grammatical processing cannot be equated with one universal routine based on English word order. Structural theories describe the organization of the representation, whereas processing theories additionally predict reading times, listening behavior, errors, and how a developing representation changes after each new word. MacDonald, Pearlmutter, and Seidenberg (1994); Jaeger and Tily (2011).
Incrementality: interpretation before the final word
Incremental comprehension means that information is used as it arrives, rather than being stored as an uninterpreted sequence until the sentence ends. The opening “The teacher who…” creates expectations about an upcoming relative clause. When the verb arrives, the comprehender relates it to the available participants and interprets the construction. Continuing words confirm, elaborate, or challenge that analysis. Laboratory measures reveal systematic costs at particular regions of sentences, demonstrating that structure-building occurs during reading and listening. Incrementality does not imply that every sentence receives an immediately complete, fully correct analysis; sometimes multiple possibilities remain active or only a provisional interpretation is available. Levy (2008); Smith and Levy (2013).
Competence, performance, and understanding
Grammatical knowledge concerns the structures a language user can represent. Processing performance concerns the mechanisms that construct and retrieve those structures under time, memory, attention, and perceptual constraints. A grammatically legitimate sentence can be difficult to process. A grammatically anomalous sentence can nevertheless support a plausible interpretation. Accuracy on a comprehension question, reading speed, and grammaticality judgment therefore measure related but nonidentical outcomes. A convincing model of sentence processing must say what its parser represents, what information it uses, when that information matters, and why some representations are difficult to establish or maintain. Ferreira, Bailey, and Ferraro (2002); Reichle (2021).
How a Sentence Is Built in Real Time
Identifying incoming words and their grammatical possibilities
Before a sentence can be interpreted, its word forms must be recognized. Recognition activates possible meanings and grammatical properties, such as whether a form is a noun or verb, what complements a verb normally takes, and which inflections are present. A frequent verb may strongly favor one construction while still permitting another. Such distributional knowledge influences the interpretation of ambiguous input. The processing of a word is affected by frequency, context, morphological structure, perceptual clarity, and the alternatives that remain compatible with the input. Lexical access and sentence structure jointly constrain the unfolding interpretation. MacDonald, Pearlmutter, and Seidenberg (1994).
Establishing phrase structure and dependencies
Words are organized into hierarchical groups. In “The old paintings in the gallery attracted visitors,” the phrase “in the gallery” modifies the paintings rather than serving as the subject of “attracted.” The comprehender must track the head of the subject phrase and associate it with the verb despite intervening material. Dependencies can also cross clauses. In “Which painting did the curator say the museum acquired?”, the question phrase must be connected with a position after “acquired.” Psycholinguistic accounts differ in whether they characterize such links through discrete syntactic operations, cue-driven retrieval, probabilistic alternatives, learned constructions, or mixtures of these mechanisms. Evidence from response-time and interference paradigms shows that distant grammatical material creates real retrieval demands. Lewis, Vasishth, and Van Dyke (2006); Gibson (1998).
Assigning thematic roles and constructing meaning
A sentence specifies events and their participants. Agent, patient, experiencer, recipient, and instrument are descriptive labels for recurring semantic roles, although their implementation differs across theories. The verb “give” creates expectations about someone giving something to someone. Animacy and plausibility help identify likely role assignments, but contextual plausibility must be evaluated together with grammatical relationships. “The evidence examined by the lawyer…” can favor a reduced-relative reading because evidence is an unlikely examiner; “The defendant examined by the lawyer…” more easily supports an initial main-clause interpretation. Experiments manipulating animacy and verb meaning have shown early effects on ambiguity resolution, challenging an absolute separation of first-pass syntax from all semantic information. Trueswell, Tanenhaus, and Garnsey (1994).
Maintaining and revising a situation model
Comprehenders also construct representations of the described events, participants, locations, and relationships. When a later clause identifies a different actor, the situation model must change. Discourse context can establish a referent before the sentence is finished; an apparently ambiguous phrase becomes easier when the setting makes one reading salient. This coordination of grammatical and situational information is demonstrable in eye movements during spoken instructions. A model restricted to sentence-internal words misses effects of the listener’s visible environment and task. Tanenhaus, Spivey-Knowlton, Eberhard, and Sedivy (1995); Eberhard, Spivey-Knowlton, Sedivy, and Tanenhaus (1995).
Ambiguity: When Grammar Allows More Than One Analysis
Temporary structural ambiguity
Words often admit several grammatical continuations. The beginning “While the musician played the song…” encourages the interpretation that the song is the object of “played.” Continuing it as “While the musician played the song was recorded” forces a new interpretation: the song is the subject of the main clause. The difficulty occurs where the sentence disconfirms the earlier attachment. A garden-path effect is the extra processing caused when an initially attractive analysis has to be reconsidered. The important scientific observation is the timing and conditions of the disruption, not the mere fact that readers sometimes make mistakes. Reichle (2021).
Why a garden path happens
A structural preference may arise from a simple attachment procedure, learned distributional expectations, lexical constraints, semantic plausibility, discourse context, or several factors together. Classic syntax-first models propose that early analysis strongly prioritizes a limited set of structural principles. Constraint-based models allow syntactic, lexical, semantic, and contextual evidence to influence competing analyses from an early point. These accounts need discriminating experiments, because an observed reading delay can be explained by different combinations of early commitment, competition, repair, and memory retrieval. The substantive question is which source of information changes the probability of an analysis and at what time. MacDonald, Pearlmutter, and Seidenberg (1994); Trueswell, Tanenhaus, and Garnsey (1994).
Revision is sometimes incomplete
Readers may detect a grammatical inconsistency yet retain a lingering interpretation. The good-enough approach proposes that ordinary comprehension can settle on meanings that are adequate for the task while remaining incomplete or inconsistent with a final syntactic analysis. In misleading passive and garden-path sentences, comprehenders sometimes report an event that is plausible but different from the literal grammatical content. Such findings establish that successful communication and fully detailed parsing are distinct outcomes. They do not establish that grammar is unused; they show that the route from form to interpretation admits heuristics and incomplete repair. Ferreira, Bailey, and Ferraro (2002).
Major Theories of Sentence Comprehension
Serial, syntax-first, and garden-path models
In strong serial approaches, the parser initially selects a structure using structural principles and revises it when new input makes the choice untenable. Explanatory virtues include precise predictions about an initial preferred analysis and the location of reanalysis costs. The main challenge is explaining early effects of verb meaning, referential context, animacy, and other sources of information. Evidence favoring early contextual influence is compatible with a range of revised serial theories, so the field evaluates specific models rather than treating every structure-first formulation as identical. Comparative accounts of parsing theories set out their competing predictions for well-controlled constructions. Reichle (2021); Trueswell, Tanenhaus, and Garnsey (1994).
Constraint-based, interactive, and experience-sensitive models
Constraint-based accounts represent grammatical alternatives as supported to different degrees by multiple information sources. Word frequency, verb preferences, argument structure, semantic fit, discourse expectations, and construction experience can jointly influence which reading becomes dominant. A person’s experience with a verb in one structure may make that construction more accessible. Such models explain graded preferences and early contextual effects but must specify how the constraints are learned, integrated, and converted into quantitative processing predictions. Simply saying that “everything interacts” is insufficient; the theory must predict when a particular manipulation changes reading, listening, or eye-movement behavior. MacDonald, Pearlmutter, and Seidenberg (1994); Trueswell, Tanenhaus, and Garnsey (1994).
Memory-based and dependency-locality accounts
A structure may be hard to comprehend because the processor must retain unresolved dependencies and later retrieve a distant element. Gibson’s locality account separates memory demands for maintaining predicted grammatical categories from integration costs when distant material is connected. Retrieval-based models describe memory as activated representations accessed by cues; similarity among candidate items can produce interference. These theories differ in their proposed source of difficulty. Distance in words, number of intervening discourse referents, similarity of distractors, and cue diagnosticity are not interchangeable quantities. Gibson (1998); Lewis, Vasishth, and Van Dyke (2006).
Probabilistic and expectation-based models
Expectation-based approaches relate processing effort to the probabilities assigned to upcoming linguistic continuations. In surprisal theory, a word that is improbable given prior context imposes more information-processing cost than a predictable one. Surprisal can be written as −log P(word | context). This formal quantity makes graded predictions that can be estimated from controlled expectations or computational language models. Levy demonstrated how probabilistic incremental parsing can connect disambiguation to processing difficulty. The theory addresses expectation and update costs, while dependency locality and retrieval interference address other possible sources of difficulty; their predictions sometimes overlap and sometimes diverge. Levy (2008); Demberg and Keller (2008).
Good-enough comprehension and interpretive heuristics
Good-enough theories focus on the representational end product: the meaning a comprehender actually retains. Heuristics favoring typical agent–patient assignments may coexist with detailed grammatical representations, and task demands alter whether errors become evident. When a comprehension question invites a superficial plausible response, errors may appear that are less common under tasks emphasizing exact syntactic relationships. This family of findings redirects attention from grammatical decoding alone toward what kind of representation is constructed for a specific goal. Ferreira, Bailey, and Ferraro (2002).
Models as testable alternatives rather than slogans
A theory earns explanatory force by predicting patterns across constructions, languages, populations, and methods. A reading-time difference does not uniquely identify a mechanism. For example, a difficult relative clause might involve longer dependencies, interference, an infrequent word sequence, and an unexpected thematic assignment. Investigators distinguish the candidates through carefully matched conditions, quantitative fits, and measures taken at the region where the theories disagree. Reviews of modeling work show that no single classic account explains every result, making explicit comparison and replication indispensable. Reichle (2021).
What Makes a Sentence Difficult to Process?
Ambiguous attachment and reanalysis
The number of possible analyses matters, but the key determinant is often their relative support. If one continuation is strongly expected, disconfirming it may require a substantial revision. If context keeps several analyses plausible, the processor may resolve uncertainty gradually. The location of measurable cost therefore depends on the structure, verb preferences, plausibility, and context in which the critical word appears. An experiment must distinguish the initial ambiguity from the moment the ambiguity is resolved. Trueswell, Tanenhaus, and Garnsey (1994).
Long-distance dependencies and retrieval interference
Compare “The reporter who questioned the witness left” with “The reporter who the witness questioned left.” The second construction often places heavier demands on the parser because the relative-clause head must fill a noncanonical role and compete with another plausible noun. English subject and object relatives reveal reliable difficulty differences, although their size depends on animacy and other constraints. Eye-movement experiments found higher costs for object relatives and demonstrated that making the main noun inanimate can reduce that cost. Such interactions support theories that combine structure with semantic and retrieval demands. Traxler, Morris, and Seely (2002).
Agreement and deceptive grammatical cues
In “The key to the cabinets is missing,” the main subject is singular although a plural noun intervenes. A mismatching verb can sometimes seem easier to process when a plural distractor is present, an effect known as agreement attraction. This does not necessarily mean that the listener has changed the grammatical number of the subject in memory. Experiments showing different patterns for grammatical and ungrammatical sentences favor accounts involving cue-based retrieval under particular conditions. Agreement phenomena are especially valuable because they distinguish grammatical constraints from the fallible memory operations used to implement them. Wagers, Lau, and Phillips (2009).
Frequency, predictability, and information
A highly predictable next word is generally processed more efficiently than an unexpected one. Studies using natural reading data have found systematic relationships between word probability and reading time, including evidence for a logarithmic relationship. Yet lexical predictability is influenced by sentence context, word frequency, and the statistical properties of the corpus or model used to estimate it. A model that predicts reading duration may accurately characterize difficulty without explaining every representational operation that causes it. The strongest claims require held-out validation and experimental contrasts. Smith and Levy (2013); Staub (2015).
Semantic plausibility and discourse context
The human system uses event knowledge to assess whether a noun is a plausible agent or object of a verb. Referential context can eliminate some structural preferences, particularly in situations where the listener is choosing among visible objects. Eye movements in the visual-world paradigm show how candidate referents become active as speech arrives and how visual context interacts with linguistic structure. These observations support immediate integration under well-defined conditions, but a visually grounded instruction differs from reading an isolated laboratory sentence. The temporal and functional consequences of context cannot be assumed identical across settings. Tanenhaus, Spivey-Knowlton, Eberhard, and Sedivy (1995).
Prosody, morphology, and modality
Speech offers stress, phrasing, pitch movement, rhythm, and timing that can guide the grouping of words. Written language provides spacing and punctuation, while signed languages convey linguistic structure in visual-manual form. Morphological marking, such as case or agreement, can signal grammatical relationships before the listener reaches a clause-final verb. A general model of sentence processing must allow these cues to carry different weights across languages and modalities. Observations from English text should be identified as such; they are an important empirical domain, not a default description of every language.
The Time Course of Understanding: What Experiments Measure
Self-paced reading
In self-paced reading, participants reveal one word or phrase at a time and their response times are recorded. Longer times at or after a critical region indicate processing disruption or additional effort, but can reflect several mechanisms. Reading pace is controlled by the task, and sentence regions may influence subsequent regions through spillover. Research designs therefore compare matched conditions and evaluate the distribution of effects rather than treating one slow response as direct evidence of reanalysis. Self-paced reading is useful for testing precisely positioned structural contrasts and dependency manipulations. Wagers, Lau, and Phillips (2009).
Eye tracking during reading
Eye tracking measures where and when a reader looks. First-fixation duration, gaze duration, skipping, regressions, and later rereading provide complementary information about the trajectory of comprehension. Earlier measures often reveal lexical and initial integration effects, whereas regressions may signal difficulty and repair, but the mapping from any one eye movement to a single processing stage is imperfect. Word length, frequency, layout, launch position, and parafoveal preview must be considered. Natural-text corpora are especially valuable for estimating the simultaneous contributions of lexical properties, syntactic dependencies, and probabilistic information. Demberg and Keller (2008); Staub (2015).
Visual-world eye tracking during listening
In the visual-world paradigm, a person hears spoken language while looking at an arrangement of objects or pictures. The observer’s gaze indicates which referents have become plausible as the linguistic input develops. Experiments with spoken instructions established that eye movements follow rapidly changing linguistic information and that the visual scene can influence syntactic interpretation. The method is powerful for referential and interactive language, while its inference depends on the available objects, task goals, and the delay between hearing a word and moving the eyes. Eberhard, Spivey-Knowlton, Sedivy, and Tanenhaus (1995); Tanenhaus, Spivey-Knowlton, Eberhard, and Sedivy (1995).
Event-related potentials: N400 and P600
Electroencephalography measures electrical activity time-locked to words or other events. The N400 is sensitive to factors associated with access to and integration of meaning, including contextual expectancy; its amplitude is not an unambiguous indicator of a single cognitive operation. The P600 is often observed with syntactic anomalies, difficult reanalysis, and some semantic violations. The historical pairing of N400 with semantics and P600 with syntax proved too simple: their functional interpretation depends on experimental design and the processing demands involved. Neural components offer constraints on processing timescales, not a one-to-one anatomical map of grammatical rules. Kutas and Federmeier (2011); Kuperberg (2007).
Brain imaging and the limits of localization
Functional MRI identifies activity associated with language processing at the network level. Contemporary research describes a language-selective network spanning frontal and temporal regions, interacting with perceptual and wider knowledge systems. Combining words and constructions contributes to the work of this network, but temporal detail is constrained by hemodynamic measures. Current discussions emphasize that rapid syntactic and semantic computations may require methods with much finer temporal resolution, including magnetoencephalography and intracranial recordings. One activated brain area is not itself proof of a unique parsing algorithm. Fedorenko, Ivanova, and Regev (2024); Murphy and Woolnough (2024).
Acceptability judgments, comprehension questions, and recall
Judgments about whether a sentence sounds grammatical can reveal linguistic preferences, but those judgments may rely on reflection and differ from spontaneous comprehension. Questions about who did what to whom probe the eventual interpretation. Recall can identify which propositions survived processing. Together these measures reveal an essential methodological distinction: online effort at a specific word, final interpretation, and explicit linguistic judgment are related outcomes with different time courses. A detailed model should explain where they converge and why they sometimes diverge. Ferreira, Bailey, and Ferraro (2002).
Grammar and Meaning: A Coordinated Process
Syntax constrains semantic composition
The sentence “The lawyer interviewed the witness” and its passive equivalent “The witness was interviewed by the lawyer” describe similar central events with different grammatical configurations. Interpreting the passive requires mapping the subject of the sentence to a nonagent role. A comprehender cannot derive that relationship from word identities alone. Yet plausibility can also shape expectations for structural alternatives. Models differ over whether such information is combined continuously or enters at different stages. Experiments must measure both sensitivity to explicit markers such as “by” and the influence of semantic fit before those markers appear. Trueswell, Tanenhaus, and Garnsey (1994); Ferreira, Bailey, and Ferraro (2002).
Prediction is broader than guessing one exact word
A listener hearing “The student opened the…” may anticipate a noun phrase or an object type without selecting one specific next word. Prediction may involve grammatical categories, meaning features, structural continuations, or particular lexical forms. Many experiments demonstrate facilitatory effects of context. A stronger theory proposing obligatory preactivation of exact words, across every situation and population, requires separate evidence. Reviews of electrophysiology and eye movements caution that prediction benefits and costs need independent specification; graded facilitation can arise without categorical commitment to one lexical item. Van Petten and Luka (2012); Staub (2015).
Context can be helpful and misleading
World knowledge accelerates plausible interpretations and can also strengthen an incorrect interpretation when a sentence has unusual syntax. A reader who expects events to unfold conventionally may rely on that expectation when assigning agents and patients. At other times, contextual information helps disambiguate early enough to prevent a garden path. The same broad source of information can thus support fluency or bias, depending on the structure and task. Good experimental inference requires distinguishing the actual syntactic analysis from the final plausible event model. Ferreira, Bailey, and Ferraro (2002); Tanenhaus, Spivey-Knowlton, Eberhard, and Sedivy (1995).
Individual Differences and Learning
Linguistic experience and construction frequency
Comprehension is shaped by exposure to particular words, constructions, and discourse settings. A structure that is frequent in one language or register may be unusual in another. Learning can adjust expectations and lexical–syntactic associations. A model claiming that all readers use exactly the same attachment preference regardless of experience must therefore face quantitative tests across populations and materials. Variation in frequency is not itself proof of a purely experience-based grammar; it indicates an empirical influence that competing theoretical accounts must explain. MacDonald, Pearlmutter, and Seidenberg (1994); Jaeger and Tily (2011).
Working memory and attention
Temporary representations must remain available while the parser connects distant elements. People vary in performance on memory tasks, but a correlation between a laboratory memory score and sentence accuracy does not identify the precise memory operation responsible. Cue-based retrieval predicts specific interference effects, while locality theories emphasize costs associated with unresolved relations. Attentional demands and reading strategies interact with these mechanisms. Cognitive resources are therefore defined more precisely when a theory identifies maintenance, retrieval, interference, or control operations instead of attributing every difficult sentence to “low working memory.” Lewis, Vasishth, and Van Dyke (2006).
Second-language processing and multilingual experience
Language learners differ in vocabulary access, familiarity with grammatical markers, exposure to constructions, proficiency, literacy, and task demands. Their processing profiles can change with practice and language experience. A delay when comprehending a rare construction in a second language indicates an interaction among knowledge, expectations, and processing resources; it does not on its own indicate a disorder. Cross-linguistic comparisons must respect different word orders, inflectional cues, and structural frequencies rather than importing the predictions of English sentence processing as universal benchmarks. Jaeger and Tily (2011).
Children, aging, and accessibility
Incremental interpretation develops as linguistic knowledge and processing skills accumulate. Age, hearing and vision, attention, task design, and reading experience may affect performance. A slower response in an unfamiliar format says little about an individual's full comprehension capacity. Accessible presentation can reduce extraneous demands, but a claim about a universally optimal sentence length or syntactic form would require population- and task-specific evidence. Studies of ordinary development and healthy variation are distinguished from specialized clinical evaluation.
What Sentence-Processing Research Says About Disorders
A person can misparse a complex sentence because of ordinary ambiguity, fatigue, unfamiliar syntax, perceptual conditions, or limited context. Such instances are features of normal cognition and are not diagnoses. Developmental language disorder, aphasia, dyslexia, and other clinical conditions have different criteria, developmental or acquired histories, and mechanisms that need dedicated evaluation. Sentence-comprehension difficulties can be among their manifestations, but isolated reading times or classroom observations do not establish a clinical label. Research on ordinary grammatical processing can inform test design and theoretical interpretation; differential diagnosis and treatment require condition-specific evidence and qualified assessment. This article maintains the scientific scope of sentence processing rather than making clinical recommendations. Fedorenko, Ivanova, and Regev (2024).
Sentence Processing and Computational Language Models
What probabilistic models can predict
Computational systems estimate possible continuations from an input sequence. Such estimates can be used to calculate surprisal and entropy at each word and then test whether they predict human reading times or neural responses. The method offers a direct quantitative connection between a formal language model and measured behavior. A 2025 review synthesizes work on large language models as instruments for investigating incremental human reading and stresses methodological questions about interpreting prediction accuracy. Kaye and Gordon (2025).
Why model fit is not mechanism identity
A model that forecasts longer reading time at a difficult word may capture some distributional expectations without using the grammatical representations, attention, working-memory mechanisms, developmental history, or embodied context of human comprehension. Conversely, a failure to predict a particular psycholinguistic effect may reflect a modeling choice or benchmark mismatch. Mechanistic similarity requires converging tests: how representations are structured, whether controlled interventions alter predictions, how the model generalizes to rare structures, and whether the same analysis survives changes in language and context. Behavioral correspondence is evidence about a measured relationship, not independent proof of human-type understanding, intentions, or subjective experience. Kaye and Gordon (2025); Fedorenko, Ivanova, and Regev (2024).
A research tool within the Artificial Era
In the Artificial Era, formal linguistic models and artificial systems provide new experimental instruments and new forms of linguistic production. A comparison of human and machine sentence processing must state which capability is being tested: probabilistic prediction, structural generalization, functional comprehension, explanation, or subjective experience. Those are separate empirical and philosophical questions. This distinction makes possible detailed investigation of computational processing while protecting the interpretation of experimental findings from claims they cannot establish. Kaye and Gordon (2025).
Evidence Quality and Open Research Problems
Converging observations and established findings
Strong evidence supports incremental interpretation; measurable lexical, structural, and contextual influences; processing difficulty at disambiguating regions; dependency and interference effects in appropriate designs; and imperfect comprehension under certain task conditions. Multiple methods converge on the claim that grammar is actively constructed during processing. Exact causal accounts of many effects remain more contested. The existence of a slowdown is well established in some constructions; whether the slowdown is best attributed to serial reanalysis, competition, low probability, or a memory operation can remain unresolved. Reichle (2021); Lewis, Vasishth, and Van Dyke (2006).
Limits of individual experiments
Materials may contain unnoticed lexical frequency differences, uneven sentence plausibility, or unnatural continuations. Online measures may have spillover, individual variability, multiple plausible analysis windows, and sensitivity to presentation. A corpus establishes statistical associations but may not separate correlated predictors. A laboratory manipulation improves causal identification for the manipulated contrast, yet may not generalize to conversation. An ERP component or regional fMRI response is a measurement of a neural process under task conditions, not a transparent readout of a single grammatical operation. Scientific confidence comes from converging designs and appropriately constrained interpretation. Demberg and Keller (2008); Kuperberg (2007).
Cross-linguistic generalization and replicability
The distribution of word order, morphology, relative-clause types, argument structure, and contextual cues varies across languages. Results from a small set of constructions should not be treated as a universal cognitive law. Research also needs adequate statistical power, preregistered comparisons when suitable, transparent model choices, and out-of-sample evaluation. The relationship between sentence complexity and communicative efficiency is promising, but broad conclusions about why languages have particular grammatical patterns require language-diverse and historically informed evidence. Jaeger and Tily (2011); Traxler, Morris, and Seely (2002).
What remains theoretically contested
The strongest ongoing disputes concern the initial priority of syntax relative to meaning, the granularity of probabilistic predictions, the contribution of retrieval interference relative to structural distance, the completeness of resulting interpretations, and the neural implementation of hierarchical structure. New methods can narrow these disagreements only when studies are designed to make different theories predict different outcomes. A model should be judged by whether it predicts detailed patterns rather than how intuitively plausible its central metaphor sounds. Levy (2008); Murphy and Woolnough (2024).
Practical Implications for Reading, Teaching, and Communication
Sentence-processing science explains why an accurate sentence can be difficult to understand on first encounter. Dense embedding, ambiguous references, delayed resolution of dependencies, unfamiliar terminology, and low-probability constructions can increase processing demands. In educational writing and public communication, it is often useful to make referents identifiable, make relationships between clauses clear, provide meaningful context, and avoid preventable ambiguity. These are applications of general mechanisms, not clinical prescriptions or claims of a universally optimal writing formula.
During reading, rereading a complex sentence can help revise an initial parse and integrate delayed information. When explaining technical material, examples that clarify who or what performs an action can reduce unnecessary competition among interpretations. In listening situations, suitable prosodic grouping and adequate time for processing may support comprehension. Such design choices should respect the audience, task, language, and medium. Processing research supports the mechanisms underlying these practices, while the effectiveness of any particular educational intervention requires its own outcomes research. Reichle (2021); Ferreira, Bailey, and Ferraro (2002).
Frequently Asked Questions
Is sentence processing the same as grammar?
Grammar refers to systematic linguistic knowledge and structural organization. Sentence processing concerns how a person uses that knowledge and other information to build an interpretation over time. A grammar describes permissible relations; a processing theory must also explain speed, uncertainty, memory costs, and errors.
Do we wait for the entire sentence before understanding it?
Ordinarily, no. Reading, listening, and eye-movement studies demonstrate incremental interpretation as the input unfolds. Later words can confirm or revise earlier analyses, and not every initial interpretation is fully detailed. Levy (2008).
Why are some grammatically correct sentences hard to read?
Difficulty can reflect temporary ambiguity, unusual structural patterns, long dependencies, interference during retrieval, unfamiliar vocabulary, or violated expectations. Different experimental contrasts identify different contributions. Grammatical correctness and processing ease are separate properties. Gibson (1998).
What is a garden-path sentence?
It is a sentence whose early words strongly invite an analysis that later words make untenable. The resulting revision can be detected as a processing cost and may sometimes leave a lingering misinterpretation. Garden-path research is one specialized case of the wider science of incremental syntax.
Does the brain predict every next word?
Context influences processing and expectations can be graded. That does not demonstrate that every word is preceded by the obligatory conscious or unconscious selection of one exact lexical item. Studies distinguish predictions for words, grammatical categories, and meanings, and debate the mechanisms responsible for facilitatory effects. Van Petten and Luka (2012).
Is slower reading evidence of a language disorder?
No single reading-time observation supports that conclusion. Complexity, experience, presentation, sensory conditions, and attention can change processing speed. Clinical language and reading disorders require separate assessments and criteria.
Do language models understand sentences the same way humans do?
Language models can provide useful quantitative predictions of human reading difficulty and can exhibit impressive grammatical behavior. Matching one behavioral distribution does not show that they rely on the same representations or processes as humans. The question of subjective experience is analytically distinct from prediction performance. Kaye and Gordon (2025).
Scientific Synthesis
Real-time sentence comprehension is the coordination of grammatical structure-building, lexical and semantic interpretation, contextual inference, and memory-based operations. The emerging representation is built incrementally and can be revised when later input conflicts with earlier commitments. Experimental evidence establishes the importance of syntax, probability, thematic fit, discourse, and retrieval, while comparative models continue to investigate their relative timing and computational implementation.
The central scientific achievement of sentence-processing research is the ability to connect precise grammatical manipulations with measurable changes in human behavior and brain activity. Progress depends on converging methods, quantitative theoretical contrasts, and cross-linguistic tests that identify which mechanisms generalize and where experience changes the form of processing. These principles provide a framework for understanding the everyday ease of comprehension as well as the predictable circumstances in which it becomes difficult.
Related Articles
• Cluster Language, Thought and Psycholinguistics — the scientific research chapter connecting grammar with other cognitive language processes.
• Word Recognition: How the Mind Identifies Spoken and Written Words — How lexical competition, frequency, and sound-to-word or print-to-word identification provide the input for sentence parsing.
• Garden-Path Sentences: Why We Misinterpret Grammar — A focused investigation of syntactic ambiguity, misanalysis, reanalysis, and experimental evidence.
• Predictive Language Processing: How the Mind Anticipates Words — Expectation, surprisal, anticipatory processing, neural evidence, and theoretical limits of prediction.
• Discourse Comprehension: How We Understand Connected Language — How sentence interpretations are integrated across utterances through coherence, reference, and inference.
• Semantic Memory: Facts, Meaning, and Knowledge — a complementary account of stored conceptual knowledge that contributes to interpretation.
• Framing Effect: How the Wording of Choices Changes Decisions — a distinct decision-making phenomenon demonstrating how linguistic formulation affects judgments.
• Dyslexia Across Languages: Orthography and Reading Development — clinical and developmental questions about reading across writing systems, treated under their own canonical intent.
References
Demberg, V., and 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
Eberhard, K. M., Spivey-Knowlton, M. J., Sedivy, J. C., and Tanenhaus, M. K. (1995). Eye movements as a window into real-time spoken language comprehension in natural contexts. Journal of Psycholinguistic Research, 24, 409–436. https://doi.org/10.1007/BF02143160
Fedorenko, E., Ivanova, A. A., and Regev, T. I. (2024). The language network as a natural kind within the broader landscape of the human brain. Nature Reviews Neuroscience, 25, 289–312. https://doi.org/10.1038/s41583-024-00802-4
Ferreira, F., Bailey, K. G. D., and Ferraro, V. (2002). Good-enough representations in language comprehension. Current Directions in Psychological Science, 11(1), 11–15. https://doi.org/10.1111/1467-8721.00158
Gibson, E. (1998). Linguistic complexity: Locality of syntactic dependencies. Cognition, 68(1), 1–76. https://doi.org/10.1016/S0010-0277(98)00034-1
Jaeger, T. F., and Tily, H. (2011). On language ‘utility’: Processing complexity and communicative efficiency. WIREs Cognitive Science, 2, 323–335. https://doi.org/10.1002/wcs.126
Kaye, N. G., and Gordon, P. C. (2025). Sentence processing by humans and machines: Large language models as a tool to better understand human reading. Psychonomic Bulletin and Review, 32, 2719–2733. https://doi.org/10.3758/s13423-025-02756-9
Kuperberg, G. R. (2007). Neural mechanisms of language comprehension: Challenges to syntax. Brain Research, 1146, 23–49. https://doi.org/10.1016/j.brainres.2006.12.063
Kutas, M., and Federmeier, K. D. (2011). Thirty years and counting: Finding meaning in the N400 component of the event-related brain potential (ERP). Annual Review of Psychology, 62, 621–647. https://doi.org/10.1146/annurev.psych.093008.131123
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., and Van Dyke, J. A. (2006). Computational principles of working memory in sentence comprehension. Trends in Cognitive Sciences, 10(10), 447–454. https://doi.org/10.1016/j.tics.2006.08.007
MacDonald, M. C., Pearlmutter, N. J., and Seidenberg, M. S. (1994). The lexical nature of syntactic ambiguity resolution. Psychological Review, 101(4), 676–703. https://doi.org/10.1037/0033-295X.101.4.676
Murphy, E., and Woolnough, O. (2024). The language network is topographically diverse and driven by rapid syntactic inferences. Nature Reviews Neuroscience, 25, 705. https://doi.org/10.1038/s41583-024-00852-8
Reichle, E. D. (2021). Models of sentence processing. In Computational Models of Reading: A Handbook, 186–275. Oxford University Press. https://doi.org/10.1093/oso/9780195370669.003.0004
Smith, N. J., and 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
Staub, A. (2015). The effect of lexical predictability on eye movements in reading: Critical review and theoretical interpretation. Language and Linguistics Compass, 9, 311–327. https://doi.org/10.1111/lnc3.12151
Tanenhaus, M. K., Spivey-Knowlton, M. J., Eberhard, K. M., and 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
Traxler, M. J., Morris, R. K., and Seely, R. E. (2002). Processing subject and object relative clauses: Evidence from eye movements. Journal of Memory and Language, 47(1), 69–90. https://doi.org/10.1006/jmla.2001.2836
Trueswell, J. C., Tanenhaus, M. K., and Garnsey, S. M. (1994). Semantic influences on parsing: Use of thematic role information in syntactic ambiguity resolution. Journal of Memory and Language, 33(3), 285–318. https://doi.org/10.1006/jmla.1994.1014
Van Petten, C., and Luka, B. J. (2012). Prediction during language comprehension: Benefits, costs, and ERP components. International Journal of Psychophysiology, 83(2), 176–190. https://doi.org/10.1016/j.ijpsycho.2011.09.015
Wagers, M. W., Lau, E. F., and Phillips, C. (2009). Agreement attraction in comprehension: Representations and processes. Journal of Memory and Language, 61(2), 206–237. https://doi.org/10.1016/j.jml.2009.04.002