What Is a Garden-Path Sentence?
A garden-path sentence is a sentence whose opening words encourage a particular grammatical interpretation, while later words require the reader or listener to abandon or substantially revise that interpretation. The expression names a measurable phenomenon in psycholinguistics: comprehension proceeds incrementally, and the developing syntactic analysis can take a temporarily misleading path. A garden path is usually grammatical under its intended final analysis. What makes it difficult is the relationship between the structure initially favored and the structure ultimately required.
Consider “While Anna dressed the baby played in the crib.” Many readers initially treat “the baby” as the object of “dressed,” forming the interpretation that Anna dressed the baby. When they reach “played,” they must reconsider the sentence: “While Anna dressed, the baby played in the crib.” The baby is the subject of the main clause, and “dressed” is used without an expressed object. The missing comma is an important cue, but the deeper issue is that the sentence prefix permits more than one structural analysis. A reader must interpret that prefix before learning which analysis the final words require.
Garden-path research asks how an initially plausible grammatical structure becomes preferred, how a later word exposes its inadequacy, how the reader attempts to repair the analysis, and why a rejected meaning sometimes continues to influence the final interpretation. The phenomenon has been studied through eye movements, word-by-word reading times, brain potentials, and comprehension judgments. Its significance reaches beyond an entertaining sentence puzzle: it reveals the timing, expectations, and limits of human language understanding. Frazier and Rayner (1982); Christianson et al. (2001); Ceháková and Chromý (2023).
How a Sentence Leads the Reader Astray
Incremental parsing: interpreting before the ending arrives
Sentence comprehension is an unfolding cognitive operation. With each incoming word, readers retrieve lexical information, assign potential grammatical categories, connect phrases, estimate semantic roles, and update expectations about the continuation. They cannot normally suspend interpretation until the entire sentence has arrived. Instead, they build a provisional representation from the available evidence. The representation is often accurate enough to support smooth comprehension, but the same efficiency makes temporary structural errors possible.
The prefix “While Anna dressed the baby” has two important properties. First, “dressed” can occur with an object, so “the baby” can initially be attached as the person being dressed. Second, “dressed” can occur without an expressed object, allowing a clause boundary before “the baby.” The continuation determines which alternative produces a coherent full sentence. A processing system that favors the common or contextually attractive first interpretation must revise its commitments when “played” arrives.
A distinction is needed between syntactic structure and event meaning. The syntactic analysis specifies which noun phrase is the subject or object of which verb and how clauses are embedded. The event interpretation specifies who did what to whom. Reassigning a noun phrase from object to subject changes grammatical relationships; removing the imagined event of Anna dressing the baby requires a change in the semantic representation as well. These updates may occur at different times and with different degrees of success. Huang and Ferreira (2021); Slattery et al. (2013).
The point of disambiguation
The disambiguating region is the word or sequence at which the competing analyses cease to be equally viable. In the Anna example, “played” forces “the baby” to be understood as the subject of a new clause. The words before “played” supported an optional-object interpretation, while the later finite verb reveals the need for a clause boundary. The reader may slow down, move the eyes backward, or take longer to answer a comprehension question after encountering this region.
Such disruption indicates processing difficulty. Its presence alone does not reveal the precise hidden operation. A long fixation might reflect detection of a conflict, retrieval of preceding words, revision of an attachment, inhibition of an incorrect meaning, or several of these operations together. Inferences about the representation itself become stronger when online measurements are combined with carefully chosen comprehension probes. A theory that explains why reading slows must also explain what the reader ultimately understands.
Local ambiguity and global ambiguity
A temporarily ambiguous sentence can have a uniquely intended interpretation after later words appear. The garden path arises when processing initially favors an alternative that the ending rules out. By contrast, “I saw the visitor with the telescope” may remain ambiguous after the sentence ends: the telescope could belong to the visitor or be used by the speaker. That is a global attachment ambiguity unless more context resolves it. Some examples involve temporary ambiguity without substantial measurable difficulty, and some difficult sentences are not ambiguous at all. Garden-path effects therefore require evidence of both a misleading initial analysis and a subsequent demand for correction.
Garden-Path Sentence Examples, With Their Grammatical Explanations
“While the man hunted the deer ran into the woods”
The initial interpretation is that a man hunted a deer. The later verb “ran” requires a different structure: “While the man hunted, the deer ran into the woods.” Here “the deer” is the subject of the main clause, and the verb “hunted” completes the subordinate clause without a direct object. This is a classic noun-phrase/zero-complement ambiguity, often abbreviated NP/Z. The first noun phrase following the subordinate verb looks like that verb’s object, even though it must eventually function as the following clause’s subject.
Christianson and colleagues used related constructions to investigate whether the erroneous event interpretation persists after readers encounter the disambiguating region. Their experiments found that an incorrect thematic relation, such as interpreting the deer as something hunted, can continue to influence comprehension even when the reader has encountered the information needed for the correct structure. This finding underpins the research program on lingering misinterpretations. Christianson et al. (2001).
“The horse raced past the barn fell”
The first five words suggest a complete main-clause proposition: the horse raced past the barn. The later “fell” requires “raced past the barn” to be reinterpreted as a reduced relative clause modifying “the horse.” An expanded version makes the intended structure clear: “The horse that was raced past the barn fell.” The horse is the subject of “fell,” while “raced past the barn” describes it.
This example shows how grammatical well-formedness and immediate processing ease can diverge. A passive reduced relative leaves out overt markers such as “that was,” and the same word form can initially be taken as an ordinary past-tense main verb. The preference depends partly on a verb’s distribution, the thematic plausibility of the noun as an agent, and the surrounding context. Experiments show that semantic information can affect processing of these ambiguities at an early stage. Trueswell, Tanenhaus, and Garnsey (1994); MacDonald, Pearlmutter, and Seidenberg (1994).
“After the musician played the piano was moved”
At first, “the piano” seems to be what the musician played. When “was moved” follows, “the piano” must become the subject of the main clause: “After the musician played, the piano was moved.” The construction again depends on the availability of an object-taking interpretation for the earlier verb. Verb bias and plausibility may make the temporary object attachment more or less attractive. The experimental classification matters: the syntactic repair is related to the Anna example, but different verbs and nouns can change the size of the garden-path effect.
“The teacher knew the answer was wrong”
“The teacher knew the answer” encourages an object reading for “the answer.” The rest of the sentence instead licenses a clausal complement: the teacher knew that the answer was wrong. Inserting “that” clarifies the intended construction. This example may produce a mild temporary ambiguity rather than a dramatic garden path, since a comprehender can entertain the clausal alternative early. It illustrates why a careful study cannot assume that every sentence with more than one possible early parse creates the same degree of disruption.
“The old man the boats”
This popular linguistic puzzle relies on an unexpected grammatical category. “The old” denotes older people, and “man” functions as a verb meaning staff or operate: older people staff the boats. It demonstrates strong lexical-category expectations and how these can obscure a grammatically possible sentence. Its mechanism differs from the NP/Z cases in which a later finite verb forces a noun phrase to change grammatical role. Different ambiguity types should be analyzed separately rather than grouped under a single explanation merely because they produce surprise.
Types of Syntactic Ambiguity That Produce Garden Paths
Noun-phrase/zero-complement ambiguity
NP/Z garden paths involve a verb that can be used with or without a direct object. The following noun phrase initially invites attachment to that verb, but later material shows that it begins a new clause. The experimental contrast often compares an ambiguous version without a comma or additional marker to a version in which the structural boundary becomes apparent earlier. Because the erroneous attachment also creates an event relation, NP/Z items are especially useful for testing whether an initially assigned thematic role lingers after revision.
Main-verb/reduced-relative ambiguity
A past-tense or participial verb form may initially be interpreted as the main verb of a clause and later as part of a reduced relative modifying an earlier noun. A sentence such as “The horse raced past the barn fell” becomes easy once the omitted passive-relative markers are supplied. The difficulty depends on grammatical possibilities, usage frequency, and semantic plausibility. It is a productive test case for comparing syntactic economy with constraint-based accounts of sentence processing.
Phrase attachment and coordination ambiguity
A prepositional phrase can modify a noun, a verb, or an event location depending on the sentence, and a coordinated expression may join phrases or entire clauses. Subsequent words or a visible situation may rule out an initially preferred attachment. Such cases should be distinguished from NP/Z because the competing structures and available cues differ. A complete explanation identifies the competing constituents, the source of the preference, and the exact word or contextual event that forces reinterpretation.
Complement and category ambiguity
Some temporary ambiguities involve whether a noun phrase is a direct object or the subject of an embedded proposition. Others involve whether a word is a noun, adjective, or verb. The processing requirements differ: complement ambiguities may be resolved by a clause marker or later verb, while lexical-category puzzles require reconsidering a word’s grammatical function. Experimental work benefits from separating these varieties rather than treating garden-path sentences as one uniform stimulus class.
The Classic Garden-Path Model: Minimal Attachment and Late Closure
The influential early garden-path account proposed that a sentence processor often builds a first syntactic structure according to economical parsing principles. Minimal attachment favors the analysis that introduces less additional phrase structure; late closure favors attaching new material to the clause or phrase currently being processed when that is grammatically possible. In an appropriate ambiguity, such preferences explain why a short, locally available interpretation is chosen before the continuation reveals a different structure.
Frazier and Rayner’s 1982 eye-movement research made a foundational contribution by measuring how readers process structural ambiguity as they encounter critical words. Reading-time differences at or near disambiguation helped establish that these commitments can be observed during reading, rather than inferred only from a final judgment. Their framework distinguished initial syntactic analysis from subsequent revision. Frazier and Rayner (1982).
Minimal attachment and late closure are theoretical proposals about parsing preferences, not universal descriptions of every sentence in every language. A structural principle must be tested against alternative accounts and against evidence that verb meaning, discourse, referential context, and experience can affect decisions early. Historical garden-path theory continues to matter because it offers precise predictions about where a conflict should occur and which structure the comprehender initially prefers.
How researchers test parsing principles
A strong experiment varies the source of structural ambiguity while controlling length, lexical material, plausibility, and the location of a disambiguating word. Ambiguous and unambiguous versions of a sentence can differ by a clause marker, auxiliary verb, or punctuation cue. The comparison should preserve the intended meaning closely enough that any difference in processing cost can be attributed to the ambiguous prefix rather than a change in subject matter. Researchers also examine whether alternative parses were available weakly before disambiguation and whether that availability affects later repair. Staub (2007).
Constraint-Based, Interactive, and Experience-Sensitive Parsing
Constraint-based accounts describe initial parsing as the outcome of interacting information sources. The syntactic possibilities offered by the grammar are only part of the evidence: comprehenders also know the frequency with which a verb occurs in a given structure, the kinds of participants a verb normally takes, and the referential possibilities available in a discourse or visible situation. These cues change the relative strength of competing interpretations, sometimes before a word formally resolves the ambiguity.
MacDonald, Pearlmutter, and Seidenberg argued that lexical representations and their experienced distribution help explain syntactic ambiguity resolution. Trueswell, Tanenhaus, and Garnsey demonstrated early semantic influences in the interpretation of reduced relatives. These studies challenged rigid versions of a syntax-first account. The results do not make syntax dispensable: grammatical structure constrains what the alternatives can be. They show that a preferred parse can arise from several jointly informative cues. MacDonald, Pearlmutter, and Seidenberg (1994); Trueswell, Tanenhaus, and Garnsey (1994).
A central methodological question is when each cue becomes influential. An eye-movement difference may reflect early selection, competition among alternatives, detection of conflict, or later recovery. A final comprehension advantage can occur even without an easily measurable initial preference. A persuasive account therefore makes stage-specific predictions rather than simply asserting that meaning or syntax “wins.” Theories should be tested on more than one construction and more than one measure of performance.
Expectation, Surprisal, and the Probability of a Continuation
A probabilistic perspective asks which continuations are expected after the words read so far. Surprisal is commonly defined as the negative logarithm of a word’s conditional probability given its preceding context. A less expected continuation carries more information and may be associated with greater processing difficulty. In a garden-path sentence, the disambiguating word can be costly precisely because the structure it requires had a low probability under the preferred interpretation. Levy developed a formal expectation-based account that made such predictions central to syntactic comprehension. Levy (2008).
Several ideas often called prediction must be kept distinct. A reader may anticipate a particular word, maintain different probabilities over possible grammatical structures, or update the interpretation of an event after receiving unexpected material. Those operations need not share a single mechanism. Moreover, the probabilities used by a computational model depend on its training data and representation, so success in predicting text does not by itself demonstrate that the model reconstructs human garden-path behavior.
A comparative study published in 2026 examined several predictors of difficulty in locally ambiguous sentences, including cloze probability, structural and lexical biases, surprisal, and plausibility. In the experimental conditions studied, cloze probability was the most consistently supported predictor of reading difficulty and comprehension accuracy, while several other predictors showed less stable relationships. This informs models of expectations without establishing a universal ranking across every language, ambiguity type, or reading task. Ceháková and Chromý (2026).
What Happens When the Reader Must Reanalyze the Sentence?
Detecting the failure of the first interpretation
When a new word cannot be integrated under an earlier attachment, the processing system must notice the mismatch and respond. Detection may be rapid, yielding prolonged first-pass reading at the disambiguating word, or may become evident only after later material is read. The same visible response can reflect several operations, including retrieval of the original phrase, suppression of a conflicting role assignment, and construction of another sentence structure. Detecting that something is wrong does not guarantee that the entire sentence will be represented correctly.
Full reconstruction and partial repair
One possible outcome is complete reanalysis: the earlier structure is abandoned or appropriately subordinated, and the new syntax and event meaning are integrated into a coherent whole. Another is partial repair, in which the final words are accommodated but some earlier relationships remain incorrect. A third is simultaneous availability of incompatible or locally coherent analyses. Reading-time evidence alone rarely distinguishes these outcomes; final paraphrases, thematic-role questions, and probes about the full sentence are essential.
Christianson and colleagues showed that early thematic-role assignments can linger after disambiguation. Later experiments gave evidence for competing syntactic representations, while still other work emphasized incomplete reconstruction of global structure or fallible retrieval. Ceháková and Chromý’s 2023 investigation in Czech reported heterogeneous outcomes, including accurate analyses, mixtures of local interpretations, and cases where the final coherent structure was missing. The scientific picture is therefore richer than a single claim that readers either “fix” or “fail to fix” a garden path. Christianson et al. (2001); Slattery et al. (2013); Huang and Ferreira (2021); Ceháková and Chromý (2023).
The difference between representation and retrieval
Suppose a reader answers “yes” when asked whether Anna dressed the baby. One explanation is that the first syntactic parse was never replaced. Another is that a correct parse was constructed, but memory retains and later retrieves the abandoned event. The question itself may cue the mistaken interpretation. Because different hidden states can lead to the same answer, researchers cannot infer the exact failure mechanism from a single yes/no response. Huang and Ferreira examined this distinction, asking whether lingering mistakes result primarily from incorrect syntax or memory processes. Their work makes the relationship between interpretation, representation, and retrieval an empirical question. Huang and Ferreira (2021).
Good-Enough Language Processing and Lingering Misinterpretations
The good-enough processing framework investigates situations in which comprehenders produce an interpretation that serves an immediate communicative purpose without fully reconstructing every syntactic relation. Language processing operates under constraints of time, attention, and relevance. An event meaning that is plausible and locally coherent may continue to guide responses even after later words favor a different grammatical arrangement. Experiments on garden-path sentences provide some of the most prominent evidence for this family of ideas.
The framework is more precise when its different predictions are separated. An initial error may persist as a competing event model; a corrected syntactic representation may coexist with an obsolete thematic role; or the parser may never construct a globally consistent structure. These predictions are related but empirically distinguishable. A 2024 review by Frances examined two decades of good-enough language-processing research and discussed how processing heuristics, task demands, and the content of final representations shape the interpretation of experimental findings. Frances (2024).
The term “good enough” does not imply carelessness or low intelligence. It names an account of processing depth and outcome under particular circumstances. In natural communication, the most relevant meaning may be obtained rapidly; in an experiment, asking detailed questions about who performed every action can expose inaccuracies in the resulting representation. The science requires both online and final-outcome measures to determine where such inaccuracies arise.
Does Rereading Correct a Garden-Path Mistake?
Readers often move their eyes backward when a sentence becomes difficult. It is tempting to treat that regression as evidence that they have returned to the precise word requiring a new parse and completed the repair. Eye movement alone does not justify that conclusion. People can reread to check their understanding, recover a missed word, or confirm an interpretation they already possess. Some backward movements are automatic consequences of processing disruption and are not targeted at the grammatical source of the error.
Christianson and colleagues tested this issue in 2024 using two eye-tracking experiments that manipulated the text available for rereading. They studied how rereading was distributed and whether it predicted more accurate comprehension after NP/Z garden paths. The results showed considerable variability, little evidence of selective targeting of syntactically decisive regions, and no consistent improvement of final interpretation from rereading in the tested conditions. The investigators concluded that rereading was frequently confirmatory rather than a reliable vehicle for structural reanalysis. Christianson et al. (2024).
The result applies to the studied materials and procedures; it does not imply that deliberately reconstructing a sentence is ineffective. A reader who identifies the second finite verb, locates the earlier clause boundary, and reassigns the noun phrase can sometimes obtain the intended meaning. That strategy directs attention to a specific structural alternative. The important distinction is between an observed backward eye movement and demonstrated successful repair. An experimental method should measure both separately.
How Garden-Path Effects Are Studied Scientifically
Eye tracking: first-pass time, regressions, and rereading
Eye tracking measures where a reader fixates, how long the eyes remain on regions of a sentence, and how often they move backward. Researchers can examine first-pass fixation duration, regressions out of a region, go-past or regression-path time, and later rereading. These measures index overlapping but distinct phases of processing. A disambiguating verb can generate difficulty during initial reading, whereas subsequent regressions may indicate checking or revision attempts.
The placement of the critical region is essential. A delay precisely at the first word incompatible with the preferred parse is stronger evidence of an online garden-path effect than a difference in total time spent on the sentence. Yet eye movements are indirect indicators of cognitive operations. A long fixation does not by itself reveal which syntactic tree is represented or whether a coherent final interpretation will emerge. Frazier and Rayner’s original work and more recent rereading experiments demonstrate why process timing and comprehension outcome must be analyzed together. Frazier and Rayner (1982); Christianson et al. (2024).
Self-paced reading and controlled word presentation
In a self-paced reading experiment, words or segments appear sequentially as a participant advances through the sentence. The time spent before revealing the next region provides a measure of processing demand. Matched ambiguous and unambiguous conditions help estimate the cost of disambiguation. This method permits tightly controlled presentation and can be adapted to different languages, but it removes the natural eye regressions that readers ordinarily make. Researchers must consider whether the procedure alters strategies or exaggerates particular kinds of difficulty.
Comprehension questions, paraphrases, and interpretation profiles
A correct answer about one grammatical relationship does not prove that all components of the sentence were reconstructed. Likewise, an incorrect answer may reflect a lingering event representation, a retrieval failure, or the wording of the question. Experiments therefore benefit from questions testing both the intended and the erroneous event, open-ended paraphrasing, and probes about different sentence regions. The 2023 Czech study used several measurement approaches to reveal differences among final representations that would have been missed by a single accuracy score. Ceháková and Chromý (2023).
Event-related brain potentials and the P600
Electroencephalography can measure neural responses time-locked to a word that conflicts with an earlier analysis. A late positive event-related potential, commonly called the P600, has frequently been associated with syntactic disruption and reanalysis. Osterhout and Holcomb’s landmark research demonstrated sensitivity of brain potentials to syntactically anomalous continuations. However, a P600-like positivity can also arise in circumstances involving semantic or thematic incongruity. It is evidence of a particular neural response pattern, not a direct reading of an internal grammatical repair or proof that comprehension ultimately succeeded. Osterhout and Holcomb (1992); Friederici (2011).
Experimental controls and statistical interpretation
A well-designed garden-path study controls or measures sentence length, lexical frequency, verb bias, plausibility, ambiguity duration, punctuation, and the position of disambiguation. Differences among experimental items matter because one verb may strongly prefer a direct object and another may frequently appear without one. Results should be evaluated across sentences as well as participants. A study reporting greater reading time at a critical word establishes a processing difference under its tested conditions; claiming that this proves a unique parsing algorithm requires additional evidence separating competing theoretical predictions.
Context, Semantic Plausibility, and Prosody
What world knowledge contributes
A sentence parser does not operate independently of the meanings of words and the situations they describe. The probability that a noun is an agent of a particular verb can influence which interpretation seems plausible. Reduced-relative sentences are especially informative: a noun strongly associated with performing the apparent main verb may encourage an agentive reading, while an unlikely agent may encourage a passive modifier reading. Experimental evidence shows that thematic-role plausibility can shape processing early. Trueswell, Tanenhaus, and Garnsey (1994).
This finding explains why familiar garden-path examples vary in their effect across readers. A grammatical ambiguity is a property of the structures licensed by the sentence prefix; the strength of the garden-path experience also depends on how strongly experience and context favor a particular candidate. Grammatical possibility and processing probability are therefore different dimensions of the same experiment.
Referential context and visual-world evidence
In conversation, listeners often hear an instruction while looking at objects in their environment. The visible scene provides potential referents and can make some syntactic analyses more relevant than others. Visual-world research demonstrated that listeners rapidly integrate linguistic information with visual and referential context while interpreting spoken language. This integration can shape temporary structural preferences and the difficulty of subsequent disambiguation. Tanenhaus et al. (1995); Eberhard et al. (1995).
These findings matter for generalization. A sentence presented on a blank laboratory screen may create a stronger or qualitatively different ambiguity than the same utterance embedded in a meaningful scene. Context can make a particular referent salient, supply expectations about actions, or render a candidate interpretation implausible. A comprehensive model has to account for processing both in carefully controlled isolated sentences and in communication with rich contextual cues.
Commas, pauses, and intonational boundaries
Written punctuation can prevent a garden path by identifying a clause boundary earlier. Compare “While Anna dressed the baby played” with “While Anna dressed, the baby played.” The comma signals that the subordinate clause has ended. Spoken language often provides analogous information through timing, intonation, stress, and prosodic grouping. These cues do not map perfectly onto punctuation and may be ambiguous themselves. Their contribution depends on the language, speech context, and grammatical construction. Friederici (2011).
Prosody makes the interpretation process genuinely multimodal: linguistic structure is constrained by the acoustic form of an utterance as well as its words. The presence of a pause does not automatically produce one analysis in every circumstance. Experiments have to manipulate and measure prosodic cues directly rather than assuming the spoken version is equivalent to a punctuated written sentence.
Individual Differences, Task Demands, and Language Experience
Readers vary in familiarity with grammatical constructions, vocabulary, reading proficiency, prior knowledge, and the amount of attention devoted to an utterance. These factors can change which interpretation initially dominates and how successfully it is reconsidered. An experienced reader may still commit a temporary error when encountering a familiar garden-path structure quickly. Conversely, a person may repair a difficult sentence after its grammatical relationships have been pointed out. Garden-path susceptibility is not a simple index of intelligence or general linguistic ability.
The task also shapes performance. A participant who expects detailed comprehension questions may invest more effort in reconstructing syntax than someone casually reading a passage. Open-ended paraphrasing can reveal different errors from yes/no questions; word-by-word presentation can influence recovery differently from natural reading. Differences between participants must be separated from differences produced by instructions, stimulus materials, or measurement procedures. These distinctions are necessary before proposing a broad psychological trait that supposedly determines garden-path accuracy.
Garden-path difficulty also belongs to ordinary language processing. A misleading sentence that momentarily confuses a healthy reader is not a symptom or screening test for dyslexia, developmental language disorder, aphasia, or an attentional disorder. Clinical evaluation requires evidence about functioning, developmental or medical history, and appropriate assessment procedures. The present article concerns mechanisms of syntactic comprehension; clinical diagnosis and intervention are separate research and care questions.
Do Garden Paths Work the Same Way in Spoken Language and Other Languages?
Speech reaches the listener incrementally, so temporary syntactic commitment can occur without written words. Listeners also receive timing, intonation, coarticulatory information, and frequently a visual or social context. Those cues may reduce ambiguity, intensify a misleading reading, or shift the point at which conflict becomes apparent. Spoken and written experiments investigate a common problem of incremental interpretation using partly different sources of information. Eberhard et al. (1995).
Cross-linguistic comparisons must take grammatical differences seriously. Languages vary in basic word order, case marking, agreement, optional arguments, morphosyntactic cues, and how relative clauses are signaled. An English NP/Z sentence cannot simply be translated word for word and expected to yield the same ambiguity in another language. Czech experiments, including work examining incomplete and competing final representations, show the value of investigating local ambiguity within a language’s own grammatical resources. Ceháková and Chromý (2023).
The broad generalization is that comprehenders build hypotheses about structure before every cue is available. More specific claims concerning minimal attachment, expectation strength, the timing of repair, and the persistence of a mistaken event must be established for each relevant construction, population, and task. A research program that considers several languages can help distinguish universal properties of incremental cognition from the learned statistics of particular language systems.
Do Garden-Path Experiments Prove a Single Parsing Algorithm?
The existence of a garden-path effect establishes that interpretation can be temporarily misleading and that later evidence can cause observable disruption. It does not uniquely select a serial parsing algorithm, an interactive parallel algorithm, a surprisal-based account, or a good-enough interpretation model. These theoretical approaches address overlapping questions at different explanatory levels, and several can predict slowing at the same disambiguating word.
A classic serial account can explain a preferred initial structure followed by revision cost. A probabilistic account can explain why an unlikely continuation is costly. A constraint-based account can explain why lexical experience and situational context change preferences from the beginning. A good-enough account can explain why an erroneous event relation may persist even after later information is encountered. The strongest experiments construct conditions in which these accounts make different predictions about early reading, regressions, and final answers.
Not all competition-based intuitions have received straightforward experimental support. Van Gompel and colleagues reported evidence against simple competition accounts in particular syntactic ambiguity conditions. This illustrates a larger methodological lesson: the intuitive idea that the reader “holds two parses in conflict” is a theoretical claim about internal representation. It must be evaluated against data rather than treated as something directly observed from a prolonged fixation. van Gompel et al. (2005).
What Research Published From 2023 Through 2026 Adds
Recent studies focus more explicitly on the final content of comprehension. In 2023, Ceháková and Chromý found that garden-path sentences in Czech could result in several forms of representation, including a coherent correct interpretation, incompatible locally coherent interpretations, or an incomplete global structure. These outcomes suggest that the binary distinction between an initial mistake and a successful correction is sometimes too coarse. Ceháková and Chromý (2023).
In 2024, Christianson and colleagues tested whether rereading was directed toward a structural repair and associated with better comprehension. Their evidence did not support that assumption in the NP/Z materials examined. A 2024 review of good-enough processing placed such findings in a broader debate about comprehension depth, task demands, and the representations remaining after difficult syntax. Christianson et al. (2024); Frances (2024).
Further work published in 2025 examined how the consequences of a garden path can extend beyond the original misanalysis and disrupt broader sentence processing. In 2026, a comparative investigation evaluated cloze probability, frequency, surprisal, and plausibility as predictors of processing difficulty. Its results favored cloze probability within the studied conditions. Collectively these studies make the research agenda more precise: researchers must identify the construction, the earliest conflicting cue, the interpretation process, and the kind of final representation that is being measured. Ceháková and Chromý (2025); Ceháková and Chromý (2026).
What Garden-Path Research Means for Clear Writing
The practical principle is to supply structural cues before the reader is encouraged to adopt an interpretation that must later be overturned. Clause punctuation, overt complementizers, and expanded relative clauses can make intended relationships explicit. “The horse that was raced past the barn fell” signals its relative-clause structure more clearly than “The horse raced past the barn fell.” “While Anna dressed, the baby played” marks the boundary that was previously concealed.
When encountering a difficult sentence, a reader can locate the disambiguating word, identify the attachment it rules out, and paraphrase the sentence with explicit subjects and clause boundaries. This is a more specific strategy than simply rereading the same phrase. The advice is consistent with the distinction between deliberate structural reconstruction and backward eye movements that do not reliably lead to improved understanding. It is a practical application of the research rather than evidence for any particular clinical intervention. Christianson et al. (2024).
In teaching, a comparison among an ambiguous sentence, a version with punctuation, and a version with the relevant clause expanded can make abstract syntax visible. Students can identify why a noun phrase changes its grammatical role and why the first event interpretation was attractive. In instructions, contracts, educational materials, and public communication, early clarity about the actor, action, and clause boundary helps readers identify the intended meaning without costly revision. Actual improvements in reading performance for a specific document or population should still be evaluated with users rather than assumed.
Computational Models, Human Language, and the Artificial Era
Computational systems can assign probabilities to word continuations, estimate surprisal, and generate explanations for alternative grammatical structures. Such capabilities are valuable in psycholinguistic research when models help construct stimuli or produce quantitative predictions to compare with human reading times and judgments. A model may predict a disambiguating word with low probability, but prediction of text and reanalysis of human-like event representations are different scientific targets.
A large language model that correctly identifies the subject of a garden-path sentence has demonstrated performance on a linguistic task. That observation alone does not establish the same perception, memory, or repair mechanisms documented in human experiments. Likewise, fluent explanation does not establish human-type understanding or subjective experience. The psychological study of garden paths can compare measurable behavior and representations, while claims about artificial subjective experience require their own independent investigation. This distinction allows computational language research to contribute to Psychology for the Artificial Era without collapsing the differences among prediction, interpretation, and experience.
Frequently Asked Questions
Are garden-path sentences grammatically incorrect?
Classic garden-path examples have an intended grammatical interpretation. They seem incorrect or confusing because an earlier phrase encourages a different analysis. Some widely circulated examples are merely awkward or ungrammatical; establishing a valid final parse is part of analyzing a genuine garden-path sentence.
Why does a comma make some garden paths easier?
A comma can mark the end of a subordinate clause before a noun phrase appears. In “While Anna dressed, the baby played,” it makes it easier to treat “the baby” as the subject of “played” rather than the object of “dressed.” It functions as an early structural cue.
Does everyone follow the same incorrect interpretation?
No. Verb bias, semantic plausibility, context, reading conditions, and individual language experience influence what a reader initially favors. A garden-path effect is measured across sentences and readers; a memorable puzzle is not proof of a universal response.
What is the difference between a garden path and ordinary ambiguity?
A garden path involves a preferred early analysis that later becomes untenable or expensive to revise. A globally ambiguous sentence can retain two viable interpretations at its ending. The two phenomena are related but address different aspects of language processing.
Why can the original wrong meaning remain in memory?
The reader may have formed an event representation before the syntax was clarified. Experiments show that this initial interpretation can continue to influence answers, though researchers debate whether the cause lies in competing grammatical structures, partial repair, or retrieval of an obsolete semantic trace. Huang and Ferreira (2021).
Does rereading guarantee that I understand the sentence?
No. In relevant eye-tracking experiments, backward eye movements were not reliable evidence of successful reanalysis. Identifying the word that forces a new grammatical structure and paraphrasing the entire sentence is a more explicit repair procedure. Christianson et al. (2024).
Can garden paths occur when listening rather than reading?
Yes. Speech also requires incremental interpretation, although intonation, timing, and referential context can change the location and strength of an ambiguity. Visual-world experiments document the rapid integration of linguistic and situational information during listening. Tanenhaus et al. (1995).
Is difficulty with garden-path sentences a sign of dyslexia or aphasia?
No. Temporary misinterpretation of a structurally misleading sentence occurs in normal comprehension. Diagnoses require assessment of broader patterns of language or reading functioning, not performance on a single experimental or recreational sentence.
Scientific Conclusions
Garden-path sentences reveal how language comprehension proceeds before grammatical certainty is available. Readers integrate words into provisional structures, strongly preferred interpretations can become incompatible with later input, and recovery requires changes in grammatical and semantic representations. Eye tracking, behavioral experiments, event-related brain potentials, and comprehension testing each illuminate different parts of this sequence.
The major open questions concern the mechanisms of initial selection, the interaction of grammar with meaning and context, the role of probabilistic expectation, the extent of structural reanalysis, and the persistence of discarded interpretations. Research through 2026 supports treating these as separable empirical problems. A scientifically adequate explanation specifies the kind of ambiguity, the information available at each word, the measured behavior, and the representation assessed after comprehension. The garden-path phenomenon is therefore a central case study in the temporal organization of human language understanding.
Related Articles
Sentence Processing: How We Understand Grammar in Real Time — The broader science of incremental parsing, structural dependencies, and real-time grammatical interpretation.
Semantic Priming: How Related Words Influence Processing — The separate experimental study of how prior meaning and associative context influence language processing.
Cluster Language, Thought and Psycholinguistics — The published research chapter connecting sentence processing to language, thought, and comprehension.
Predictive Processing in Perception: Theory, Evidence, and Debate — A separate analysis of general perceptual prediction, evidence, and competing accounts.
Selective Attention: How the Mind Filters Information — General attention and cognitive control as neighboring mechanisms.
Semantic Memory: Facts, Meaning, and Knowledge — The representation and retrieval of conceptual knowledge, a distinct memory topic.
References
Ceháková, M., and Chromý, J. (2023). Garden-path sentences and the diversity of their (mis)representations. PLOS ONE, 18(7), e0288817. https://doi.org/10.1371/journal.pone.0288817.
Ceháková, M., and Chromý, J. (2025). Beyond initial misanalysis: The disruptive nature of garden-path sentences processing. Discourse Processes, 62(10), 723–743. https://doi.org/10.1080/0163853X.2025.2513210.
Ceháková, M., and Chromý, J. (2026). Cloze, frequency, surprisal, or plausibility? A comparative analysis of predictors for local ambiguity resolution. Cognitive Science, 50(4), e70208. https://doi.org/10.1111/cogs.70208.
Christianson, K., Dempsey, J., Tsiola, A., Deshaies, S.-E. M., and Kim, N. (2024). Retracing the garden-path: Nonselective rereading and no reanalysis. Journal of Memory and Language, 137, 104515. https://doi.org/10.1016/j.jml.2024.104515.
Christianson, K., Hollingworth, A., Halliwell, J. F., and Ferreira, F. (2001). Thematic roles assigned along the garden path linger. Cognitive Psychology, 42(4), 368–407. https://doi.org/10.1006/cogp.2001.0752.
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(6), 409–436. https://doi.org/10.1007/BF02143160.
Frances, C. (2024). Good enough processing: What have we learned in the 20 years since Ferreira et al. (2002)? Frontiers in Psychology, 15, 1323700. https://doi.org/10.3389/fpsyg.2024.1323700.
Frazier, L., and Rayner, K. (1982). Making and correcting errors during sentence comprehension: Eye movements in the analysis of structurally ambiguous sentences. Cognitive Psychology, 14(2), 178–210. https://doi.org/10.1016/0010-0285(82)90008-1.
Friederici, A. D. (2011). The brain basis of language processing: From structure to function. Physiological Reviews, 91(4), 1357–1392. https://doi.org/10.1152/physrev.00006.2011.
Huang, Y., and Ferreira, F. (2021). What causes lingering misinterpretations of garden-path sentences: Incorrect syntactic representations or fallible memory processes? Journal of Memory and Language, 121, 104288. https://doi.org/10.1016/j.jml.2021.104288.
Levy, R. (2008). Expectation-based syntactic comprehension. Cognition, 106(3), 1126–1177. https://doi.org/10.1016/j.cognition.2007.05.006.
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.
Osterhout, L., and Holcomb, P. J. (1992). Event-related brain potentials elicited by syntactic anomaly. Journal of Memory and Language, 31(6), 785–806. https://doi.org/10.1016/0749-596X(92)90039-Z.
Slattery, T. J., Sturt, P., Christianson, K., Yoshida, M., and Ferreira, F. (2013). Lingering misinterpretations of garden path sentences arise from competing syntactic representations. Journal of Memory and Language, 69(2), 104–120. https://doi.org/10.1016/j.jml.2013.04.001.
Staub, A. (2007). The return of the repressed: Abandoned parses facilitate syntactic reanalysis. Journal of Memory and Language, 57(2), 299–323. https://doi.org/10.1016/j.jml.2006.09.001.
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.
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 Gompel, R. P. G., Pickering, M. J., Pearson, J., and Liversedge, S. P. (2005). Evidence against competition during syntactic ambiguity resolution. Journal of Memory and Language, 52(2), 284–307. https://doi.org/10.1016/j.jml.2004.11.003.